RDA-ABSA

A framework for Relation-aware Data Augmentation in Aspect-Based Sentiment Analysis (ABSA).

Overview

RDA-ABSA (Relation-aware Data Augmentation for Aspect-Based Sentiment Analysis) is a pipeline designed to enhance ABSA performance through targeted data augmentation. This repository provides tools to generate high-quality augmented data, train augmentation models, and evaluate ABSA performance.

📋 Environment Setup

Prerequisites

• Python 3.8+
• PyTorch-compatible GPU (recommended for faster training)

Installation

Install all required dependencies with: pip install

• scikit-learn # For text vectorization, topic modeling, and evaluation metrics
• transformers # For Hugging Face models (tokenizers, pre-trained LLMs)
• torch # PyTorch deep learning framework
• tqdm # For progress bars in long-running tasks
• numpy # For numerical operations and array handling
• pandas # For data manipulation and processing

🚀 Experimental Workflow

Follow these steps to execute the pipeline:

1. Initial Data Augmentation

Generate candidate augmented texts from the original training data using DPO (Direct Preference Optimization) principles: python dpo_augmentation.py

2. Reward Scoring

Evaluate the quality of augmented texts using two reward models to assign scores:

python reward1.py

python reward2.py

3. Preference Dataset Construction

Create a preference-aligned dataset by ranking augmented texts based on their reward scores: python preference_data.py

4. Train the Data Augmentation LLM

Fine-tune a language model to generate high-quality augmented texts using the preference dataset.

5. Generate Enhanced Training Data

Use the trained augmentation model to produce final augmented data for ABSA training: python absa_augmentation.py

6. Train the ABSA Model

Fine-tune the ABSA model on the combined original + augmented training data.

7. Evaluate the ABSA Model

Assess the performance of the trained ABSA model on the test set: python absa_evaluation.py

📊 Expected Outputs

• Augmented datasets in JSON format
• Trained models saved in Hugging Face format
• Evaluation metrics including accuracy and F1-score

💡 Notes

• Adjust hyperparameters based on your specific dataset characteristics
• The number of augmentations can be tuned based on the size of your original dataset
• All paths can be customized according to your file system structure