datamokotow/hybrid-late-interaction-bm25

When Does Lexical Search Help Late-Interaction Retrieval?

This repository contains the paper, code, configurations, and result artifacts for a study of lexical+semantic candidate generation for late-interaction reranking.

The central finding is that BM25 helps late-interaction retrieval most as part of a hybrid first-stage candidate generator. On seven BEIR-style datasets, ModernColBERT reranking over lexical+semantic RRF top-100 candidates improves average nDCG@10 from 0.3826 to 0.4211 compared with reranking BM25 top-100 candidates. The gain is explained by candidate recall: average Recall@100 rises from 0.5262 for BM25 to 0.6743 for the hybrid first stage.

Paper

• PDF: paper/main.pdf
• LaTeX source: paper/main.tex
• Bibliography: paper/references.bib

Main Result

System	nDCG@10	Recall@100
BM25 first stage	0.2956	0.5262
Semantic first stage	0.3831	0.6738
Lexical+semantic RRF first stage	0.3606	0.6743
ModernColBERT rerank of BM25 top-100	0.3826	0.5262
ModernColBERT rerank of RRF top-100	0.4211	0.6743

Complete per-dataset tables are available in results/paper8_chunked/paper_tables.md.

Artifact Contents

• paper/: arXiv-style paper PDF and LaTeX source.
• experiments/: configs and scripts for dataset preparation, retrieval, fusion, reranking, analysis, and table compilation.
• results/paper8_chunked/: metrics, tables, and analysis summaries for the completed run.
• artifacts/paper_main/: command sheet, manifest, and result snapshot.

Raw BEIR datasets and generated indexes are not included. They can be recreated with the provided scripts.

Reproduction

Install dependencies:

uv sync

Prepare datasets:

uv run python experiments/scripts/prepare_paper_data.py \
  --config experiments/configs/paper_main.yaml

Run the main experiment:

uv run python experiments/scripts/run_paper_main.py \
  --config experiments/configs/paper_main.yaml

Compile the paper table:

uv run python experiments/scripts/compile_paper_tables.py \
  --result-dir results/paper_main \
  --out results/paper_main/paper_tables

The exact command sheet is in artifacts/paper_main/COMMANDS.md.

Citation

If you use this artifact, cite the repository and paper PDF:

@misc{acharya2026lexical_late_interaction,
  title={When Does Lexical Search Help Late-Interaction Retrieval? Hybrid Candidate Generation for Multi-Vector Reranking},
  author={Acharya, Rutvik},
  year={2026},
  howpublished={Hugging Face dataset artifact},
  url={https://huggingface.co/datasets/datamokotow/hybrid-late-interaction-bm25}
}