Sanjithganesp/LegalBiasDataset

Legal Case Dataset Loader

This project provides a Python-based pipeline to load and access pre-processed legal case datasets stored in the Hugging Face datasets format. The datasets are intended for Natural Language Processing (NLP) tasks such as token classification, masked language modeling, and span extraction.

---

Dataset Overview

The datasets are stored in Arrow format using Hugging Face's datasets library. There are four main datasets:

1. masked

   • Contains masked versions of legal case texts.
   • Useful for masked language modeling (MLM) tasks.
   • Each example may contain fields like case_id, masked_text, and tokenized input_ids.

2. extracted

   • Contains raw extracted text from legal case PDFs.
   • Useful for text analysis, summarization, and feature extraction.
   • Typical fields: case_id, text.

3. spans

   • Contains span information indicating important sections or phrases in the legal text.
   • Useful for tasks such as named entity recognition (NER) or section classification.
   • Typical fields: case_id, spans (list of start/end positions).

4. token_labels

   • Contains token-level labels for NLP tasks.
   • Useful for token classification tasks like NER or legal argument labeling.
   • Typical fields: case_id, tokens, labels.

---

Prerequisites

Ensure you have Python installed (>=3.8) and the required packages.

Install the datasets library from Hugging Face:

pip install datasets

Optional but recommended for NLP pipelines:

pip install transformers torch

---

Project Structure

arrow_dataset/
├── masked/
├── extracted/
├── spans/
└── token_labels/

Each folder contains a dataset saved in Hugging Face Arrow format.

---

Loading the Datasets

Use the datasets library to load the datasets from disk:

from datasets import load_from_disk

# Load each dataset folder
masked_dataset = load_from_disk("arrow_dataset/masked")
extracted_dataset = load_from_disk("arrow_dataset/extracted")
spans_dataset = load_from_disk("arrow_dataset/spans")
token_labels_dataset = load_from_disk("arrow_dataset/token_labels")

# Example access
print("Masked Example:", masked_dataset[0])
print("Token Labels Example:", token_labels_dataset[0])

---

Example Data

Masked Dataset Example

{
  "case_id": "Sat_Paul_And_Others_vs_State_Of_Punjab_And_Another_on_9_March_2010",
  "masked_text": "The accused ___ entered the premises...",
  "input_ids": [101, 2003, 1037, 7099, 6251, 102]
}

Token Labels Example

{
  "case_id": "Sat_Paul_And_Others_vs_State_Of_Punjab_And_Another_on_9_March_2010",
  "tokens": ["The", "accused", "___", "entered", "the", "premises", "..."],
  "labels": [0, 1, 2, 0, 0, 1, 0]
}

---

Iterating Through Dataset

You can iterate through all examples for preprocessing or exporting:

for example in masked_dataset:
    case_id = example['case_id']
    masked_text = example['masked_text']
    print(f"Case ID: {case_id}\nMasked Text: {masked_text}\n")

You can also convert the dataset to pandas DataFrame for easier analysis:

import pandas as pd

df_masked = masked_dataset.to_pandas()
print(df_masked.head())

---

Use Cases

• Legal NLP Modeling

  • Token classification (e.g., labeling legal entities)
  • Masked language modeling (MLM)
  • Span extraction or section identification
  • Summarization and legal text analysis

• Data Analysis

  • Exploratory Data Analysis (EDA) on legal cases
  • Frequency analysis of legal phrases or tokens

---

Notes

• Ensure that the dataset directories exist under arrow_dataset/.
• The Hugging Face Arrow format is efficient for large datasets and can handle memory-mapped loading.
• For tokenization and model training, you can use Hugging Face transformers with datasets directly.

---

References

• Hugging Face Datasets Documentation
• Hugging Face Transformers Documentation

---