Interpretable sentiment analysis based on sentiment words’ syntax information
Qingqing Zhao, Huaping Zhang, Jianyun Shang · 2022
In recent years, with the vigorous development of deep learning, the pre-trained models such as Bert and GPT have been brilliant, and the sentiment analysis task has made increasingly outstanding achievements. The sentimental accuracy of model recognition is getting higher and higher, and the related application fields are also getting wider and wider. However, because deep learning is a black box model, its internal decision-making mechanism is not transparent to users, and it can’t reasonably explain the output of the model, which brings great limitations to the application of sentiment analysis. In this paper, we integrate the syntax tree based on sentiment words into the embedding module and the attention module of the interpretable sentiment model, and filter the evidence tokens output by the model to achieve the interpretability of sentiment analysis. The model is validated on the DuTrust dataset, and the experiment proves the validity of sentiment words' syntax in interpretable sentiment analysis.