Legal Document Retrieval Solution for SOICT

Overview

This solution provides a robust and efficient system for Legal Document Retrieval that addresses the complexities of processing legal terms and leverages Sentence Transformer models for enhanced semantic understanding. The primary focus is on specialized text processing techniques tailored for legal language and fine-tuning Sentence Transformer models to achieve high accuracy in legal information retrieval tasks.

---

Features

• Legal Term Processing:

  • Handles long, jargon-heavy sentences common in legal documents.
  • Normalizes legal terms and abbreviations for consistent processing.
  • Supports multilingual legal texts for global applications.

• Semantic Search:

  • Uses Sentence Transformer models fine-tuned on legal datasets.
  • Provides accurate retrieval of legal clauses, statutes, and case laws based on user queries.

• Fine-tuning:

  • Enhances pre-trained Sentence Transformer models to specialize in legal document retrieval.
  • Employs pairwise ranking and classification to improve query-document relevance.

• Customizable Workflow:

  • Modular components allow integration with existing document management systems.
  • Configurable processing pipelines to handle domain-specific requirements.

---

Components

1. Data Preprocessing

• Legal Term Normalization:
• Tokenization:
• Stopword Handling:
• Named Entity Recognition (NER):

---

2. Model Fine-Tuning

• Dataset:

  • Prepares a legal corpus with annotated query-document pairs.
  • Example structure:
    • Queries: Short legal questions or issues.
    • Documents: Relevant statutes, precedents, or legal provisions.

• Training Objective:

  • Minimizes ranking loss to align query-document embeddings.

• Evaluation Metrics:

  • Mean Reciprocal Rank (MRR).
  • Normalized Discounted Cumulative Gain (nDCG).
  • Precision@k.

---

3. Search Pipeline

• BM25s for Initial Retrieval:

  • Quickly identifies a candidate set of documents using lexical similarity.

• Semantic Reranking:

  • Refines results using fine-tuned Sentence Transformer embeddings for semantic relevance.

• Query Expansion:

  • Suggests related legal terms to enhance retrieval scope.

---

4. Notes

Since it is kinda time-consuming for me to solo-coding this, while current time is peak season for my company's projects. I decided to cancel the competition, but hopefully you guys can see something useful from the pipeline i created.

Here i added bm25s which convinced to be faster than the origin bm25, i did train the SupSimCSE-VoVanPhuc for like 2 or 3 days but it turned out, they did not allow to use. However, you can do ensemble with sbert and alibaba sentence transformers with adding in vocabs.

I dont try to use the Generation part in RAG for the fact that i dont have a good feeling that it is a good option for me.

More things that can be further added that using vllm to create a second chance question for those exist more than one each query. I also highlighted the intention of the question so that the answer can be more specific during the search.

And yeap, you can give me a like or stars iusecase. Even raising an issue, if any in my code seemed confusing. Thank you boys <3