Charge prediction model enhanced by supplying legal knowledge

Legal charge prediction aims to determine charges based on the fact description of a case. Existing works take case descriptions as the sole input to determine charges, and ignore the significant role of external legal knowledge related to charges. In this paper, we propose a charge prediction model that simultaneously leverages historical precedents and external knowledge from a legal knowledge base. Specifically, we design a similar case matching module, which matches historical precedents with a given case via cosine similarity. According to the charge labels of similar cases, legal knowledge aligned with each charge is obtained. Furthermore, in the charge prediction module, we propose a gated attention to selectively retain and integrate the obtained knowledge into the given case. By this way, model can identify the criminal constitutive information in the case for charge prediction. Experimental results on CAIL2018 dataset demonstrate that the introduction of legal knowledge can improve the charge prediction performance.