Archives for posts with tag: Semantic computing

Résumé

Le but de ce texte est de présenter une vue générale des limites de l’IA contemporaine et de proposer une voie pour les dépasser. L’IA a accompli des progrès considérables depuis l’époque des Claude Shannon, Alan Turing et John von Neumann. Néanmoins, de nombreux obstacles se dressent encore sur la route indiquée par ces pionniers. Aujourd’hui l’IA symbolique se spécialise dans la modélisation conceptuelle et le raisonnement automatique tandis que l’IA neuronale excelle dans la catégorisation automatique. Mais les difficultés rencontrées aussi bien par les approches symboliques que neuronales sont nombreuses. Une combinaison des deux branches de l’IA, bien que souhaitable, laisse encore non résolus les problèmes du cloisonnement des modèles et les difficultés d’accumulation et d’échange des connaissances. Or l’intelligence humaine naturelle résout ces problèmes par l’usage du langage. C’est pourquoi je propose que l’IA adopte un modèle calculable et univoque du langage humain, le Métalangage de l’Économie de l’Information (IEML pour Information Economy MetaLanguage), un code sémantique de mon invention. IEML a la puissance d’expression d’une langue naturelle, il possède la syntaxe d’un langage régulier, et sa sémantique est univoque et calculable parce qu’elle est une fonction de sa syntaxe. Une architecture neuro-sémantique basée sur IEML allierait les forces de l’IA neuronale et de l’IA symbolique classique tout en permettant l’intégration des connaissances grâce à un calcul interopérable de la sémantique. De nouvelles avenues s’ouvrent à l’intelligence artificielle, qui entre en synergie avec la démocratisation du contrôle des données et l’augmentation de l’intelligence collective.
La fin du texte contient des références bibliographiques et des liens pour approfondir.

Art: Emma Kunz

Introduction

Examinons d’abord comment le terme “intelligence artificielle” (IA) est utilisé dans la société en général, par exemple par les journalistes et les publicitaires. L’observation historique montre que l’on a tendance à classer dans l’intelligence artificielle les applications considérées comme “avancées” à l’époque où elles apparaissent. Mais quelques années plus tard ces mêmes applications seront le plus souvent réinterprétées comme appartenant à l’informatique ordinaire. Par exemple, la reconnaissance optique de caractères, perçue comme de l’IA à l’origine, est aujourd’hui considérée comme normale et silencieusement intégrée dans de nombreux logiciels. Une machine capable de jouer aux échecs était célébrée comme un exploit technique jusqu’aux années 1970, mais l’on peut aujourd’hui télécharger un programme d’échecs gratuit sur son smartphone sans que nul ne s’en étonne. De plus, selon que l’IA est en vogue (comme aujourd’hui) ou déconsidérée (comme dans les années 1990-2000), les efforts de marketing mettront ce terme en avant ou le remplaceront par d’autres. Par exemple, les “systèmes experts” des années 1980 deviennent les anodines “règles d’affaire” des années 2000. C’est ainsi que des techniques ou des concepts identiques changent de dénomination selon les modes, rendant la perception du domaine et de son évolution particulièrement opaque.

Quittons maintenant le vocabulaire du journalisme ou du marketing pour nous intéresser à la discipline académique. L’intelligence artificielle désigne depuis les années 1950 la branche de l’informatique qui se préoccupe de modéliser et de simuler l’intelligence humaine dans son ensemble plutôt que de résoudre tel ou tel problème particulier. La modélisation informatique de l’intelligence humaine est un but scientifique légitime qui a eu et continuera à avoir des retombées théoriques et pratiques considérables. Néanmoins, échaudés par les prévisions enthousiastes, mais démenties par les faits, des débuts de la discipline, la plupart des chercheurs du domaine ne croient pas que l’on construira bientôt des machines intelligentes autonomes. Beaucoup de recherches dans ce domaine – ainsi que la plupart des applications pratiques – visent d’ailleurs une augmentation de la cognition humaine plutôt que sa reproduction mécanique. Par opposition au programme de recherche orienté vers la construction d’une intelligence artificielle générale autonome, j’ai défendu dans mon livre La Sphère Sémantique l’idée d’une intelligence artificielle au service de l’intelligence collective et du développement humain. Je poursuis ici cette ligne de pensée.

D’un point de vue technique, l’IA se partage en deux grandes branches: symbolique et statistique. Un algorithme d’IA statistique “apprend” à partir des données qu’on lui fournit. Il simule donc (imparfaitement, nous le verrons plus bas), la dimension inductive du raisonnement humain. Par contraste, l’IA symbolique n’apprend pas à partir des données, mais dépend de la formalisation logique de la connaissance d’un domaine par des ingénieurs. Comparée à l’IA statistique, elle demande donc en principe une quantité plus importante de travail intellectuel humain. Un algorithme d’IA symbolique applique aux données les règles qu’on lui a données. Il simule donc plutôt la dimension déductive du raisonnement humain. Je vais successivement passer en revue ces deux grandes branches de l’IA, en m’attachant plus particulièrement à souligner leurs limites.

L’IA statistique et ses limites

L’IA neuronale

La branche statistique de l’IA entraîne des algorithmes à partir d’énormes masses de données pour les rendre capable de reconnaître des formes visuelles, sonores, linguistiques ou autres. C’est ce que l’on appelle l’apprentissage automatique ou machine learning. Lorsque l’on parle d’IA en 2021, c’est généralement pour désigner ce type de technique. On l’a vu, l’IA statistique économise le travail humain si on la compare à l’IA symbolique. Il suffit de fournir à un algorithme d’apprentissage automatique un ensemble de données d’entraînement pour qu’un programme de reconnaissance de formes s’écrive tout seul. Si l’on donne par exemple à une IA statistique des millions d’images de canards accompagnées d’étiquettes précisant que l’image représente un canard, elle apprend à reconnaître un canard et, à l’issue de son entraînement, elle sera capable de coller elle-même l’étiquette “canard” sur une image non catégorisée de ce volatile. Personne n’a expliqué à la machine comment reconnaître un canard : on s’est contenté de lui fournir des exemples. La traduction automatique répond au même principe : on donne à une IA statistique des millions de textes dans une langue A accompagnés de leur traduction dans une langue B. Entraîné sur ces exemples, le système apprend à traduire un texte de la langue A dans la langue B. C’est ainsi que fonctionnent des algorithmes de traduction automatique comme DeepL ou Google Translate. Pour prendre un exemple dans un autre domaine, l’IA statistique utilisée pour conduire les “véhicules autonomes” fonctionne également en appariant deux ensembles de données : des images de la route sont mises en correspondance avec des actions telles qu’accélérer, freiner, tourner, etc. En somme, l’IA statistique établit une correspondance (mapping) entre un ensemble de données et un ensemble d’étiquettes (cas de la reconnaissance de forme) ou bien entre deux ensembles de données (cas de la traduction ou des véhicules autonomes). Elle excelle donc dans les tâches de catégorisation, de reconnaissance de forme et d’appariement réflexe entre données perceptives et données motrices. 

Dans sa version la plus perfectionnée, l’IA statistique repose sur des modèles de réseaux neuronaux qui simulent grossièrement le mode d’apprentissage du cerveau. On parle d’apprentissage “profond” (deep learning en anglais) pour qualifier ces modèles parce qu’ils reposent sur plusieurs couches superposées de neurones formels. Les réseaux neuronaux représentent le sous-domaine le plus complexe et le plus avancé de l’IA statistique. L’intelligence artificielle de type neuronal existe depuis l’origine de l’informatique, comme l’illustrent les recherches de McCulloch dans les années 1940 et 50, de Franck Rosenblatt et Marvin Minsky dans les années 1950 et de von Fœrster dans les  années 1960 et 70. D’importants travaux dans ce domaine ont également été menés dans les années 1980, impliquant notamment David Rumelhart et Geoffrey Hinton, mais toutes ces recherches ont eu peu de succès pratique avant les années 2010.

Outre certains perfectionnements scientifiques des modèles, deux facteurs indépendants des progrès de la théorie expliquent que les réseaux neuronaux soient de plus en plus utilisés : la disponibilité d’énormes masses de données et l’augmentation de la puissance de calcul. À partir de la seconde décennie du XXIe siècle, les organisations s’engagent dans la transformation numérique et une part croissante de la population mondiale utilise le Web. Tout cela génère de gigantesques flux de données. Les informations ainsi produites sont traitées par les grandes plateformes numériques dans des centres de données (le “cloud“) qui concentrent une puissance de calcul inouïe. Au début du XXIe siècle, les réseaux neuronaux étaient implémentés par des processeurs conçus à l’origine pour le calcul graphique, mais les centres de données des grandes plateformes utilisent maintenant des processeurs spécialement destinés à l’apprentissage neuronal. C’est ainsi que des modèles théoriques intéressants, mais peu pratiques, du XXe siècle sont soudain devenus pertinents au XXIe siècle au point de soutenir une nouvelle industrie.

Des rendements décroissants

Néanmoins, après les avancées foudroyantes des années 2010 en matière d’apprentissage automatique par les réseaux neuronaux, les progrès semblent marquer le pas depuis quelques années. En effet, pour obtenir des performances marginalement meilleures, il faut désormais multiplier par plusieurs ordres de grandeur la taille des ensembles de données et la puissance de calcul utilisée pour entraîner les modèles. Nous avons déjà atteint l’époque des rendements cognitifs décroissants pour l’IA neuronale. Il est donc temps de s’interroger sur les limites de cet ensemble de techniques et d’envisager sérieusement un changement de paradigme. 

Les principaux problèmes portent sur la qualité des données d’entraînement, l’absence de modélisation causale, le caractère inexplicable des résultats, l’absence de généralisation, la cécité par rapport au sens des données et les difficultés d’accumulation et d’intégration des connaissances.

La qualité des données d’entraînement

Un ingénieur de Google aurait déclaré plaisamment: “Chaque fois que nous licencions un linguiste, notre performance en traduction automatique s’améliore”. Mais bien que l’IA statistique soit réputée peu gourmande en travail humain, les risques de biais et d’erreurs soulignés par des utilisateurs de plus en plus sourcilleux poussent à mieux sélectionner les données d’entraînement et à les étiqueter d’une manière plus soigneuse. Or cela demande du temps et de l’expertise humaine, bien qu’il s’agisse précisément des facteurs que l’on espérait éliminer.

L’absence d’hypothèses causales explicites

Tous les cours de statistiques commencent par une mise en garde contre la confusion entre corrélation et causalité. Une corrélation entre A et B ne prouve pas que A est la cause de B. Il peut s’agir d’une coïncidence, ou bien B peut être la cause de A, ou bien un facteur C non pris en compte par le recueil de données est la véritable cause de A et B, sans parler de toutes les relations systémiques complexes imaginables impliquant A et B. Or l’apprentissage automatique repose sur des appariements de données, c’est-à-dire sur des corrélations. La notion de causalité est étrangère à l’IA statistique, comme à de nombreuses techniques d’analyse de données massives, bien que des hypothèses causales interviennent souvent de manière implicite dans le choix des ensembles de données et de leur catégorisation. En somme, l’IA neuronale contemporaine n’est pas capable de distinguer les causes des effets. Pourtant, quand on utilise l’IA pour l’aide à la décision et plus généralement pour s’orienter dans des domaines pratiques, il est indispensable de posséder des modèles causaux explicites, car les actions efficaces doivent bel et bien intervenir sur les causes. Dans une démarche scientifique intégrale, les mesures statistiques et les hypothèses causales s’inspirent et se contrôlent réciproquement. Ne considérer que les corrélations statistiques relève d’une dangereuse hémiplégie cognitive. Quant à la pratique répandue qui consiste à garder ses théories causales implicites, elle interdit de les relativiser, de les comparer avec d’autres théories, de les généraliser, de les partager, de les critiquer et de les perfectionner.

Des résultats inexplicables 

Le fonctionnement des réseaux neuronaux est opaque. Des millions d’opérations transforment de manière incrémentale la force des connexions dans des assemblées de neurones comportant des centaines de couches. Comme leurs résultats ne peuvent être expliqués ni justifiés de manière conceptuelle, c’est-à-dire sur un mode compréhensible par des humains, il est difficile de faire confiance à ces modèles. Cette absence d’explication devient inquiétante lorsque les machines prennent des décisions financières, judiciaires, médicales ou liés à la conduite de véhicules autonomes, sans parler des applications militaires. Pour surmonter cet obstacle, et parallèlement au développement de l’éthique de l’intelligence artificielle, de plus en plus de chercheurs explorent le nouveau champ de recherche de “l’IA explicable” (explainable AI).

L’absence de généralisation. 

L’IA statistique se présente à première vue comme une forme de raisonnement inductif, c’est-à-dire comme une capacité à inférer des règles générales à partir d’une multitude de cas particuliers. Pourtant, les systèmes d’apprentissage automatique contemporains ne parviennent pas à généraliser au-delà des limites des données d’entraînement qui leur ont été fournies. Non seulement nous – les humains – sommes capables de généraliser à partir de quelques exemples, alors qu’il faut des millions de cas pour entraîner des machines, mais nous pouvons abstraire et conceptualiser ce que nous avons appris tandis que l’apprentissage automatique ne parvient pas à extrapoler et encore moins à conceptualiser. Il reste au niveau d’un apprentissage purement réflexe, étroitement circonscrit par l’espace des exemples qui l’ont alimenté.

La cécité au sens

Alors que les performances en traduction ou en écriture automatique (tel qu’illustré par le programme GPT3) progressent, les machines ne parviennent toujours pas à comprendre le sens des textes qu’elles traduisent ou rédigent. Leurs réseaux neuronaux ressemblent au cerveau d’un perroquet mécanique capable d’imiter des performances linguistiques sans avoir la moindre idée du contenu des textes. La succession des mots dans une langue ou leur correspondance d’une langue à l’autre sont bien maîtrisées, mais les textes “reconnus” n’alimentent pas de représentations utilisables des situations ou des domaines de connaissance dont ils traitent. 

Les difficultés d’accumulation et d’intégration des connaissances par l’IA statistique 

Privée de concepts, l’IA statistique parvient difficilement à accumuler des connaissances. A fortiori, l’intégration de savoirs de divers champs d’expertise semble hors de portée. Cette situation ne favorise pas les échanges de connaissances entre machines. Il faut donc souvent recommencer à zéro pour chaque nouveau projet. Signalons néanmoins l’existence de modèles de traitement des langues naturelles comme BERT qui sont pré-entraînés sur des données générales et qu’il est ensuite possible de spécialiser dans des domaines particuliers. Une forme limitée de capitalisation est donc atteignable. Mais il reste impossible d’intégrer directement à un système neuro-mimétique l’ensemble des connaissances objectives accumulées par l’humanité depuis quelques siècles.

L’IA symbolique et ses limites

La branche symbolique de l’IA correspond à ce qui a été successivement appelé dans les soixante-dix dernières années: réseaux sémantiques, systèmes à base de règles, bases de connaissances, systèmes experts, web sémantique et, plus récemment, graphes de connaissance. Depuis ses origines dans les années 1940-50, une bonne partie de l’informatique appartient de fait à l’IA symbolique. 

L’IA symbolique code la connaissance humaine de manière explicite sous forme de réseaux de relations entre catégories et de règles logiques donnant prise au raisonnement automatique. Ses résultats sont donc plus facilement explicables que ceux de l’IA statistique. 

Les difficultés d’accumulation et d’intégration des connaissances par l’IA symbolique

L’IA symbolique fonctionne bien dans les micromondes fermés des jeux ou des laboratoires, mais se trouve rapidement dépassée dans les environnements ouverts qui ne répondent pas à un petit nombre de règles strictes. La plupart des programmes d’IA symbolique utilisés dans des environnements de travail réels ne résolvent de problèmes que dans un domaine étroitement limité, qu’il s’agisse de diagnostic médical, de dépannage de machines, de conseil en investissement ou autre. Un “système expert” fonctionne de fait comme un médium d’encapsulation et de distribution d’un savoir-faire particulier, qui peut être distribué partout où on en a besoin. La compétence pratique devient alors disponible même en l’absence de l’expert humain. 

À la fin des années 1980, à la suite d’une série de promesses inconsidérées suivies de déceptions, commence ce que l’on a appelé “l’hiver” de l’intelligence artificielle (toutes tendances confondues). Pourtant, les mêmes procédés continuent à être utilisés pour résoudre le même type de problèmes. On a seulement renoncé au programme de recherche général dans lequel ces méthodes s’inscrivaient. C’est ainsi qu’au début du XXIe siècle, les règles d’affaires des logiciels d’entreprise et les ontologies du Web Sémantique ont succédé aux systèmes experts des années 1980. Malgré les changements de nom, il est aisé de reconnaître dans ces nouvelles spécialités les procédés de la bonne vieille IA symbolique. 

À partir du début du XXIe siècle, le “Web sémantique” s’est donné pour finalité d’exploiter les informations disponibles dans l’espace ouvert du Web. Afin de rendre les données lisibles par les ordinateurs, on organise les différents domaines de connaissance ou de pratique en modèles cohérents. Ce sont les “ontologies”, qui ne peuvent que reproduire le cloisonnement logique des décennies précédentes, malgré le fait que les ordinateurs soient maintenant beaucoup plus interconnectés.

Malheureusement, nous retrouvons dans l’IA symbolique les mêmes difficultés d’intégration et d’accumulation des connaissances que dans l’IA statistique. Ce cloisonnement entre en opposition avec le projet originel de l’intelligence artificielle comme discipline scientifique, qui veut modéliser l’intelligence humaine en général et qui tend normalement vers une accumulation et une intégration des connaissances mobilisables par les machines.

Malgré le cloisonnement de ses modèles, l’IA symbolique est cependant un peu mieux lotie que l’IA statistique en matière d’accumulation et d’échange. Un nombre croissant d’entreprises, à commencer par les grandes compagnies du Web, organisent leurs bases de données au moyen d’un graphe de connaissance constamment amélioré et augmenté. Par ailleurs, Wikidata offre un bon exemple de graphe de connaissance ouvert grâce auquel une information lisible aussi bien par les machines que par les humains s’accumule progressivement. Néanmoins, chacun de ces graphes de connaissance est organisé selon les finalités – toujours particulières – de ses auteurs, et ne peut être réutilisable facilement pour d’autres fins. Ni l’IA statistique, ni l’IA symbolique ne possèdent les propriétés de recombinaison fluide que l’on est en droit d’attendre des modules d’une intelligence artificielle au service de l’intelligence collective.

