Research On Short Text Emotion Classification Based On Capsule Neural Network

With the development of social networks and e-commerce platforms,a large number of users post their opinions on hot events and comments on products on the Internet every day.A large part of the information is in the form of short text,which hide the emotional information of users.Mining the emotional information of short text can help the government quickly understand people's views on hot events,and also help enterprises understand users' preferences and adjust business strategies.Therefore,the emotional classification of short texts is of great value.In this paper,we first study the sentiment classification of short text,and then study the emotion classification based on sentiment classification.In the stage of short text sentiment classification,a dynamic global-local attention network based on capsule is proposed.High-quality global features are extracted using capsule neural networks.In order to explore the relationship between global features and local features,a globallocal attention mechanism was designed.In the emotional classification stage of short text,a global-local enhancement network is proposed.In order to improve the quality of features,a group-wise enhancement mechanism has been introduced.To solve the problem of information loss in traditional pool operation,a global-local pooling mechanism based on capsules was proposed.The main work and innovative achievements are as follows:(1)In short text sentiment classification task,capsule neural network is introduced to extract high quality global features.The dynamic routing process in capsule network can dynamically adjust the weight of input features and mine more hidden information.(2)Most sentiment classification methods regard global features and local features of text as two independent parts,while ignoring the relationship between them.To explore the relationship between global features and local features,a global-local attention mechanism is designed,which can dynamically adjust the importance of local features by using global features.(3)In the task of emotional classification for short text,the group enhancement mechanism was introduced to improve the quality of phrase-level and sentence-level features.Group-wise enhancement mechanism can effectively enhance the important features and weaken the unimportant features.(4)In the task of emotional classification for short text,aiming at the information loss problem in the traditional pool operation,a global-local pooling mechanism based on capsules is designed,which can filter out the phrase-level features which are more closely related to the global features.