IRIT_IRIS_A at SemEval-2023 Task 6: Legal Rhetorical Role Labeling Supported by Dynamic-Filled Contextualized Sentence Chunks

Proceedings of the The 17th International Workshop on Semantic Evaluation (SemEval-2023), pages 905–912 July 13-14, 2023 ©2023 Association for Computational Linguistics IRIT_IRIS_A at SemEval-2023 Task 6: Legal Rhetorical Role Labeling Supported by Dynamic-Filled Contextualized Sentence Chunks Alexandre G. de Lima1,2,3, Jose G. Moreno3, Eduardo H. da S. Aranha2 1Instituto Federal do Rio Grande do Norte, Natal, Brazil 2Universidade Federal do Rio Grande do Norte, Natal, Brazil 3Institut de Recherche en Informatique de Toulouse, UMR 5505 CNRS, Toulouse, France alexandre.lima@ifrn.edu.br, jose.moreno@irit.fr, eduardoaranha@dimap.ufrn.br Abstract This work presents and evaluates an approach to efficiently leverage the context exploitation ability of pre-trained Transformer models as a way of boosting the performance of models tackling the Legal Rhetorical Role Labeling task. The core idea is to feed the model with sentence chunks that are assembled in a way that avoids the insertion of padding tokens and the truncation of sentences and, hence, obtain better sentence embeddings. The achieved re- sults show that our proposal is efficient, despite its simplicity, since models based on it over- come strong baselines by 3.76% in the worst case and by 8.71% in the best case. 1 Introduction SemEval-2023 Task 6 (LegalEval) focuses on the understanding of texts from the legal domain and it comprises three sub-tasks: Legal Rhetorical Role Labeling, Legal Named Entity Recognition, and Court Judgment Prediction with Explanation (Modi et al., 2023). The Legal Rhetorical Role Labeling (Legal RRL) task aims to assign semantic roles to the sentences from a legal document. Exam- ples of such roles are facts, arguments, and rulings. The exact roles are defined by the documents’ na- ture and by the following utilization of the labeled sentences. Legal documents are usually long and complex, so the task is relevant because the labeled sentences assist the document interpretation and the extraction of information. The labeled sentences are useful for downstream tasks, such as summa- rization and judgment prediction. The task is not trivial even for professionals from the legal domain, mainly because of its inherent subjectivity. The Legal RRL sub-task dataset comprises judgment documents written in English from the Indian legal system. The documents’ sentences were manually labeled by students from Indian law universities, whose each sentence was assigned with one of 13 pre-defined rhetorical roles (Modi et al., 2023). The Legal RRL can be tackled as a sentence classification task, and this approach requires the utilization of an encoder to convert text sentences into numeric vectors. Recent Legal RRL works exploit pre-trained Transformer models to encode sentences. However, they do not fully leverage the context exploitation ability of Transformer models: sentences are usually encoded in isolation and the inter-sentence context is lost (Savelka et al., 2020; Aragy et al., 2021), the inter-sentence context is yielded by a Recurrent Neural Network (RNN) (Li et al., 2021; Bhattacharya et al., 2021) or by using data augmentation (de Lima et al., 2022), or only cover facts instead of the full range of rhetorical roles (de Lima et al., 2023). This work presents the approach exploited by …

Content truncated to 500 words