Deep Neural Networks Based End-to-End Discourse Parsing

Discourse relations refer to the logical relationship between two text units,such as cause,contrast and so on.Discourse parsing is given free texts as input and returns discourse relations contained in the texts.It can benefit many downstream natural language processing(NLP)applications,such as QA systems,statistical machine translation(SMT),automatic summarization.The previous work mainly focuses on the subtasks of discourse parsing,such as identifying discourse connectives,labeling arguments and classifying the sense of explicit or implicit discourse relations.The first part of this work is to build a well-designed English discourse parser.Based on Lin et al.[7]'s discourse parser,more effective features and components are proposed in our parser,and we achieved the first place in CoNLL-2015 Shared Task—Shallow Discourse Parsing.In order to build an explicit discourse parser for practical application,we construct two different classifiers with different feature sets for two arguments respectively,which significantly improves the performance of the arguments extraction.This part of work has been published at KSEM 2015 conference,and we won the best paper award.Due to the lack of Chinese discourse corpus,only few study has been carried out on Chinese discourse relations.With the release of CDTB corpus,the second part of this work is to analyze the different discourse characteristics between Chinese and English in depth,and builds an effective Chinese discourse parser.This system ranks second in CoNLL-2016 Shared Task-Multilingual Shallow Discourse Parsing-Implicit discourse recognition is the bottleneck of the whole discourse parser.Previous work often extracts lots of linguistic features combined with machine learning algorithms to identify the implicit discourse relations,which requires experts' domain-specific knowledge and has poor generalization performance.With the development of deep learning in NLP,the third part of this work firstly proposes to use convolutional neural network to do implicit discourse relations identification,and this system ranks second in CoNLL-2016 Shared Task.Furthermore,this work proposes an attention neural network to further improve the performance.However,lots of training data are required for training well-performed neural networks.In order to solve the limitation of annotation data,the fourth part of this work proposes a multi-task neural network to further improve the performance of implicit discourse recognition by using a large number of synthetic implicit discourse relations.The third and fourth part of this work has now been submitted for EMNLP 2017.In this paper,lots of experiments have been carried out for Chinese and English discourse paring.The experimental results show the effectiveness of our end-to-end English and Chinese discourse parsers,and using the atteantion neural network and multi-task neural network can help improve the performance of Chinese and English implicit discourse relations recognition.