Rajaraman, K., & Veeramani, H. (2024). Semantists at LegalLens-2024: Data-efficient Training of LLM’s for Legal Violation Identification. Proceedings of the Natural Legal Language Processing Workshop 2024, 355–360. https://doi.org/10.18653/v1/2024.nllp-1.31
Abstract:
In this paper, we describe our system for
LegalLens-2024 Shared Task on automatically
identifying legal violations from unstructured
text sources. We participate in Subtask B,
called Legal Natural Language Inference (LNLI), that aims to predict the relationship between a given premise summarizing a class action complaint and a hypothesis from an online
media text, indicating any association between
the review and the complaint. This task is challenging as it provides only limited labelled data.
In our work, we adopt LLM based methods
and explore various data-efficient learning approaches for maximizing performance. In the
end, our best model employed an ensemble of
LLM’s fine-tuned on the task-specific data, and
achieved a Macro F1 score of 78.5% on test
data, and ranked 2nd among all teams submissions.
License type:
Attribution 4.0 International (CC BY 4.0)
Funding Info:
This research / project is supported by the Agency for Science, Technology and Research - Epidemic Preparedness Horizontal Co-ordinating Office (EP HTCO)
Grant Reference no. : FY22_CF_HTCO SEED_EP_IDLabs_C22714020