L’IA symbolique est gourmande en travail intellectuel humain

On a bien tenté d’enfermer toute la connaissance humaine dans une seule ontologie afin de permettre une meilleure interopérabilité, mais alors la richesse, la complexité, l’évolution et les multiples perspectives du savoir humain sont effacées. Sur un plan pratique, les ontologies universelles – voire celles qui prétendent formaliser l’ensemble des catégories, relations et règles logiques d’un vaste domaine – deviennent vite énormes, touffues, difficiles à comprendre et à maintenir pour l’humain qui est amené à s’en occuper. Un des principaux goulets d’étranglement de l’IA symbolique est d’ailleurs la quantité et la haute qualité du travail humain nécessaire à modéliser un domaine de connaissance, aussi étroitement circonscrit soit-il. En effet, il est non seulement nécessaire de lire la documentation, mais il faut encore interroger et écouter longuement plusieurs experts du domaine à modéliser. Acquis par l’expérience, les savoirs de ces experts s’expriment le plus souvent par des récits, des exemples et par la description de situations-types. Il faut alors transformer une connaissance empirique de style oral en un modèle logique cohérent dont les règles doivent être exécutables par un ordinateur. En fin de compte, le raisonnement des experts sera bien automatisé, mais le travail “d’ingénierie de la connaissance” d’où procède la modélisation ne peut pas l’être.

Position du problème: quel est le principal obstacle au développement de l’IA?

Vers une intelligence artificielle neuro-symbolique

Il est maintenant temps de prendre un peu de recul. Les deux branches de l’IA – neuronale et symbolique – existent depuis le milieu du XXe siècle et elles correspondent à deux styles cognitifs également présents chez l’humain. D’une part, nous avons la reconnaissance de formes (pattern recognition) qui correspond à des modules sensorimoteurs réflexes, que ces derniers soient appris ou d’origine génétique. D’autre part, nous avons une connaissance conceptuelle explicite et réfléchie, souvent organisée en modèles causaux et qui peut faire l’objet de raisonnements. Comme ces deux styles cognitifs fonctionnent ensemble dans la cognition humaine, il n’existe aucune raison théorique pour ne pas tenter de les faire coopérer dans des systèmes d’intelligence artificielle. Les bénéfices sont évidents et, en particulier, chacun des deux sous-systèmes peut remédier aux problèmes rencontrés par l’autre. Dans une IA mixte, la partie symbolique surmonte les difficultés de conceptualisation, de généralisation, de modélisation causale et de transparence de la partie neuronale. Symétriquement, la partie neuronale amène les capacités de reconnaissance de forme et d’apprentissage à partir d’exemples qui font défaut à l’IA symbolique. 

Aussi bien d’importants chercheurs en intelligence artificielle que de nombreux observateurs avertis de la discipline poussent dans cette direction d’une IA hybride. Par exemple, Dieter Ernst a récemment défendu une “intégration entre les réseaux neuronaux, qui excellent dans la classification des perceptions et les systèmes symboliques, qui excellent dans l’abstraction et l’inférence”. Emboîtant le pas à Gary Marcus, les chercheurs en IA Luis Lamb et Arthur D’avila Garcez ont récemment publié un article en faveur d’une IA neuro-symbolique dans laquelle des représentations acquises par des moyens neuronaux seraient interprétées et traitées par des moyens symboliques. Il semble donc que l’on ait trouvé une solution au problème du blocage de l’IA : il suffirait d’accoupler intelligemment les branches symbolique et statistique plutôt que de les maintenir séparées comme deux programmes de recherche en concurrence. D’ailleurs, ne voit-on pas les grandes compagnies du Web, qui mettent en avant l’apprentissage automatique et l’IA neuronale dans leurs efforts de relations publiques, développer plus discrètement en interne des graphes de connaissance pour organiser leur mémoire numérique et donner sens aux résultats des réseaux neuronaux? Mais avant de déclarer la question réglée, réfléchissons encore un peu aux données du problème.

Cognition animale et cognition humaine

Pour chacune des deux branches de l’IA, nous avons dressé une liste des obstacles qui se dressent sur le chemin menant vers une intelligence artificielle moins fragmentée, plus utile et plus transparente. Or nous avons trouvé un même inconvénient des deux côtés: le cloisonnement logique, les difficultés d’accumulation et d’intégration. Réunir le neuronal au symbolique ne nous aidera pas à surmonter cet obstacle puisque ni l’un ni l’autre n’en sont capables. Pourtant, les sociétés humaines réelles peuvent transformer des perceptions muettes et des savoir-faire issus de l’expérience en connaissances partageables. À force de dialogue, un spécialiste d’un domaine finit par se faire comprendre d’un spécialiste d’un autre domaine et va peut-être même lui enseigner quelque chose. Comment reproduire ce type de performances cognitives dans des sociétés de machines? Qu’est-ce qui joue le rôle intégrateur du langage naturel dans les systèmes d’intelligence artificielle?

Bien des gens pensent que, le cerveau étant le support organique de l’intelligence, les modèles neuronaux sont la clé de sa simulation. Mais de quelle intelligence parle-t-on? N’oublions pas que tous les animaux ont un cerveau, or ce n’est pas l’intelligence du moucheron ou de la baleine que l’IA veut simuler, mais celle de l’humain. Et si nous sommes “plus intelligents” que les autres animaux (au moins de notre point de vue) ce n’est pas à cause de la taille de notre cerveau. L’éléphant possède un plus gros cerveau que l’Homme en termes absolus et le rapport entre la taille du cerveau et celle du corps est plus grand chez la souris que chez l’humain. C’est principalement notre capacité linguistique, notamment supportée par les aires de Broca, Wernicke et quelques autres (uniques à l’espèce humaine), qui distingue notre intelligence de celle des autres vertébrés supérieurs. Or ces modules de traitement du langage ne sont pas fonctionnellement séparés du reste du cerveau, ils informent au contraire l’ensemble de nos processus cognitifs, y compris nos compétences techniques et sociales. Nos perceptions, nos actions, nos émotions et nos communications sont codées linguistiquement et notre mémoire est largement organisée par un système de coordonnées sémantiques fourni par le langage.

Fort bien, dira-t-on. Simuler les capacités humaines de traitement symbolique, y compris la faculté linguistique, n’est-ce pas précisément ce que l’IA symbolique est censée faire? Mais alors comment se fait-il qu’elle soit cloisonnée en ontologies distinctes, qu’elle peine à assurer l’interopérabilité sémantique de ses systèmes et qu’elle ne parvienne si difficilement à accumuler et à échanger les connaissances? Tout simplement parce que, malgré son nom de “symbolique”, l’IA ne dispose toujours pas d’un modèle calculable du langage. Depuis les travaux de Chomsky, nous savons calculer la dimension syntaxique des langues, mais leur dimension sémantique reste hors de portée de l’informatique. Afin de comprendre cette situation, il est nécessaire de rappeler quelques éléments de sémantique.

La sémantique en linguistique

Du point de vue de l’étude scientifique du langage, la sémantique d’un mot ou d’une phrase se décompose en deux parties, mélangées dans la pratique, mais conceptuellement distinctes: la sémantique linguistique et la sémantique référentielle. En gros, la sémantique linguistique s’occupe des relations entre les mots alors que la sémantique référentielle traite de la relation entre les mots et les choses.

La sémantique linguistique ou sémantique mot-mot. Un symbole linguistique (mot ou phrase) possède généralement deux faces: le signifiant, qui est une image visuelle ou acoustique et le signifié qui est un concept ou une catégorie générale. Par exemple, le signifiant “arbre”, a pour signifié : “végétal ligneux, de taille variable, dont le tronc se garnit de branches à partir d’une certaine hauteur”. La relation entre signifiant et signifié étant fixée par la langue, le signifié d’un mot ou d’une phrase se définit comme un nœud de relations avec d’autres signifiés. Dans un dictionnaire classique, chaque mot est situé par rapport à d’autres mots proches (le thésaurus) et il est expliqué par des phrases (la définition) utilisant des mots eux-mêmes expliqués par d’autres phrases, et ainsi de suite de manière circulaire. Un dictionnaire classique relève principalement de la sémantique linguistique. Les verbes et les noms communs (par exemple: arbre, animal, organe, manger) représentent des catégories qui sont elles-mêmes connectées par un dense réseau de relations sémantiques telles que: “est une partie de”, “est un genre de”, “appartient au même contexte que”, “est la cause de”, “est antérieur à”, etc. Nous ne pouvons penser et communiquer à la manière humaine que parce que nos mémoires collectives et personnelles sont organisées par des catégories générales connectées par des relations sémantiques.

La sémantique référentielle ou sémantique mot-chose. Par contraste avec la sémantique linguistique, la sémantique référentielle fait le pont entre un symbole linguistique (signifiant et signifié) et un référent (un individu réel). Lorsque je dis que “les platanes sont des arbres”, je précise le sens conventionnel du mot “platane” en le mettant en relation d’espèce à genre avec le mot “arbre” et je ne mets donc en jeu que la sémantique linguistique. Mais si je dis que “Cet arbre-là, dans la cour, est un platane”, alors je pointe vers un état de chose réel, et ma proposition est vraie ou fausse. Ce second énoncé met évidemment en jeu la sémantique linguistique puisque je dois d’abord connaître le sens des mots et la grammaire du français pour la comprendre. Mais s’ajoute à la dimension linguistique une sémantique référentielle puisque l’énoncé se rapporte à un objet particulier dans une situation concrète. Certains mots, comme les noms propres, n’ont pas de signifiés. Leur signifiant renvoie directement à un référent. Par exemple, le signifiant “Alexandre le Grand” désigne un personnage historique et le signifiant “Tokyo” désigne une ville. Par contraste avec un dictionnaire ordinaire, qui définit des concepts ou des catégories, un dictionnaire encyclopédique contient des descriptions d’individus réels ou fictifs pourvus de noms propres tels que divinités, héros de roman, personnages et événements historiques, objets géographiques, monuments, œuvres de l’esprit, etc. Sa principale fonction est de répertorier et de décrire des objets externes au système d’une langue. Il enregistre donc une sémantique référentielle.

Nota bene: Une catégorie est une classe d’individus, une abstraction. Il peut y avoir des catégories d’entités, de process, de qualités, de quantités, de relations, etc. Les mots “catégorie” et “concept” sont ici traités comme des synonymes.

La sémantique en IA

En informatique, les références ou individus réels (les réalités dont on parle) deviennent les données alors que les catégories générales deviennent les rubriques, champs ou métadonnées qui servent à classer et retrouver les données. Par exemple, dans la base de données d’une entreprise, “nom de l’employé”, “adresse” et “salaire” sont des catégories ou métadonnées tandis que “Tremblay”, “33 Boulevard René Lévesques” et “65 K$ / an” sont des données. Dans ce domaine technique, la sémantique référentielle correspond au rapport entre données et métadonnées et la sémantique linguistique au rapport entre les métadonnées ou catégories organisatrices, qui sont généralement représentées par des mots ou de courtes expressions linguistiques. 

Dans la mesure ou la finalité de l’informatique est d’augmenter l’intelligence humaine, elle doit notamment nous aider à donner sens aux flots de données numériques et à en tirer le maximum de connaissances utiles pour l’action. À cet effet, nous devons catégoriser correctement les données – c’est-à-dire mettre en œuvre une sémantique mot-chose – et organiser les catégories selon des relations pertinentes, qui nous permettent d’extraire des données toutes les connaissances utiles pour l’action – ce qui correspond à la sémantique mot-mot.

En discutant le sujet de la sémantique en informatique, nous devons nous souvenir que les ordinateurs ne voient pas spontanément un mot ou une phrase comme un concept en relation déterminée avec d’autres concepts dans le cadre d’une langue, mais seulement comme des suites de lettres, des “chaînes de caractères”. C’est pourquoi les relations entre les catégories qui semblent évidentes aux humains et qui relèvent de la sémantique linguistique, doivent être ajoutées – le plus souvent à la main – dans une base de données si l’on veut qu’un programme en tienne compte.

Examinons maintenant dans quelle mesure l’IA symbolique modélise la sémantique. Si l’on considère les ontologies du “Web Sémantique” (le standard en IA symbolique), on découvre que les sens des mots et des phrases n’y dépendent pas de la circularité auto-explicative de la langue (comme dans un dictionnaire classique), mais d’un renvoi à des URI (Uniform Resource Identifiers) qui fonctionne sur le mode de la sémantique référentielle (comme un dictionnaire encyclopédique). Au lieu de reposer sur des concepts (ou catégories) déjà donnés dans une langue et qui se présentent dès l’origine comme des nœuds de relations avec d’autres concepts, les échafaudages du Web sémantique s’appuient sur des concepts définis séparément les uns des autres au moyen d’identifiants uniques. La circulation du sens dans un réseau de signifiés est escamotée au profit d’une relation directe entre signifiant et référent, comme si tous les mots étaient des noms propres. En l’absence d’une sémantique linguistique fondée sur une grammaire et un dictionnaire communs, les ontologies restent donc cloisonnées. En somme, l’IA symbolique contemporaine n’a pas accès à la pleine puissance cognitive et communicative du langage parce qu’elle ne dispose pas d’une langue, mais seulement d’une sémantique référentielle rigide.

Pourquoi l’IA n’utilise-t-elle pas les langues naturelles – avec leur sémantique linguistique inhérente – pour représenter les connaissances? La réponse est bien connue : parce que les langues naturelles sont ambiguës. Un mot peut avoir plusieurs sens, un sens peut s’exprimer par plusieurs mots, les phrases ont plusieurs interprétations possibles, la grammaire est élastique, etc. Comme les ordinateurs ne sont pas des êtres incarnés et pleins de bon sens, comme nous, ils ne sont pas capables de désambiguïser correctement les énoncés en langue naturelle. Pour ses locuteurs humains, une langue naturelle étend un filet de catégories générales prédéfinies qui s’expliquent mutuellement. Ce réseau sémantique commun permet de décrire et de faire communiquer aussi bien les multiples situations concrètes que les différents domaines de connaissance. Mais, du fait des limitations des machines, l’IA ne peut pas faire jouer ce rôle à une langue naturelle. C’est pourquoi elle reste aujourd’hui fragmentée en micro-domaines de pratiques et de connaissance, chacun d’eux avec sa sémantique particulière.

L’automatisation de la sémantique linguistique pourrait ouvrir de nouveaux horizons de communication et de raisonnement à l’intelligence artificielle. Pour traiter la sémantique linguistique, l’IA aurait besoin d’une langue standardisée et univoque, d’un code spécialement conçu à l’usage des machines, mais que les humains pourraient aisément comprendre et manipuler. Cette langue permettrait enfin aux modèles de se connecter et aux connaissances de s’accumuler. En somme, le principal obstacle au développement de l’IA est l’absence d’un langage commun calculable. C’est précisément le problème résolu par IEML, qui possède la capacité d’exprimer le sens, comme les langues naturelles, mais dont la sémantique est non ambiguë et calculable, comme un langage mathématique. L’utilisation d’IEML rendra l’IA moins coûteuse en efforts humains, plus apte à traiter le sens et la causalité, et surtout, capable d’accumuler et d’échanger des connaissances.

Une solution basée sur un codage de la sémantique

Le métalangage de l’économie de l’information

Beaucoup de progrès en informatique viennent de l’invention d’un système de codage pertinent rendant l’objet codé (nombre, image, son, etc.) facilement calculable par une machine. Par exemple, le codage binaire pour les nombres et le codage en pixels ou en vecteurs pour les images. C’est pourquoi je me suis attaché à la conception d’un code qui rende la sémantique linguistique calculable. Cette langue artificielle, IEML (Information Economy MetaLanguage) possède une grammaire régulière et un dictionnaire compact de trois mille mots. Des catégories complexes peuvent être construites en combinant les mots selon les règles de la grammaire. Les catégories complexes peuvent à leur tour être utilisées pour en définir d’autres, et ainsi de suite récursivement. Une des parties les plus difficiles de la conception d’IEML a été de trouver le plus petit ensemble de mots à partir duquel n’importe quelle catégorie pourrait être construite. 

Sur un plan linguistique, IEML possède la même capacité expressive qu’une langue naturelle. Elle peut donc traduire n’importe quelle autre langue. C’est d’autre part une langue univoque : ses expressions n’ont qu’un seul sens. Enfin, sa sémantique est calculable. Cela signifie que son dictionnaire et ses règles de grammaire suffisent à déterminer automatiquement le sens de ses expressions (ce qui n’est pas le cas des langues naturelles). Soulignons qu’IEML n’est pas une ontologie universelle, mais bel et bien une langue qui permet d’exprimer n’importe quelle ontologie ou classification particulière. 

Sur un plan mathématique, IEML est un langage régulier au sens de Chomsky : c’est une algèbre. Elle est donc susceptible de toutes sortes de traitements et de transformations automatiques. 

Sur un plan informatique, comme nous le verrons plus en détail ci-dessous, le métalangage donne prise à un langage de programmation de réseaux sémantiques et supporte le système d’indexation d’une base de connaissances.

L’éditeur IEML

Le métalangage de l’économie de l’information est défini par sa grammaire et son dictionnaire, que l’on trouvera en consultant le site intlekt.io. Mais la langue ne suffit pas. Nous avons besoin d’un outil numérique facilitant son écriture, sa lecture et son utilisation: l’éditeur IEML. 

L’éditeur IEML sert à produire et à explorer des modèles de données. Cette notion de “modèle” englobe les réseaux sémantiques, les systèmes de métadonnées sémantiques, les ontologies, les graphes de connaissances et les systèmes d’étiquettes pour catégoriser des données d’entraînement. L’éditeur contient un langage de programmation permettant d’automatiser la création de nœuds (les catégories) et de liens (les relations sémantiques entre catégories). Ce langage de programmation est de type déclaratif, c’est-à-dire qu’il ne demande pas à son utilisateur d’organiser des flots d’instructions conditionnelles, mais seulement de décrire les résultats à obtenir.

