一种面向机器阅读理解系统的问题意图识别方法

Method for Recognizing Question Intent for Machine Reading Comprehension System

随着深度学习不断发展,基于问答系统的机器阅读理解能力不断提高并达到了较高水平,但当用户提出模棱两可的问题,机器阅读理解模型在回答时仍存在偏差。为优化模型,提出了一种创新方法,即候选修正问题生成(CRQG)。CRQG专门进行输入问题的修正,并生成多个最优问题候选集,便于用户在候选集中选择最接近其表达意图的问题,从而在阅读理解中获得较高精准度的答案。为实现该方法,同时设计了一个轻量级候选问题生成模型(LCQGM),该模型融入问题与段落两种复制机制,用于描述在段落中损失的用户输入问题信息。实验结果表明,该模型有效地提升了机器阅读理解的精确性。

With the continuous development of deep learning,the machine reading comprehension ability based on question answering systems has been continuously improved and reached a high level.However,when users ask ambiguous questions,there are still deviations in the answers provided by the machine reading comprehension model.In order to optimize the model,this paper proposes an innovative method called candidate revised question generation(CRQG),which specifically modifies the input question and generates multiple optimal question candidate sets,allowing users to choose the question that most closely matches their intention from the candidate set,thereby obtaining a more accurate answer in reading comprehension.To implement this method,this paper also designs a lightweight candidate question generation model(LCQGM),which incorporates two replication mechanisms for questions and paragraphs to describe the loss of user input question information in paragraphs.Experimental results show that this model effectively improves the accuracy of machine reading comprehension.