Research On Sarcasm Detection Incorporating Adversarial Learning Framework And Sentiment Knowledge Base

With the rapid development of Internet,more and more people have expressed their opinions on the Internet.Existing statistical learning based sentiment analysis systems normally may correctly classify the sentiment polarity of text.However,these systems usually fail to predict the sentiment polarity of sarcastic text.Sarcasm is a type of complicated language phenomenon which translates literally positive or negative discourse into opposite emotional polarities.Currently,most existing methods of sarcasm detection may be camped into rule-based,machine learning based,and deep learning based approaches.Both rule-based and machine learning based methods rely on domain expert rules and artificial designed features.Deep learning based approach can effectively improve the generalization performance of the model,but requires large-scale high-quality annotation data.Target to the problems of existing methods,this thesis studies the sarcasm detection method based on adversarial learning framework and sentiment knowledge base.To address the lack of high-quality annotated sarcasm data,we propose a sarcasm detection model based on adversarial learning frameworks.The base model of sarcasm detection is based on Convolution Neural Network(CNN)with attention mechanism.Based on this,adversarial-example based learning framework is investigated.This framework generates pseudo samples during the training phase to enhance the robustness of the classifier.Next,domain adaptation based adversarial learning framework is investigated.It leverages cross-domain satiric data to further improve the performance of sarcasm detection.Furthermore,we investigate the incorporation of adversarial samples based and domain adaptation based adversarial learning frameworks.The experimental results on the three public IAC sarcasm detection datasets(Generic?Hyperbole?Rhetorical Questions)show that adversarial learning framework brings more than 3% F1 performance improvement.The adversarial learning frameworks improve the performance at the training level,but ignore the linguistic expression characteristics of sarcasm itself.Based on the external sentiment knowledge base and two intuitive linguistic characteristics(semantic consistency and emotional consistency) of sarcasm,we introduce text similarity matrix and emotional interaction channel and propose sarcasm detection model based on sentiment knowledge bases.The experimental results show that the proposed model achieves more than 7% F1 improvement on the Hyperbole dataset compared to the baseline model.On the Generic and Rhetorical Questions datasets,the proposed model brings more than2% and 3% F1 improvement,respectively.Finally,we investigate the incorporation of sarcasm detection model based on sentiment knowledge base and adversarial learning framework.The final experimental results demonstrate that the incorporated method achieves more than 4% and 5% F1 improvement on Hyperbole and Rhetorical Questions datasets compared to the baseline model,respectively.