Mode d’utilisation de l’éditeur

Comment se sert-on de l’éditeur? 

  1. Le modélisateur répertorie les catégories qui vont servir de conteneurs (ou de cases-mémoire) aux différents types de données. S’il a besoin de catégories qui ne correspondent à aucun des 3000 mots du dictionnaire IEML il les crée au moyen de phrases.
  2. Il programme ensuite les relations sémantiques qui vont connecter les données catégorisées. Les relations, définies par des phrases, ont un contenu sémantique aussi varié que nécessaire. Leurs propriétés mathématiques (réflexivité, symétrie, transitivité) sont spécifiées. Des instructions conditionnent l’établissement des relations à la présence de signifiants ou de valeurs de données déterminées à certaines adresses syntaxiques des catégories.
  3. Une fois les données catégorisées, le programme tisse automatiquement le réseau de relations sémantiques qui va leur donner sens. Des fonctions de fouille de données, d’exploration hypertextuelle et de visualisation des relations par tables et par graphes permettent aux utilisateurs finaux d’explorer le contenu modélisé.

Avantages

Plusieurs traits fondamentaux distinguent l’éditeur IEML des outils contemporains qu’on utilise pour modéliser les données: les catégories et relations sont programmables, les modèles obtenus sont interopérables et transparents.

Catégories et relations sont programmables. La structure régulière d’IEML permet de générer les catégories et de tisser les relations de manière fonctionnelle ou automatique au lieu de les créer une par une. Cette propriété fait gagner au modélisateur un temps considérable. Le temps gagné par l’automatisation de la création des catégories et des relations compense largement le temps passé à coder les catégories en IEML, et cela d’autant plus qu’une fois créées, les nouvelles catégories et relations peuvent être échangées entre les utilisateurs. 

Les modèles sont interopérables. Tous les modèles se ramènent au même dictionnaire de trois mille mots et à la même grammaire. Les modèles sont donc interopérables, c’est-à-dire qu’ils peuvent facilement fusionner ou échanger des catégories et des sous-modèles. Chaque modèle reste adapté à un contexte particulier, mais les modèles peuvent désormais se comparer, s’interconnecter et s’intégrer.

Les modèles sont transparents. Bien qu’ils soient codés en IEML, les modèles rédigés au moyen de l’éditeur IEML sont lisibles en langue naturelle. De plus, les catégories et relations se présentent comme des mots ou des phrases. Or les mots sont expliqués par leurs relations avec les autres mots du dictionnaire et le sens des phrases est défini par les mots qui les composent selon une grammaire régulière. Toutes les catégories et toutes les relations sont donc explicitement définies, ce qui rend les modèles clairs aussi bien pour les modélisateurs que pour les utilisateurs finaux et adéquats aux principes d’éthique et de transparence contemporains.

Au prix d’un bref apprentissage, l’éditeur peut être utilisé par des non-informaticiens qui ne connaissent pas la langue IEML. Seule la grammaire (simple et régulière) doit être maîtrisée, les mots IEML étant représentés en langues naturelles. L’éditeur IEML pourrait être utilisé dans les écoles et ouvrir la voie à une démocratisation de la maîtrise des données.

L’architecture neuro-sémantique

Figure 1: Une architecture Neuro-sémantique pour l’IA

Je vais maintenant proposer une architecture de système d’IA basée sur IEML. Cette architecture (schématisée dans la figure 1) est évidemment un cas particulier d’architecture neuro-symbolique, mais je la nomme neuro-sémantique afin de souligner qu’elle résout le problème du calcul de la sémantique et de l’interopérabilité sémantique entre systèmes. 

Les neurones sensorimoteurs

Le module d’entrée est occupé par des réseaux de neurones sensoriels, qui ont été entraînés par des exemples de données catégorisées en IEML. On doit distinguer plusieurs types de données d’entraînement (texte, image, sons, etc.) d’où résultent plusieurs types de réseaux de neurones. Les données catégorisées par les neurones sensoriels sont transmis à la base de connaissance sémantique. Si l’on détecte des incohérences, des erreurs ou des biais, il faut évidemment revoir les données d’entraînement ou réviser leur conceptualisation. Le système doit donc comprendre une boucle de dialogue entre les annotateurs de données qui entraînent les réseaux de neurones et les ingénieurs qui gèrent la base de connaissance.

En sortie, des réseaux de neurones moteurs transforment des données catégorisées en données qui commandent des actions, telles que rédaction de texte, synthèse d’image, émission vocale, instructions envoyées à des effecteurs (robots), etc. Ces neurones moteurs sont entraînés sur des exemples qui apparient des données catégorisées en IEML et des données motrices. Là encore, les données d’entraînement et les réseaux de neurones doivent être distinguées selon leurs types.

La mémoire et le traitement sémantique

La base de connaissance est organisée par un réseau sémantique. Elle est donc de préférence supportée par une base de données de graphes (graph database). Sur le plan de l’interface, cette base de connaissance se présente comme une encyclopédie hypertextuelle du domaine dont elle traite. Elle autorise aussi la programmation de simulations et de divers tableaux de bord pour la veille et le renseignement.

L’éditeur IEML évoqué à la section précédente peut servir à d’autres tâches qu’à la modélisation. Il permet en effet de conditionner les opérations d’écriture-lecture les plus variées à la présence de contenus sémantiques situés à certaines adresses syntaxiques. Lorsqu’ils sont codés en IEML les concepts deviennent les variables d’une algèbre, ce qui n’est évidemment pas le cas lorsqu’elles sont exprimés en langue naturelle. C’est pourquoi des transformations sémantiques peuvent être programmées et calculées. Cette programmation sémantique ouvre la voie non seulement aux raisonnements logiques classiques auxquels les moteurs d’inférence de l’IA symbolique nous ont habitué depuis des décennies, mais aussi à d’autres formes de raisonnement automatique. Puisqu’en IEML la sémantique est une image fonctionnelle de la syntaxe, il devient possible d’automatiser le raisonnement analogique de type “A est à B ce que C est à D”.  D’autres d’opérations sémantiques peuvent également être programmées, telles que sélection et fouille ; substitution, insertion ou effacement ; extraction de sous-réseaux sémantiques pertinents ; résumé ou développement ; inversion, allusion, atténuation ou amplification ; extraction ou projection de structures narratives, et ainsi de suite.

Quelques applications 

Quelques applications évidentes de notre architecture d’IA neuro-sémantique sont : l’intégration de données, l’aide à la décision à partir de modèles causaux, la gestion des connaissances, la compréhension et le résumé de texte, la génération de texte contrôlée (contrairement aux systèmes de type GPT3 dont le texte n’est pas contrôlé), les chatbots et la robotique. Je vais maintenant brièvement commenter deux exemples d’usage : la compréhension de texte et la génération de texte contrôlée. 

Concernant la génération de texte contrôlée, imaginons en entrée des données de télémétrie, des informations comptables, des examens médicaux, des résultats de tests de connaissance, etc. On peut alors concevoir en sortie des textes narratifs en langue naturelle synthétisant le contenu des flux de données d’entrée : diagnostics médicaux, bulletins scolaires, rapports, conseils, etc. Quant à la compréhension de texte, elle suppose d’abord la catégorisation automatique du contenu du document présenté en entrée du système. Dans un deuxième temps, le modèle sémantique extrait du texte est inscrit dans la mémoire du système de manière à s’intégrer aux connaissances déjà acquises. En somme, des systèmes d’intelligence artificielle pourraient accumuler des connaissances à partir de la lecture automatique de documents. À supposer qu’IEML soit adopté, les systèmes d’intelligence artificielle deviendraient non seulement capables d’accumuler des connaissances, mais de les intégrer en modèles cohérents et de les échanger. Il s’agit évidemment là d’une perspective à long terme qui exigera des efforts coordonnés.

Conclusion: vers un tournant humaniste en IA

Sans langage, nous n’aurions accès ni au questionnement, ni au dialogue, ni au récit. La langue est simultanément un médium de l’intelligence personnelle – il est difficile de penser sans dialogue intérieur – et de l’intelligence collective. La plupart de nos connaissances ont été accumulées et transmises par la société sous forme linguistique. Vu le rôle de la parole dans l’intelligence humaine, Il est surprenant qu’on ait espéré atteindre une intelligence artificielle générale sans disposer d’un modèle calculable du langage et de sa sémantique. La bonne nouvelle est que nous en avons finalement un. Même si l’architecture neuro-sémantique ici proposée ne débouche pas directement sur une intelligence artificielle générale, elle autorise au moins la construction d’applications capables de traiter le sens des textes ou des situations. Elle permet aussi d’envisager un marché des données privées étiquetées en IEML qui stimulerait, s’il en était besoin, le développement de l’apprentissage statistique. Elle devrait aussi supporter une mémoire publique collaborative qui serait particulièrement utile dans les domaines de la recherche scientifique, de l’éducation et de la santé.

La multiplicité des langues, des systèmes de classification, des points de vue disciplinaires et des contextes pratiques cloisonne aujourd’hui la mémoire numérique. Or la communication des modèles, la comparaison critique des points de vue et l’accumulation des connaissances sont essentiels à la cognition symbolique humaine, une cognition indissolublement personnelle et collective. L’intelligence artificielle ne pourra durablement augmenter la cognition humaine qu’à la condition d’être interopérable, cumulable, intégrable, échangeable et distribuée. C’est dire qu’on ne fera pas de progrès notable en intelligence artificielle sans viser en même temps une intelligence collective capable de se réfléchir et de se coordonner dans la mémoire mondiale. L’adoption d’une langue calculable fonctionnant comme système universel de coordonnées sémantiques – une langue facile à lire et à écrire permettant de tout dire comme de distinguer les nuances – ouvrirait de nouvelles voies à l’intelligence collective humaine, y compris sous la forme d’une interaction immersive multimédia dans le monde des idées. En ce sens, la communauté des utilisateurs d’IEML pourrait inaugurer une nouvelle époque de l’intelligence collective.

L’IA contemporaine, majoritairement statistique, a tendance à créer des situations où les données pensent à notre place et à notre insu. Par contraste, je propose de développer une IA qui aide les humains à prendre le contrôle intellectuel des données pour en extraire un sens partageable de manière durable. IEML nous permet de repenser la finalité et le mode d’action de l’IA d’un point de vue humaniste, point de vue pour qui le sens, la mémoire et la conscience personnelle doivent être traités avec le plus grand sérieux.

NOTES ET RÉFÉRENCES

Sur les origines de l’IA
L’expression “Intelligence artificielle” fut utilisée pour la première fois en 1956, lors d’une conférence du Dartmouth College à Hanover, New Hampshire. Participaient notamment à cette conférence l’informaticien et chercheur en sciences cognitives Marvin Minsky (Turing Award 1969) et l’inventeur du langage de programmation LISP John McCarthy (Turing Award 1971).

Sur l’augmentation cognitive
L’augmentation cognitive (plutôt que l’imitation de l’intelligence humaine) était l’orientation principale de nombre des pionniers de l’informatique et du Web. Voir par exemple :
– Bush, Vannevar. “As We May Think.” Atlantic Monthly, July 1945.
– Licklider, Joseph. “Man-Computer Symbiosis.” IRE Transactions on Human Factors in Electronics, 1, 1960, 4-11.
– Engelbart, Douglas. Augmenting Human Intellect. Technical Report. Stanford, CA: Stanford Research Institute, 1962.
– Berners-Lee, Tim. Weaving the Web. San Francisco: Harper, 1999.

Sur l’histoire de l’IA neuronale
Beaucoup de gens connaissent Geoffrey Hinton, Yann Le Cun et Yoshua Benjio comme les fondateurs de l’IA neuronale contemporaine. Mais l’IA neuronale commence dès les années 40 du XXe siècle. Je fournis ci-dessous une brève bibliographie.
– McCulloch, Warren, and Walter Pitts. “A Logical Calculus of Ideas Immanent in Nervous Activity.” Bulletin of Mathematical Biophysics, 5, 1943: 115-133. 
– McCulloch, Warren. Embodiments of Mind. Cambridge, MA: MIT Press, 1965.)
– Lévy, Pierre. “L’Œuvre de Warren McCulloch.” Cahiers du CREA, 7, 1986, p. 211-255.
– Frank Rosenblatt est l’inventeur du Perceptron, qui peut être considéré comme le premier système d’apprentissage automatique basé sur un réseau neuro-mimétique. Voir son livre Principles of Neurodynamics: Perceptrons and the Theory of Brain Mechanisms, publié en 1962 par Spartan Books.
– Le mémoire de doctorat de 1954 de Marvin Minsky était intitulé: “Theory of neural-analog reinforcement systems and its application to the brain-model problem.”
– Minsky critiquera le perceptron de Frank Rosenblatt dans son livre Perceptrons de 1969 (MIT Press) écrit avec Seymour Papert et poursuivra par la suite le programme de recherche de l’IA symbolique.
– Toujours de Minsky, The Society of Mind (Simon and Schuster, 1986) résume bien son approche de la cognition humaine comme une émergence à partir de l’interaction de multiples modules cognitifs aux fonctions variées.
– Foerster, Heinz von. Observing Systems: Selected Papers of Heinz von Foerster. Seaside, CA: Intersystems Publications, 1981.
– Von Fœrster était directeur du Biological Computer Laboratory. Voir Lévy, Pierre. “Analyse de contenu des travaux du Biological Computer Laboratory (BCL).” In Cahiers du CREA, 8, 1986, p. 155-191.
– McClelland, James L., David E. Rumelhart and PDP research group. Parallel Distributed Processing: Explorations in the Microstructure of Cognition. 2 vols. Cambridge, MA: MIT Press, 1986.
– Rumelhart, David E.; Hinton, Geoffrey E.; Williams, Ronald J. (9 October 1986). “Learning representations by back-propagating errors”. Nature. 323 (6088): 533–536. Hinton a été reconnu pour ses travaux pionniers par un prix Turing obtenu avec Yann LeCun et Joshua Benjio en 2018.

La critique de l’IA statistique
Ce texte reprend quelques-uns des arguments avancés par des chercheurs comme Judea Pearl, Gary Marcus et Stephen Wolfram.
– Judea Pearl, a reçu le prix Turing en 2011 pour ses travaux sur la modélisation de la causalité en IA. Il a  écrit avec Dana Mackenzie, The Book of Why, The new science of cause and effect, Basic books, 2019.
– Voir l’article séminal de Gary Marcus de 2018 “Deep learning, a critical appraisal” https://arxiv.org/pdf/1801.00631.pdf?u (Consulté le 8 août 2021)
– Voir aussi le livre de Gary Marcus, écrit avec Ernest Davis, Rebooting AI: Building Artificial Intelligence We Can Trust, Vintage, 2019.
– Stephen Wolfram est l’auteur du logiciel Mathematica et du moteur de recherche Wolfram Alpha. Voir son entretien pour Edge.org de 2016 “AI and the future of civilisation” https://www.edge.org/conversation/stephen_wolfram-ai-the-future-of-civilization Consulté le 8 août 2021.
– Outre les travaux de Judea Pearl sur l’importance de la modélisation causale en IA, rappelons les thèses du philosophe Karl Popper sur les limites du raisonnement inductif et des statistiques. Voir, en particulier, de Karl Popper, Objective Knowledge: An Evolutionary Approach. Oxford: Clarendon Press, 1972.

Sur l’IA neuronale contemporaine
– Sur BERT, voir: https://en.wikipedia.org/wiki/BERT_(language_model) Consulté le 8 août 2021.
– Voir le récent rapport du Center for Research on Foundation Models (CRFM) at the Stanford Institute for Human-Centered Artificial Intelligence (HAI), intitulé On the Opportunities and Risks of Foundation Models et qui commence ainsi: “AI is undergoing a paradigm shift with the rise of models (e.g., BERT, DALL-E, GPT-3) that are trained on broad data at scale and are adaptable to a wide range of downstream tasks.” https://arxiv.org/abs/2108.07258
– Sur Open AI https://openai.com/blog/gpt-3-apps/ et https://www.technologyreview.com/2020/08/22/1007539/gpt3-openai-language-generator-artificial-intelligence-ai-opinion/ Sites visités le 16 août 2021.

Sur l’IA symbolique contemporaine
– L’intégration des connaissances existantes dans les systèmes d’IA est un des principaux objectifs du ​​”Wolfram Language” de Stephen Wolfram. Voir https://www.wolfram.com/language/principles/ consulté le 16 août 2021.
– Sur le Web sémantique, voir le site https://www.w3.org/standards/semanticweb/# et https://en.wikipedia.org/wiki/Semantic_Web Consultés le 8 août 2021
–  Sur Wikidata: https://www.wikidata.org/wiki/Wikidata:Main_Page Consulté le 16 août 2021.
– Sur le projet Cyc de Douglas Lenat : ​​https://en.wikipedia.org/wiki/Cyc Consulté le 8 août 2021.

Sur la perspective neuro-symbolique
– “AI Research and Governance Are at a Crossroads” by Dieter Ernst. https://www.cigionline.org/articles/ai-research-and-governance-are-crossroads/ Consulté le 8 août 2021.
–  Neurosymbolic AI: The 3rd Wave, Artur d’Avila Garcez and Luıs C. Lamb, Décembre, 2020 (https://arxiv.org/pdf/2012.05876.pdf) Consulté le 8 août 2021.
– Voir le récent rapport de L’université de Stanford “100 Year Study on AI” qui identifie le courant neuro-symbolique comme une des clés de l’avancement de la discipline.  https://ai100.stanford.edu/ Consulté le 20 septembre 2021.

Sur l’interopérabilité sémantique
– Tous les éditeurs de métadonnées sémantique prétendent à l’interopérabilité, mais il s’agit généralement d’une interopérabilité des formats de fichiers, cette dernière étant effectivement assurée par les standards du Web sémantique (XML, RDF, OWL, etc.). Mais je parle dans ce texte d’interopérabilité des modèles sémantiques proprement dits (on parle de concepts: les catégories et leurs relations). Donc ne pas confondre interopérabilité sémantique et l’interopérabilité des formats. Voir sur ce point: https://pierrelevyblog.com/2021/04/03/esquisse-dun-modele-daffaire-pour-un-changement-de-civilisation/
– Si nécessaire, les modèles rédigés au moyen de l’éditeur IEML peuvent être exportés dans les formats standards de métadonnées sémantiques tels que RDF et JSON-LD.

