Predicting Chinese Irony Intensity Using Deep Learning

In the sentiment analysis of text,there are two main types of research.It is Categorical Sentiment analysis and Dimensional Sentiment analysis respectively.The study of Dimensional Sentiment analysis can provide more fine-grained sentiment information.At present,the lack of Chinese dimensional affective corpus and the low accuracy of dimensional sentiment analysis,which make difficulties in applying dimensional approach for Chinese text.In the study of the sentiment analysis,an ironic expression is a special that implies the opposite of literal meaning by using the exaggerated modal words.Contextual information and background knowledge are necessary to understand its intentions.So,it is hard to predict Chinese Irony intensity because of the specialty of the structure in Irony expression.In order to solve the problem,we carry out the study of Chinese Irony intensity prediction in Valence-Arousal-Irony three-dimensional space of a multi-dimensional Chinese Irony corpus.Firstly,the study use CNN,LSTM and their combinations as based model to evaluate the multi-dimensional Chinese Irony corpus for prediction of VAI ratings,because the method using deep learning can extract feature automatically.The analysis for modeling the data in the multi-dimensional Chinese Irony corpus and propose three model to predict irony as follows:(1)Multi-dimensional Linear Adjustment model for Irony prediction(MDLA).Firstly,the relationship between valence,arousal and irony dimensions are analyzed and visualized.And then,the MDLA uses the multi-dimensional relations captured by linear regression to refine the predictive results output by the neural networks.The experimental results show that MDLA model show better performance than baseline model.(2)Multi-dimensional Feature Combination model for Irony Prediction(MDFC).The adjustment of predictive result output by treating each dimension independently can cause some loss of dimensional relation between VAI.It can influence the performance.Therefore,the study use the deep learning neural network model to generate the most important feature vector of the text,and then use our multi-dimensional feature combination layer to get the affective vector for Irony prediction.It can extract more relation feature for sentiment prediction to improve the performance.The experimental results show that MDFC model achieve better performance than MDLA model.(3)Multi-dimensional Relational model for Irony Prediction(MDR).The model combine the advantage of MDLA model and MDFC model.Firstly,the study use the MDFC model to combine the feature vector in all dimensions,calculate the affective vector,to predict the VAI ratings.And then,using the MDLA model to refine the predictive results.The experimental results show that MDR model achieve best predictive result,which achieve better performance than MDLA and MDFC model.In the pre-trained word vector produced by Glove and Word2vec,the result on evaluation method including Mean absolute error(MAE)and Pearson correlation coefficient(r)achieve better performance.The MAE improved for 18%and 14.5%,the r improved for 99.6%and 72.7%,respectively.