Prompting techniques for Natural Language Generation in the Medical Domain

Rossini, Martina (2023) Prompting techniques for Natural Language Generation in the Medical Domain. [Laurea magistrale], Università di Bologna, Corso di Studio in Artificial intelligence [LM-DM270]
Documenti full-text disponibili:
[img] Documento PDF (Thesis)
Disponibile con Licenza: Salvo eventuali più ampie autorizzazioni dell'autore, la tesi può essere liberamente consultata e può essere effettuato il salvataggio e la stampa di una copia per fini strettamente personali di studio, di ricerca e di insegnamento, con espresso divieto di qualunque utilizzo direttamente o indirettamente commerciale. Ogni altro diritto sul materiale è riservato

Download (1MB)


Generating automatic explanations for the predictions of machine learning models has long since been a major challenge in Artificial Intelligence, especially when considering sensible domains like healthcare. In this thesis, we approach the problem of generating a fluent, natural language justification for a medical diagnosis given all the information in the case description and the disease's symptomatology. We treat this problem as a data-to-text task and solve it using prompting techniques for natural language generation. In particular, we propose two architectural modifications to standard Prefix Tuning called Layer Dependency and Prefix Pooling; we evaluate their results, comparing with current state-of-the-art methods for the data-to-text task, on both a general-purpose benchmark (WebNLG) and on a dataset of clinical cases and relative explanations built as part of the ANTIDOTE project. Results show that layer dependency boosts the generation capabilities of our models when training on a limited computational budget, while Prefix Pooling is a valid dynamic prompting technique that achieves performances on par with the current state-of-the-art without requiring any additional information to be associated with the input. Finally, we note how, for our domain and in the context of the ANTIDOTE project, interpreting the explanation-generation task as data-to-text is a viable approach which produces high-quality justifications.

Tipologia del documento
Tesi di laurea (Laurea magistrale)
Autore della tesi
Rossini, Martina
Relatore della tesi
Correlatore della tesi
Corso di studio
Ordinamento Cds
Parole chiave
Natural Language Generation,Data-to-text,Prompting,Pretrained Language Models,Deep Learning,Medical Domain
Data di discussione della Tesi
23 Marzo 2023

Altri metadati

Statistica sui download

Gestione del documento: Visualizza il documento