Sur Chomsky et la syntaxe
– Chomsky, Noam. Syntaxic Structures. The Hague and Paris: Mouton, 1957.
– Chomsky, Noam, and Marcel-Paul Schützenberger. “The Algebraic Theory of Context-Free Languages.” In Computer Programming and Formal Languages. Ed. P. Braffort and D. Hirschberg. Amsterdam: North Holland, 1963. p. 118-161.
– Pour une approche plus philosophique, voir Chomsky, Noam. New Horizons in the Study of Language and Mind. Cambridge, UK: Cambridge UP, 2000.
– Voir aussi mon article sur les fondements linguistiques d’IEML.

Sur les noms propres
– J’adopte ici en gros la position de Saul Kripke, suivie par la majorité des philosophes et grammairiens. Voir, de Saul Kripke, Naming and Necessity, Oxford, Blackwell, 1980. Trad. fr. La logique des noms propres, Paris, Minuit, 1982, (trad. P. Jacob et F. Recanati).
– Voir ma récente entrée de blog à ce sujet.

Pierre Lévy sur IEML
– “Toward a Self-referential Collective Intelligence: Some Philosophical Background of the IEML Research Program.” Computational Collective Intelligence, Semantic Web, Social Networks and Multiagent Systems, ed. Ngoc Than Nguyen, Ryszard Kowalczyk and Chen Shyi-Ming, First International Conference, ICCCI, Wroclaw, Poland, Oct. 2009, proceedings, Berlin-Heidelberg-New York: Springer, 2009, p. 22-35. 
– “The IEML Research Program: From Social Computing to Reflexive Collective Intelligence.” In Information Sciences, Special issue on Collective Intelligence, ed. Epaminondas Kapetanios and Georgia Koutrika, vol. 180, no. 1, Amsterdam: Elsevier, 2 Jan. 2010, p. 71-94.
– Les considérations philosophiques et scientifiques qui m’ont mené à l’invention d’IEML ont été amplement décrites dans La Sphère sémantique. Computation, cognition, économie de l’information. Hermes-Lavoisier, Paris / Londres 2011 (400 p.). Trad. anglaise: The Semantic Sphere. Computation, cognition and information economy. Wiley, 2011. Ce livre contient une nombreuse bibliographie.
– Les principes généraux d’IEML sont résumés dans: https://intlekt.io/ieml/ (consulté le 17 août 2021).
– Sur la grammaire d’IEML, voir: https://intlekt.io/ieml-grammar/ (consulté le 17 août 2021).
– Sur le dictionnaire d’IEML, voir: https://intlekt.io/ieml-dictionary/ (consulté le 17 août 2021).
– Pour une exposition des principes linguistiques à la base d’IEML, voir: https://intlekt.io/the-linguistic-roots-of-ieml/ (consulté le 17 août 2021).

Autres références pertinentes de Pierre Lévy
L’intelligence collective, pour une anthropologie du cyberespace, La Découverte, Paris, 1994. Traduction en anglais par Robert Bonono: Collective Intelligence Perseus Books, Cambridge MA, 1997.
 – “Les systèmes à base de connaissance comme médias de transmission de l’expertise” (knowledge based systems as media for transmission of expertise), in Intellectica  (Paris) special issue on “Expertise and cognitive sciences”, ed. Violaine Prince. 1991. p. 187 to 219.
– J’ai analysé en détail le travail d’ingénierie de la connaissance sur plusieurs cas dans mon livre De la programmation considérée comme un des beaux-arts, La Découverte, Paris, 1992.

Vassili Kandinsky: Circles in a Circle

A Scientific Language

IEML is an acronym for Information Economy MetaLanguage. IEML is the result of many years of fundamental research under the direction of Pierre Lévy, fourteen years of which were funded by the Canadian federal government through the Canada Research Chair in Collective Intelligence at the University of Ottawa (2002-2016). In 2020, IEML is the only language that has the following three properties:

– it has the expressive power of a natural language;

– it has the syntax of a regular language;

– its semantics is unambiguous and computable, because it is aligned with its syntax.

In other words, it is a “well-formed symbolic system”, which comprises a bijection between a set of relations between signifieds, or meanings (a language) and a set of relations between signifiers (an algebra) and which can be manipulated by a set of symmetrical and automatic operations. 

On the basis of these properties, IEML can be used as a concept coding system that solves the problem of semantic interoperability in an original way, lays the foundations for a new generation of artificial intelligence and allows collective intelligence to be reflexive. IEML complies with Web standards and can be exported in RDF. IEML expressions are called USLs (Uniform Semantic Locators). They can be read and translated into any natural language. Semantic ontologies – sets of IEML expressions linked by a network of relations – are interoperable by design. IEML provides the coordinate system of a common knowledge base that feeds both automatic reasoning and statistical calculations. In sum, IEML fulfills the promise of the Semantic Web through its computable meaning and interoperable ontologies. IEML’s grammar consists of four layers: elements, words, sentences and texts. Examples of elements and words can be found at https://dev.intlekt.io/.

Elements

The semantic elements are the basic building blocks, or elementary concepts, from which all language expressions are composed. A dictionary of about 5000 elements translated into natural languages is given with IEML and shared among all its users. Semantic interoperability comes from the fact that everyone shares the same set of elements whose meanings are fixed. The dictionary is organized into tables and sub-tables related to the same theme and the elements are defined reciprocally through a network of explicit semantic relations. IEML allows the design of an unlimited variety of concepts from a limited number of elements. 

Exemple of an elements paradigm in the IEML dictionary

The user does not have to worry about the rules from which the elements are constructed. However, they are regularly generated from six primitive symbols forming the “layer 0” of the language, and since the generative operation is recursive, the elements are stratified on six layers above layer 0.

Words

Using the elements dictionary and grammar rules, users can freely model a field of knowledge or practice within IEML. These models can be original or translate existing classifications, ontologies or semantic metadata.

The basic unit of an IEML sentence is the word. A word is a pair composed of two small sets of elements: the radical and the inflection. The choice of radical elements is free, but inflection elements are selected from a closed list of elements tables corresponding to adverbs, prepositions, postpositions, articles, conjugations, declensions, modes, etc. (see “auxiliary morphemes” in https://dev.intlekt.io/)

Each word or sentence corresponds to a distinct concept that can be translated, according to its author’s indications and its grammatical role, as a verb (encourage), a noun (courage), an adjective (courageous) or an adverb (bravely). 

Sentences 

The words are distributed on a grammatical tree composed of a root (verbal or nominal) and eight leaves corresponding to the roles of classical grammar: subject, object, complement of time, place, etc. 

The nine grammatical roles

Nine grammatical roles

The Root of the sentence can be a process (a verb), a substance, an essence, an affirmation of existence… 

The Initiator is the subject of a process, answering the question “who?” He can also define the initial conditions, the first motor, the first cause of the concept evoked by the root.

The Interactant corresponds to the object of classical grammar. It answers the question “what”. It also plays the role of medium in the relationship between the initiator and the recipient. 

The Recipient is the beneficiary (or the victim) of a process. It answers the questions “for whom, to whom, towards whom?”. 

The Time answers the question “when?”. It indicates the moment in the past, the present or the future and gives references as to anteriority, posteriority, duration, date and frequency. 

The Place answers the question “where?”. It indicates the location, spatial distribution, pace of movement, paths, paths, spatial relationships and metaphors. 

The Intention answers the question of finality, purpose, motivation: “for what”, “to what end?”It concerns mental orientation, direction of action, pragmatic context, emotion or feeling.

The Manner answers the questions “how?” and “how much?”. It situates the root on a range of qualities or on a scale of values. It specifies quantities, gradients, measurements and sizes. It also indicates properties, genres and styles.

The Causality answers the question “why? It specifies logical, material and formal determinations. It describes causes that have not been specified by the initiator, the interactant or the recipient: media, instruments, effects, consequences. It also describes the units of measurement and methods. It may also specify rules, laws, reasons, points of view, conditions and contracts.

For example: Robert (initiator) offers (root-process) a (interactant) gift to Mary (recipient) today (time) in the garden (place), to please her (intention), with a smile (manner), for her birthday (causality). 

Junctions 

IEML allows the junction of several words in the same grammatical role. This can be a logical connection (and, or inclusive or exclusive), a comparison (same as, different from), an ordering (larger than, smaller than…), an antinomy (but, in spite of…), and so on.

Layers of complexity

Grammatical roles of a complex sentence

A word that plays one of the eight leaf roles at complexity layer 1 can play the role of secondary root at a complexity layer 2, and so on recursively up to layer 4.

Literals

IEML strictly speaking enables only general categories or concepts to be expressed. It is nevertheless possible to insert numbers, units of measurement, dates, geographical positions, proper names, etc. into a sentence, provided they are categorized in IEML. For example t.u.-t.u.-‘. [23] means ‘number: 23’. Individual names, numbers, etc. are called literals in IEML.

Texts 

Relations 

A semantic relationship is a sentence in a special format that is used to link a source node (element, word, sentence) to a target node. IEML includes a query language enabling easy programming of semantic relationships on a set of nodes. 

By design, a semantic relationship makes the following four points explicit.

1. The function that connects the source node and the target node.

2. The mathematical form of the relation: equivalence relationship, order relationship, intransitive symmetrical relationship or intransitive asymmetrical relationship.

3. The kind of context or social rule that validates the relationship: syntax, law, entertainment, science, learning, etc.

4. The content of the relationship: logical, taxonomic, mereological (whole-part relationship), temporal, spatial, quantitative, causal, or other. The relation can also concern the reading order or the anaphora.

The (hyper) textual network

An IEML text is a network of semantic relationships. This network can describe linear successions, trees, matrices, cliques, cycles and complex subnetworks of all types.

An IEML text can be considered as a theory, an ontology, or a narrative that accounts for the dataset it is used to index.

We can define a USL as an ordered (normalized) set of triples of the form : (a source node, a target node, a relationship sentence).  A set of such triples describes a semantic network or IEML text. 

The following special cases should be noted:

– A network may contain only one sentence.

– A sentence may contain only one root to the exclusion of other grammatical roles.

– A root may contain only one word (no junction).

– A word may contain only one element.

******* 

In short, IEML is a language with computable semantics that can be considered from three complementary points of view: linguistics, mathematics and computer science. Linguistically, it is a philological language, i.e. it can translate any natural language. Mathematically, it is a topos, that is, an algebraic structure (a category) in isomorphic relation with a topological space (a network of semantic relations). Finally, on the computer side, it functions as the indexing system of a virtual database and as a programming language for semantic networks.

More than 60% of the human population is connected to the Internet, most sectors of activity have switched to digital and software drives innovation. Yet Internet standards and protocols were invented at a time when less than one percent of the population was connected. It is time to use the data flows, the available computing power and the possibilities of interactive communication for human development… and to solve the serious problems we are facing. That is why I will launch soon a major international project – comparable to the construction of a cyclotron or a voyage to Mars – aiming at an augmentation of the Internet in the service of collective intelligence.

This project has several interrelated objectives: 

  • Decompartmentalize digital memory and ensure its semantic (linguistic, cultural and disciplinary) interoperability.
  • Open up indexing modes and maximize the diversity of interpretations of the digital memory.
  • Make communication between machines, but also between humans and machines, more fluid in order to enforce our collective mastery of the Internet of Things, intelligent cities, robots, autonomous vehicles, etc.
  • Establish new forms of modeling and reflexive observation of human collective intelligence on the basis of our common memory.

IEML

The technical foundation of this project is IEML (Information Economy MetaLanguage), a semantic metadata system that I invented with support from the Canadian federal government. IEML has :

  • the expressive power of a natural language, 
  • the syntax of a regular language, 
  • calculable semantics aligned with its syntax.

IEML is exported in RDF and is based on Web standards. IEML concepts are called USLs (Uniform Semantic Locators). They can be read and translated into any natural language. Semantic ontologies – sets of USLs linked by a network of relationships – are interoperable by design. IEML establishes a virtual knowledge base that feeds both automatic reasoning and statistical calculations. In short, IEML fulfills the promise of the Semantic Web through its computable meaning and interoperable ontologies.

For a short description of the IEML grammar, click here.

Intlekt

The URLs system and the http standard only become useful through a browser. Similarly, the new IEML-based semantic addressing system for the Internet requires a special application, called Intlekt, whose technical project manager is Louis van Beurden. Intlekt is a collaborative and distributed platform that supports concept editing, data curation and new forms of search, data mining and data visualization. 

Intlekt empowers the edition and publishing of semantic ontologies – sets of linked concepts – related to a field of practice or knowledge. These ontologies can be original or translate existing semantic metadata such as: thesauri, documentary languages, ontologies, SKOS taxonomies, folksonomies, sets of tags or hashtags, keywords, column and row headings, etc. Published semantic ontologies augment a dictionary of concepts, which can be considered as an open meta-ontology

Intlekt is also a data curation tool. It enables editing, indexing in IEML and publishing data collections that feed a common knowledge base. Eventually, statistical algorithms will be used to automate the semantic indexing of data.

Finally, Intlekt exploits the properties of IEML to allow new forms of search, automatic reasoning and simulation of complex systems.

Special applications can be imagined in many areas, like:

  • the preservation of cultural heritage, 
  • research in the humanities (digital humanities), 
  • education and training
  • public health, 
  • informed democratic deliberation, 
  • commercial transactions, 
  • smart contracts, 
  • the Internet of things, 
  • and so on…

And now, what?

Where do we stand on this project in the summer of 2020? After many tests over several years, IEML’s grammar has stabilized, as well as the base of morphemes of about 5000 units which enables any concept to be built at will. I tested positively the expressive possibilities of the language in several fields of humanities and earth sciences. Nevertheless, at the time of writing, the latest state of the grammar is not yet implemented. Moreover, to obtain a version of Intlekt that enables the semantic ontology editing, data curation and data mining functions described above, a team of several programmers working for one year is needed. In the coming months, the friends of IEML will be busy pursuing this critical mass. 

Come and join us!

For more information, see: https://pierrelevyblog.com/my-research-in-a-nutshell/ and https://pierrelevyblog.com/my-research-in-a-nutshell/the-basics-of-ieml/

The coronavirus pandemic has and will continue to have catastrophic effects not only in terms of physical health and mortality, but also in the areas of mental health and the economy, with social, political and cultural consequences that are difficult to calculate. Already it can be said that the scale of suffering and destruction is approaching that of a world war.

If there was still need, we are progressing in the awareness of the unity and physical continuity of a planetary human population sharing a common environment. The public space has shifted to the virtual and everyone is participating in communication through social media. Major web platforms and online services have seen a considerable increase in their use and digital communication infrastructures are at the limit of their capacity. Distance medicine, education, work and commerce have become commonplace, heralding a profound change in habits and skills, but also the possibility of limiting pollution and carbon emissions. The Internet is more than ever a part of essential services and even human rights. To provide solutions to this multifaceted crisis, new forms of collective intelligence are bypassing official institutions and national barriers, particularly in the scientific and health fields.

At the same time, conflicts of interpretation, information wars and propaganda battles are intensifying. False news – also viral – is pouring in from all sides, adding to the confusion and panic. Shameful or malicious manipulation of data accompanies ideological, cultural or national disputes in the midst of a global geopolitical reorganization. Global and local exchanges are rebalancing in favour of the latter. Political power is increasing at all levels of government with a remarkable merging of intelligence, police and medical services instrumented by digital communications and artificial intelligence. In the interests of public health and national security, the universal geolocation of individuals by mobile phone, bracelet or ring is on the horizon. Automatic identification by facial recognition or heartbeat will do the rest. 

To balance these trends, we need greater transparency of scientific, political and economic powers. The automatic analysis of data flows must become an essential skill taught in schools because it now conditions the understanding of the world. Learning and analytical resources must be shared and open to all free of charge. An international and cross-linguistic harmonization of semantic metadata systems would help to process and compare data and support more powerful forms of collective intelligence than those we know today.

With a crown of thorns on his bloody skull, humanity enters a new era.

IEML (the Information Economy Meta Language) has four main directions of research and development in 2019: in mathematics, data science, linguistics and software development. This blog entry reviews them successively.

1- A mathematical research program

I will give here a philosophical description of the structure of IEML, the purpose of the mathematical research to come being to give a formal description and to draw from this formalisation as much useful information as possible on the calculation of relationships, distances, proximities, similarities, analogies, classes and others… as well as on the complexity of these calculations. I had already produced a formalization document in 2015 with the help of Andrew Roczniak, PhD, but this document is now (2019) overtaken by the evolution of the IEML language. The Brazilian physicist Wilson Simeoni Junior has volunteered to lead this research sub-program.

IEML Topos

The “topos” is a structure that was identified by the great mathematician Alexander Grothendieck, who “is considered as the re-founder of algebraic geometry and, as such, as one of the greatest mathematicians of the 20th century” (see Wikipedia).

Without going into technical details, a topos is a bi-directional relationship between, on the one hand, an algebraic structure, usually a “category” (intuitively a group of transformations of transformation groups) and, on the other hand, a spatial structure, which is geometric or topological. 

In IEML, thanks to a normalization of the notation, each expression of the language corresponds to an algebraic variable and only one. Symmetrically, each algebraic variable corresponds to one linguistic expression and only one. 

Topologically, each variable in IEML algebra (i.e. each expression of the language) corresponds to a “point”. But these points are arranged in different nested recursive complexity scales: primitive variables, morphemes of different layers, characters, words, sentences, super-phrases and texts. However, from the level of the morpheme, the internal structure of each point – which comes from the function(s) that generated the point – automatically determines all the semantic relationships that this point has with the other points, and these relationships are modelled as connections. There are obviously a large number of connection types, some very general (is contained in, has an intersection with, has an analogy with…) others more precise (is an instrument of, contradicts X, is logically compatible with, etc.).

The topos that match all the expressions of the IEML language with all the semantic relationships between its expressions is called “The Semantic Sphere”.

Algebraic structure of IEML

In the case of IEML, the algebraic structure is reduced to 

  • 1. Six primitive variables 
  • 2. A non-commutative multiplication with three variables (substance, attribute and mode). The IEML multiplication is isomorphic to the triplet ” departure vertex, arrival vertex, edge ” which is used to describe the graphs.
  • 3. A commutative addition that creates a set of objects.

