Sentiment Analysis Method for Web Short Texts Based on Fusion Features

Abstract

Danmaku is a special kind of short text, highly associated with video content, with few features and sparse semantics. Existing methods only consider the text itself and are not suitable for sentiment analysis of danmaku. To solve the above problems, a dataset of time-based synchronized videos for annotation is firstly constructed. Then, a dual-channel sentiment analysis method based on text and time is proposed. The text channel uses ERNIE and TextCNN to extract the deep semantic features of words and characters of danmaku, which introduces external knowledge and enhances the feature representation; the temporal features associate danmaku with the video content; after feature fusion, the BiLSTM combined with attention mechanism is used for sentiment classification. The experiment results show that the method is better than the mainstream models and can be effectively applied to the sentiment analysis of danmaku.