基于知识蒸馏的隐式篇章关系识别

Knowledge Distillation Based Implicit Discourse Relation Recognition

由于缺少连接词信息,隐式篇章关系识别模型需要基于两个论元(子句或者句子)的语义来推导它们之间的篇章关系,但目前性能还比较低。对于语料标注人员而言,隐式篇章关系的标注是很困难的,他们通常先插入一个合适的连接词用于辅助隐式篇章关系的标注。基于上述情况,文中提出了一种基于知识蒸馏的隐式篇章关系识别方法,其目的是利用语料标注时插入的连接词信息来提高识别的性能。具体地,先构建一个连接词增强的模型用于融合连接词信息,然后基于知识蒸馏的方式把连接词增强模型学到的知识迁移到隐式篇章关系识别模型中。实验结果表明,在常用的PDTB数据集上,所提方法取得了比同类基准方法更好的识别性能。

Due to the lack of connectives,implicit discourse relation recognition models infer the semantic relations(e.g.,causal)between two arguments(clauses or sentences)based on their semantics.The performance of these models is still relatively low.It is also very difficult for corpus annotators to annotate implicit discourse relations.They usually insert an appropriate connective to assist the annotation of an implicit discourse relation instance.Considering the above,a knowledge distillation based method is proposed for implicit discourse relation recognition to take use of the connectives inserted during corpus annotating.Specifically,a connective-enhanced model is constructed to integrate the connective information,and then the integrated connective information is transferred to the implicit discourse relation recognition model via knowledge distillation.Experimental results on the commonly used PDTB dataset show that the proposed method achieves better performance than the baselines.