基于多阶段推理的多跳机器阅读理解

Multi-hop machine reading comprehension based on multi-stage reasoning

多跳机器阅读理解需要从相关文档中提取关键线索,并以此推理出问题的答案。随着自然语言技术的进步,该任务吸引了越来越多研究者的关注,但很少注重关键线索之间直接的信息交互。因此,本文提出了一种基于多阶段推理的多跳阅读理解模型结构,基于逐步精炼信息的思路,通过基于PLM的深度学习模型,依次提取与问题有关的相关文档集、支持句集以及最终的答案片段,并在答案预测模型中提出多级关系图卷积网络,以实现线索信息的充分交互。在HotpotQA数据集上进行的实验表明,与基准模型相比答案片段EM(Exact Match)值提升了24.26%,F1值提升了23.86%。

Multi-hop machine reading comprehension involves extracting key clues from relevant documents and inferring the answer to a given question based on those clues.With the advancement of natural language processing technology,this task has attracted increasing attention from researchers.However,little attention has been paid to direct interaction between key clues.Therefore,this paper proposes a multi-stage reasoning-based multi-hop reading comprehension model structure.Following the idea of progressively refining information,it sequentially extracts relevant document sets,supporting sentence sets,and final answer fragments based on a PLM-based deep learning model that is related to the question.In addition,a multi-level relational graph convolution network is proposed in the answer prediction model to facilitate sufficient interaction between key information.Experiments conducted on the HotpotQA dataset demonstrate that,compared to the baseline model,the proposed model achieves an improvement of 24.26%in Exact Match(EM)value and 23.86%in F1 value for answer fragment prediction.