🚀 Margin-MSE Trained DistilBert for Dense Passage Retrieval
We offer a retrieval-trained DistilBert-based model (we name the architecture BERT_Dot). This model is trained with Margin-MSE using a 3 teacher BERT_Cat (concatenated BERT scoring) ensemble on MSMARCO-Passage.
This model instance can be utilized to re-rank a candidate set or directly for a vector index based dense retrieval. The architecture is a 6-layer DistilBERT, without any architecture additions or modifications (we only change the weights during training). To obtain a query/passage representation, we pool the CLS vector. We employ the same BERT layers for both query and passage encoding, which yields better results and reduces memory requirements.
If you're interested in learning more about our simple yet effective knowledge distillation method for efficient information retrieval models across various student architectures used in this model instance, check out our paper: https://arxiv.org/abs/2010.02666 🎉
For more details, training data, source code, and a minimal usage example, please visit: https://github.com/sebastian-hofstaetter/neural-ranking-kd
✨ Features
Effectiveness on MSMARCO Passage & TREC-DL'19
We trained our model on the MSMARCO standard ("small"-400K query) training triples with knowledge distillation, using a batch size of 32 on a single consumer-grade GPU (11GB memory).
For re-ranking, we used the top-1000 BM25 results.
MSMARCO-DEV
Property |
BM25 |
Margin-MSE BERT_Dot (Re-ranking) |
Margin-MSE BERT_Dot (Retrieval) |
MRR@10 |
.194 |
.332 |
.323 |
NDCG@10 |
.241 |
.391 |
.381 |
Recall@1K |
.868 |
.868 (from BM25 candidates) |
.957 |
TREC-DL'19
For MRR and Recall, we use the recommended binarization point of the graded relevance of 2. This might skew the results when compared to other binarization point numbers.
Property |
BM25 |
Margin-MSE BERT_Dot (Re-ranking) |
Margin-MSE BERT_Dot (Retrieval) |
MRR@10 |
.689 |
.862 |
.868 |
NDCG@10 |
.501 |
.712 |
.697 |
Recall@1K |
.739 |
.739 (from BM25 candidates) |
.769 |
For more baselines, information, and analysis, please see the paper: https://arxiv.org/abs/2010.02666
Limitations & Bias
- The model inherits social biases from both DistilBERT and MSMARCO.
- The model is only trained on relatively short passages of MSMARCO (avg. 60 words length), so it might struggle with longer text.
📄 License
Citation
If you use our model checkpoint, please cite our work as:
@misc{hofstaetter2020_crossarchitecture_kd,
title={Improving Efficient Neural Ranking Models with Cross-Architecture Knowledge Distillation},
author={Sebastian Hofst{\"a}tter and Sophia Althammer and Michael Schr{\"o}der and Mete Sertkan and Allan Hanbury},
year={2020},
eprint={2010.02666},
archivePrefix={arXiv},
primaryClass={cs.IR}
}