Legal_try at SemEval-2023 Task 6: Voting Heterogeneous Models for Entities identification in Legal Documents

Proceedings of the The 17th International Workshop on Semantic Evaluation (SemEval-2023), pages 1282–1286 July 13-14, 2023 ©2023 Association for Computational Linguistics Legal_try at SemEval-2023 Task 6: Voting Heterogeneous Models for Entities identification in Legal Documents Junzhe Zhao and Yingxi Wang and Nicolay Rusnachenko and Huizhi Liang Newcastle University, Newcastle Upon Tyne, England {j.zhao33,Y.wang318,nicolay.rusnachenko,huizhi.liang}@ncl.ac.uk Abstract Named Entity Recognition (NER) is a cru- cial subtask of Natural Language Processing (NLP) that involves identifying and categoriz- ing named entities in text. The resulting an- notation enables unstructured natural language texts for various NLP tasks, such as informa- tion retrieval, question answering, and machine translation. NER is essential in the legal do- main as an initial stage for extracting relevant entities. However, legal texts contain domain- specific named entities, including applicants, defendants, courts, statutes, and articles, ren- dering standard named entity recognizers in- compatible with legal documents. This paper proposes an approach that combines multiple model results through a voting mechanism to identify unique entities in legal texts. This study’s primary focus is extracting named enti- ties from legal texts in the context of SemEval- 2023 Task 6, Sub-task B: Legal Named Entities Extraction (L-NER). The goal is to create a le- gal NER system for unique entity annotation in legal documents. Our experiments’ results and our system’s implementation are published and accessible online1. 1 Introduction and Related Work The significant growth of legal texts, particularly in highly populated countries, leads to a consider- able burden on the workload of competent judges. Automated text structurization is the first and es- sential step for other downstream tasks that may support judges’ work. The distinct nature of the Indian legal process and the terminology employed in legal texts are primary reasons why the SemEval- 2023 competition (Modi et al., 2023) encourages research in this area (Kalamkar et al., 2022). Initially, researchers focused on rule-based meth- ods, such as Ralph Grishman’s 1995 NER sys- tem (Grishman and Sundheim, 1996), which de- pended on manually created rules and patterns to 1https://github.com/SuperEDG/Legal_NER identify named entities in text. Machine learn- ing approaches were subsequently employed, with the Hidden Markov Model (HMM) (Zhou and Su, 2002) as one of the earliest methods modeling the probability distribution of named entity labels given a sequence of words. It laid the founda- tion for further research, culminating in the de- velopment of neural networks for NER, includ- ing the application of Long Short-Term Memory (LSTM) (Hochreiter and Schmidhuber, 1997; Ham- merton, 2003), Bidirectional LSTM (Huang et al., 2015), Convolutional Neural Networks (CNN), and Iterated Dilated CNN (IDCNN) (Strubell et al., 2017). The recent emergence of transform- ers (Vaswani et al., 2017) has significantly im- pacted the NLP field, with promising results from NER models based on transformers. Numerous research areas are active in NER, including meth- ods combining BERT, CRF, and BiLSTM (Dai et al., 2019; Jiang et al., 2019). These models can be trained on labeled NER datasets to recognize named entities in new texts. Researchers (Chalkidis et al., 2019, 2020) inves- tigate …

Content truncated to 500 words