Enhancing Retrieval-Based Question Answering

Abstract

This thesis delves into the critical domains of retrieval-based question-answering (QA) tasks, particularly in enhancing the quality of retrieved texts. Traditional IR methods often generate noisy text, hindering the accuracy of subsequent answer extraction. While neural models have been employed to re-rank retrieved passages, this work delves deeper into improving retrieval efficiency and precision. In the open-domain QA domain, where the system tackles any user-posed question, the thesis introduces a novel Passage Ranker model. This model surpasses existing approaches by incorporating local-context information through cross-passage interaction. Unlike prior methods, it leverages the initial ranking provided by search engines and utilizes tailored attention mechanisms for more accurate confidence score calculation. Furthermore, semantic role labeling (SRL) is integrated into the passage reader, enabling it to better grasp contextual semantics. Extensive evaluations demonstrate the significant superiority of this model compared to recent baselines. For long-context multiple-choice question answering, the thesis proposes Options Aware Dense Retrieval (OADR) as a novel approach. OADR addresses the challenges of reasoning over extensive textual sources by utilizing query-options embeddings. This innovative strategy aims to align with the embeddings of the ”Oracle query” (query paired with the correct answer), allowing for better identification of crucial evidence spans essential for accurate answer selection. Experiments on the QuALITY benchmark dataset validate the effectiveness of OADR, showcasing its superior performance and accuracy compared to established baselines.