ViLQA: Vietnamese Legal Question Answering

Introduction

ViLQA is a project focused on fine-tuning Question Answering models for Vietnamese legal text. This project aims to enhance the performance of models in answering questions based on Vietnamese legal documents, particularly on extractive question answering datasets.

What is Extractive Question Answering?

Extractive Question Answering (EQA) is a Natural Language Processing task where the model is given a question and a context (usually a paragraph or a document) and must extract the answer directly from the context. The key assumption in EQA is that the answer is a continuous span of text within the given context.

Here's an example of Extractive Question Answering:

Fig.1: An example of Extractive Question Answering

In this example:

• Context: The legal text about the crime of appropriating or destroying relics of fallen soldiers
• Question: "Chiếm đoạt di vật của tử sĩ có thể bị phạt tù lên đến bao nhiêu năm?" (How many years of imprisonment can one face for appropriating relics of fallen soldiers?)
• Answer: "07 năm" (07 years)

The model must extract the correct answer span from the given context based on the question.

Key Features

• Fine-tuning Question Answering models on the Vietnamese Extractive Question Answering dataset.
• Support for advanced language models (BERT, PhoBERT, XLM_RoBERTa, ViT5, BARTPho,...).
• Easy integration with popular tools and libraries.
• Customizable parameters for difference purposes.

Project Structure

This project implements two approaches to Extractive Question Answering:

1. MRC (Machine Reading Comprehension): This approach extracts the answer (span text) directly from the given context.
2. AG (Answer Generation): This approach generates the answer based on the context and question.

The code for each approach is organized in separate directories (MRC/ and AG/) for clarity and maintainability.

ViEQA/
│
├── MRC/
|   ├── config.py                 # Project configuration file
|   ├── train.py                  # Main script for fine-tuning
|   ├── requirements.txt          # List of required Python libraries
|   ├── data/
|   │   ├── __init__.py           # Init module for data
|   │   ├── data_processing.py    # Data processing script
|   │   └── dataset/
|   │       └── ALQAC.csv         # Vietnamese dataset for QA
|   ├── models/
|   │   ├── __init__.py           # Init module for models
|   │   └── eqa_model.py          # Model for Extractive Question Answering
|   └── utils/
|       ├── __init__.py           # Init module for utils
|       └── metrics.py            # Script for calculating model metrics
└── AG/
    ├── data/
    │   └── ALQAC.csv             # Vietnamese dataset for QA
    ├── data_processor.py         # Handles data preprocessing and formatting
    ├── config.py                 # Contains configuration settings for the model and training process
    ├── utils.py                  # Utility functions for text processing and evaluation metrics
    ├── model.py                  # Defines the model architecture and loading functions
    ├── train.py                  # Implements the training loop and related functions
    ├── evaluate.py               # Contains functions for model evaluation and prediction generation
    └── config.py                 # The main script to run the entire pipeline

System Requirements

• Python 3.7 or higher
• Python libraries listed in requirements.txt

Problem Formulation

1. Machine Reading Comprehension (MRC)

Given:

• A context C = [c1, c2, ..., cn], where ci are tokens in the context
• A question Q = [q1, q2, ..., qm], where qi are tokens in the question

The goal is to find:

• Start index s and end index e in C, so the span [cs, cs+1, ..., ce] answers the question Q.

2. Answer Generation (AG)

Given:

• A context C = [c1, c2, ..., cn], where ci are tokens in the context
• A question Q = [q1, q2, ..., qm], where qi are tokens in the question

The goal of our Answer Generation (AG) approach is to:

• Generate an answer A = [a1, a2, ..., ak], where ai are tokens in the generated answer

Installation Guide

I. Clone the repository:

git clone https://github.com/ntphuc149/ViLQA.git

II. Choose the approach to EQA

2.1. With MRC approach

cd ViLQA/EQA

2.2 Create a virtual environment (recommended):

python3 -m venv venv
source venv/bin/activate

2.3. Install the dependencies:

pip install -r requirements.txt

2.2. With AG approach

cd ViLQA/AG

2.2 Create a virtual environment (recommended):

python3 -m venv venv
source venv/bin/activate

2.3. Install the dependencies:

pip install -r requirements.txt

2.4. Download the Vietnamese spaCy model:

python -m spacy download vi_core_news_lg

Usage Instructions

With MRC approach

1. Configure the project:

Update the parameters in config.py to suit your dataset and requirements.

2. Fine-tune and evaluate the model:

Run the following command to start fine-tuning and evaluate the model:

python train.py

With AG approach

1. Configure the project:

You can modify the Config class in config.py to adjust various parameters for each model.

2. Fine-tune and evaluate the model:

To run the entire pipeline (data processing, training, and evaluation) with a specific model:

python main.py <model_name>

Supported models:

• ViT5: python main.py vit5
• BARTPho: python main.py bartpho

Contribution

We welcome contributions to this project. Please create a pull request or open an issue to discuss your ideas for improvement.

License

This project is licensed under the MIT License. See the LICENSE file for details.