\title{\underline{L}e\underline{G}en: End-to-end \underline{Leg}al Information Extraction using \underline{Gen}erative Models}

Anonymous

Abstract: Despite the rapid growth in access to digital devices, the new users of the devices, especially in developing countries like India, are not able to access information on their rights and entitlements, jobs and livelihood, healthcare, education, etc. as the information is in the form of very long, complex sentences and heavy in legal parlance. Open information extraction techniques can be used to convert unstructured legal text into triples of the form \triple{subject, relation, object} in a domain-independent manner. However, the legal text is long and complex which calls for extracting structure beyond triples, also called complex information extraction. This paper proposes a generative approach to perform complex information extraction from legal statements. We achieve this by encoding legal statements as trees to capture their complex structure and semantics. This end-to-end modelling reduces the propagation of errors across complicated pipelines. We experimented with multiple generative architectures to conclude that our proposed approach reports up to 14.7 \% gain on an Indian Legal benchmark and is competitive on open information extraction benchmarks.

Paper Type: long

Research Area: Information Extraction

Languages Studied: English