metadata
license: mit
This model is the passage encoder of ANCE-Tele trained on TriviaQA, described in the EMNLP 2022 paper "Reduce Catastrophic Forgetting of Dense Retrieval Training with Teleportation Negatives". The associated GitHub repository is available at https://github.com/OpenMatch/ANCE-Tele.
ANCE-Tele only trains with self-mined negatives (teleportation negatives) without using additional negatives (e.g., BM25, other DR systems) and eliminates the dependency on filtering strategies and distillation modules.
| NQ (Test) | R@5 | R@20 | R@20 |
|---|---|---|---|
| ANCE-Tele | 76.9 | 83.4 | 87.3 |
@inproceedings{sun2022ancetele,
title={Reduce Catastrophic Forgetting of Dense Retrieval Training with Teleportation Negatives},
author={Si, Sun and Chenyan, Xiong and Yue, Yu and Arnold, Overwijk and Zhiyuan, Liu and Jie, Bao},
booktitle={Proceedings of EMNLP 2022},
year={2022}
}