Research On Semantic Enhancement For Implicit Discourse Relationship Recognition

The purpose of implicit discourse relationship recognition(IDRR)is to distinguish whether there are some logical relationships between two discourse units(DUs)without obvious logical clues,then classify these logical relationships into more specific categories.However,most IDRR models rely on complex and numerous layers of neural networks to obtain hidden semantic information.Alternatively,this paper proposes simple and effective semantic enhancement approaches to help obtain higher quality semantic information,thus improving the final classification performance.This paper focuses on the research of IDRR from the following aspects:(1)Implicit discourse relation recognition as SRL-enhanced connective generation.IDRR is usually formalized as a classification task in previous studies.In this paper we alternatively propose a novel approach which regards IDRR as a text generation task that directly generates connectives given discourse units.Then the generated connectives are mapped exactly into one discourse relation without ambiguity.In particular,we first design a connective-to-relation table which converts unambiguous connectives to their corresponding discourse relations.Then this paper introduces two different connective substitution strategies to replace ambiguous connectives in training instances with an unambiguous one.Finally,IDRR is regarded as a sequence-to-sequence task whose targetside sequence consists of DUs enhanced with semantic role labels and a connective in between.Experimental results show that this proposed approach is simple and effective.(2)Implicit discourse relation recognition with multi-view contrastive learning.Previous researches on IDRR usually focus on designing effective discourse encoders.Differently,to better enhance the semantic representations of DUs,this paper introduces contrastive learning into IDRR so as to obtain representations of DUs with more differentiation.Specifically,this paper first uses a lightweight IDRR classification model.Then,to better learn representations of DUs,this paper explores the application of three different contrastive learning methods in IDDR from multiple views,including instancelevel,batch-level,and group-level.Finally,this paper combines above three multi-view contrastive learning objectives for better IDRR.Experimental results show that the proposed approach can effectively improve the performance of IDRR with only slightly increasing training time and introducing small additional parameters.(3)Constrained multi-layer contrastive learning for implicit discourse relationship recognition.This paper proposes coupling label-and instance-centered contrastive learning approach.Then this paper adopts this contrastive learning approach to outputs of the multilayer modules in IDRR model.In addition,this paper proposes a multi-level constraint contrastive learning approach to compensate the redundancy from the application of contrastive learning between multi modules,that is,the contrastive loss of higher layers in the baseline model should be smaller than that of lower layers.Experimental results show that our approach achieves a new state of the art.Experimental results show that the proposed approach achieves competitive performance compared to existing cutting-edge approaches.