A Study Of Conversational Implicature-computing Models And Applications Based On Bayesian Theories

The notion of conversational implicature,the implicit meaning implied by the utterance beyond the literal,is one of the single most important ideas in pragmatics.Since conversational implicature usually requires understanding the speaker’s beliefs and intentions in highly context-dependent or ambiguous conditions,for natural language processing,the calculation of conversational implicature not only needs to process the information of language itself,but also requires to have the ability to reason about background knowledge,context and intention.Currently existing intelligent systems are relatively weak in dealing with conversational implicature.A new generation of deep learning and data-driven natural language processing systems are developed without incorporating pragmatic theories related to conversational implicature.Not only do they not make full use of existing theoretical resources,but their performance and interpretability are also limited.Therefore,this dissertation takes conversational implicature as the research object,attempting to model the process of expression and inference of conversational implicature from the perspective of probability and computer science,to give computer the ability to infer the implicature preliminarily.Conversational principles are explained according to the underlying probability,and then the rationality of conversational principles is interpreted from a computational point of view.This research aims to establish the computational pragmatic models of conversational implicature.The whole process can be divided into three stages:model preparation,model building,result analysis and application according to pre-modelling,modeling and post-modelling.In the model preparation stage,this thesis needs to explain the necessity and feasibility of using Bayesian theories to establish the computational models of conversational implicature.Regarding the necessity,the second chapter points out that there are some problems in the current research on the computational pragmatics of conversational implicature,and the third chapter lists the main advantages of exploring conversational implicature based on Bayesian theories.Combining them can show that Bayesian theories play an irreplaceable role in the study of conversational implicature by other theoretical perspectives.Regarding the feasibility,the third chapter puts forward four basic ideas in Bayesian theories,and explains how they fit with conversational implicature.In the model building stage,the thesis describes the interlocutors’expression and inference process of the conversational implicature by establishing computational models based on Bayesian theories.Chapter Four first establishes a Bayesian inference model of scalar implicature,which describes the speaker’s judgment on the state of the world in the cognitive context and the possibility of producing a specific utterance in a given state by using prior and conditional probability,respectively.After calculation,the model can obtain the posterior probability based on the intention judged by the utterance,to obtain the scalar implicature.Based on the above process,the probabilistic analysis of scalar implicature and the simple man-machine dialogue are implemented by programs.Besides,Chapter Four describes the logical structure in the expression and inference of particularized conversational implicature,and its conclusion is that when the speaker produces utterances based on background knowledge and intention according to deductive,inductive,and abductive logics,the hearer’s inference is abductive,abductive,and deductive,respectively.After summarization,it is found that the logic of the hearer’s utterance inference is obtained by fixing one premise and exchanging the position of another premise and conclusion based on the expression logic of the speaker’s utterance.Finally,Chapter Four builds a Bayesian belief network model of particularized conversational implicature,which describes the listener’s inference process of implicatures in pope question as a response,tautology as a response,idiom as a response,simile as a response,and relevance.Based on the logical structure,the model stores the cognitive context and utterance information into the corresponding nodes in the graphs,and determines the probability of each answer to the context utterance by drawing a directed acyclic graph,designing a reasoning algorithm,and constructing a conditional probability table to get the final conversational implicature.In the stage of result analysis and application,first,based on the inference probabilities of implicature output by the two Bayesian models in Chapter Four,the underlying probability mechanism for the implementation of conversational principles and criteria is described.The probability values illustrate why communicators need to adopt these principles and maxims in the expression and inference of conversational implicature,that is,the choice supported by pragmatic principles and maxims tends to have the greatest posterior or output probability.Then,based on the responses of the response node“CU”to the contextual utterance in the Bayesian belief network model of particularized conversational implicature in Chapter Four,Chapter Five defines categories consistent and inconsistent with the contextual utterance as“the types of conversational implicature”,and uses them as labels or features to apply them to sentiment analysis and text classification to further discover new rules or patterns.The experimental results of sentiment analysis show that there is a clear statistical dependence between the type of conversational implicature and sentiment score.The utterance source used to distinguish the sentiment score is the response utterance,and whether the sentimental score of the conversational implicature is consistent with that of the literal meaning is related to the selected methods in sentiment analysis.The results of text classification show that if based on the Optimal Feature Sets F₁ and F₂,the text classification of conversational implicature can be performed based on lexical features,and the results of text classification will not be different due to different contextual utterances or the types of conversational implicature.Finally,Chapter Five illustrates that the models proposed in this study can solve the challenges of non-literal,individual variability,context sensitivity,uncertainty,ostensive-inference and dynamics brought by conversational implicature to computational pragmatics at the micro level and technical details.The solutions to the above challenges are given,which solve the difficulties encountered by computational pragmatics in conversational implicature to a certain extent.This study can provide reference for the further development of computational pragmatics and the integration of interpretable conversational implicature modules in natural language processing.