A Multi-feature Emotion Classification Model Based on LDA Subject Target Words

Abstract

In recent years, a variety of sentiment classification models based on deep neural networks have emerged. Most of the existing models are trained based on word embedding, or rely on expensive word-level annotation, or use sentence-level annotation only. However, some important linguistic phenomena and resources have not been fully studied. Aiming at the linguistic phenomenon that a sentence may have multiple sentiments and different target words may have different sentiments, this thesis proposes a multi-feature sentiments classification model based on LDA. The model automatically extracts the subject target words through LDA, screens the global sentiment features of sentences, extracts the local sentiment features of sentences with the external sentiment vocabulary, and integrates various features with the sentiments classification model. A series of experiments on three datasets show that the multi-feature model is effective. The introduction of LDA can not only reduce the demand for labeled target words, improve the accuracy of sentiments classification, but also more accurately analyze the internal emotional trend of public opinion events.