End-to-End Multi-task Learning for Allusion Detection in Ancient Chinese Poems

Much efforts have been devoted to research about ancient Chinese poems. However, tasks around allusions, a fundamental element of ancient Chinese poetry, has received little attention. To mitigate this gap, we introduce three allusion tasks: allusion entity recognition (AER), allusion source identification (ASI), and allusion entity classification (AEC). For each task, we create a large corpus extracted from allusion dictionary. We explore the performance of two learning strategies: single-task model and allusion hierarchical multi-task learning (AHMTL) model. Compared with the single-task model, experimental results show that the AHMTL model improves each task's overall performance by formulating relationship between tasks. In addition, poem readability, a downstream task of allusion tasks, is combined to gain improvement in the F1-score by 1.4%.