基于汉语篇章框架语义分析的阅读理解问答研究

摘要

Answer-sentence retrieval and answer extraction are two core techniques in Reading Comprehension (RC ).This paper proposed a new method of answer-sentence retrieval and answer extraction for Chinese RC based on discourse frame semantic parsing.At the stage of retrieving answer sentences,frame similarity,frame-frame relations and discourse frame-frame relations are employed in the new method.Specifically,frame similarity is used to compute the semantic scenarios (i.e.frames)similarities between questions and reading materials;frame-frame relations and discourse frame-frame relations are utilized to obtain answer sentences from the perspective of semantic relevance.After that,frame similarity,defined null instantiation, and frame-frame relations are applied in the process of answer extraction.Among them,frame similarity is used to extract the frame element which serves as question role from answer sentences with semantic similarities and the extracted frame element is treated as the final answer;defined null instantiation is utilized to locate the exact position of the missing semantic role in discourses and the missing semantic roles is considered to be the final answer;because frame-frame relations can establish the relation between frame elements,they are used to extract the highly relevant frame element and the highly relevant frame element is seen as the final answer.To evaluate the method,552 RC questions in 15 different fields are utilized,and our method performs better than some traditional similarity-based methods in answer-sentence retrieval.Compared with the baseline experiment based on frame semantic similarity,the addition of discourse-level frame semantic features like frame-to-frame relations,discourse-frame-frame relations and defined null instantiation clues,gains better results in answer-sentences retrieval and answer extraction.%答案句检索和答案抽取是阅读理解中的两个核心技术。针对汉语阅读理解，该文提出一种新的基于篇章框架语义分析的答案句检索和答案抽取方法。答案句检索是基于框架相似性、框架关系及篇章框架关系来实现。其中，基于框架相似性的方法是通过计算背景材料与问句之间语义场景（框架）的相似度来进行答案句检索；基于框架关系和篇章框架关系的方法可以从语义相关角度获得与问句语义相关的答案句。在答案抽取时，提出基于框架语义相似性、有定零形式线索及框架关系的答案抽取方法。基于框架语义相似性可以从语义相似的答案句中抽取出充当问句疑问角色的框架元素作为答案；有定零形式线索能够在篇章范围定位答案句中充当答案的缺失语义成分；框架关系则能够通过建立框架元素之间的关系，抽取相关度高的框架元素作为答案。针对15个领域的552个阅读理解问题，该方法在答案句检索时相比传统基于相似度的方法能够获得更好的答案句检索结果；相比基于框架相似性的 Baseline 实验，加入篇章框架关系、框架关系及有定零形式线索的篇章级框架语义特征，能够获得更优的答案句检索与答案抽取结果。