File size: 4,251 Bytes
d084283 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 |
# SimLM: Pre-training with Representation Bottleneck for Dense Passage Retrieval
paper available at [https://arxiv.org/pdf/2207.02578](https://arxiv.org/pdf/2207.02578)
code available at [https://github.com/microsoft/unilm/tree/master/simlm](https://github.com/microsoft/unilm/tree/master/simlm)
## Paper abstract
In this paper, we propose SimLM (Similarity matching with Language Model pre-training), a simple yet effective pre-training method for dense passage retrieval.
It employs a simple bottleneck architecture that learns to compress the passage information into a dense vector through self-supervised pre-training.
We use a replaced language modeling objective, which is inspired by ELECTRA,
to improve the sample efficiency and reduce the mismatch of the input distribution between pre-training and fine-tuning.
SimLM only requires access to unlabeled corpus, and is more broadly applicable when there are no labeled data or queries.
We conduct experiments on several large-scale passage retrieval datasets, and show substantial improvements over strong baselines under various settings.
Remarkably, SimLM even outperforms multi-vector approaches such as ColBERTv2 which incurs significantly more storage cost.
## Results on MS-MARCO passage ranking task
| Model | dev MRR@10 | dev R@50 | dev R@1k | TREC DL 2019 nDCG@10 | TREC DL 2020 nDCG@10 |
|--|---|---|---|---|---|
| RocketQAv2 | 38.8 | 86.2 | 98.1 | - | - |
| coCondenser | 38.2 | 86.5 | 98.4 | 71.7 | 68.4 |
| ColBERTv2 | 39.7 | 86.8 | 98.4 | - | - |
| **SimLM (this model)** | 41.1 | 87.8 | 98.7 | 71.4 | 69.7 |
## Usage
Get embeddings from our fine-tuned model:
```python
import torch
from transformers import AutoModel, AutoTokenizer, BatchEncoding, PreTrainedTokenizerFast
from transformers.modeling_outputs import BaseModelOutput
def l2_normalize(x: torch.Tensor):
return torch.nn.functional.normalize(x, p=2, dim=-1)
def encode_query(tokenizer: PreTrainedTokenizerFast, query: str) -> BatchEncoding:
return tokenizer(query,
max_length=32,
padding=True,
truncation=True,
return_tensors='pt')
def encode_passage(tokenizer: PreTrainedTokenizerFast, passage: str, title: str = '-') -> BatchEncoding:
return tokenizer(title,
text_pair=passage,
max_length=144,
padding=True,
truncation=True,
return_tensors='pt')
tokenizer = AutoTokenizer.from_pretrained('intfloat/simlm-base-msmarco-finetuned')
model = AutoModel.from_pretrained('intfloat/simlm-base-msmarco-finetuned')
model.eval()
with torch.no_grad():
query_batch_dict = encode_query(tokenizer, 'what is qa')
outputs: BaseModelOutput = model(**query_batch_dict, return_dict=True)
query_embedding = l2_normalize(outputs.last_hidden_state[0, 0, :])
psg1 = 'Quality assurance (QA) is a process-centered approach to ensuring that a company or organization is providing the best possible products or services. It is related to quality control, which focuses on the end result, such as testing a sample of items from a batch after production.'
psg1_batch_dict = encode_passage(tokenizer, psg1)
outputs: BaseModelOutput = model(**psg1_batch_dict, return_dict=True)
psg1_embedding = l2_normalize(outputs.last_hidden_state[0, 0, :])
psg2 = 'The Super Bowl is typically four hours long. The game itself takes about three and a half hours, with a 30 minute halftime show built in.'
psg2_batch_dict = encode_passage(tokenizer, psg2)
outputs: BaseModelOutput = model(**psg2_batch_dict, return_dict=True)
psg2_embedding = l2_normalize(outputs.last_hidden_state[0, 0, :])
# Higher cosine similarity means they are more relevant
print(query_embedding.dot(psg1_embedding), query_embedding.dot(psg2_embedding))
```
## Citation
```bibtex
@article{Wang2022SimLMPW,
title={SimLM: Pre-training with Representation Bottleneck for Dense Passage Retrieval},
author={Liang Wang and Nan Yang and Xiaolong Huang and Binxing Jiao and Linjun Yang and Daxin Jiang and Rangan Majumder and Furu Wei},
journal={ArXiv},
year={2022},
volume={abs/2207.02578}
}
```
|