/ 76
- Dettagli
- Categoria: Area Scientifica
- Scritto da Giacomo Grandi
- Visite: 646
PubbliTesi - La Tesi
Early Diagnosis of Dementia: The Contribution of Language Models’ Perplexity
Scheda Sintetica
Autore: Giacomo Grandi
Relatore: Daniele Paolo Radicioni
Università: Università degli Studi di Torino
Facoltà: Dipartimento di Informatica
Corso: Laurea Magistrale in Informatica
Data di Discussione: 17/04/2024
Voto: 110 cum laude
Disciplina: Tecnologie del Linguaggio Naturale
Tipo di Tesi: di Ricerca
Altri Relatori: Davide Colla, Matteo Delsanto
Lingua: Inglese
Grande Area: Area Scientifica
Dignità di Stampa: Si
Settori Interessati: medico, economico
Descrizione:
In this work we explore how language models can be employed to analyze language and discriminate between subjects affected by mental disorders (across a broad spectrum, falling within the realm of dementia) and healthy subjects, using the perplexity metric. Perplexity has been conceived as an intrinsic measure for evaluating language models (how suitable a given language model is for predicting a text sequence or, equivalently, how well a sequence of words fits into a specific language model). We conducted an experimentation on a dataset of interviews with both healthy subjects and people affected by dementia, employing different language models such as N-grams and GepPeTto, a language model for Italian and based on GPT-2. Our best-performing models achieved very high accuracy and competitive F-scores compared to the state-of-the-art, both in categorizing subjects with dementia and those in the control group.