This algebraic structure is used to construct the following functions and levels of variables…

1. Functions using primitive variables, called “morpheme paradigms”, have as inputs morphemes at layer n and as outputs morphemes at layer n+1. Morpheme paradigms include additions, multiplications, constants and variables and are visually presented in the form of tables in which rows and columns correspond to certain constants.

2. “Character paradigms” are complex additive functions that take morphemes as inputs and characters as outputs. Character paradigms include a group of constant morphemes and several groups of variables. A character is composed of 1 to 5 morphemes arranged in IEML alphabetical order. (Characters may not include more than five morphemes for cognitive management reasons).

3. IEML characters are assembled into words (a substance character, an attribute character, a mode character) by means of a multiplicative function called a “word paradigm”. A word paradigm intersects a series of characters in substance and a series of characters in attribute. The modes are chosen from predefined auxiliary character paradigms, depending on whether the word is a noun, a verb or an auxiliary. Words express subjects, keywords or hashtags. A word can be composed of only one character.

4. Sentence building functions assemble words by means of multiplication and addition, with the necessary constraints to obtain grammatical trees. Mode words describe the grammatical/semantic relationships between substance words (roots) and attribute words (leaves). Sentences express facts, proposals or events; they can take on different pragmatic and logical values.

5. Super-sentences are generated by means of multiplication and addition of sentences, with constraints to obtain grammatical trees. Mode sentences express relationships between substance sentences and attribute sentences. Super-sentences express hypotheses, theories or narratives.

6. A USL (Uniform Semantic Locator) or IEML text is an addition (a set) of words, sentences and super-sentences. 

Topological structure of IEML: a semantic rhizome

Static

The philosophical notion of rhizome (a term borrowed from botany) was developed on a philosophical level by Deleuze and Guattari in the preface to Mille Plateaux (Minuit 1980). In this Deleuzo-Guattarian lineage, by rhizome I mean here a complex graph whose points or “vertices” are organized into several levels of complexity (see the algebraic structure) and whose connections intersect several regular structures such as series, tree, matrix and clique. In particular, it should be noted that some structures of the IEML rhizome combine hierarchical or genealogical relationships (in trees) with transversal or horizontal relationships between “leaves” at the same level, which therefore do not respect the “hierarchical ladder”. 

Dynamic

We can distinguish the abstract, or virtual, rhizomatic grid drawn by the grammar of the language (the sphere to be dug) and the actualisation of points and relationships by the users of the language (the dug sphere of chambers and galleries).  Characters, words, sentences, etc. are all chambers in the centre of a star of paths, and the generating functions establish galleries of “rhizomatic” relationships between them, as many paths for exploring the chambers and their contents. It is therefore the users, by creating their lexicons and using them to index their data, communicate and present themselves, who shape and grow the rhizome…

Depending on whether circuits are more or less used, on the quantity of data or on the strength of interactions, the rhizome undergoes – in addition to its topological transformations – various types of quantitative or metric transformations. 

* The point to remember is that IEML is a language with calculable semantics because it is also an algebra (in the broad sense) and a complex topological space. 

* In the long term, IEML will be able to serve as a semantic coordinate system for the information world at large.

2 A research program in data science

The person in charge of the data science research sub-program is the software engineer (Eng. ENSIMAG, France) Louis van Beurden, who holds also a master’s degree in data science and machine translation from the University of Montréal, Canada. Louis is planning to complete a PhD in computer science in order to test the hypothesis that, from a data science perspective, a semantic metadata system in IEML is more efficient than a semantic metadata system in natural language and phonetic writing. This doctoral research will make it possible to implement phases A and B of the program below and to carry out our first experiment.

Background information

The basic cycle in data science can be schematized according to the following loop:

  • 1. selection of raw data,
  • 2. pre-processing, i.e. cleaning data and metadata imposition (cataloguing and categorization) to facilitate the exploitation of the results by human users,
  • 3. statistical processing,
  • 4. visual and interactive presentation of results,
  • 5. exploitation of the results by human users (interpretation, storytelling) and feedback on steps 1, 2, 3

Biases or poor quality of results may have several causes, but often come from poor pre-treatment. According to the old computer adage “garbage in, garbage out“, it is the professional responsibility of the data-scientists to ensure the quality of the input data and therefore not to neglect the pre-processing phase where this data is organized using metadata.

Two types of metadata can be distinguished: 1) semantic metadata, which describes the content of documents or datasets, and 2) ordinary metadata, which describes authors, creation dates, file types, etc. Let us call “semantic pre-processing” the imposition of semantic metadata on data.

Hypothesis

Since IEML is a univocal language and the semantic relationships between morphemes, words, sentences, etc. are mathematically computable, we assume that a semantic metadata system in IEML is more efficient than a semantic metadata system in natural language and phonetic writing. Of course, the efficiency in question is related to a particular task: search, data analysis, knowledge extraction from data, machine learning, etc.

In other words, compared to a “tokenization” of semantic metadata in phonetic writing noting a natural language, a “tokenization” of semantic metadata in IEML would ensure better processing, better presentation of results to the user and better exploitation of results. In addition, semantic metadata in IEML would allow datasets that use different languages, classification systems or ontologies to be de-compartmentalized, merged and compared.

Design of the first experience

The ideal way to do an experiment is to consider a multi-variable system and transform only one of the system variables, all other things being equal. In our case, it is only the semantic metadata system that must vary. This will make it easy to compare the system’s performance with one (phonetic tokens) or the other (semantic tokens) of the semantic metadata systems.

  • – The dataset of our first experience encompasses all the articles of the Sens Public scientific journal.
  • – Our ordinary metadata are the author, publication date, etc.
  • – Our semantic metadata describe the content of articles.
  •     – In phonetic tokens, using RAMEAU categories, keywords and summaries,
  •     – In IEML tokens by translating phonetic tokens.
  • – Our processes are “big data” algorithms traditionally used in natural language processing 
  •     – An algorithm for calculating the co-occurrences of keywords.
  •     – A TF-IDF (Term Frequency / Inverse Document Frequency) algorithm that works from a word / document matrix.
  •     – A clustering algorithm based on “word embeddings” of keywords in articles (documents are represented by vectors, in a space with as many dimensions as words).
  • – A user interface will offer a certain way to access the database. This interface will be obviously adapted to the user’s task (which remains to be chosen, but could be of the “data analytics” type).
  • Result 1 corresponds to the execution of the “machine task”, i.e. the establishment of a connection network on the articles (relationships, proximities, groupings, etc.). We’ll have to compare….
  •     – result 1.1 based on the use of phonetic tokens with 
  •     – result 1.2 based on the use of IEML tokens.
  • Result 2 corresponds to the execution of the selected user-task (data analytics, navigation, search, etc.). We’ll have to compare….
  •     – result 2.1, based on the use of phonetic tokens, with 
  •     – result 2.2, based on the use of IEML tokens.

Step A: First indexing of a database in IEML

Reminder: the data are the articles of the scientific journal, the semantic metadata are the categories, keywords and summaries of the articles. From the categories, keywords and article summaries, a glossary of the knowledge area covered by the journal is created, or a sub-domain if it turns out that the task is too difficult. It should be noted that in 2019 we do not yet have the software tools to create IEML sentences and super-phrases that allow us to express facts, proposals, theories, narratives, hypotheses, etc. Phrases and super-phrases, perhaps accessible in a year or two, will therefore have to wait for a later phase of the research.

The creation of the glossary will be the work of a project community, linked to the editors of Sens-Public magazine and the Canada Research Chair in Digital Writing (led by Prof. Marcello Vitali-Rosati) at the Université de Montréal (Digital Humanities). Pierre Lévy will accompany this community and help it to identify the constants and variables of its lexicon. One of the auxiliary goals of the research is to verify whether motivated communities can appropriate IEML to categorize their data. Once we are satisfied with the IEML indexing of the article database, we will proceed to the next step.

Step B: First experimental test

  • 1. The test is determined to measure the difference between results based on phonetic tokens and results based on IEML tokens. 
  • 2. All data processing operations are carried out on the data.
  • 3. The results (machine tasks and user tasks) are compared with both types of tokens.

The experiment can eventually be repeated iteratively with minor modifications until satisfactory results are achieved.

If the hypothesis is confirmed, we proceed to the next step

Step C: Towards an automation of semantic pre-processing in IEML.

If the superior efficiency of IEML tokens for semantic metadata is demonstrated, then there will be a strong interest in maximizing the automation of IEML semantic pre-processing

The algorithms used in our experiment are themselves powerful tools for data pre-processing, they can be used, according to methods to be developed, to partially automate semantic indexing in IEML. The “word embeddings” will make it possible to study how IEML words are correlated with the natural language lexical statistics of the articles and to detect anomalies. For example, we will check if similar USLs (a USL is an IEML text) point to very different texts or if very different texts have similar USLs. 

Finally, methods will be developed to use deep learning algorithms to automatically index datasets in IEML.

Step D: Research and development perspective in Semantic Machine Learning

If step C provides the expected results, i.e. methods using AI to automate the indexing of data in IEML, then big data indexed in IEML will be available.  As progress will be made, semantic metadata may become increasingly similar to textual data (summary of sections, paragraphs, sentences, etc.) until translation into IEML is achieved, which remains a distant objective.

The data indexed in IEML could then be used to train artificial intelligence algorithms. The hypothesis that machines learn more easily when data is categorized in IEML could easily be validated by experiments of the same type as described above, by comparing the results obtained from training data indexed in IEML and the results obtained from the same data indexed in natural languages.

This last step paves the way for a better integration of statistical AI and symbolic AI (based on facts and rules, which can be expressed in IEML).

3 A research program in linguistics, humanities and social sciences

Introduction

The semiotic and linguistic development program has two interdependent components:

1. The development of the IEML metalanguage

2. The development of translation systems and bridges between IEML and other sign systems, in particular… 

  •     – natural languages,
  •     – logical formalisms,
  •     – pragmatic “language games” and games in general,
  •     – iconic languages,
  •     – artistic languages, etc.

This research and development agenda, particularly in its linguistic dimension, is important for the digital humanities. Indeed, IEML can serve as a system of semantic coordinates of the cultural universe, thus allowing the humanities to cross a threshold of scientific maturity that would bring their epistemological status closer to that of the natural sciences. Using IEML to index data and to formulate assumptions would result in….

  • (1) a de-silo of databases used by researchers in the social sciences and humanities, which would allow for the sharing and comparison of categorization systems and interpretive assumptions;
  • (2) an improved analysis of data.
  • (3) The ultimate perspective, set out in the article “The Role of the Digital Humanities in the New Political Space” (http://sens-public.org/article1369.html in French), is to aim for a reflective collective intelligence of the social sciences and humanities research community. 

But IEML’s research program in the perspective of the digital humanities – as well as its research program in data science – requires a living and dynamic semiotic and linguistic development program, some aspects of which I will outline here.

IEML and the Meaning-Text Theory

IEML’s linguistic research program is very much based on the Meaning-Text theory developed by Igor Melchuk and his school. “The main principle of this theory is to develop formal and descriptive representations of natural languages that can serve as a reliable and convenient basis for the construction of Meaning-Text models, descriptions that can be adapted to all languages, and therefore universal. ”(Excerpt translated from the Wikipedia article on Igor Melchuk). Dictionaries developed by linguists in this field connect words according to universal “lexical functions” identified through the analysis of many languages. These lexical functions have been formally transposed into the very structure of IEML (See the IEML Glossary Creation Guide) so that the IEML dictionary can be organized by the same tools (e.g. Spiderlex) as those of the Meaning-Text Theory research network. Conversely, IEML could be used as a pivot language – or concept description language – *between* the natural language dictionaries developed by the network of researchers skilled in Meaning-Text theory.

Construction of specialized lexicons in the humanities and social sciences

A significant part of the IEML lexicon will be produced by communities having decided to use IEML to mark out their particular areas of knowledge, competence or interaction. Our research in specialized lexicon construction aims to develop the best methods to help expert communities produce IEML lexicons. One of the approaches consists in identifying the “conceptual skeleton” of a domain, namely its main constants in terms of character paradigms and word paradigms. 

The first experimentation of this type of collaborative construction of specialized lexicons by experts will be conducted by Pierre Lévy in collaboration with the editorial team of the Sens Public scientific journal and the Canada Research Chair in Digital Textualities at the University of Montréal (led by Prof. Marcello Vitali-Rosati). Based on a determination of their economic and social importance, other specialized glossaries can be constructed, for example on the theme of professional skills, e-learning resources, public health prevention, etc.

Ultimately, the “digital humanities” branch of IEML will need to collaboratively develop a conceptual lexicon of the humanities to be used for the indexation of books and articles, but also chapters, sections and comments in documents. The same glossary should also facilitate data navigation and analysis. There is a whole program of development in digital library science here. I would particularly like to focus on the human sciences because the natural sciences have already developed a formal vocabulary that is already consensual.

Construction of logical, pragmatic and narrative character-tools

When we’ll have a sentence and super-phrase editor, it is planned to establish a correspondence between IEML – on the one hand – and propositional calculus and first order logics – on the other hand –. This will be done by specifying special character-tools to implement logical functions. Particular attention will be paid to formalizing the definition of rules and the declaration that “facts” are true in IEML. It should be noted in passing that, in IEML, grammatical expressions represent classes, sets or categories, but that logical individuals (proper names, numbers, etc.) or instances of classes are represented by “literals” expressed in ordinary characters (phonetic alphabets, Chinese characters, Arabic numbers, URLs, etc.).

In anticipation of practical use in communication, games, commerce, law (smart contracts), chatbots, robots, the Internet of Things, etc., we will develop a range of character-tools with illocutionary force such as “I offer”, “I buy”, “I quote”, “I give an instruction”, etc.

Finally, we will making it easier for authors of super-sentences by developing a range of character-tools implementing “narrative functions”.

4 A software development program

A software environment for the development and public use of the IEML language

Logically, the first multi-user IEML application will be dedicated to the development of the language itself. This application is composed of the following three web modules.

  • 1. A morpheme editor that also allows you to navigate in the morphemes database, or “dictionary”.
  • 2. A character and word editor that also allows navigation in the “lexicon”.
  • 3. A navigation and reading tool in the IEML library as a whole, or “IEML database” that brings together the dictionary and lexicon, with translations, synonyms and comments in French and English for the moment.

The IEML database is a “Git” database and is currently hosted by GitHub. Indeed, a Git database makes it possible to record successive versions of the language, as well as to monitor and model its growth. It also allows large-scale collaboration among teams capable of developing specific branches of the lexicon independently and then integrating them into the main branch after discussion, as is done in the collaborative development of large software projects. As soon as a sub-lexicon is integrated into the main branch of the Git database, it becomes a “common” usable by everyone (according to the latest General Public License version.

Morpheme and word editors are actually “Git clients” that feed the IEML database. A first version of this collaborative read-write environment should be available in the fall of 2019 and then tested by real users: the editors of the Scientific Journal “Sens Public” as well as other participants in the University of Montréal’s IEML seminar.

The following versions of the IEML read/write environment should allow the editing of sentences and texts as well as literals that are logical individuals not translated into IEML, such as proper names, numbers, URLs, etc.

A social medium for collaborative knowledge management

A large number of applications using IEML can be considered, both commercial and non-commercial. Among all these applications, one of them seems to be particularly aligned with the public interest: a social medium dedicated to collaborative knowledge and skills management. This new “place of knowledge” could allow the online convergence of the missions of… 

  • – museums and libraries, 
  • – schools and universities, 
  • – companies and administrations (with regard to their knowledge creation and management dimension), 
  • – smart cities, employment agencies, civil society networks, NGO, associations, etc.

According to its general philosophy, such a social medium should…

  • – be supported by an intrinsically distributed platform, 
  • – have the simplicity – or the economy of means – of Twitter,
  • – ensure the sovereignty of users over their data,
  • – promote collaborative processes.

The main functions performed by this social medium would be:

  • – data curation (reference and categorization of web pages, edition of resource collections), 
  • – teaching offers and learning demands,
  • – offers and demands for skills, or employment market.

IEML would serve as a common language for

  • – data categorization, 
  • – description of the knowledge and skills, 
  • – the expression of acts within the social medium (supply, demand, consent, publish, etc.)
  • – addressing users through their knowledge and skills.

Three levels of meaning would thus be formalized in this medium.

  • (1) The linguistic level in IEML  – including lexical and narrative functions – formalizes what is spoken about (lexicon) and what is said (sentences and super-phrases).
  • – (2) The logical – or referential – level adds to the linguistic level… 
  •     – logical functions (first order logic and propositional logic) expressed in IEML using logical character-tools,
  •     – the ability of pointing to references (literals, document URLs, datasets, etc.),
  •     – the means to express facts and rules in IEML and thus to feed inference engines.
  • – (3) The pragmatic level adds illocutionary functions and users to the linguistic and logical levels.
  •     – Illocutionary functions (thanks to pragmatic character-tools) allow the expression of conventional acts and rules (such as “game” rules). 
  •     – The pragmatic level obviously requires the consideration of players or users, as well as user groups.
  •     – It should be noted that there is no formal difference between logical inference and pragmatic inference but only a difference in use, one aiming at the truth of propositions according to referred states of things, the other calculating the rights, obligations, gains, etc. of users according to their actions and the rules of the games they play.

The semantic profiles of users and datasets will be arranged according to the three levels that have just been explained. The “place of knowledge” could be enhanced by the use of tokens or crypto-currencies to reward participation in collective intelligence. If successful, this type of medium could be generalized to other areas such as health, democratic governance, trade, etc.

I put forward in this paper a vision for a new generation of cloud-based public communication service designed to foster reflexive collective intelligence. I begin with a description of the current situation, including the huge power and social shortcomings of platforms like Google, Apple, Facebook, Amazon, Microsoft, Alibaba, Baidu, etc. Contrasting with the practice of these tech giants, I reassert the values that are direly needed at the foundation of any future global public sphere: openness, transparency and commonality. But such ethical and practical guidelines are probably not powerful enough to help us crossing a new threshold in collective intelligence. Only a disruptive innovation in cognitive computing will do the trick. That’s why I introduce “deep meaning” a new research program in artificial intelligence, based on the Information Economy  MetaLanguage (IEML). I conclude this paper by evoking possible bootstrapping scenarii for the new public platform.

