Ali Behlooeidolatsaraei
Unanswerable Question Detection.
Rel. Luca Cagliero, Alessia Intini. Politecnico di Torino, Corso di laurea magistrale in Data Science And Engineering, 2026
Abstract
Visual Large Language Models (VLLMs) are increasingly used to answer questions over Visually Rich Documents (VRDs), but most research assumes every question has an answer in the document. In practice this is false, and a model that fabricates an answer instead of recognizing this is at least as dangerous as one that answers incorrectly. VRD-UQA (Napolitano et al., submitted to AAAI 2026) introduced a controlled corruption methodology for studying this problem, evaluating twelve VLLMs zero-shot and leaving supervised fine-tuning as future work. This thesis takes up that gap, testing whether fine-tuning Qwen2.5-VL-7B-Instruct on a balanced set of answerable and unanswerable VRD questions improves unanswerable-question detection over the zero-shot baseline.
A training set of about 4,000 questions is built from DUDE by combining its OCR annotations with document element types and positions detected via DocLayout-YOLO, supporting entity, element, and layout corruptions
Relatori
Anno Accademico
Tipo di pubblicazione
Numero di pagine
Informazioni aggiuntive
Corso di laurea
Classe di laurea
Aziende collaboratrici
URI
![]() |
Modifica (riservato agli operatori) |
