NotaBene е електронно списание за философски и политически науки. Повече за нас
Abstract: Artificial Intelligence methodologies for natural language processing has steadily converged towards heavy deployment of Large Language Models. The underlying operational principle of these models relies on mining massive sets of text corpuses, where the semantic elicitation between the entities of the text is interpreted as statistical interdependency (correlation) between words located close to each other in a sentence. The problem with this approach is that probable semantic connection could only be established, if such connections exist in the already mentioned large text corpuses, furthermore, in order for models further to generalize over unseen examples, this plethora of training examples must be provided, in a vastly jumbled and diverse texts, which is inherently problematic, if the model has to be updated with new correlations, since new text corpuses has to be carefully selected or synthetically generated in order not to strongly weaken initial correlations. As a rule of thumb, in contrast to the artificial large language models, humans, need but a few examples, in order to embed a possible semantic model that governs any new concept, independently of the informational structure of the examples, be that text, images, graphs, etc. Although, human cognition is a constantly proliferated domain of research, most researchers would agree that humans are able to elicit semantic, not only by learning interdependencies between entities, but also semantic relationship rules that governs those entities. This knowledge about the rules and entities is then constantly updated throughout our life. This paper explores one possibility of creating such hybrid computational model, where semantic relationships rules are created out of graph structured data and are used to improve semantic interpretation in natural language processing by unification of two learning paradigms on graph textual representation, to improve semantic interpretation.
Keywords: Artificial Intelligence, natural language processing, hybrid models, graph, graph neural, network.