TA-LJP: Term-Aware Legal Judgment Prediction

Legal Judgment Prediction (LJP) is a crucial task in the field of legal artificial intelligence. It leverages the fact description of a case to automatically render a verdict, deriving judgment results (including legal articles, charges, and penalty terms). Current LJP methods are overly simplistic in integrating legal articles and charge definitions into case fact representations, neglecting attention to key legal elements information such as legal concepts and terminology, resulting in the omission of key legal elements. Simultaneously, they overlook the sentencing range information contained in legal articles, often leading to judgment results that exceed the statutory penalty terms. In light of this, we propose a novel LJP method—TA-LJP (Term-Aware Legal Judgment Prediction). This method effectively fuses legal articles (or charge definitions) with case fact representations step by step through an improved multi-level fusion module, enhancing the weights of key legal elements to highlight their modeling and effectively extracting sentencing range information from legal articles to further strengthen case fact representations, thereby improving the overall performance of the LJP task. TA-LJP consists of three main stages: Firstly, to fully model key legal elements when integrating legal articles and charge definitions into fact representations, legal articles (or charge definitions) are incrementally integrated through an improved multi-level fusion module, finely increasing the weights of key legal elements to initially enhance case fact representations. Subsequently, sentencing range information from legal articles is extracted and effectively utilized to further strengthen case fact representations. Finally, the enhanced fact representations are used to predict the legal articles, charges, and penalty terms of the case. Experimental results on LAIC2021 datasets demonstrate that TA-LJP exhibits distinct advantages in LJP, particularly in the penalty term prediction task, achieving a relative improvement of 3.02% compared to the best baseline method.