Prova de Doutoramento do aluno Anthony Andrea Faustine

Título da Tese: Characterizing Individuals through Language using Fuzzy Fingerprints and Language Models

Local da Prova: Anfiteatro PA-3 (Piso -1 do Pavilhão de Matemática) do IST

Abstract: Language serves as a vital conduit of personal identity, enabling individuals to project and express their self-image through subtle linguistic choices. The markers present in textual communication, ranging from pronoun usage and emotive expressions to more complex syntactic structures, offer an expressive channel through which researchers can infer aspects such as identity, profile information, and personality traits. Recent advances in computational linguistics have leveraged such patterns, particularly through language models, to map language features to dimensions of individual characterization. While these models achieve strong predictive performance, they do not reveal which internal representations support decisions about individuals. A compact and interpretable way to capture such distinctions is provided by Fuzzy Fingerprints (FFPs), a fuzzy-based classification framework that assigns membership degrees to the most relevant features. Yet, in their original form, FFPs often lack the contextual depth that modern language models provide. Moreover, the scarcity and complexity of high-quality labeled data, particularly datasets enriched with profiles, personality, or other individual-specific metadata, limit the robustness and generalizability of these models. In this thesis, we investigate how language can be used as a tool to identify and describe individuals while addressing these challenges. To this end, we propose a framework that unites the interpretability and compactness of FFPs with the representational capabilities of language models. By integrating the strengths of both methods, the approach yields competitive performance in authorship attribution and textual speaker identification while enabling insights into model decisions through compact representations. We further study optimization strategies for these FFPs and introduce a differentiable formulation that allows compact FFPs to be learned jointly with language models. Furthermore, this thesis reformulates existing dialogue resources for profile detection, introduces a new dataset for profile generation from speaker utterances, and contributes a multilingual personality corpus linking open-ended self-narratives to Mini-IPIP item-level responses and Big Five trait scores. Collectively, this work aims to advance the field of computational linguistics by providing foundations, methods, and resources for more interpretable and better grounded modeling of individual characterization through language.

Fonte: Departamento de Engenharia Informática (DEI). Horário, local exato e preço na página oficial.