Knowledge-guided Transformer for Joint Topic and Sentiment Analysis of Chinese Classical Poetry

Abstract The analyses of the topic and sentiment are essential for understanding Chinese classical poetry and historical culture. Existing works fail to consider the lexical knowledge mined from poem annotations, which partly contains some information about the topic and sentiment. In addition, most works ignore the interdependence and diversity of the topic and sentiment in one poem. Hence, in this paper, we propose a Knowledge-guided Transformer Model (KTM) for joint multiple topic and sentiment analysis of Chinese classical poetry. Specifically, we first respectively construct two lexical dictionaries for the topic and sentiment based on the poem annotations. Then we take full advantage of the lexical dictionaries with a knowledge-based mask-transformer to represent poems. Furthermore, considering the correlations between the topic and sentiment, our model jointly classifies the multiple topics and sentiments in Chinese classical poetry by stacking the two subtasks. Considering there is no public dataset, we release a new Chinese classical poetry dataset CCPD for joint multiple topic and sentiment analysis. Extensive experiments demonstrate that our model achieves state-of-the-art performance on both topic and sentiment analyses, especially on tail labels.