ABSA Data Augmentation Framework

This repository provides data augmentation pipelines for Aspect-Based Sentiment Analysis (ABSA).
It supports two complementary approaches:

1. Agentic Pipeline (LangGraph + Ollama) – generates sentences with aspect–polarity pairs using a generator + evaluator agent.
2. Prompting Pipeline (Naive Generation) – directly prompts an LLM to produce aspect–polarity sentences without explicit validation.

The augmented data is used with the InstructABSA framework for training and evaluation.

---

🚀 Features

• Augment datasets for ABSA in the Restaurant domain
• Two strategies:
  • Agentic: validated samples, slower but higher quality.
  • Prompting: faster generation, noisier but scalable.
• Uses local LLMs with Ollama and Hugging Face Transformers.
• Seamlessly integrates with InstructABSA for downstream experiments.

---

⚙️ Setup

1. Clone the repo

git clone https://github.com/mohamad7395/Thesis.git
cd absa-augmentation

2. Install requirements

pip install -r requirements.txt

3. Install Ollama

Follow instructions from Ollama

4. Pull required models

ollama pull qwen2.5:14b
ollama pull llama3:8b-instruct

📜 Usage

Agentic Pipeline (LangGraph + Ollama)

Run the controlled agent-based data generation:

python run_agent.py

Prompting Pipeline (Naive LLM prompts)

python run_prompting.py

Workflow

SemEval Dataset
        │
        ├── Agentic Pipeline  ──> Augmented Data (validated)
        └── Prompting Pipeline ─> Augmented Data (naive)
        
Augmented Data ──> InstructABSA ──> Model Training & Evaluation

📂 Project Structure & Additions

This project builds on top of the original InstructABSA codebase.
On top of the baseline implementation, we added new scripts and generated datasets to support data augmentation experiments.

• Augmented Datasets
  Located in:

  InstructABSA/Dataset/Generated
  

• Experiment Results
  Located in:

  Thesis/All Results
  

• Experiment Scripts Additional Python scripts for running automated experiments are stored in:

  InstructABSA/Research
  

  These paths reflect the extended functionality for generating augmented data and evaluating it within the InstructABSA framework.

📑 Documentation

Presentation