Implicit Discourse Relation Recognition Based On Deep Neural Networks

Implicit discourse relation recognition is one of the most important task of discourse parsing in natural language processing.The main purpose of this task is to determine the discourse relation of two sentences or clauses in a discourse.Usually,these classified sentences or clauses are called arguments.Discourse relation reflects the semantic coherence between the arguments,such as temporal relation,expansion relation,comparison relation,and so on.It is very meaningful to correctly classify the discourse relations since it can help the machines understand the long texts,and help some downsteam tasks such as machine translations and text abstractions.Implicit discourse relation recognition has no connective and other surface feature for easy classificaiton,and it heavily relies on semantic un-derstanding of texts.Only with semantic understanding of the arguments,can we classify the relation correctly.In this article,we proposed a deep enhance representation model,which can incorporate more information by representations of several levels,and produce a semantic representation re-flecting the discourse relations.This better representation can make the classifier more easily get the semantic information to improve the classifi-caiton performance.Besides this,another severe problem is lacking data.It is very hard to get argument pairs of implicit discourse relations and to annotate them,so the data are not sufficient.The standard dataset PDTB 2.0 has only tens of thousands of data.So we want to fully utilize these data.So in this article,we proposed a memory component to memorize the training data and help the classificaiton.The memory component store the semantic representation of argument pair produced by an encoder of a relation classifier.It use this representation as a key and also store the corresponding relaiton.During testing,the model can retrieve similar intances in training set by the semantic representation,and get their discourse relations to help the classificaiton of test intances.In this work we verified these methods by doing experiments on standard dataset.These methods can improve the performances of this task,and obtain the state-of-the-art results.