DeepCorp:一种基于上下文和操作角色的需求实体共指检测网络

DEEPCORP: A DETECTION NETWORK FOR REQUIREMENTS ENTITY COREFERENCE BASED ON CONTEXT AND REQUIREMENTS OPERATION ROLE

  • 摘要: 自动化检测需求实体共指(EC),对于需求质量的一致性分析十分重要。现有方法往往是利用编辑距离或词嵌入等完成EC检测,其在没有大量专家标注数据的情况下和在捕获需求语句较为复杂的语义信息方面表现不佳。提出一种用于EC检测的新型深度网络DeepCorp(Deep & Context-wise & Requirements Operation Role Network),通过引入实体上下文和需求操作角色信息,使用多层感知机隐式融合嵌入表示来实现需求实体深层次的语义表达,从而进行实体语义相似性判断。在公开需求文档仓库上的实验表明,DeepCorp可达到96.72%准确率、96.67%召回率和96.69%F1,相较于现有方法平均提升1.27%。

     

    Abstract: Automatically detecting the requirement entity coreference (EC) is very important for the consistency analysis of the requirement quality. Existing methods often use edit distance or word embedding to complete EC detection, which does not perform well in capturing complex semantic information of requirement sentences without a large number of expert-labeled data. This paper proposes a new type of deep network Deep & Context-wise & Requirements Operation Role Network (DeepCorp) for EC detection. It introduced entity context and requirement operation role information, and multi-layer perceptron (MLP) was used to implicitly fuse embedded representation to realize the in-depth semantic expression of the requirement entity, so as to judge the semantic similarity of the entity. Experiments on the open requirement document repository show that DeepCorp can achieve the precision of 96.72%, the recall of 96.67% and the F1 value of 96.69%, which has an average increase of 1.27% compared with the existing methods.

     

/

返回文章
返回