Jus Mundi at SemEval-2023 Task 6: Using a Frustratingly Easy Domain Adaption for a Legal Named Entity Recognition System

Proceedings of the The 17th International Workshop on Semantic Evaluation (SemEval-2023), pages 1783–1790 July 13-14, 2023 ©2023 Association for Computational Linguistics Jus Mundi at SemEval-2023 Task 6: Using a Frustratingly Easy Domain Adaption for a Legal Named Entity Recognition System Luis Adrián Cabrera-Diego and Akshita Gheewala Jus Mundi / 30 rue de Lisbonne, 75008, Paris, France {a.cabrera,a.gheewala}@jusmundi.com Abstract In this work, we present a Named Entity Recog- nition (NER) system that was trained using a Frustratingly Easy Domain Adaptation (FEDA) over multiple legal corpora. The goal was to create a NER capable of detecting 14 types of legal named entities in Indian judgments. Be- sides the FEDA architecture, we explored a method based on overlapping context and aver- aging tensors to process long input texts, which can be beneficial when processing legal docu- ments. The proposed NER reached an F1-score of 0.9007 in the sub-task B of Semeval-2023 Task 6, Understanding Legal Texts. 1 Introduction In this paper, we present Jus Mundi’s1 participation in Task 6 of Semeval-2023, LegalEval: Under- standing Legal Texts (Modi et al., 2023). Specif- ically, Jus Mundi participated in the sub-task B, Legal Named Entities Extraction, which consisted of creating a Named Entity Recognition (NER) system for Indian judgment documents in English. The goal was to predict multiple types of entities from both the judgment preamble and the judg- ment body. Our participation consisted of a model trained using a Frustratingly Easy Domain Adap- tation (FEDA) algorithm (Daumé III, 2007; Kim et al., 2016; Cabrera-Diego et al., 2021b). The FEDA algorithm, as its name indicates, is a type of domain adaption method created by Daumé III (2007). The objective of using a FEDA algorithm is to learn common and domain-specific patterns between multiple datasets (Daumé III, 2007), but it can be used also to mix multiple datasets despite not having the same tagset, as seen in Cabrera-Diego et al. (2021b). Moreover, a FEDA algorithm simplifies many of the questions that gen- erally arise while exploring other types of domain adaptation and transfer learning techniques. For 1https://jusmundi.com instance, with a FEDA algorithm, it is not neces- sary to determine which layers should be frozen, fine-tuned, or substituted. The rest of the paper is organized as follows. In Section 2, we introduce the background for the proposed work. This is followed by the methodol- ogy in Section 3. The data and the experimental settings are described in Section 4 and Section 5, re- spectively. In Section 6, we present the results and discuss them. Finally, the conclusions and future work are detailed in Section 7. 2 Related Work Named entity recognition (NER) is a fundamen- tal Natural Language Processing (NLP) task that consists of identifying entities that semantically represent elements such as locations, organizations, and people (Li et al., 2020). While the applications and advantages of using NER systems are well known, in many domains and languages, it is a task that continues to be a challenge. This is the case within the legal domain, where the extent …

Content truncated to 500 words