The rise of platforms

At the end of the 20th century, one percent of the human population was connected to the Internet. In 2017, more than half the population is connected. Most of the users interact in social media, search information, buy products and services online. But despite the ongoing success of digital communication, there is a growing dissatisfaction about the big tech companies – the “Silicon Valley” – who dominate the new communication environment.

The big techs are the most valued companies in the world and the massive amount of data that they possess is considered the most precious good of our time. Silicon Valley owns the big computers: the network of physical centers where our personal and business data are stored and processed. Their income comes from their economic exploitation of our data for marketing purposes and from their sales of hardware, software or services. But they also derive considerable power from the knowledge of markets and public opinions that stems from their information control.

The big cloud companies master new computing techniques mimicking neurons when they learn a new behavior. These programs are marketed as deep learning or artificial intelligence even if they have no cognitive autonomy and need some intense training by humans before becoming useful. Despite their well known limitations, machine learning algorithms have effectively augmented the abilities of digital systems. Deep learning is now used in every economic sector. Chips specialized in deep learning are found in big data centers, smartphones, robots and autonomous vehicles. As Vladimir Putin rightly told young Russians in his speech for the first day of school in fall 2017: “Whoever becomes the leader in this sphere [of artificial intelligence] will become the ruler of the world”.

The tech giants control huge business ecosystems beyond their official legal borders and they can ruin or buy competitors. Unfortunately, the big tech rivalry prevents a real interoperability between cloud services, even if such interoperability would be in the interest of the general public and of many smaller businesses. As if their technical and economic powers were not enough, the big tech are now playing into the courts of governments. Facebook warrants our identity and warns our family and friends that we are safe when a terrorist attack or a natural disaster occurs. Mark Zuckerberg states that one of Facebook’s mission is to insure that the electoral process is fair and open in democratic countries. Google Earth and Google Street View are now used by several municipal instances and governments as their primary source of information for cadastral plans and other geographical or geospatial services. Twitter became an official global political, diplomatic and news service. Microsoft sells its digital infrastructure to public schools. The kingdom of Denmark opened an official embassy in Silicon Valley. Cryptocurrencies independent from nation states (like Bitcoin) are becoming increasingly popular. Blockchain-based smart contracts (powered by Ethereum) bypass state authentication and traditional paper bureaucracies. Some traditional functions of government are taken over by private technological ventures.

This should not come as a surprise. The practice of writing in ancient palace-temples gave birth to government as a separate entity. Alphabet and paper allowed the emergence of merchant city-states and the expansion of literate empires. The printing press, industrial economy, motorized transportation and electronic media sustained nation-states. The digital revolution will foster new forms of government. Today, we discuss political problems in a global public space taking advantage of the web and social media and the majority of humans live in interconnected cities and metropoles. Each urban node wants to be an accelerator of collective intelligence, a smart city. We need to think about public services in a new way. Schools, universities, public health institutions, mail services, archives, public libraries and museums should take full advantage of the internet and de-silo their datasets. But we should go further. Are current platforms doing their best to enhance collective intelligence and human development? How about giving back to the general population the data produced in social media and other cloud services, instead of just monetizing it for marketing purposes ? How about giving to the people access to cognitive powers unleashed by an ubiquitous algorithmic medium?

Information wants to be open, transparent and common

We need a new kind of public sphere: a platform in the cloud where data and metadata would be our common good, dedicated to the recording and collaborative exploitation of memory in the service of our collective intelligence. The core values orienting the construction of this new public sphere should be: openness, transparency and commonality

Firstly openness has already been experimented in the scientific community, the free software movement, the creative commons licensing, Wikipedia and many more endeavors. It has been adopted by several big industries and governments. “Open by default” will soon be the new normal. Openness is on the rise because it maximizes the improvement of goods and services, fosters trust and supports collaborative engagement. It can be applied to data formats, operating systems, abstract models, algorithms and even hardware. Openness applies also to taxonomies, ontologies, search architectures, etc. A new open public space should encourage all participants to create, comment, categorize, assess and analyze its content.

Then, transparency is the very ground for trust and the precondition of an authentic dialogue. Data and people (including the administrators of a platform), should be traceable and audit-able. Transparency should be reciprocal, without distinction between the rulers and the ruled. Such transparency will ultimately be the basis for reflexive collective intelligence, allowing teams and communities of any size to observe and compare their cognitive activity

Commonality means that people will not have to pay to get access to this new public sphere: all will be free and public property. Commonality means also transversality: de-silo and cross-pollination. Smart communities will interconnect and recombine all kind of useful information: open archives of libraries and museums, free academic publications, shared learning resources, knowledge management repositories, open-source intelligence datasets, news, public legal databases…

From deep learning to deep meaning

This new public platform will be based on the web and its open standards like http, URL, html, etc. Like all current platforms, it will take advantage of distributed computing in the cloud and it will use “deep learning”: an artificial intelligence technology that employs specialized chips and algorithms that roughly mimic the learning process of neurons. Finally, to be completely up to date, the next public platform will enable blockchain-based payments, transactions, contracts and secure records

If a public platform offers the same technologies as the big tech (cloud, deep learning, blockchain), with the sole difference of openness, transparency and commonality, it may prove insufficient to foster a swift adoption, as is demonstrated by the relative failures of Diaspora (open Facebook) and Mastodon (open Twitter). Such a project may only succeed if it comes up with some technical advantage compared to the existing commercial platforms. Moreover, this technical advantage should have appealing political and philosophical dimensions.

No one really fancies the dream of autonomous machines, specially considering the current limitations of artificial intelligence. Instead, we want an artificial intelligence designed for the augmentation of human personal and collective intellect. That’s why, in addition to the current state of the art, the new platform will integrate the brand new deep meaning technology. Deep meaning will expand the actual reach of artificial intelligence, improve the user experience of big data analytics and allow the reflexivity of personal and collective intelligence.

Language as a platform

In a nutshell, deep learning models neurons and deep meaning models language. In order to augment the human intellect, we need both! Right now deep learning is based on neural networks simulation. It is enough to model roughly animal cognition (every animal species has neurons) but it is not refined enough to model human cognition. The difference between animal cognition and human cognition is the reflexive thinking that comes from language, which adds a layer of semantic addressing on top of neural connectivity. Speech production and understanding is an innate property of individual human brains. But as humanity is a social species, language is a property of human societies. Languages are conventional, shared by members of the same culture and learned by social contact. In human cognition, the categories that organize perception, action, memory and learning are expressed linguistically so they may be reflected upon and shared in conversations. A language works like the semantic addressing system of a social virtual database.

But there is a problem with natural languages (english, french, arabic, etc.), they are irregular and do not lend themselves easily to machine understanding or machine translation. The current trend in natural language processing, an important field of artificial intelligence, is to use statistical algorithms and deep learning methods to understand and produce linguistic data. But instead of using statistics, deep meaning adopts a regular and computable metalanguage. I have designed IEML (Information Economy MetaLanguage) from the beginning to optimize semantic computing. IEML words are built from six primitive symbols and two operations: addition and multiplication. The semantic relations between IEML words follow the lines of their generative operations. The total number of words do not exceed 10 000. From its dictionary, the generative grammar of IEML allows the construction of sentences at three layers of complexity: topics are made of words, phrases (facts, events) are made of topics and super-phrases (theories, narratives) are made of phrases. The higher meaning unit, or text, is a unique set of sentences. Deep meaning technology uses IEML as the semantic addressing system of a social database.

Given large datasets, deep meaning allows the automatic computing of semantic relations between data, semantic analysis and semantic visualizations. This new technology fosters semantic interoperability: it decompartmentalizes tags, folksonomies, taxonomies, ontologies and languages. When on line communities categorize, assess and exchange semantic data, they generate explorable ecosystems of ideas that represent their collective intelligence. Take note that the vision of collective intelligence proposed here is distinct from the “wisdom of the crowd” model, that assumes independent agents and excludes dialogue and reflexivity. Just the opposite : deep meaning was designed from the beginning to nurture dialogue and reflexivity.

The main functions of the new public sphere

deepmeaning

In the new public sphere, every netizen will act as an author, editor, artist, curator, critique, messenger, contractor and gamer. The next platform weaves five functions together: curation, creation, communication, transaction and immersion.

By curation I mean the collaborative creation, edition, analysis, synthesis, visualization, explanation and publication of datasets. People posting, liking and commenting content on social media are already doing data curation, in a primitive, simple way. Active professionals in the fields of heritage preservation (library, museums), digital humanities, education, knowledge management, data-driven journalism or open-source intelligence practice data curation in a more systematic and mindful manner. The new platform will offer a consistent service of collaborative data curation empowered by a common semantic addressing system.

Augmented by deep meaning technology, our public sphere will include a semantic metadata editor applicable to any document format. It will work as a registration system for the works of the mind. Communication will be ensured by a global Twitter-like public posting system. But instead of the current hashtags that are mere sequences of characters, the new semantic tags will self-translate in all natural languages and interconnect by conceptual proximity. The blockchain layer will allow any transaction to be recorded. The platform will remunerate authors and curators in collective intelligence coins, according to the public engagement generated by their work. The new public sphere will be grounded in the internet of things, smart cities, ambient intelligence and augmented reality. People will control their environment and communicate with sensors, software agents and bots of all kinds in the same immersive semantic space. Virtual worlds will simulate the collective intelligence of teams, networks and cities.

Bootstrapping

This IEML-based platform has been developed between 2002 and 2017 at the University of Ottawa. A prototype is currently in a pre-alpha version, featuring the curation functionality. An alpha version will be demonstrated in the summer of 2018. How to bridge the gap from the fundamental research to the full scale industrial platform? Such endeavor will be much less expensive than the conquest of space and could bring a tremendous augmentation of human collective intelligence. Even if the network effect applies obviously to the new public space, small communities of pioneers will benefit immediately from its early release. On the humanistic side, I have already mentioned museums and libraries, researchers in humanities and social science, collaborative learning networks, data-oriented journalists, knowledge management and business intelligence professionals, etc. On the engineering side, deep meaning opens a new sub-field of artificial intelligence that will enhance current techniques of big data analytics, machine learning, natural language processing, internet of things, augmented reality and other immersive interfaces. Because it is open source by design, the development of the new technology can be crowdsourced and shared easily among many different actors.

Let’s draw a distinction between the new public sphere, including its semantic coordinate system, and the commercial platforms that will give access to it. This distinction being made, we can imagine a consortium of big tech companies, universities and governments supporting the development of the global public service of the future. We may also imagine one of the big techs taking the lead to associate its name to the new platform and developing some hardware specialized in deep meaning. Another scenario is the foundation of a company that will ensure the construction and maintenance of the new platform as a free public service while sustaining itself by offering semantic services: research, consulting, design and training. In any case, a new international school must be established around a virtual dockyard where trainees and trainers build and improve progressively the semantic coordinate system and other basic models of the new platform. Students from various organizations and backgrounds will gain experience in the field of deep meaning and will disseminate the acquired knowledge back into their communities.

Emission de radio (Suisse romande), 25 minutes en français.

Sémantique numérique et réseaux sociaux. Vers un service public planétaire, 1h en français

You-Tube Video (in english) 1h

 

 

Abstract

IEML is an artificial language that allows the automatic computing of (a) the semantic relationships internal to its texts and of (b) the semantic relationships between its texts. Such an innovation could have a positive impact on the development of human collective intelligence. While we are currently limited to logical and statistical analytics, semantic coding could allow large scale computing on the meaning of data, provided that these data are categorized in IEML. Moreover “big data” algorithms are currently monopolized by big companies and big governemnts. But according to the perspective adopted here, the algorithmic tools of the future will put data-anaytics, machine learning and reflexive collective intelligence in the hands of the majority of Internet users.
I will first describe the main components of an algorithm (code, operators, containers, instructions), then I will show that the growth of the algorithmic medium has been shaped by innovations in coding and containers addressing. The current limitations of the web (absence of semantic interoperability and statistical positivism) could be overcomed by the invention of a new coding system aimed at making the meaning computable. Finally I will describe the cognitive gains that can be secured from this innovation.

This paper has been published by Spanda Journal special issue on “Creativity & Collective Enlightenment”,  VI, 2, December 2015, p. 59-66

Our communications—transmission and reception of data—are based on an increasingly complex infrastructure for the automatic manipulation of symbols, which I call the algorithmic medium because it automates the transformation of data, and not only their conservation, reproduction and dissemination (as with previous media). Both our data-centric society and the algorithmic medium that provides its tools are still at their tentative beginnings. Although it is still hard to imagine today, a huge space will open up for the transformation and analysis of the deluge of data we produce daily. But our minds are still fascinated by the Internet’s power of dissemination of messages, which has almost reached its maximum.
In the vanguard of the new algorithmic episteme, IEML (or any other system that has the same properties) will democratize the categorization and automatic analysis of the ocean of data. The use of IEML to categorize data will create a techno-social environment that is even more favourable for collaborative learning and the distributed production of knowledge. In so doing, it will contribute to the emergence of the algorithmic medium of the future and reflect collective intelligence in the form of ecosystems of ideas.
This text begins by analyzing the structure and functioning of algorithms and shows that the major stages in the evolution of the new medium correspond to the appearance of new systems for encoding and addressing data: the Internet is a universal addressing system for computers and the Web, a universal addressing system for data. However, the Web, in 2016, has many limitations. Levels of digital literacy are still low. Interoperability and semantic transparency are sorely lacking. The majority of its users see the Web only as a big multimedia library or a means of communication, and pay no attention to its capacities for data transformation and analysis. As for those concerned with the processing of big data, they are hindered by statistical positivism. In providing a universal addressing system for concepts, IEML takes a decisive step toward the algorithmic medium of the future. The ecosystems of ideas based on this metalanguage will give rise to cognitive augmentations that are even more powerful than those we already enjoy.

What is an algorithm?

To help understand the nature of the new medium and its evolution, let us represent as clearly as possible what an algorithm is and how it functions. In simplified explanations of programming, the algorithm is often reduced to a series of instructions or a “recipe.” But no series of instructions can play its role without the three following elements: first, an adequate encoding of the data; second, a well-defined set of reified operators or functions that act as black boxes; third, a system of precisely addressed containers capable of recording initial data, intermediate results and the end result. The rules—or instructions—have no meaning except in relation to the code, the operators and the memory addresses.
I will now detail these aspects of the algorithm and use that analysis to periodize the evolution of the algorithmic medium. We will see that the major stages in the growth of this medium are precisely related to the appearance of new systems of addressing and encoding, both for the containers of data and for the operators. Based on IEML, the coming stage of development of the algorithmic medium will provide simultaneously a new type of encoding (semantic encoding) and a new system of virtual containers (semantic addressing).

Encoding of data

For automatic processing, data must first be encoded appropriately and uniformly. This involves not only binary encoding (zero and one), but more specialized types of encoding such as encoding of numbers (in base two, eight, ten, sixteen, etc.), that of characters used in writing, that of images (pixels), that of sounds (sampling), and so on.

Operators

We must then imagine a set of tools or specialized micro-machines for carrying out certain tasks on the data. Let us call these specialized tools “operators.” The operators are precisely identified, and they act in a determined or mechanical way, always the same way. There obviously has to be a correspondence or a match between the encoding of the data and the functioning of the operators.
The operators were first identified insider computers: they are the elementary electronic circuits that make up processors. But we can consider any processor of data—however complex it is—as a “black box” serving as a macro-operator. Thus the protocol of the Internet, in addressing the computers in the network, at the same time set up a universal addressing system for operators.

Containers

In addition to a code for the data and a set of operators, we have to imagine a storehouse of data whose basic boxes or “containers” are completely addressed: a logical system of recording with a smooth surface for writing, erasing and reading. It is clear that the encoding of data, the operations applied to them and the mode of recording them—and therefore their addressing—must be harmonized to optimize processing.
The first addressing system of the containers is internal to computers, and it is therefore managed by the various operating systems (for example, UNIX, Windows, Apple OS, etc.). But at the beginning of the 1990s, a universal addressing system for containers was established above that layer of internal addressing: the URLs of the World Wide Web.

Instructions

The fourth and last aspect of an algorithm is an ordered set of rules—or a control mechanism—that organizes the recursive circulation of data between the containers and the operators. The circulation is initiated by a data flow that goes from containers to the appropriate operators and then directs the results of the operations to precisely addressed containers. A set of tests (if . . . , then . . .) determines the choice of containers from which the data to be processed are drawn, the choice of operators and the choice of containers in which the results are recorded. The circulation of data ends when a test has determined that processing is complete. At that point, the result of the processing—a set of encoded data—is located at a precise address in the system of containers.

The growth of the new medium

To shape the future development of the algorithmic medium, we have to first look at its historical evolution.

Automatic calculation (1940-1970)

From when can we date the advent of the algorithmic medium? We might be tempted to give its date of birth as 1937, since it was in that year that Alan Turing (1912-1954) published his famous article introducing the concept of the universal machine, that is, the formal structure of a computer. The article represents calculable functions as programs of the universal machine, that is, essentially, algorithms. We could also choose 1945, because in June of that year, John von Neumann (1903-1957) published his “First draft of a report on the EDVAC,” in which he presented the basic architecture of computers: 1) a memory containing data and programs (the latter encoding algorithms), 2) an arithmetic, logical calculation unit and 3) a control unit capable of interpreting the instructions of the programs contained in the memory. Since the seminal texts of Alan Turing and John von Neumann represent only theoretical advances, we could date the new era from the construction and actual use of the first computers, in the 1950s. It is clear, however, that (in spite of the prescience of a few visionaries ) until the end of the 1970s, it was still hard to talk about an algorithmic medium. One of the main reasons is that the computers at that time were still big, costly, closed machines whose input and output interfaces could only be manipulated by experts. Although already in its infancy, the algorithmic medium was not yet socially prevalent.
It should be noted that between 1950 and 1980 (before Internet connections became the norm), data flows circulated mainly between containers and operators with local addresses enclosed in a single machine.

