Research On Machine Reading Comprehension Based On Frame Semantic Representation

The rapid development of Artificial Intelligence(AI)puts forward higher requirements for Natural Language Understanding(NLU).Machine Reading Comprehension(MRC)requires machines to read and understand text passages,and answer relevant questions about it,which is regarded as an effective way to measure language understanding.In recent years,with the rapid development of deep learning technology and the emergence of pre-training language models,the performance of MRC has been greatly improved.However,for questions that cannot be answered by only utilizing the information contained in the sentence itself,and need to be solved by integrating human prior knowledge,MRC methods only relying on sufficient trained data is no longer applicable.Therefore,integrating human experience knowledge into text representation model is an effective way to break through the limitation of only analyzing the text,which contributes to realize deep semantic analysis of MRC.Frame Net,as a typical semantic knowledge base,describes complex linguistic phenomena from the perspective of cognition,and provides a rich semantic description system for NLU.Thus it could be potentially leveraged to better understand sentences and has a unique advantage in improving the ability of language semantic understanding.Aiming at collaborative reasoning of language and semantic knowledge in MRC,the thesis investigates MRC methods from data and knowledge fusion representation and reasoning,multi-source knowledge representation and reasoning,and the mechanism of real exam scenarios based on Frame semantic representation.Main results are obtained as follows:(1)This paper has established MRC model based on data and Frame semantics fusion representation and reasoning,which aims at integrating Frame semantic knowledge into text semantic space.By combining text information and Frame semantic knowledge with neural network to improve the semantic understanding of natural language.Specially,frame semantic knowledge includes Frames,Frame Elements(FEs),and frame hierarchical structure.For Frame representation,this paper proposes 3 different Frame representation models,i.e.,Lexical Units Aggregation Model,Lexical Units Attention Model and Frame Relation Attention Model,which take full advantage of Lexical Units and Frame-to-Frame relations to model Frames.Frame-based Sentence Representation model integrates multi-Frame semantic information to obtain richer semantic aggregation for better sentence representation.In addition,Frame representation method for MRC explicitly leverages the text information and Frame semantic knowledge to assist the understanding of semantic scenario.For FEs representation,this paper proposes a novel Multi-Perspective Frame Element Representation method,which models FEs from three perspectives,i.e.,FE definition,Frame,and FE-to-FE relations,to improve fine-grained semantic describing ability of MRC.For multi-level Frame semantics representation,this paper proposes Frame-based Multi-level Semantics Representation model,which leverages Frame to extract multi-level semantics from sentences to display the different level semantic information of text.In addition,an attention mechanism is used to fuse different level semantic information,and reveal the semantic interaction of text,which can improve the semantic understanding ability of MRC.(2)This paper has established MRC model based on multi-source knowledge fusion representation and reasoning,which has the ability to express the deep semantics and structure of text.By taking full advantage of the complementarity in different knowledge to improve the multi-source knowledge joint reasoning ability of MRC.Frame semantics describes the semantic scene information of text,aiming to reveal the implicit semantic knowledge contained in the text and the internal thinking rules of the text.Syntax reflects the organization of the text.Frame semantics and syntax describe the semantic information of the text from content and structure perspectives.However,there are differences in the organization and presentation of different knowledge.Therefore,this paper proposes a MRC model that integrates syntax and Frame semantics,which aims at analyzing the semantic relevance between Frame semantics and other knowledge,and studying multi-source knowledge fusion method of language cognitive space and knowledge space for language comprehension.Specifically,a multi-source knowledge fusion representation method maps Frame semantics and syntax into the same semantic space,and further combines knowledge and text information with deep learning models to achieve knowledge-driven text semantics understanding and improve the multi-source knowledge joint reasoning ability of MRC.(3)This paper has studied MRC method for real exam scenes.For the characteristics of real exam scenes,i.e.,high complexity,strong comprehensiveness and great difficulty,this paper studies the MRC method of college entrance examination based on Chinese Frame Net(CFN),which cooperates with CFN and other semantic knowledge to improve the comprehensive analysis capability of MRC from multi-level and multi-dimension.The answer of college entrance examination involves explicit information,implicit information,micro semantics and macro semantics of text,which requires system to have the ability to understand and analyze the text from different levels and dimensions.Therefore,this paper proposes a MRC method for college entrance examination based on CFN,which cooperates with CFN and other semantic knowledge to mine the implicit information of the text,and improves the ability of MRC model to describe the micro and macro semantics of the text from lexical level and sentence level.Specifically,for lexical semantic analysis,this paper studies the automatic construction algorithm of lexical semantic relations,such as synonymous relations,relevance relations,semantic scene relations and antonym relations,to analyze the semantic relations between Chinese words.For sentence semantic analysis,this paper studies the sentence semantic relevance calculation method based on multi-dimensional voting strategy,which models the text semantic information from its own characteristics,similarity and CFN scene relevance dimensions.It also maps CFN knowledge into text semantic space,and makes full use of explicit and implicit information to depict the semantic relationship of sentences.In short,this paper aims to improve the comprehensive analysis and understanding ability of MRC by analyzing the text semantic from multi-level and multi-dimensional.