Abstract: Document-based question answering (DBQA) generally relies solely on the one-way matching relationship between documents and questions to locate evidence spans and generate answers.However,capturing concise evidence spans is difficult when facing semantic challenges such as distant interference and multiple answer words.To address this issue,an evidence span prediction model ESP-BSA based on a bidirectional superposition attention mechanism is proposed.Firstly,the implicit interaction of text semantics is enriched by cross-matching the question with the text.Secondly,soft evidence label pairs are designed based on the heterogeneity of evidence distribution to represent the forward and backward evidence scores.Finally,the evidence scores at each position in the bidirectional stacked sequence are superposed to obtain evidence spans that better meet the contextual requirements.Experimental results demonstrate that the proposed model improves the precision of evidence span prediction and the accuracy of question answering in complex contexts,as evidenced by respective improvements in Span-F1 and Span-EM evaluation metrics compared to baseline models.

Key words: document-based question answering (DBQA), evidence span, attention mechanism, bidirectional superposition, soft evidence label