In this paper, a deep reinforcement learning (DRL) question-answering algorithm oriented to legal semantic analysis is proposed and applied to legal consulting robot, aiming at solving the problems of low efficiency and high cost of traditional legal consulting mode, insufficient semantic understanding ability and poor cross-domain adaptability of existing legal question-answering system. In this study, the legal semantic analysis is modeled as markov decision processes, and the key elements in the legal text are accurately extracted through the dynamic attention mechanism, and the cross-domain case association and judgment prediction are realized by using DRL technology. In addition, through attention visualization and logical chain display, the interpretability and user trust of answer generation are improved. The experimental results show that this method is significantly superior to the existing models in the accuracy of legal entity analysis, with an average accuracy of 84.1%. In cross-domain transfer learning, when the training data in the target field accounts for less than 15%, the transfer efficiency is improved by 41.2% compared with the existing legal fine-tuning model LawBERT. At the same time, this method is significantly superior to the current legal question answering system SOTA (LegalQA) in terms of decision-making quality and interpretability indicators, such as article citation accuracy, logical chain integrity and user trust, and is close to the level of manual experts. This research provides efficient and interpretable decision support for the legal consulting robot, and promotes the efficient allocation of judicial resources and the construction of "digital rule of law".
Tianbin Liao (Sun,) studied this question.