CrownKira / GEC-with-MoE-CoT-and-Rubrics

An innovative GEC system that harnesses the power of Mixture of Experts, Chain-of-Thought (CoT) and rubric-guided evaluations to enhance correction quality and accuracy. Built on the foundation of advanced reasoning models like GPT-3.5-turbo-1106 and integrating teacher-student model feedback.

GEC-with-MoE-CoT-and-Rubrics

GEC System Architecture

Overview

Our GEC system architecture employs a novel combination of methodologies to enhance the accuracy and diversity of grammatical error corrections:

• Mixture of Experts: Utilizes a specialized approach where each "student" model is an expert on one of four datasets—A, B, C, or N—achieved through fine-tuning on each specific dataset.
• Chain-of-Thought Prompting: Utilizes structured reasoning prompts to guide the model towards more logical and contextually relevant corrections.
• Rubric-Guided Evaluations: Incorporates specific evaluation criteria, ensuring that corrections adhere to predefined quality standards.
• In-Context Learning: Leverages the capabilities of models with advanced reasoning abilities, notably GPT-3.5-turbo-1106, for dynamic learning within contextual boundaries.
• Teacher and Student Model Integration: Combines the insights of teacher models in evaluating student corrections, fostering a rich pool of diverse and insightful corrections.

Installation

Steps

1. Clone the Repository:

   git clone [e-mail]:CrownKira/GEC-with-MoE-CoT-and-Rubrics.git

2. Navigate to the Project Directory:

   cd GEC-with-MoE-CoT-and-Rubrics

3. Create and Activate the Virtual Environment:

   python3 -m venv venv
   source venv/bin/activate

4. Install Required Python Packages:

   pip3 install -r requirements.txt

5. Download the Necessary spaCy Language Model:

   python3 -m spacy download en_core_web_sm

Usage

1. Configure Environment Variables: Obtain the .env file from me or set your own environment variables as needed.

2. Run the Main Script: Replace main.py with the actual name of your script.

   python3 main.py

3. Check Outputs: Navigate to the outputs directory to access corrected text files and CSV outputs.