LBNet: A Model for Judicial Reading Comprehension

Hao Liu, Jungang Xu

Abstract: In this paper, a new model for judicial reading comprehension called LBNet that combines an end-to-end network with a BERT structure is proposed, which aims to answer questions from a given passage in judicial files. Firstly, BERT is used to extract the representation of the passage and the question, and. self-matching attention mechanism is introduced to refine the representation by matching the passage against itself, which can effectively encode information from the whole passage. In the question and answer model, the pointer networks is used to locate the positions of answers from the passages. Experimental results on the CAIL2019 datasets (Chinese Judicial Reading Comprehension), show that our model can achieve good results.