The Internet and personal computers (1970-1995)

A new trend emerged in the 1970s and became dominant in the 1980s: the interconnection of computers. The Internet protocol (invented in 1969) won out over its competitors in addressing machines in telecommunication networks. This was also the period when computing became personal. The digital was now seen as a vector of transformation and communication of all symbols, not only numbers. The activities of mail, telecommunications, publishing, the press, and radio and television broadcasting began to converge.
At the stage of the Internet and personal computers, data processed by algorithms were always stored in containers with local addresses, but—in addition to those addresses—operators now had universal physical addresses in the global network. Consequently, algorithmic operators could “collaborate,” and the range of types of processing and applications expanded significantly.

The World Wide Web (1995-2020)

It was only with the arrival of the Web, around 1995, however, that the Internet became the medium of most communication—to the point of irreversibly affecting the functioning of the traditional media and most economic, political and cultural institutions.
The revolution of the Web can be explained essentially as the creation of a universal system of physical addresses for containers. This system, of course, is URLs. It should be noted that—like the Internet protocol for operators—this universal system is added to the local addresses of the containers of data, it does not eliminate them. Tim Berners-Lee’s ingenious idea may be described as follows: by inventing a universal addressing system for data, he made possible the shift from a multitude of actual databases (each controlled by one computer) to a single virtual database for all computers. One of the main benefits is the possibility of creating hyperlinks among any of the data of that universal virtual database: “the Web.”
From then on, the effective power and the capacity for collaboration—or inter-operation—between algorithms increased and diversified enormously, since both operators and containers now possessed universal addresses. The basic programmable machine became the network itself, as is shown by the spread of cloud computing.
The decade 2010-2020 is seeing the beginning of the transition to a data-centric society. Indeed, starting with this phase of social utilization of the new medium, the majority of interactions among people take place through the Internet, whether purely for socialization or for information, work, research, learning, consumption, political action, gaming, watches, and so on. At the same time, algorithms increasingly serve as the interface for relationships between people, relationships among data, and relationships between people and data. The increase in conflicts around ownership and free accessibility of data, and around openness and transparency of algorithms, are clear signs of a transition to a data-centric society. However, in spite of their already decisive role, algorithms are not yet perceived in the collective consciousness as the new medium of human communication and thought. People were still fascinated by the logic of dissemination of previous media.
The next stage in the evolution of the algorithmic medium—the semantic sphere based on IEML—will provide a conceptual addressing system for data. But before we look at the future, we need to think about the limitations of the contemporary Web. Indeed, the Web was invented to help solve problems in interconnecting data that arose around 1990, at a time when one percent of the world’s population (mainly anglophone) was connected. But now in 2014, new problems have arisen involving the difficulties of translating and processing data, as well as the low level of digital literacy. When these problems become too pronounced (probably around 2020, when more than half the world’s population will be connected), we will be obliged to adopt a conceptual addressing system on top of the layer of physical addressing of the WWW.

The limitations of the Web in 2016

The inadequacy of the logic of dissemination

From Gutenberg until the middle of the twentieth century, the main technical effect of the media was the mechanical recording, reproduction and transmission of the symbols of human communication. Examples include printing (newspapers, magazines, books), the recording industry, movies, telephone, radio and television. While there were also technologies for calculation, or automatic transformation of symbols, the automatic calculators available before computers were not very powerful and their usefulness was limited.
The first computers had little impact on social communication because of their cost, the complexity of using them and the small number of owners (essentially big corporations, some scientific laboratories and the government administrations of rich countries). It was only beginning in the 1980s that the development of personal computing provided a growing proportion of the population with powerful tools for producing messages, whether these were texts, tables of numbers, images or music. From then on, the democratization of printers and the development of communication networks among computers, as well as the increased number of radio and television networks, gradually undermined the monopoly on the massive dissemination of messages that had traditionally belonged to publishers, professional journalists and the major television networks. This revolution in dissemination accelerated with the arrival of the World Wide Web in the mid-1990s and blossomed into the new kind of global multimedia public sphere that prevails now at the beginning of the twenty-first century.
In terms of the structure of social communication, the essential characteristic of the new public sphere is that it permits anyone to produce messages, to transmit to a community without borders and to access messages produced and transmitted by others. This freedom of communication is all the more effective since its exercise is practically free and does not require any prior technical knowledge. In spite of the limits I will describe below, we have to welcome the new horizon of communication that is now offered to us: at the rate at which the number of connections is growing, almost all human beings in the next generation will be able to disseminate their messages to the entire planet for free and effortlessly.
It is certain that automatic manipulation—or transformation—of symbols has been practiced since the 1960s and 1970s. I have also already noted that a large proportion of personal computing was used to produce information and not only to disseminate it. Finally, the major corporations of the Web such as Google, Amazon, eBay, Apple, Facebook, Twitter and Netflix daily process huge masses of data in veritable “information factories” that are entirely automated. In spite of that, the majority of people still see and use the Internet as a tool for the dissemination and reception of information, in continuity with the mass media since printing and, later, television. It is a little as if the Web gave every individual the power of a publishing house, a television network and a multimedia postal service in real time, as well as access to an omnipresent global multimedia library. Just as the first printed books—incunabula—closely copied the form of manuscripts, we still use the Internet to achieve or maximize the power of dissemination of previous media. Everyone can transmit universally. Everyone can receive from anywhere.
No doubt we will have to exhaust the technical possibilities of automatic dissemination—the power of the media of the last four centuries—in order to experience and begin to assimilate intellectually and culturally the almost unexploited potential of automatic transformation—the power of the media of centuries to come. That is why I am again speaking of the algorithmic medium: to emphasize digital communication’s capacity for automatic transformation. Of course, the transformation or processing power of the new medium can only be actualized on the basis of the irreversible achievement of the previous medium, the universal dissemination or ubiquity of information. That was nearly fully achieved at the beginning of the twenty-first century, and coming generations will gradually adapt to automatic processing of the massive flow of global data, with all its unpredictable cultural consequences. There are at this time three limits to this process of adaptation: users’ literacy, the absence of semantic interoperability and the statistical positivism that today governs data analysis.

The problem of digital literacy

The first limit of the contemporary algorithmic medium is related to the skills of social groups and individuals: the higher their education level (elementary, secondary, university), the better developed their critical thinking, the greater their mastery of the new tools for manipulation of symbols and the more capable they are of turning the algorithmic medium to their advantage. As access points and mobile devices increase in number, the thorny question of the digital divide is less and less related to the availability of hardware and increasingly concerns problems of print literacy, media literacy and education. Without any particular skills in programming or even in using digital tools, the power provided by ordinary reading and writing is greatly increased by the algorithmic medium: we gain access to possibilities for expression, social relationships and information such as we could not even have dreamed of in the nineteenth century. This power will be further increased when, in the schools of the future, traditional literacy, digital literacy and understanding of ecosystems of ideas are integrated. Then, starting at a very young age, children will be introduced to categorization and evaluation of data, collection and analysis of large masses information and programming of semantic circuits.

The absence of semantic interoperability

The second limit is semantic, since, while technical connection is tending to become universal, the communication of meaning still remains fragmented according to the boundaries of languages, systems of classification, disciplines and other cultural worlds that are more or less unconnected. The “semantic Web” promoted by Tim Berners-Lee since the late 1990s is very useful for translating logical relationships among data. But it has not fulfilled its promise with regard to the interoperability of meaning, in spite of the authority of its promoter and the contributions of many teams of engineers. As I showed in the first volume of The Semantic Sphere, it is impossible to fully process semantic problems while remaining within the narrow limits of logic. Moreover, the essentially statistical methods used by Google and the numerous systems of automatic translation available provide tools to assist with translation, but they have not succeeded any better than the “semantic Web” in opening up a true space of translinguistic communication. Statistics are no more effective than logic in automating the processing of meaning. Here again, we lack a coding of linguistic meaning that would make it truly calculable in all its complexity. It is to meet this need that IEML is automatically translated into natural languages in semantic networks.

Statistical positivism

The general public’s access to the power of dissemination of the Web and the flows of digital data that now result from all human activities confront us with the following problem: how to transform the torrents of data into rivers of knowledge? The solution to this problem will determine the next stage in the evolution of the algorithmic medium. Certain enthusiastic observers of the statistical processing of big data, such as Chris Anderson, the former editor-in-chief of Wired, were quick to declare that scientific theories—in general!—were now obsolete. In this view, we now need only flows of data and powerful statistical algorithms operating in the computing centres of the cloud: theories—and therefore the hypotheses they propose and the reflections from which they emerge—belong to a bygone stage of the scientific method. It appears that numbers speak for themselves. But this obviously involves forgetting that it is necessary, before any calculation, to determine the relevant data, to know exactly what is being counted and to name—that is, to categorize—the emerging patterns. In addition, no statistical correlation directly provides causal relationships. These are necessarily hypotheses to explain the correlations revealed by statistical calculations. Under the guise of revolutionary thought, Chris Anderson and his like are reviving the old positivist, empiricist epistemology that was fashionable in the nineteenth century, according to which only inductive reasoning (that is, reasoning based solely on data) is scientific. This position amounts to repressing or ignoring the theories—and therefore the risky hypotheses based on individual thought—that are necessarily at work in any process of data analysis and that are expressed in decisions of selection, identification and categorization. One cannot undertake statistical processing and interpret its results without any theory. Once again, the only choice we have is to leave the theories implicit or to explicate them. Explicating a theory allows us to put it in perspective, compare it with other theories, share it, generalize from it, criticize it and improve it. This is even one of the main components of what is known as critical thinking, which secondary and university education is supposed to develop in students.
Beyond empirical observation, scientific knowledge has always been concerned with the categorization and correct description of phenomenal data, description that is necessarily consistent with more or less formalized theories. By describing functional relationships between variables, theory offers a conceptual grasp of the phenomenal world that make it possible (at least partially) to predict and control it. The data of today correspond to what the epistemology of past centuries called phenomena. To extend this metaphor, the algorithms for analyzing flows of data of today correspond to the observation tools of traditional science. These algorithms show us patterns, that is, ultimately, images. But the fact that we are capable of using the power of the algorithmic medium to observe data does not mean we should stop here on this promising path. We now need to use the calculating power of the Internet to theorize (categorize, model, explain, share, discuss) our observations, without forgetting to make our theorizing available to the rich collective intelligence.
In their 2013 book on big data, Viktor Mayer-Schonberger and Kenneth Cukier, while emphasizing the distinction between correlation and causality, predicted that we would take more and more interest in correlations and less and less in causality, which put them firmly in the empiricist camp. Their book nevertheless provides an excellent argument against statistical positivism. Indeed, they recount the very beautiful story of Matthew Maury, an American naval officer who in the mid-nineteenth century compiled data from log books in the official archives to establish reliable maps of winds and currents. Those maps were constructed from an accumulation of empirical data. But with all due respect for Cukier and Mayer-Schonberger, I would point out that such an accumulation would never have been useful, or even feasible, without the system of geographic coordinates of meridians and parallels, which is anything but empirical and based on data. Similarly, it is only by adopting a system of semantic coordinates such as IEML that we will be able to organize and share data flows in a useful way.
Today, most of the algorithms that manage routing of recommendations and searching of data are opaque, since they are protected trade secrets of major corporations of the Web. As for the analytic algorithms, they are, for the most part, not only opaque but also beyond the reach of most Internet users for both technical and economic reasons. However, it is impossible to produce reliable knowledge using secret methods. We must obviously consider the contemporary state of the algorithmic medium to be transitory.
What is more, if we want to solve the problem of the extraction of useful information from the deluge of big data, we will not be able to eternally limit ourselves to statistical algorithms working on the type of organization of digital memory that exists in 2016. We will sooner or later, and the sooner the better, have to implement an organization of memory designed from the start for semantic processing. We will only be able to adapt culturally to the exponential growth of data—and therefore transform these data into reflected knowledge—through a qualitative change of the algorithmic medium, including the adoption of a system of semantic coordinates such as IEML.

The semantic sphere and its conceptual addressing (2020…)

It is notoriously difficult to observe or recognize what does not yet exist, and even more, the absence of what does not yet exist. However, what is blocking the development of the algorithmic medium—and with it, the advent of a new civilization—is precisely the absence of a universal, calculable system of semantic metadata. I would like to point out that the IEML metalanguage is the first, and to my knowledge (in 2016) the only, candidate for this new role of a system of semantic coordinates for data.
We already have a universal physical addressing system for data (the Web) and a universal physical addressing system for operators (the Internet). In its full deployment phase, the algorithmic medium will also include a universal semantic code: IEML. This system of metadata—conceived from the outset to optimize the calculability of meaning while multiplying its differentiation infinitely—will open the algorithmic medium to semantic interoperability and lead to new types of symbolic manipulation. Just as the Web made it possible to go from a great many actual databases to one universal virtual database (but based on a physical addressing system), IEML will make it possible to go from a universal physical addressing system to a universal conceptual addressing system. The semantic sphere continues the process of virtualization of containers to its final conclusion, because its semantic circuits—which are generated by an algebra—act as data containers. It will be possible to use the same conceptual addressing system in operations as varied as communication, translation, exploration, searching and three-dimensional display of semantic relationships.
Today’s data correspond to the phenomena of traditional science, and we need calculable, interoperable metadata that correspond to scientific theories and models. IEML is precisely an algorithmic tool for theorization and categorization capable of exploiting the calculating power of the cloud and providing an indispensable complement to the statistical tools for observing patterns. The situation of data analysis before and after IEML can be compared to that of cartography before and after the adoption of a universal system of geometric coordinates. The data that will be categorized in IEML will be able to be processed much more efficiently than today, because the categories and the semantic relationships between categories will then become not only calculable but automatically translatable from one language to another. In addition, IEML will permit comparison of the results of the analysis of the same set of data according to different categorization rules (theories!).

Algo-medium

FIGURE 1 – The four interdependent levels of the algorithmic medium

When this symbolic system for conceptual analysis and synthesis is democratically accessible to everyone, translated automatically into all languages and easily manipulated by means of a simple tablet, then it will be possible to navigate the ocean of data, and the algorithmic medium will be tested directly as a tool for cognitive augmentation—personal and social—and not only for dissemination. Then a positive feedback loop between the collective testing and creation of tools will lead to a take-off of the algorithmic intelligence of the future.
In Figure 1, the increasingly powerful levels of automatic calculation are represented by rectangles. Each level is based on the “lower” levels that precede it in order of historical emergence. Each level is therefore influenced by the lower levels. But, conversely, each new level gives the lower levels an additional socio-technical determination, since it uses them for a new purpose.
The addressing systems, which are represented under the rectangles, can be considered the successive solutions—influenced by different socio-technical contexts—to the perennial problem of increasing the power of automatic calculation. An addressing system thus plays the role of a step on a stairway that lets you go from one level of calculation to a higher level. The last addressing system, that of metadata, is supplied by IEML or any other system of encoding of linguistic meaning that makes that meaning calculable, exactly as the system of pixels made images manipulable by means of algorithms.

The cognitive revolution of semantic encoding

We know that the algorithmic medium is not only a medium of communication or dissemination of information but also, especially, a ubiquitous environment for the automatic transformation of symbols. We also know that a society’s capacities for analysis, synthesis and prediction are based ultimately on the structure of its memory, and in particular its system for encoding and organizing data. As we saw in the previous section, the only thing the algorithmic medium now in construction lacks to become the matrix of a new episteme that is more powerful than today’s, which has not yet broken its ties to the typographical era, is a system of semantic metadata that is equal to the calculating power of algorithms.

Memory, communication and intuition

It is now accepted that computers increase our memory capacities, in which I include not only capacities for recording and recall, but also those for analysis, synthesis and prediction. The algorithmic medium also increases our capacities for communication, in particular in terms of the breadth of the network of contacts and the reception, transmission and volume of flows of messages. Finally, the new medium increases our capacities for intuition, because it increases our sensory-motor interactions (especially gestural, tactile, visual and sound interactions) with large numbers of people, documents and environments, whether they are real, distant, simulated, fictional or mixed. These augmentations of memory, communication and intuition influence each other to produce an overall augmentation of our field of cognitive activity.
Semantic encoding, that is, the system of semantic metadata based on IEML, will greatly increase the field of augmented cognitive activity that I have described. It will produce a second level of cognitive complexity that will enter into dynamic relationship with the one described above to give rise to algorithmic intelligence. As we will see, semantic coding will generate a reflexivity of memory, a new perspectivism of intellectual intuition and an interoperability of communication.

Reflexive memory

The technical process of objectivation and augmentation of human memory began with the invention of writing and continued up to the development of the Web. But in speaking of reflexive memory, I go beyond Google and Wikipedia. In the future, the structure and evolution of our memory and the way we use it will become transparent and open to comparison and critical analysis. Indeed, communities will be able to observe—in the form of ecosystems of ideas—the evolution and current state of their cognitive activities and apply their capacities for analysis, synthesis and prediction to the social management of their knowledge and learning. At the same time, individuals will become capable of managing their personal knowledge and learning in relation to the various communities to which they belong. So much so that this reflexive memory will enable a new dialectic—a virtuous circle—of personal and collective knowledge management. The representation of memory in the form of ecosystems of ideas will allow individuals to make maximum use of the personal growth and cross-pollination brought about by their circulation among communities.

Perspectivist intellectual intuition

Semantic coding will give us a new sensory-motor intuition of the perspectivist nature of the information universe. Here we have to distinguish between the conceptual perspective and the contextual perspective.
The conceptual perspective organizes the relationships among terms, sentences and texts in IEML so that each of these semantic units can be processed as a point of view, or a virtual “centre” of the ecosystems of ideas, organizing the other units around it according to the types of relationships it has with them and their distance from it.
In IEML, the elementary units of meaning are terms, which are organized in the IEML dictionary (optimized for laptops + Chrome) in paradigms, that is, in systems of semantic relationships among terms. In the IEML dictionary, each term organizes the other terms of the same paradigm around it according to its semantic relationships with them. The different paradigms of the IEML dictionary are in principle independent of each other and none has precedence over the others a priori. Each of them can, in principle, be used to filter or categorize any set of data.
The sentences, texts and hypertexts in IEML represent paths between the terms of various paradigms, and these paths in turn organize the other paths around them according to their relationships and semantic proximity in the ecosystems of ideas. It will be possible to display this cascade of semantic perspectives and points of view using three-dimensional holograms in an immersive interactive mode.
Let us now examine the contextual perspective, which places in symmetry not the concepts within an ecosystem of ideas, but the ecosystems of ideas themselves, that is, the way in which various communities at different times categorize and evaluate data. It will thus be possible to display and explore the same set of data interactively according to the meaning and value it has for a large number of communities.
Reflexive memory, perspectivist intuition, interoperable and transparent communication together produce a cognitive augmentation characteristic of algorithmic intelligence, an augmentation more powerful than that of today.

