VTCC-NLP at SemEval-2023 Task 6:Long-Text Representation Based on Graph Neural Network for Rhetorical Roles Prediction

Proceedings of the The 17th International Workshop on Semantic Evaluation (SemEval-2023), pages 1121–1126 July 13-14, 2023 ©2023 Association for Computational Linguistics VTCC-NLP at SemEval-2023 Task 6: Long-Text Representation Based on Graph Neural Network for Rhetorical Roles Prediction Huu Hiep Nguyen† and Hoang Ngo† and Khac-Hoai Nam Bui∗ Viettel Cyberspace Center, Viettel Group, Vietnam {hiepnh1307,ngoviethoang735}@gmail.com, nambkh@viettel.com.vn Abstract Rhetorical Roles (RR) prediction is to predict the label of each sentence in legal documents, which is regarded as an emergent task for le- gal document understanding. In this study, we present a novel method for the RR task by exploiting the long context representation. Specifically, legal documents are known as long texts, in which previous works have no ability to consider the inherent dependencies among sentences. In this paper, we propose GNNRR (Graph Neural Network for Rhetorical Roles Prediction), which is able to model the cross- information for long texts. Furthermore, we de- velop multitask learning by incorporating label shift prediction (LSP) for segmenting a legal document. The proposed model is evaluated on the SemEval 2023 Task 6 - Legal Eval Un- derstanding Legal Texts for RR sub-task. Ac- cordingly, our method achieves the top 4 in the public leaderboard of the sub-task. Our source code is available for further investigation1. 1 Introduction SemEval 2023 Task 6 Legal Eval Understanding Legal Texts (Modi et al., 2023) aims to automati- cally process some intermediate tasks to boost the productivity of legal systems. The task is split into three sub-tasks, which are Rhetorical Roles (RR), Legal Named Entity Recognition (L-NER), and Court Judgement Prediction with Explanation (CJPE). The purpose of the RR prediction task is to standardize an unstructured corpus so that it is easier for the machine to automatically understand legal documents. Participants are given a dataset including long and unstructured legal documents. Each document is segmented into sentences using automatic tools (Kalamkar et al., 2022a). Seg- mented units could be classified into 13 labels, including 12 semantic labels and a NONE label † Equal contribution; ∗ Corresponding author 1https://github.com/hiepnh137/SemEval2023-Task6- Rhetorical-Roles for those not belonging to any semantic role. The RR task is considered as the sequence classifica- tion task with single-label multiple classes, where participants are required to predict the suitable la- bels for a series of consecutive text segments in those 13 labels. Traditional approach (Saravanan et al., 2008) uses CRF with hand-crafted features to segment the document into seven different roles. (Bhattacharya et al., 2019) proposed BiLSTM- CRF model with sentence embeddings constructed from sent2vec to label rhetorical roles in Indian Supreme Court documents. A multitask learning- based model developed by (Malik et al., 2022) uses label shift information to predict labels. (Kalamkar et al., 2022b) created a RR corpus of English Indian legal documents and proposed the model based on SciBERT-HSLN architecture (Brack et al., 2021). Despite the success of the aforementioned models, there are still two remaining challenges: long-range dependence between sentences and label ambiguity. It is because legal documents are quite long. There- fore, it is hard to …

Content truncated to 500 words