Sentiment Analysis And Affective Text Generation Of Social Media Based On Generative Adversarial Nets

Social media serves as an important platform for people to express feelings and lead off emotions.The sentiment detection based on social media data is meant to analyze the attitude and preference of the users which is useful in many related tasks,such as public opinion analysis,traffic condition analysis,advertisement recommendation and safety monitoring.On the other hand,with the rapid development of the artificial intelligence,affective text generation has boosted into a new research hotspot and social media provides affective text generation with abundant affective data.The affective text generation can improve the quality of human-computer interaction,benefitting comment generation,data augmentation,dialogue generation tasks and so on.One of the common challenges in sentiment detection and affective text generation is the affective gap.Based on the previous studies,our work finds that the Generative Adversarial Nets(GAN)is able to capture the overall association of data distribution;the feature of the discriminative model is effective in improving classification result;and the reinforcement learning can help the GAN in gaining high-dimensional semantic information.Considering these characteristics,we have studied how to get robust sentiment representations in sentiment detection task and how to add abstract affective features into affective text generation task based on the GAN.The main contributions are as follows:1.In the sentiment detection task,our work focuses on solving two problems:the affective gap and the labor-intensive data annotation.We have brought out a method based on the feature fusion of the feature of the discriminative model and the feature of deep learning model to increase the robustness of sentiment representation which is contribute to eliminating the affective gap of the sentiment representation and the real sentiment.In order to add affective information of the GAN model,our work uses a training set annotated by emoticons.The experimental results shows that the features of the discriminative model are useful in improving the performance of the sentiment analysis model.To the best of our knowledge,this work is the first one to perform sentiment detection based on GAN features.2.In the affective text generation task,we mainly discuss two problems:first,the affective information is not adequate in text generation;and second,the grammaticality and expressions of emotion is incompatible.We reinforce the affective learning ability relying on the affective feedback within a LeakGAN model,which is constrained by the affective label.Our work has designed several different affective feedback strategies and tested the influence of the strategies on balancing the expressions of emotion and the grammaticality.The experimental results implies that our method is able to endow the GAN with the ability to generate sentiment-oriented text.To the best of our knowledge,this is the first study that applies GAN to generate the Chinese with emotion.3.Facing the problem of the high-cost and limited coverage of data in traffic condition analysis task,we have proposed a low-cost,wi de-cove rage,and real-time traffic condition analysis system based on our first part of work.The method is tested in both the time and the space dimensions.And the results have proved the effectiveness and feasibility of our method.