Interoperable and transparent communication

The interoperability of communication will first concern the semantic compatibility of various theories, disciplines, universes of practices and cultures that will be able to be translated into IEML and will thus become not only comparable but also capable of exchanging concepts and operating rules without loss of their uniqueness. Semantic interoperability will also cover the automatic translation of IEML concepts into natural languages. Thanks to this pivot language, any semantic network in any natural language will be translated automatically into any other natural language. As a result, through the IEML code, people will be able to transmit and receive messages and categorize data in their own languages while communicating with people who use other languages. Here again, we need to think about cultural interoperability (communication in spite of differences in conceptual organization) and linguistic interoperability (communication in spite of differences in language) together; they will reinforce each other as a result of semantic coding.

Emergence

Emergence happens through an interdependant circulation of information between two levels of complexity. A code translates and betrays information in both directions: bottom-up and top-down.

Nature

According to our model, human collective intelligence emerges from natural evolution. The lower level of quantic complexity translates into a higher level of molecular complexity through the atomic stabilization and coding. There are no more than 120 atomic elements that explain the complexity of matter by their connections and reactions. The emergence of the next level of complexity – life – comes from the genetic code that is used by organisms as a trans-generational memory. Communication in neuronal networks translates organic life into conscious phenomena, including sense data, pleasure and pain, desire, etc. So emerges the animal life. Let’s note that organic life is intrinsically ecosystemic and that animals have developed many forms of social or collective intelligence. The human level emerges through the symbolic code : language, music, images, rituals and all the complexity of culture. It is only thank to symbols that we are able to conceptualize phenomena and think reflexively about what we do and think. Symbolic systems are all conventional but the human species is symbolic by nature, so to speak. Here, collective intelligence reaches a new level of complexity because it is based on collaborative symbol manipulation.

Culture

[WARNING: the next 5 paragraphs can be found in “collective intelligence for educators“, if you have already read them, go to the next slide: “algorithmic medium”] The above slide describes the successive steps in the emergence of symbolic manipulation. As for the previous slide, each new layer of cultural complexity emerges from the creation of a coding system.

During the longest part of human history, the knowledge was only embedded in narratives, rituals and material tools. The first revolution in symbolic manipulation is the invention of writing with symbols endowed with the ability of self-conservation. This leads to a remarquable augmentation of social memory and to the emergence of new forms of knowledge. Ideas were reified on an external surface, which is an important condition for critical thinking. A new kind of systematic knowledge was developed: hermeneutics, astronomy, medicine, architecture (including geometry), etc.

The second revolution optimizes the manipulation of symbols like the invention of the alphabet (phenician, hebrew, greek, roman, arab, cyrilic, korean, etc.), the chinese rational ideographies, the indian numeration system by position with a zero, paper and the early printing techniques of China and Korea. The literate culture based on the alphabet (or rational ideographies) developed critical thinking further and gave birth to philosophy. At this stage, scholars attempted to deduce knowledge from observation and deduction from first principles. There was a deliberate effort to reach universality, particularly in mathematics, physics and cosmology.

The third revolution is the mecanization and the industrialization of the reproduction and diffusion of symbols, like the printing press, disks, movies, radio, TV, etc. This revolution supported the emergence of the modern world, with its nation states, industries and its experimental mathematized natural sciences. It was only in the typographic culture, from the 16th century, that natural sciences took the shape that we currently enjoy: systematic observation or experimentation and theories based on mathematical modeling. From the decomposition of theology and philosophy emerged the contemporary humanities and social sciences. But at this stage human science was still fragmented by disciplines and incompatible theories. Moreover, its theories were rarely mathematized or testable.

We are now at the beginning of a fourth revolution where an ubiquitous and interconnected infosphere is filled with symbols – i.e. data – of all kinds (music, voice, images, texts, programs, etc.) that are being automatically transformed. With the democratization of big data analysis, the next generations will see the advent of a new scientific revolution… but this time it will be in the humanities and social sciences. The new human science will be based on the wealth of data produced by human communities and a growing computation power. This will lead to reflexive collective intelligence, where people will appropriate (big) data analysis and where subjects and objects of knowledge will be the human communities themselves.

Algo-medium

Let’s have a closer look at the algorithmic medium. Four layers have been added since the middle of the 20th century. Again, we observe the progressive invention of new coding systems, mainly aimed at the addressing of processors, data and meta-data.

The first layer is the invention of the automatic digital computer itself. We can describe computation as « processing on data ». It is self-evident that computation cannot be programmed if we don’t have a very precise addressing system for the data and for the specialized operators/processors that will transform the data. At the beginning these addressing systems were purely local and managed by operating systems.

The second layer is the emergence of a universal addressing system for computers, the Internet protocol, that allows for exchange of data and collaborative computing across the telecommunication network.

The third layer is the invention of a universal system for the addressing and displaying of data (URLs, http, html). Thank to this universal addressing of data, the World Wide Web is a hypertextual global database that we all create and share. It is obvious that the Web has had a deep social, cultural and economic impact in the last twenty years.

The construction of the algorithmic medium is ongoing. We are now ready to add a fourth layer of addressing and, this time, it will be a universal addressing system for semantic metadata. Why? First, we are still unable to resolve the problem of semantic interoperability across languages, classifications and ontologies. And secondly, except for some approximative statistical and logical methods, we are still unable to compute semantic relations, including distances and differences. This new symbolic system will be a key element to a future scientific revolution in the humanities and social sciences, leading to a new kind of reflexive collective intelligence for our species. Moreover, it will pave the way for the emergence of a new scientific cosmos – not a physical one but a cosmos of the mind that we will build and explore collaboratively. I want to strongly underline here that the semantic categorization of data will stay in the hands of people. We will be able to categorize the data as we want, from many different point of views. All that is required is that we use the same code. The description itself will be free.

Algo-intel

Let’s examine now the future emerging algorithmic intelligence. This new level of symbolic manipulation will be operated and shared in a mixed environment combining virtual worlds and augmented realities. The two lower levels of the above slide represent the current internet: an interaction between the « internet of things » and the « clouds » where all the data converge in an ubiquitous infosphere… The two higher levels, the « semantic sensorium » and the « reflexive collective intelligence » depict the human condition that will unfold in the future.

The things are material, localized realities that have GPS addresses. Here we speak about the smart territories, cities, buildings, machines, robots and all the mobile gadgets (phones, tablets, watches, etc.) that we can wear. Through binary code, the things are in constant interaction with the ubiquitous memory in the clouds. Streams of data and information processing reverberate between the things and the clouds.

When the data will be coded by a computable universal semantic addressing system, the data in the clouds will be projected automatically into a new sensorium. In this 3D, immersive and dynamic virtual environment we will be able to explore through our senses the abstract relationships between the people, the places and the meaning of digital information. I’m not speaking here of a representation, reproduction or imitation of the material space, like, for example, in Second Life. We have to imagine something completely different: a semantic sphere where the cognitive processes of human communities will be modeled. This semantic sphere will empower all its users. Search, knowledge exploration, data analysis and synthesis, collaborative learning and collaborative data curation will be multiplied and enhanced by the new interoperable semantic computing.

We will get reflexive collective intelligence thank to a scientific computable and transparent modeling of cognition from real data. This modeling will be based on the semantic code, that provides the « coordinate system » of the new cognitive cosmos. Of course, people will not be forced to understand the details of this semantic code. They will interact in the new sensorium through their prefered natural language (the linguistic codes of the above slide) and their favorite multimedia interfaces. The translation between different languages and optional interface metaphors will be automatic. The important point is that people will observe, analyze and map dynamically their own personal and collective cognitive processes. Thank to this new reflexivity, we will improve our collaborative learning processes and the collaborative monitoring and control of our physical environments. And this will boost human development!

Collective-Intelligence

The above slide represents the workings of a collective intelligence oriented towards human development. In this model, collective intelligence emerges from an interaction between two levels: virtual and actual. The actual is addressed in space and time while the virtual is latent, potential or intangible. The two levels function and communicate through several symbolic codes. In any coding system, there are coding elements (signs), coded references (things) and coders (being). This is why both actual and virtual levels can be conceptually analysed into three kinds of networks: signs, beings and things.

The actual human development can be analysed into a sphere of messages (signs), a sphere of people (beings) and a sphere of equipments – this last word understood in the largest possible sense – (things). Of course, the three spheres are interdependent.

The virtual human development is analysed into a sphere of knowledge (signs), a sphere of ethics (being) and a sphere of power (things). Again, the three spheres are interdependent.

Each of the six spheres is further analysed into three subdivisions, corresponding to the sub-rows on the slide. The mark S (sign) points to the abstract factors, the mark B (being) indicates the affective dimensions and the mark T (thing) shows the concrete aspects of each sphere.

All the realities described in the above table are interdependent following the actual/virtual and the sign/being/thing dialectics. Any increase of decrease in one « cell » will have consequences in other cells. This is just an example of the many ways collective intelligence will be represented, monitored and made reflexive in the semantic sensorium…

To dig into the philosophical concept of algorithmic intelligence go there

E-sphere-copie

An IEML paradigm projected onto a sphere.

Communication presented at The Future of Text symposium IV at the Google’s headquarters in London (2014).

Symbolic manipulation accounts for the uniqueness of human cognition and consciousness. This symbolic manipulation is now augmented by algorithms. The problem is that we still have not invented a symbolic system that could fully exploit the algorithmic medium in the service of human development and human knowledge.

E-Cultural-revolutions

The slide above describes the successive steps in the augmentation of symbolic manipulation.

The first revolution is the invention of writing with symbols endowed with the ability of self-conservation. This leads to a remarquable augmentation of social memory and to the emergence of new forms of knowledge.

The second revolution optimizes the manipulation of symbols like the invention of the alphabet (phenician, hebrew, greek, roman, arab, cyrilic, korean, etc.), the chinese rational ideographies, the indian numeration system by position with a zero, paper and the early printing techniques of China and Korea.

The third revolution is the mecanization and the industrialization of the reproduction and diffusion of symbols, like the printing press, disks, movies, radio, TV, etc. This revolution supported the emergence of the modern world, with its nation states, industries and its experimental mathematized natural sciences.

We are now at the beginning of a fourth revolution where an ubiquitous and interconnected infosphere is filled with symbols – i.e. data – of all kinds (music, voice, images, texts, programs, etc.) that are being automatically transformed. With the democratization of big data analysis, the next generations will see the advent of a new scientific revolution… but this time it will be in the humanities and social sciences.

E-Algorithmic-medium

Let’s have a closer look to the algorithmic medium. Four layers have been added since the middle of the 20th century.

– The first layer is the invention of the automatic digital computer itself. We can describe computation as « processing on data ». It is self-evident that computation cannot be programmed if we don’t have a very precise addressing system for the data and for the specialized operators/processors that will transform the data. At the beginning these addressing systems were purely local and managed by operating systems.

– The second layer is the emergence of a universal addressing system for computers, the Internet protocol, that allowed for exchange of data and collaborative computing across the telecommunication network.

– The third layer is the invention of a data universal addressing and displaying system (http, html), welcoming a hypertextual global database: the World Wide Web. We all know that the Web has had a deep social, cultural and economic impact in the last fifteen years.

– The construction of this algorithmic medium is ongoing. We are now ready to add a fourth layer of addressing and, this time, we need a universal addressing system for metadata, and in particular for semantic metadata. Why? First, we are still unable to resolve the problem of semantic interoperability across languages, classifications and ontologies. And secondly, except for some approximative statistical and logical methods, we are still unable to compute semantic relations, including distances and differences. This new symbolic system will be a key element to a future scientific revolution in the humanities and social sciences leading to a new kind of reflexive collective intelligence for our species. There lies the future of text.

E-IEML-math2

My version of a universal semantic addressing system is IEML, an artificial language that I have invented and developped over the last 20 years.

IEML is based on a simple algebra with six primitive variables (E, U, A, S, B, T) and two operations (+, ×). The multiplicative operation builds the semantic links. This operation has three roles: a depature node, an arrival node and a tag for the link. The additive operation gathers several links to build a semantic network and recursivity builds semantic networks with multiple levels of complexity: it is « fractal ». With this algebra, we can automatically compute an internal network corresponding to any variable and also the relationships between any set of variables.

IEML is still at the stage of fundamental research but we now have an extensive dictionary – a set of paradigms – of three thousand terms and grammatical algorithmic rules that conform to the algebra. The result is a language where texts self-translate into natural language, manifest as semantic networks and compute collaboratively their relationships and differences. Any library of IEML texts then self-organizes into ecosystems of texts and data categorized in IEML will self-organize according to their semantic relationships and differences.

E-Collective-intel2

Now let’s take an example of an IEML paradigm, the paradigm of “Collective Intelligence in the service of human development” for instance, where we will grasp the meaning of the primitives and in which way they are being used.

-First, let’s look at the dialectic between virtual (U) and actual (A) human development represented by the rows.

-Then, the ternary dialectic between sign (S), being (B) and thing (T) are represented by the columns.

-The result is six broad interdependent aspects of collective intelligence corresponding to the intersections of the rows (virtual/actual) and columns (sign/being/thing).

– Each of these six broad aspects of CI are then decomposed into three sub-aspects corresponding to the sign/being/thing dialectic.

The semantic relations (symmetries and inclusions) between the terms of a paradigm are all explicit and therefore computable. All IEML paradigms are designed with the same principles as this one, and you can build phrases by assembling the terms through multiplications and additions.

Fortunatly, fundamental research is now finished. I will spend the next months preparing a demo of the automatic computing of semantic relations between data coded in IEML. With tools to come…

E-Future-text2

Human-dev-CI

E = Emptiness, U = Virtual, A = Actual, S = Sign, B = Being, T = Thing


This presentation of IEML dates back to 2014. A complete and up to date presentation can be found at https://intlekt.io/ieml/

The algorithmic medium

Before the algorithmic medium was the typographical medium (printing press, broadcasting) that industrialized and automated the reproduction of information. In the new algorithmic medium, information is, de facto, ubiquitous and automation now concentrates on the transformation of information.

The algorithmic medium is built from three interdependent components: the Web as a universal database (big data), the Internet as a universal computer (cloud), and the algorithms in the hands of people.

IEML (the Information Economy MetaLanguage) has been designed to exploit the full potential of the new algorithmic medium.

IEML, who and what is it for?

It would have been impossible to have designed IEML before the automatic-computing era and, a fortiori, to implement and use it. IEML was designed for digital natives, and built to take advantage of the new pervasive social computing supported by big data, the cloud and open algorithms.

IEML is a language

IEML is an artificial language that has the expressive power of any natural language (like English, French, Russian, Arabic, etc.). In other words, you can say in IEML whatever you want and its opposite, with varying degrees of precision.

IEML is an inter-linguistic semantic code

We can describe IEML as a sort of pivot language. Its reading/writing interface pops up in the the natural language that you want with an IEML text that self-translates in that specific language.

IEML is a semantic metadata system

IEML was also designed as a tagging system supporting semantic interoperability. Its main use is data categorization. As a universal system addressing concepts, IEML can complement the universal addressing of data on the Web and of processors on the Internet.

IEML is a programming language

An IEML text programs the construction of a semantic network in natural languages and it computes its relations and its semantic differences with other texts.

IEML is a symbolic system

As with any other symbolic systems, IEML is a result from the interaction of three interdependent layers of linguistic complexity: a syntax, semantics and pragmatics.

EN-C-14-MMOM

IEML syntax

IEML syntax is an algebraic topology: this means that a complex network of relations (topology) is coded by an algebraic expression.

IEML Algebra

IEML algebra is based on six basic variables {E, U, A, S, B, T} and two operations {+, ×}. The multiplication builds links (node A, node B, tag) and the addition operation creates graphs by connecting the links. The results of any algebraic operation can be used as a basis for new operations. This recursivity allows the construction of successive layers of complexity.

A computable Topology

Each distinct variable of the IEML algebra corresponds to a distinct graph. Given a set of variables, their relations and their semantic differences are computable.

EN-D-10-MMu_MMu

IEML semantics

As it is projected onto an algebraic topology, IEML’s semantics becomes computable.

The semantic projection onto an algebraic topology

– An IEML script normalizes the notation of an algebraic expression.
– The IEML dictionary is organized as a set of paradigms, a paradigm being a semantic network of terms. Each IEML term can be translated in natural languages.
– With IEML operations {+, ×} and its recursivity, the IEML grammar allows the construction of morphemes, words, clauses, phrases, complex propositions, texts and hypertexts.

The grammatical algorithms

Embedded in IEML, any grammatical algorithms can compute:
– the intra-textual semantic network corresponding to an IEML text
– the translation of an IEML semantic network into any chosen natural language
– the inter-textual semantic network and the semantic differences corresponding to any set of IEML texts.

IEML pragmatics

IEML pragmatics is oriented towards self-organization and reflexive collective intelligence.

A new approach to data and social networks

When data are categorized in IEML, they self-organize into semantic networks and automatically compute their semantic relations and differences. Moreover, when communities engage in collaborative data curation using IEML, what they get in return is a simulated image of their collective intelligence process.

Modeling ideas as dynamic texts

We can model our collective intelligence into an evolving ecosystem of ideas. In this framework, an idea can be defined as the assembly of a concept, an affect, a percept (a sensory-motor image) and a social context. In a dynamic text, the concept is represented by an IEML text, the affect by credits (positive or negative), the percept by a multimedia dataset and the social context as an author (a player) a community (a semantic game) and a time-stamp.

Automatic computing of dynamic hypertexts

Thanks to IEML grammatical algorithms, any set of dynamic texts self-organizes into a dynamic hypertext that represents an ecosystem of ideas in the form of an immersive simulation. Now, a reflexive collective intelligence can emerge from a collaborative data curation.