Language Device And Implicit Emotion Identification Based On Language Granularity

As an import pragmatic concept in natural language processing technology,language device is a tool for helping humans to understand the material world and strengthening the expressiveness,and which is commonly used in literary works.From the perspective of pragmatics,texts carved with language device convey emotions in an implicit and obscure way without emotional words.The expressions pose a huge challenge to automatically understand emotions of texts.In order to resolve the puzzle,this paper studies the identification methods of language device and implicit emotion under multiple language granularities,which provide important technical supports for intelligent education and other fields.Due to resource shortages and immature technologies in this study,we construct three knowledge bases and develop various methods under paragraph,sentence and word granularities.The main contents and innovations in this paper are as follows:(1)Knowledge bases construction of language device and emotion.Resource shortages,incomplete category coverage and inconsistent classification standards are serious problems in existing knowledge bases of language device and implicit emotion.In order to deal with these problems,this paper constructs three knowledge bases of language device and emotion for language style,rhetoric and metaphor from three language granularities,respectively.Specifically,we analyze textual characteristics about implicit emotion expressions caved by language device,collect raw corpus,define annotated standards,and perform the kappa consistency test in order to ensure the quality of knowledge bases.The constructed knowledge bases provide solid data supports for related method researches.In addition,a statistical analysis is carried out on knowledge bases to mine data features about language device and emotion expressions,which provide some new ideas for subsequent methods.(2)Language style discrimination method based on hierarchical structure of labels.Language style is an important language device on paragraph granularity.Due to various categories and complex hierarchy structure of language style,a discrimination method based on hierarchical structure of labels is studied.Firstly,a hierarchical clustering method is applied to obtain the labels hierarchical structure of language style.Based on the structure,a SVM hierarchical classification method is designed to discriminate the language style of texts.Experimental results on constructed knowledge base show that,compared to the manually constructed hierarchy structure of language style,the automatically constructed structure is dynamically adjusted based on the associated knowledge between data tags,which is more conducive to correctly distinguish the language style.In addition,the hierarchical classification method is more flexible than flat classification,and is a more appropriate way to the answer of appreciative examination questions in literary works.(3)Rhetoric and emotion identification method based on multi-task learning.Rhetoric is an important language device and a significant expression way of implicit emotions on sentence granularity.Rhetoric is formed by flexible patterns,and there is a correlation between rhetorical and emotional categories.In order to take advantage of the features,we thus propose a multi-task learning framework that can encode the categorical correlation between tasks to improve the performances of rhetoric and emotion identification.The model captures the semantic and syntactic information of sentences via semantic and syntactic representation methods,respectively.Base on the representations,we design a gate mechanism based classification module and an emotion classifier,thus obtaining rhetorical and emotional correlation distribution representations,respectively.Then,a multi-label prediction by incorporating correlation representations is proposed,so as to get the rhetorical and emotional label sets of sentences.On constructed knowledge base,experimental results validate the benefit of the proposed model over state-of-the-art baselines for rhetoric and emotion identification tasks.(4)Metaphor and emotion identification method based on scenes distribution information of words.Metaphor is an important language device and a significant expression way of implicit emotions on word granularity.From the perspective of cognition,the essence of a metaphor is the understanding and experiencing of abstract domains(target scenes)in terms of specific domains(source scenes).Metaphorical expressions tend to have a stronger emotional effect,and emotional content is formed by compositing and interacting the meanings of the source and target emotional scenes in the metaphors.This paper identifies metaphor words and emotions of sentences by a pipeline method.Firstly,a novel semantic scenes inconsistency based neural sequence labeling approach is developed to identify metaphorical words,thus mining the word pairs about“source scene-target scene”in metaphorical sentences.Then,a metaphorical emotion identification method by compositing emotional scenes(MET-ES)is proposed.This method constructs emotional representations of sentences by utilizing the emotional scenes of words,and depicts the emotional reasons by fusing emotional representations of word pairs,which are conducive to the model in order to correctly identify the emotions of sentences.Experimental results on constructed knowledge base validate the effectiveness of the proposed model over state-of-the-art baselines for metaphor and emotion identification tasks.In addition,the roles of semantic scenes inconsistent and emotional scenes for metaphor and emotion identification tasks are verified via case studies.To sum up,this paper focuses on the problems and challenges in language device and implicit emotion identification tasks,constructs knowledge bases and carries out corresponding scientific researches from multiple language granularities.The knowledge bases and methods enrich relevant research fields,and provide data and technical supports for in-telligent education and other fields.