Scholarlink

There is a pressing need to build more Indian legal datasets to develop fair models for the Indian judiciary. To meet this need, we addressed the unstructured nature of judicial documents. We developed a systematic method for preprocessing these documents, extracting relevant information, and summarizing the lawyers’ arguments on both sides. This paper focuses on automating these processes for 25,089 bail orders from district courts of Andhra Pradesh. We evaluated several approaches for information extraction, such as spaCy NER, OpenNyAI, and GPT-3.5-turbo model. We found that the GPT-3.5-turbo model with recall of 0.98, F1-score of 0.99, and precision of 1 outperformed other tools because LLMs are trained on diverse data and are better generalized. This paper shows the potential of using LLMs in extracting relevant information, summarizing, and understanding legal information. While the results are favorable, more work is needed to expand the corpus by handling the noisy scanned bail orders with stamps, signatures, and stains.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Toward An Efficient Automated Legal Entity Extraction and Document Summarization: A Case Study on Bail Orders from South Indian District Courts

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Keerthi Lingam,
  • Suresh Chittineni

摘要

There is a pressing need to build more Indian legal datasets to develop fair models for the Indian judiciary. To meet this need, we addressed the unstructured nature of judicial documents. We developed a systematic method for preprocessing these documents, extracting relevant information, and summarizing the lawyers’ arguments on both sides. This paper focuses on automating these processes for 25,089 bail orders from district courts of Andhra Pradesh. We evaluated several approaches for information extraction, such as spaCy NER, OpenNyAI, and GPT-3.5-turbo model. We found that the GPT-3.5-turbo model with recall of 0.98, F1-score of 0.99, and precision of 1 outperformed other tools because LLMs are trained on diverse data and are better generalized. This paper shows the potential of using LLMs in extracting relevant information, summarizing, and understanding legal information. While the results are favorable, more work is needed to expand the corpus by handling the noisy scanned bail orders with stamps, signatures, and stains.