A framework for closed domain question answering systems in the low data regime.

• Ano de defesa: 2022
• Autor(a) principal: Carmo, Vinícius Cleves de Oliveira
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: eng
• Instituição de defesa: Biblioteca Digitais de Teses e Dissertações da USP
• Palavras-chave em Português: Aprendizado computacional; Information retrieval; Machine learning; Neural networks; Question answering systems; Recuperação de informação; Redes neurais; Sistemas de questões e respostas
• Link de acesso: https://www.teses.usp.br/teses/disponiveis/3/3141/tde-24052023-152815/

Question Answering (QA) systems that operate over textual databases aim at improving traditional information retrieval systems; while the latter recover a number of relevant documents from a document pool, the former can also find and present direct answers to users. Recent improvements on QA have been based on deep neural networks; such networks require large volumes of labeled data for training. Most existing datasets target general knowledge and, even though there are a few datasets for specific domains (such as biomedicine), for most domains there is no labeled, or easy to label, dataset available. This creates an obstacle for the development of domain-specific QA systems. We propose a framework for developing domain-specific QA systems by leveraging unsupervised learning so as to avoid the costs related to large scale dataset labeling. The contributions of this work are twofold. First, we apply domain-adaptive pretraining to improve out-of-domain performance of reading comprehension and question answering systems. This technique achieves state-of-the-art results on two Reading Comprehension datasets, and it exceeds the performance of state-of-the-art domain adaptation techniques in the literature by a significant margin: 2.3 exact match and 5.2 F1-score on BioASQ. Then, we propose a framework for domain-specific question answering in the low data regime. For document retrieval, we apply a combination of BM25 along with a custom text processing pipeline. We find that, in a low data setting, statistical document retrieval models outperform neural models as the data on the desired domain differ from the data used for training. For answer selection, we apply a neural reader trained with domain-adaptive pretraining to improve generalization on the desired domain. We also perform a case study by applying the proposed framework to the offshore engineering domain.