Less is More: Pre-train a Strong Text Encoder for Dense Retrieval Using a Weak Decoder
Please check the official repository for more details and updates.
Fine-tuning on Marco passage/doc ranking tasks and NQ tasks
| MSMARCO Dev Passage Retrieval |
MRR@10 |
Recall@1k |
| BM25 warmup checkpoint |
0.329 |
0.953 |
| ANCE Passage checkpoint |
0.334 |
0.961 |
| MSMARCO Document Retrieval |
MRR@10 (Dev) |
MRR@10 (Eval) |
| ANCE Document (FirstP) checkpoint |
0.394 |
0.362 |
| NQ Task |
Top-1 |
Top-5 |
Top-20 |
Top-100 |
MRR@20 |
P@20 |
| DPR checkpoint |
46.1 |
68.8 |
80.4 |
87.1 |
56.2 |
20.1 |
| ANCE NQ checkpoint |
52.5 |
73.1 |
83.1 |
88.7 |
61.5 |
22.5 |
Citation
If you find SEED-Encoder useful for your work, please cite the following paper:
@article{lu2021less,
title={Less is More: Pre-training a Strong Siamese Encoder Using a Weak Decoder},
author={Lu, Shuqi and Xiong, Chenyan and He, Di and Ke, Guolin and Malik, Waleed and Dou, Zhicheng and Bennett, Paul and Liu, Tieyan and Overwijk, Arnold},
journal={arXiv preprint arXiv:2102.09206},
year={2021}
}
