nclu_team at SemEval-2023 Task 6: Attention-based Approaches for Large Court Judgement Prediction with Explanation

Proceedings of the The 17th International Workshop on Semantic Evaluation (SemEval-2023), pages 270–274 July 13-14, 2023 ©2023 Association for Computational Linguistics nclu_team at SemEval-2023 Task 6: Attention-based Approaches for Large Court Judgement Prediction with Explanation Nicolay Rusnachenko Newcastle University Newcastle Upon Tyne England nicolay.rusnachenko@ncl.ac.uk Thanet Markchom University of Reading Reading, England t.markchom@pgr.reading.ac.uk Huizhi Liang Newcastle University Newcastle Upon Tyne England huizhi.liang@ncl.ac.uk Abstract Legal documents tend to be large in size. In this paper, we provide an experiment with attention- based approaches complemented by certain document processing techniques for judgment prediction. For the prediction of explanation, we consider this as an extractive text summa- rization problem based on an output of (1) CNN with attention mechanism and (2) self- attention of language models. Our extensive experiments show that treating document end- ings at first results in a 2.1% improvement in judgment prediction across all the models. Ad- ditional content peeling from non-informative sentences allows an improvement of explana- tion prediction performance by 4% in the case of attention-based CNN models. The best sub- missions achieved 8th and 3rd ranks on judg- ment prediction (C1) and prediction with expla- nation (C2) tasks respectively among 11 partic- ipating teams. The results of our experiments are published1. Judicial process involves a lot of challenges to perform a quick and correct case prediction. The significant growth of cases, especially in highly populated countries, significantly leverages the ca- pacity of competent judges’ work. The necessity of an automatic assistance system is crucial and becomes a main reason why SemEval-2023 Legal- Eval (Modi et al., 2023) competition promotes the studies in this area (Kalamkar et al., 2022; Ma- lik et al., 2021). Court judgment prediction (CJP) problem is separated into two parts: judgment pre- diction (“Accepted” or “Denied”) and prediction with explanation (CJPE) (Malik et al., 2021). Within the last few years, deep learning tech- niques have had significant breakthroughs. Among many significant achievements, it is worth high- lighting the appearance of the attention mecha- nism which plays a crucial part in the text gen- erative models commonly found across the whole 1https://github.com/nicolay-r/ SemEval2023-6C-cjp-explanation-with-attention natural language processing (NLP) domain. Ini- tial studies proposed this mechanism to address a long input sequence of neural machine translation problem (Bahdanau et al., 2014), with further ap- plications in other NLP domains, including text classification (Shen and Huang, 2016; Zhou et al., 2016). The appearance of self-attention (Vaswani et al., 2017) proposes state-of-the-art results for a large set of NLP tasks. Self-attention becomes a backbone component of further models (Devlin et al., 2019), with the appearance of target-oriented transformers. Once weights are visualized, the at- tention mechanism serves as information for further analysis of what was considered for making a deci- sion. Consequently, our contribution in this paper is conducting experiments with document processing techniques in a combination with attention-based mechanisms for judgment prediction explanation to complement the findings of the task paper (Malik et al., 2021). This paper is organized as follows. In Section 1, we describe attention mechanisms and …

Content truncated to 500 words