Using Attention Penalty Term And Reinforcement Learning To Implement Multi-label Emotion Classification

With the vigorous development of the Internet,more and more netizens make their comments on social platforms,which contain a large amount of emotional information.The analysis of the emotions contained in the text can provide support for many tasks such as public opinion analysis,and has high application value.Since each text posted by users on social platforms usually contains a variety of emotions and these emotions are correlated with each other,capturing such correlation can help to accurately identify the emotions in the text,so text emotion analysis should use a multi-label classification method which can model correlation among emotion.The current deep learning methods based on sequence-to-sequence model can better capture the correlation among emotion labels than earlier works.However,when multiple emotions co-exist in the text,decoder of these methods paying attention to the text information in similar position at different times,then tends to generate semantically similar emotion labels,makes it difficult to identify various emotions with large semantic differences in the text.In addition,the cross-entropy loss function adopted by these methods depends on the order of emotion labels,but the existing emotion labels in the datasets have no specific order of labeling,which leads to the problem of wrong penalty.In view of the above problems,this paper works as follows: Firstly,since existing methods are difficult to identify multiple emotions with large semantic differences in text,this paper proposes a multi-label emotion classification method with attention penalty term.The sequence-to-sequence model is used to model the high-order correlation among labels,and the position difference among the emotion labels is captured by the attention penalty term.The experimental results show that,compared with the best existing methods,the method of introducing attention penalty term improves the classification performance on both Chinese and English datasets.Secondly,in order to solve the problem of wrong penalty caused by label order,this paper proposes a multi-label emotion classification method based on reinforcement learning,which transforms the sequence generation process into the set generation process by using the delayed reward so that multiple emotion labels can be regarded as a set.The experimental results show that,compared with the method proposed by the existing latest work,the method with reinforcement learning has a further improvement in the classification performance on Chinese and English datasets.