Niloofar Bani
Analysis of Large Language Models' Narrative Construction Across Languages: A Comparative Study of English and Persian Event Descriptions.
Rel. Juan Carlos De Martin, Marco Rondina. Politecnico di Torino, Corso di laurea magistrale in Ingegneria Informatica (Computer Engineering), 2026
|
Preview |
PDF (Tesi_di_laurea)
- Tesi
Licenza: Creative Commons Attribution Non-commercial No Derivatives. Download (1MB) | Preview |
Abstract
As large language models (LLMs) increasingly act as primary intermediaries for information, understanding how they construct narratives around politically contested events is critical. Even when instructed to remain neutral, model outputs can shape public understanding through subtle differences in selection, emphasis, attribution, and lexical framing. This thesis investigates how narrative construction varies across ChatGPT, Mistral, and Qwen when describing three politically contested protest contexts: anti-ICE demonstrations in the United States, the December 2025 uprisings in Iran, and the Askatasuna social center eviction protests in Turin, Italy. Responses were collected directly from model web interfaces in both English and Persian using four distinct prompt types: a neutral descriptive request, a protester-focused critical prompt, an authority-focused critical prompt, and a request for news headline generation.
Every model–case–language–prompt configuration was run ten times, producing a dataset of 720 responses
Relatori
Anno Accademico
Tipo di pubblicazione
Numero di pagine
Corso di laurea
Classe di laurea
Aziende collaboratrici
URI
![]() |
Modifica (riservato agli operatori) |
