Tesi etd-01162025-153636 |
Link copiato negli appunti
Tipo di tesi
Tesi di laurea magistrale
Autore
MUSSO, ALICE
URN
etd-01162025-153636
Titolo
Large Language Model e la Competenza Lessicale: il Caso delle Relazioni Iperonimiche
Dipartimento
FILOLOGIA, LETTERATURA E LINGUISTICA
Corso di studi
LINGUISTICA E TRADUZIONE
Relatori
relatore Prof. Lenci, Alessandro
Parole chiave
- competenza lessicale
- iponimia
- large language models
- semantica
Data inizio appello
07/02/2025
Consultabilità
Completa
Riassunto
Negli ultimi anni, lo studio delle capacità dei modelli neurali del linguaggio (LLM) ha acquisito crescente importanza. Infatti, con l’evoluzione dai modelli one task-one model ai moderni modelli basati sull’architettura transformers, capaci di affrontare molteplici task di NLP, si tende a percepire una forte competenza linguistica. Tuttavia, sebbene i modelli transformers siano in grado di svolgere molteplici compiti di NLP, non è possibile affermare che i modelli possiedano una effettiva competenza linguistica. Per comprendere il linguaggio naturale, un sistema deve possedere sia l'abilità referenziale che inferenziale. Questa tesi analizza la capacità dei modelli di comprendere le relazioni paradigmatiche, con focus su quelle iperonimiche. A tal fine, sono stati progettati tre esperimenti volti a verificare se i modelli riconoscano le proprietà distintive delle relazioni iperonimiche: asimmetria e transitività. I risultati ottenuti evidenziano performance generalmente inferiori al 60% di accuratezza, indicando una rappresentazione limitata delle relazioni iperonimiche. Ciò dimostra che i LLM creano rappresentazioni semantiche basate su co-occorrenze lessicali, mancando di una conoscenza strutturata del lessico.
In recent years, the study of the capabilities of large language models (LLMs) has gained increasing importance. With the evolution from one task-one model approaches to modern transformer-based architectures capable of handling multiple NLP tasks, there is a tendency to perceive strong linguistic competence in these models. However, while transformer models can perform various NLP tasks, it cannot be claimed that they possess true linguistic competence. To comprehend natural language, a system must exhibit both referential and inferential abilities. This thesis examines the ability of LLMs to understand paradigmatic relations, focusing on hypernymic relationships. To this end, three experiments were designed to test whether the models recognize the distinctive properties of hypernymic relations: asymmetry and transitivity. The results show overall performance below 60% accuracy, highlighting a limited representation of hypernymic relations. This demonstrates that LLMs create semantic representations based on lexical co-occurrences, lacking a structured knowledge of the lexicon.
In recent years, the study of the capabilities of large language models (LLMs) has gained increasing importance. With the evolution from one task-one model approaches to modern transformer-based architectures capable of handling multiple NLP tasks, there is a tendency to perceive strong linguistic competence in these models. However, while transformer models can perform various NLP tasks, it cannot be claimed that they possess true linguistic competence. To comprehend natural language, a system must exhibit both referential and inferential abilities. This thesis examines the ability of LLMs to understand paradigmatic relations, focusing on hypernymic relationships. To this end, three experiments were designed to test whether the models recognize the distinctive properties of hypernymic relations: asymmetry and transitivity. The results show overall performance below 60% accuracy, highlighting a limited representation of hypernymic relations. This demonstrates that LLMs create semantic representations based on lexical co-occurrences, lacking a structured knowledge of the lexicon.
File
Nome file | Dimensione |
---|---|
TesiAliceMusso1.pdf | 2.63 Mb |
Contatta l’autore |