Topic Term Clustering Based on Semi-supervised Co-occurrence Graph and Its Application in Chinese Judgement Documents

Abstract

With the continuous progress of society and the improvement of people’s awareness of law, the number of cases accepted by the people’s court of China continues to run at a high level in recent years. In the increasingly severe situation of the contradiction of “many cases but few people”, how to ensure judicial justice and improve the efficiency of the trial has become an urgent problem to be solved. Under the new litigation system, courts organize debates around controversial issues, which are the core of the conflicts between the parties in the case. Similar controversial issues play an important role in identifying similar cases. Similar cases also play an important role in improving judges’ trial efficiency, promoting similar cases to be judged similarly, and ensuring the uniformity of law application. However, in the process of identifying controversial issues, it is not only affected by the uncertainty of law and facts, but also affected by the judge’s discretion and factors outside the case. So it is difficult to format controversial issues and inappropriate to judge the similarity by the consistency of controversial issues. Meanwhile, controversial issues data is subject to power law distribution, and its types are hard to be exhausted, which further aggravates the difficulty of manual annotation. Machine learning is an appropriate method to identify similar groups in the case of huge amount of controversial issues data. In this paper, a semi-supervised short text clustering algorithm is proposed to identify the homogeneous groups in controversial issues. In this algorithm, a graph model is constructed to discover the closely connected term groups, which are used as the clustering topics, and controversial issues are classified according to the topic term groups. In addition, the algorithm incorporates prior knowledge of law to improve the performance of the algorithm. This algorithm can capture semantic similarity in controversial issues, automatically induce the topic term groups of controversial issues’ categories, flexibly adjust the number of categories, and quickly get the clustering result, so as to promote the identification and retrieval of the similar case.

This work was partly supported by National Key R &D Program of China under Grant 2020YFC0832400. This work was partly supported by Key R &D Program of Sichuan Province under Grant 2021YFS0397.