Research On Sarcasm Detection In Social Media Text

Sarcasm detection,which aims to identify whether a sentence is sarcastic or non-sarcastic,is a sub-task in natural language processing.Sarcasm often expresses the op-posite of literal meaning and conveys negative sentiment.Recently,sarcasm detection has attracted extensive attention,due to its great significance to analyzing the sarcasm-contained text for sentiment polarity and mine user’s opinions.Thus,sarcasm detection is efficient in practice.With user-generated data ballooning on online social media and the rapid develop-ment of deep learning techniques,sarcasm detection task has achieved promising results.However,there are still challenges remained to solve,such as:(1)The current approaches are limited by the quality and quantity of sentiment annotation of public sarcasm datasets while modeling the correlation between sarcasm detection and sentiment analysis.(2)Most current models fail to consider context semantics while capturing intra-sentence incongruity.(3)Few approaches model the dependency between sarcasm and common-sense knowledge.(4)Few approaches explore the role of sentiment information between context and content in sarcasm detection task.Therefore,to address such problems,we focus on sarcasm with and without context,and conduct research and empirical analysis on several key technologies with respect to sentiment,semantics,commonsense knowl-edge,and context.The main contributions of our work are summarized as follows:Firstly,to address the sentiment annotation of public sarcasm datasets as well as the issue of lack of context semantics modeling while capturing intra-sentence incongruity,a multi-task learning framework is proposed to model sentiment clues and employs soft sentiment annotation without human involvement,and integrates semantic information while modeling context incongruity.Experimental results on datasets show that the pro-posed model yields better performance on sarcasm detection on different domain datasets with the help of sentiment clues and incongruity with semantics.Secondly,to address the lack of modeling the dependency between commonsense knowledge and sarcasm,we propose a commonsense knowledge-aware heterogeneous graph attention network framework.The proposed approach models both text and rel-evant commonsense knowledge,which leverages commonsense knowledge as well as context semantic information to help understand the implied sentiment behind literal text.Experimental results on the benchmark datasets show that the proposed model is superior to the baseline models with the help of commonsense knowledge and context semantics.Thirdly,aiming at the problem of not considering the sentiment correlation between context and target content as well as the dependence between sarcasm and world knowl-edge,we propose a sarcasm detection framework based on contextual sentiment and ex-ternal knowledge in context.This framework captures the sentiment information of con-text and target content respectively and leverages the relevant external knowledge to help understand the potential semantics of the dialogue.Experimental results on the bench-mark datasets show that the performance of the proposed model is better than the baseline models that only use conversational context.This paper conducts research on key technologies of sarcasm detection in social media text from three aspects: semantics,sentiment,and dependence on commonsense knowledge and context,which the approaches based on the three kinds of information effectively improve the performance of sarcasm detection in corresponding scenarios.In future work,we will explore more refined sarcasm features,investigate sarcasm detec-tion in multi-modal scenarios,and research the role of user interaction in social network structure on sarcasm detection.