README.md

---
license: mit
language:
- en
library_name: transformers
base_model: microsoft/deberta-v3-base
tags:
- question-answering
- information-retrieval
- reranking
- query-expansion
- bm25
- deberta-v3
---

# EAR query reranker checkpoints

Official scorer checkpoints for **Expand, Rerank, and Retrieve: Query
Reranking for Open-Domain Question Answering** (Findings of ACL 2023).

EAR is a query Expansion And Reranking method. It samples diverse query
expansions and trains a reranker to select expansions that improve passage
retrieval.

## Contents

The repository preserves the original experiment directory names:

| Path | Dataset | Variant | Files |
| --- | --- | --- | ---: |
| best_nq/vanilla | Natural Questions | EAR-RI | 3 |
| best_nq/wtop1 | Natural Questions | EAR-RD | 3 |
| best_trivia/vanilla | TriviaQA | EAR-RI | 3 |
| best_trivia/wtop1 | TriviaQA | EAR-RD | 4 |

Each directory contains answer, sentence, and title scorer checkpoints.
best_trivia/wtop1 additionally preserves the historical
scorer-answer-best.bin file from the original release.

These are raw PyTorch checkpoint/state-dict files used by the EAR codebase.
They are not standalone AutoModel.from_pretrained repositories.

## Usage

Use these checkpoints with the official code:

    git clone https://github.com/voidism/EAR
    cd EAR

For EAR-RI:

    bash one_pass_eval_ri.sh nq /path/to/best_nq/vanilla
    bash one_pass_eval_ri.sh trivia /path/to/best_trivia/vanilla

For EAR-RD:

    bash one_pass_eval_rd.sh nq /path/to/best_nq/wtop1
    bash one_pass_eval_rd.sh trivia /path/to/best_trivia/wtop1

The original release used Python 3.7.13, PyTorch 1.10.1,
Transformers 4.24.0, Tokenizers 0.11.1, Pyserini, and Weights & Biases.
The rerankers were trained from microsoft/deberta-v3-base.

SOURCE_MANIFEST.sha256 contains a SHA-256 checksum for every checkpoint.
The source archive was models_ear.tar.gz with SHA-256
d756d111404fe8859fd094e313f1e2b95c489691228dfb05044c6471fe31c819.

No training or evaluation dataset is included in this model repository.

## Links

- Paper: https://aclanthology.org/2023.findings-acl.768/
- Code: https://github.com/voidism/EAR
- arXiv: https://arxiv.org/abs/2305.17080

## Citation

    @inproceedings{chuang-etal-2023-expand,
      title = {Expand, Rerank, and Retrieve: Query Reranking for Open-Domain Question Answering},
      author = {Chuang, Yung-Sung and Fang, Wei and Li, Shang-Wen and Yih, Wen-tau and Glass, James},
      booktitle = {Findings of the Association for Computational Linguistics: ACL 2023},
      year = {2023},
      pages = {12131--12147},
      doi = {10.18653/v1/2023.findings-acl.768},
      url = {https://aclanthology.org/2023.findings-acl.768/}
    }