Research On Data-driven Shallow Discourse Parsing

The goal of shallow discourse parsing is to get the shallow discourse structure of an article by analyzing the logical and semantic relations between sentences,such as juxtaposition relations,progressive relations,transition relations,causality relations and so on.This discourse structure generate by shallow discourse parsing can be applied to many natural language processing tasks,e.g.machine translation,automatic document summarization,question answering system etc.The shallow discourse parsing task can be divided into multiple sub-tasks,including identification of shallow discourse connectives,recognition of the relation between the shallow discourse and the argument(Arg1 sentence locator,Arguments word locator),shallow discourse explicit relation classification,shallow discourse implicit relation classification.Based on studies of previous work,this paper try to make progress from the following perspectives.(1)Shallow discourse parsing system based on traditional discrete features.We summarize previous work on shallow discourse parsing and carefully design a complete analysis system of shallow discourse parsing based on traditional discrete features.For the Arg1 sentence locator,we propose a window based sequence labeling methods according to the argument of sequence constraint,find the Arg1 position in discourse.We employ some effective new features and improve the accuracy of whole argument recognition.(2)Implicit relation classification of shallow discourse structures based on neural networks.The implicit relation classification of shallow discourse structure is the key problem in the analysis of shallow discourse parsing.Different from the explicit relation classification,implicit relation lacks explicit connectives as an effective indication of the logical and semantic relations implied between the two sentences.So the implicit relation classification only explore the logical semantic relations imply by the two asyndetic arguments.Following the previous methods implicit relation classification using deep learning,this paper attempts and implements a method based on neural network with self-attention mechanism.(3)Implicit relation classification of shallow discourse structure based on syntactic information.Considering that raw text is not enough to provide sufficient information,this paper presents a method that takes automatic syntactic information as extra input of the model and provides additional information,enhancing the performance of the model.Based on the traditional machine learning method and deep learning method,try to improves the performance of shallow discourse structure analysis system and implicit relation classification,and achieves some preliminary results.We expect these research results to be applied to other natural language processing tasks in future.