A Study On Discourse Coherence Based On Event Chain

The study of discourse coherence has been playing an important role in both language learning and information retrieval and it has been the hot spot in text linguistics in recent years. Theories concerning discourse coherence took on multiple and diverse perspectives. The most representative theories are Cohesion and Register(Halliday&Hasan,1976), Macrostructure(Van Dijk,1977), Rhetorical Structure Theory(Mann&Thompson,1988), Thematic Progression Theory(Danes&Fries,1974), Centering Theory(Grosz et al.,1986), and among others. These theories underlie computational models which can be categorized into three kinds from the aspect of representation:relational model, logic model and entity model. We may find these models have some defects:some cannot be realized within the realm of current computing technology; others have relatively lax linguistic assumption foundation. Therefore we propose a novel discourse theory and model based on event chain to reinforce linguistic assumption on the models and simultaneously utilize the mature computing technology at the level of sentences and events. We hold the idea that a coherent text consists of events and all the events within the text constitute the event chain(s) according to certain semantic relations. Event chains are based on lexical chains and this ensures the computing workability in practice.This dissertation consists of8chapters. The major content and viewpoints are as follows:Chapter1:Introduction. This chapter consists of6sections. It includes background information on discourse coherence, the status quo of discourse coherence research at home and abroad, the problems related to discourse coherence, our research method, the major content of this dissertation and the layout of the dissertation.Chapter2:an overview on discourse coherence theories.This chapter gives a brief review on major discourse coherence theories and their computing workability.Chapter3:an overview on event theories and event structure. On the basis of miscellaneous event theories, we define event as the concept denoted by the predicate within a sentence. Each event has its structure from its attribute of time and space. Traditional event research is restricted on syntax at sentence level and it has not been applied into the analysis of discourse structure. However, we hold that the semantic radiation of the predicate is not only restrained within a sentence but transcends the sentence and penetrates the discourse as well. We address this configuration as event chain. Event chain determines the degree of discourse coherence. Meanwhile, the well-formedness of events enables their wide application in natural language processing and thus provides technology warrant to discourse coherence. The semantic structure of a discourse assumes a hierarchy:macrostructure comes at the top, and microstructure follows. The third level is the entity relation, with lexical semantics at the bottom. Macrostructure is denoted by mirco-semantic unit---events and their relations. Events fall into6categories:activity event, relation event, cognition event, physiology event, statement event and existence event. Event relations fall into11categories:coordinate, subsequent, interpretive, selective, progressive, transitional, conditional, hypothetical, consequent, purposely and coreferent.Chapter4:Event chain model. This chapter is to build event chain model from the aspect of computation. We propose the representation method to depict events and event chains within a discourse. We adopt predicate-argument structure in the analysis of event structure and event relations in the analysis of event chains.Chapter5:The creation of event chain within the text. This chapter consists of5sections:section1is on the buildup of lexical chain, section2the buildup of events, event chains and event relations; section3on how to calculate coherence based on event chain; section4on the assessment strategy of event chain; section5concludes the chapter.Chapter6:The construction of discourse coherence resources. This chapter consists of4sections. Section1is the overview of discourse coherence resources. We comment on the annotation principles of English discourse coherence resources and Chinese discourse coherence resources. Section2is to annotate Chinese discourses from the aspect of events and event chains. It includes:the selection of the discourse, the annotation principle, quality control, annotation efficiency and annotation platform. We download news materials from internet and use them for annotation. Detailed criteria for annotation are drawn up and annotation platform is set up. Section3displays the partial outcomes of the annotation. Section4discusses the problems in the course of annotation. Section5summarises this chapter.Chapter7:The application of event chain. Event chain can be applied in language teaching and learning and information retrieval. Section1is concerning the application of event chain into Chinese essay-writing for overseas students. Section2is concerning how to utilize event chain to generate summary. Section3summarises this chapter.Chapter8:The conclusion.The orginal and creative viewpoints of this dissertation can be summarized as follows:(1) Theoretically, we adapt the research of discourse coherence into an analytical problem based on computability and apply it into the evaluation of student writing and sentence-ranking.(2) From the aspect of representation mechanism, we propose event chain as the novel discourse representation method and consider it as the necessary conditons for discourse coherence. This mechanism not only duplicates the semantic content of the discourse, but ensures the computability in discourse as well.(3) From the aspect of analytical model, we propose event chain model be used to tackle deep semantic analysis on large-scale Chinese texts.(4) From the aspect of resource building, we annotate large-scale event chains within the discourse. This annotated resource can be used as training set and test set for natural language processing. It can also be used in semantic parsing for the general purpose.