A linguística saussuriana aplicada ao processamento de linguagem natural

Abstract

This dissertation proposes a dialogue between Ferdinand de Saussure's linguistic theory and the field of Natural Language Processing (NLP), exploring potential theoretical intersections between the Theory of Value (ToV), along with the Saussurean assumptions that frame it, and Distributional Semantics (DS), the latter widely applied in NLP for semantic tasks. In the first part, we outline a theory of signification based on Saussure that presents evident articulable elements with DS, thereby supporting the hypothesis that both the Theory of Value and Distributional Semantics share a common core—namely, the idea that language and meaning are products of a system of differential relations. Accordingly, we examine how the notion of relation—and the concepts derived from it, such as system, sign, value, and meaning—may contribute to the construction of a robust theory of signification potentially useful for semantic modeling, similarity analysis, and probability calculations in NLP tasks. Based on the analysis of the relations that comprise the linguistic system, we investigate how linguistic values emerge and generate meaning through differences of similia and dissimilia. In the second part, we examine the epistemological connections between Saussurean linguistics and DS, grounded in the ideas of Harris and Firth, and analyze word embeddings generated by the Word2Vec model from a Saussurean perspective, seeking evidence of possible articulations between Saussure and NLP. In this regard, our findings indicate that the construction of meaning in DS-based language models appears to mirror the differential and relational principles that Saussure developed for the functioning of language in the early 20th century. Additionally, we discuss a DS application in a project at the University of Geneva, highlighting the usefulness of Saussurean concepts for the semantic modeling of historical texts. In summary, the results indicate that Saussurean linguistics offers significant theoretical support for NLP, enabling a deeper understanding of the processes of signification in vector-based language models. In light of this, we argue for the possible validity of our main investigative hypothesis—namely, the idea that Saussurean principles remain relevant in contemporary linguistics while also providing a solid foundation for the advancement of new language technologies, reaffirming the contemporary relevance of Saussurean linguistics in the context of 21st-century Natural Language Processing.