Extraction of Legal Document Elements Based on LSTM

Legal documents, such as contracts, court decisions, and statutes, contain key features or elements that contain information related to parties, dates, clauses, and obligations which are useful for legal analysis and legal automation. Identifying these elements is very challenging for a variety of reasons including the complexity of sentence structures, the use of specialized vocabulary that is domain-specific, and a variety of long-range dependencies present in legal text. Typical machine learning approaches have struggled at learning and preserving contextual information in arbitrary-long documents. Therefore, this research proposes extraction from legal documents based on stacked long short-term memory (SLSTM) networks. The stacked architecture enables the model to learn hierarchical and deep contextual patterns, facilitating meaningful extraction for both short- and long-term dependencies. In this submission, we used domain-specific datasets utilizing the COLIEE (Competition on Legal Information Extraction and Entailment) dataset as well as the Indian Supreme Court Judgments corpus to explore applications of the model and evaluate model performance. The text data utilized was pre-processed with several steps including tokenization, lemmatization, and word embeddings with a specialized domain vocabulary. Our experimental findings demonstrate that our proposed model outperformed the baseline architectures and achieved an accuracy of 96.4%.