The deep integration of deep learning and the judicial field has led to the continuous development of judicial intelligence. Judicial intelligence can not only assist judicial practitioners to improve their work efficiency, but also better serve the public and promote judicial convenience. Liability determination is the main basic task of judicial intelligence, but with the increase in the number of cases, staff will spend a lot of time dealing with similar cases, reducing work efficiency and consuming energy, so this paper combines natural language processing technology, Recommendation strategy and deep learning, research and implement a road traffic responsibility identification system based on semantic understanding and similar cases. To a certain extent, it can assist sentencing decision-making and standardize judgment standards. This paper firstly studies and analyzes related technologies, then identifies similar cases, uses triples to extract text keywords, and then uses gensim library and text2vec library to calculate text similarity, and then uses D-S evidence theory to compare the above two methods. The fusion of the similarity calculation results, and combined with the road traffic responsibility identification system for application, in which the main function of the evidence theory is to complete the similar case recommendation and similarity calculation through the case description. The establishment of a road traffic responsibility identification system has the important value of improving the predictability of judicial activities and realizing formal and substantive justice.