A Comparative Study on Criminal Judgement Search Models Using Natural Language Processing - Focusing on the motive for the crime -.
Master's thesis, Hallym University.
Judgement in a Criminal Case Natural Language Processing Similar Case Matching Judgement Classification
With the advent of the Fourth Industrial Revolution, artificial intelligence technology has developed, and the data to be processed has increased exponentially. Efficient search is required to process a lot of data in a short time. This is also required in the legal tech field, which targets a large number of legal documents. Legal Tech is a field that applies technology to the legal field and provides services related to search, analysis, and writing. In the case of search for similar cases based on specific cases, it is being studied in various countries such as Singapore and Australia, but it is still insignificant in Korea. In searching for such judgments, text data is digitized through Natural Language Processing technology based on text judgment data, allowing computers to understand and analyze natural languages. Currently, similar case searches are provided bym private services such as Lbox and Big Case, but it seems difficult for the performance of the search engine to meet the level required by users, and various studies conducted in Korea do not specify the meaning of similar cases in the legal field. In this study, a similar case derivation experiment was conducted using various embedding models and similarity measurement methods for criminal judgments. However, despite using various models, it was confirmed that the similar events derived from each model were different. Accordingly, humans directly intervened, and the results of similar events were evaluated by ranking between similar events based on the reference event. However, similar cases were evaluated through people, but it was confirmed that half of the various reference cases did not obtain the consent of the majority of people. Through this, we found that simply deriving similar cases based on the embedding model is not suitable for criminal rulings. Therefore, as a result of statistical analysis of the reasons people wrote when evaluating similar events, it was found that 'motivation' was considered the most among the various criteria for screening similar events. After that, to derive similar cases, the judgment was classified by the motives described in the crime facts. Thereafter, a dataset consisting of keywords and sentences with the corresponding motive was constructed. Based on the corresponding dataset, we learned the machine learning family Decision Tree, Random Forest, SVM model, and the transformer model, KoBERT and KLUE/bert model. As a result, Random Forest as a keyword-based dataset and KLUE/bert as a sentence-based dataset performed the best.