PhD Thesis Seminar - Nazanin Jafari, NLP-based Approaches for Improved Information Integrity
Content
Speaker:
Abstract:
Preserving information integrity is a broad and challenging problem that spans multiple dimensions, including the accuracy, reliability, and consistency of information. In this thesis, I focus on three connected challenges: improving automatic fact-checking for complex domains, improving content moderation practices in ways that account for human well-being, and developing more comprehensive factuality evaluation methods for large language models (LLMs).
The first part of the thesis addresses hate speech moderation and the emotional burden placed on human moderators who are repeatedly exposed to harmful content. I propose a human-centered moderation approach based on target substitution with LLMs, designed to reduce emotional distress while preserving moderation accuracy and the original meaning of the content. The second part focuses on robust zero-shot claim verification, especially in scientific and biomedical domains where existing methods often struggle with reasoning and generalization. To address this, I introduce a retrieval- and reasoning-augmented framework that transforms complex evidence into concise factual statements, making claim–evidence relationships more explicit and improving verification robustness and interpretability.
The final part of the thesis introduces a recall-oriented metric for evaluating the factuality of long-form LLM outputs. Unlike prior approaches that focus mainly on precision, this framework also accounts for omitted but relevant facts and incorporates importance-aware scoring based on salience and query relevance.
Advisor:
James Allan