re2g reranker trex

Introduction

The RE2G-RERANKER-TREX model, developed by IBM, enhances passage reranking for information retrieval tasks. It integrates neural initial retrieval with reranking to improve results, utilizing a framework that combines retrieval, reranking, and generation. This system is built using a BART-based sequence-to-sequence architecture and aims to provide significant performance improvements in various NLP tasks like slot filling, question answering, fact-checking, and dialog.

Architecture

Re2G employs a reranking approach that merges retrieval results from multiple methods like BM25 and neural retrieval systems. This method allows the combination of different sources with incomparable scores, selecting the top-k documents from the union of these results. The architecture is designed to enhance the initial retrieval results and improve the generation of correct outputs through additional training.

Training

The training process involves a novel variation of knowledge distillation, which trains the initial retrieval, reranker, and generation components jointly using ground truth target sequences. This approach has led to significant improvements in various tasks, as evidenced by its performance on the KILT leaderboard. The training, evaluation, and inference code is available on the RE2G GitHub repository.

Guide: Running Locally

1. Clone the Repository:

   • Use the following command to clone the GitHub repository:

     git clone https://github.com/IBM/kgi-slot-filling
     cd kgi-slot-filling
     

2. Set Up Environment:

   • Ensure you have Python and PyTorch installed. Create a virtual environment and install necessary dependencies:

     python -m venv venv
     source venv/bin/activate  # On Windows use `venv\Scripts\activate`
     pip install -r requirements.txt
     

3. Run the Reranker:

   • Adapt and execute the reranker_apply.py script to apply the model to your data:

     python reranker/reranker_apply.py
     

4. Cloud GPU Suggestions:

   • For optimal performance, especially with large datasets, consider using cloud GPU services like AWS EC2, Google Cloud Platform, or Azure to accelerate the training and inference processes.

License

The RE2G-RERANKER-TREX model is licensed under the Apache License 2.0, which allows for open-source use and distribution.