Multi-stage Retrieval using RM3 or SPLADE and T5

An end-to-end Search Engine that can index documents for two-stage retrieval. The system focuses on a multi-stage retrieval architecture with query expansion using SPLADE or RM3 and BM25 for retrieval, and the T5 text-to-text transformer for re-ranking. The proposed framework was evaluated on the Complex Document and Entity Collection (CODEC), which consists of a corpus of social science domains across History, Economics and Politics. CODEC also defines a document ranking and an entity ranking task which align with each other to improve document ranking through entity query expansion and topic modelling.

Getting Started

• Fork (Optional) and clone the repository.

git clone --recurse-submodules https://github.com/<username>/multi-stage-retrieval-using-splade-and-t5

• Initialise a virtual environment (e.g. venv) and install pre-requisites.

# create a new env (from the repo root)
python3 -m venv venv

# activate env for unix/linux
source venv/bin/activate    

# activate env for windows
./source/Scripts/activate

# install pre-requisites
pip install -r requirements.txt

Try the Search Engine

• Make sure you download the whole corpus and save it as CODEC/corpus/codec_documents.jsonl.

• Start the API Server (Will automatically build the index, or load pre-built index).

python app.py

• Open the localhost:8000 URL in the browser. Select and Model you want to use and type the query in the search box.

Try the Experiments

• Make sure you download the whole corpus and save it as CODEC/corpus/codec_documents.jsonl.

• Run the following script to show all experiment results and save results to ./results/metrics.csv.

python eval.py