LegalATLE: an active transfer learning framework for legal triple extraction

Haiguang Zhang, Yuanyuan Sun, Bo Xu, Hongfei Lin

Abstract: Recently, the rich content of Chinese legal documents has attracted considerable scholarly attention. Legal Relational Triple Extraction which is a critical way to enable machines to understand the semantic information presents a significant challenge in Natural Language Processing, as it seeks to discern the connections between pairs of entities within legal case texts. This challenge is compounded by the intricate nature of legal language and the substantial expense associated with human annotation. Despite these challenges, existing models often overlook the incorporation of cross-domain features. To address this, we introduce LegalATLE, an innovative method for legal Relational Triple Extraction that integrates active learning and transfer learning, reducing the model’s reliance on annotated data and enhancing its performance within the target domain. Our model employs active learning to prudently assess and select samples with high information value. Concurrently, it applies domain adaptation techniques to effectively transfer knowledge from the source domain, thereby improving the model’s generalization and accuracy. Additionally, we have manually annotated a new theft-related triple dataset for use as the target domain. Comprehensive experiments demonstrate that LegalATLE outperforms existing efficient models by approximately 1.5%, reaching 92.90% on the target domain. Notably, with only 4% and 5% of the full dataset used for training, LegalATLE performs about 10% better than other models, demonstrating its effectiveness in data-scarce scenarios.