Research On Emotion Recognition For User-Generated Content

In recent years,people have actively participated in the Internet environment through community-based websites and social networks,resulting in a large amount of usergenerated content.Emotion recognition for user-generated content can be applied to areas such as human-computer interaction and product strategy adjustment,which is helpful for the development of advanced artificial intelligence and has practical significance.It has become a hot issue.However,user-generated content has the characteristics of short content,sparse semantics,and fast update iteration.Although existing research on emotion recognition is relatively mature,it still has certain limitations when applied to specific scenarios of user-generated content.In view of the above problems,the main work of this article can be summarized into the following four aspects:1.An experiment was designed to discover the problems of the existing methods in user-generated text emotion recognition scenarios and explore the impact of noise(content without emotion)in user-generated content on the effect of emotion recognition experiments.In this thesis,typical models based on traditional machine learning,and deep learning are selected,and the three experimental schemes proposed in the article are tested on a unified experimental basis.It is found through experiments that noise is not conducive to single emotion recognition,but an appropriate proportion of noise is helpful for multiple emotion recognition.In addition,each model has its own suitable application scenarios,but when applied to user-generated content,it is difficult for them to capture the emotional clues contained in the context of user-generated content.And for situations where multiple emotions are contained in the same user-generated content,it is often not accurately recognized and the training efficiency is low.2.This thesis proposes a recurrent gate model that can identify a single emotion and its corresponding expression phrase simultaneously.In this thesis,bilateral long and short memory models are combined with the attention mechanism to capture emotional expression phrases in the context,and make them and the possible emotions in the text mutually gain through loop iteration,which addresses the shortcomings of existing methods that cannot capture emotional clues in context.The experiment results show that the model is better than the benchmark models in the recognition of emotional expression phrases and single emotion recognition,and the model works best when the number of iterations is 5.3.This thesis proposes a multiple emotion recognition model suitable for the characteristics of user-generated content.The model is based on a convolutional neural networks and replaces the pooling layer with a self-attention mechanism,which not only avoids the loss of multiple emotional elements in the pooling operation to filter features,but also allows the model to converge faster.In addition,this thesis visualizes the convolution windows based on the self-attention weight value,and can obtain key emotional expression phrases.The experiment results show that our model performs better than the benchmark models and is robust to noise.4.This thesis proposes a practical application scheme based on the method proposed in this thesis,and substitutes specific application cases for application and analysis,which proves the practical significance of the models proposed in this thesis.