File size: 943 Bytes
4cfa74c 67e9b2a 301be7e 67e9b2a |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 |
---
license: mit
---
This model is the **query** encoder of ANCE-Tele trained on TriviaQA, described in the EMNLP 2022 paper ["Reduce Catastrophic Forgetting of Dense Retrieval Training with Teleportation Negatives"](https://arxiv.org/pdf/2210.17167.pdf). 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}
}
``` |