As a widely used rhetorical device in classical Chinese poetic composition, allusion holds significant value for literary research. Recognizing them in classical Chinese poems is an important prerequisite and foundation for the analysis and research of classical Chinese literature. However, the absence of a standardized modeling framework to effectively recognize allusions in classical poems has presented challenges. Existing methods are often constrained by the limitations of labels or fall short of achieving accurate direct recognition. In response to this issue, we introduce a novel allusion detection approach based on self-supervised contrastive learning. This method leverages weakly supervised signals from the text to comprehend the relationship between classical poems and allusions. To verify the effectiveness of our method, we collected a dataset containing 1025 allusions in classical Chinese literature and 14,016 poems that quoted them, and conducted experiments on this dataset. Experimental results demonstrate a substantial enhancement over baseline models (0.80 vs. 0.87), with the model exhibiting the ability to directly match allusion information found in poems. Our contributions include: (1) a novel SCAD model designed for the recognition of allusions in classical poems, which deciphers the allusion patterns in Chinese classical poems by establishing “Poem-Allusion” text pairs, showcasing remarkable Zero-Shot capability for unregistered allusions; and (2) an innovative allusion knowledge mining technique utilizing the LIME algorithm to validate SCAD’s interpretability, and based on the analysis results, we achieve the mining of allusion words and the construction of an allusion knowledge base, facilitating comprehensive scholarly exploration in this domain.
Shi et al. (Fri,) studied this question.