🎉 A Text Retrieval Approach Using BM25+ and BERT Reranking 🎉

Manos Chatzakis, Hind El-Bouchrifi and Lluka Stojollari

{emmanouil.chatzakis, hind.elbouchrifi, lluka.stojollari}@epfl.ch

Distributed Information Systems, Text Retrieval Project, EPFL

Context

This repository contains a solution to a Kaggle competition of Distributed Information Systems course of EPFL. The competition is about Text Retrieval, where models are automatically evaluated by their document retrieval and reranking power.

Approach

We describe our approach in our report (report/Report.pdf), as well us our final submission notebook (src/kaggle.ipynb). Briefly, we utilize BM25+ to retrieve a big set of relevant documents and rerank them using BERT Sentence embeddings.

Contents

• src/

  • models/
    • Code for all implemented information retrieval models (TF-iDF, Doc2vec, DSSM, BM25)
  • preprocessors/
    • Code for all preprocessing and query expansion techniques
  • scripts/
    • Pipelines for model training, text preprocessing, query expansion
  • testing/ - Examples of model usage
  • IR_Utils.ipynb: Utility functions for IR
  • kaggle.ipynb: Final competition notebook, implementing our final best scoring approach

• report/

  • report.pdf: A brief report describing our approach

• data/

  • Test and training query splits

Data Availability

Test and training query splits are present under data/ directory. Every other needed file is available at Google Drive