A Multi-label Joint Classification Research for Themes and Sentiments of Chinese Classical Poetry

Classical Chinese poetry is a carrier for the ancients to record their lives and express emotions. Automatically classifying emotions and themes can provide a deeper understanding of ancient Chinese culture. However, previous studies have not fully exploited the knowledge of themes and sentiments in poetry annotations nor considered the correlations between the theme and sentiment. Hence, this paper proposes a knowledge-guided masked transformer model for the multi-label joint classification of themes and sentiments in Chinese classical poetry. Specifically, we first construct two lexical dictionaries for the theme and sentiment based on the poem annotations. Then we combine the lexical dictionaries with a knowledge-based mask-transformer to represent poems at a fine-grained level. Furthermore, considering the correlations between the theme and sentiment, our model jointly classifies the multiple themes and sentiments in Chinese classical poetry by stacking the two subtasks. Experiments on our new-released Chinese classical poetry dataset CCPD demonstrate the effectiveness of the proposed model.