中文复述问句生成技术研究

Research on Chinese question paraphrase generation techniques for question answering

自动问答系统允许用户以自然语言进行提问,问题的形式多样、结构复杂,对系统的理解能力提出了极高要求。问句复述生成技术可将提出的复杂问句改写成一系列与之语义相同但形式不同的问句,避免了用户提问的不规范,可大大降低系统对问句的理解和处理难度,对于提升自动问答系统的效果有着重要意义。本文提出了一种基于模板匹配的复述问句生成方法,该方法可有效保留问句的结构特征和语义特征。引入功能标签,突出问句的结构特征;引入依存关系,提高了问句模板的泛化性能;引入候选排序,大幅提升了生成结果的准确率。通过与已有的生成方法进行对比试验,证实了该方法的有效性。

The question answering system allows the user to ask questions in natural language,the form of the query is diverse and the structure is complex,the system's understanding ability is challenged.The paraphrase generation technique can be used to paraphrase the complicated question into a series of different questions with the same semantics,which could avoid the non-standard of the user's question and reduce the difficulty of understanding and processing,and improve the performance of QA system.This paper proposes a method of question paraphrase generation based on template matching.By using the method,the structural and semantic features of the question could be effectively preserved.The functional labels can highlight the structural features;the dependency can improve the generalization of the question template;the candidate ranking can greatly enhance the precision of the generation.The experimental results prove the efficiency of this method.