Effective Knowledge Distillation on Reasoning Chains for Aspect-Based Sentiment and Emotion Analysis

Overview

This repository provides a framework for knowledge distillation in aspect-based sentiment and emotion analysis. A LLM (teacher) generates step-by-step reasoning traces and annotations for unlabelled text, which are then used to train a smaller student model to learn both intermediate reasoning steps and final predictions.

Framework

• Teacher-Student Knowledge Distillation:

  • The teacher (large language model) annotates unlabelled data with reasoning chains for five tasks:
    1. Aspect extraction
    2. Syntactic parsing
    3. Opinion extraction
    4. Sentiment classification
    5. Emotion classification
  • The student (smaller model) is trained on these annotations to mimic the teacher's reasoning and predictions.

• Pipeline Visualization:

  • 
  • Example teacher annotation:

Getting Started

Installation

1. Clone the repository:

   git clone https://github.com/onatakca/synchain-absa-emotion
   cd synchain-absa-emotion

2. Create and activate a virtual environment:

   python -m venv venv
   source venv/bin/activate

3. Install dependencies:

   pip install -r requirements.txt

Usage

1. Teacher Model Annotation

Run the teacher model to annotate unlabelled data:

python scripts/annotation/annotate.py

2. Student Model Training

Train the student model using the generated annotations and configuration files:

python scripts/modeling/knowledge_distillation.py

Configuration Files

Configuration files for training are located in:

• scripts/modeling/configs/

Prompts and Emotion Labels

Prompts and emotion label definitions are in:

• scripts/qwen_model/prompts.py

Data

• Input chunks for teacher annotation:
  • data/input_data/chunks_for_teacher_model_ann/
• Teacher annotated outputs (Qwen25-32b-Instruct):
  • data/output_data/Qwen25-32b-instruct_annotation/