Research On Knowledge Transfer Based Implicit Discourse Relation Recognition

Implicit discourse relation recognition is a basic and important task in natural language processing,with the purpose of automatically inferring the semantic relation(e.g.,causality,etc.)between two arguments(clauses or sentences)under the absence of connective information,which is the bottleneck of discourse structure analysis.In recent years,a series of implicit discourse relation recognition methods based on deep learning have emerged under the in-depth development and successful practice of deep learning technology in the domain of natural language processing.It is one of the current research hotspots that this kind of method achieves better recognition performances than the traditional method based on human-made features.To deal with the problem that the existing implicit discourse recognition model based on deep learning is prone to overfitting when training in relatively few discourse relation corpora annotated manually,this paper studies knowledge transfer based implicit discourse recognition methods that make full use of existing resources to improve the performance of recognition.The content of the research mainly includes the following two aspects:1)A Knowledge Distillation based Implicit Discourse Relation Recognition Method.Annotating staff usually inserts an agreeable connective to assist the annotating of implicit discourse relations.According to the above situation,a knowledge distillation based implicit discourse relation recognition method is proposed,which aims to improve the recognition performance by using the inserted connective information in corpus annotation.To be specific,a connective enhanced model is constructed to fuse the connective information,and then the knowledge learned from the connective enhanced model is transferred to the implicit discourse relation recognition model based on knowledge distillation.Experimental results show that the proposed method obtains better recognition performance than the similar benchmark methods on the commonly used English PDTB dataset.2)A Mutual Learning Based Chinese-English Implicit Discourse Relation Recognition Method.There is a certain amount of discourse relation annotating corpus in both Chinese and English.Some differences in the defined categories of discourse relation respectively exist between them,but they can significantly enhance with each other.Therefore,this paper proposes a mutual learning based Chinese-English implicit discourse relation recognition method,which is to improve the recognition performance of implicit discourse relation in both Chinese and English by exploiting both Chinese and English discourse corpus.Chinese-oriented and English-oriented implicit discourse relation recognition models are first constructed.Then,the two models are trained jointly based on mutual learning to achieve mutual promotion in Chinese and English annotated corpus.Experimental results demonstrate that the proposed method gains better recognition performance than the baselines on both the English PDTB dataset and Chinese CDTB dataset.