BERT-whitening

This is the Pytorch implementation of "Whitening Sentence Representations for Better Semantics and Faster Retrieval".

BERT-whitening is very practical in text semantic search, in which the whitening operation not only improves the performance of unsupervised semantic vector matching, but also reduces the vector dimension, which is beneficial to reduce memory usage and improve retrieval efficiency for vector search engines, e.g., FAISS.

This method was first proposed by Jianlin Su in his blog[1].

Reproduce the experimental results

Preparation

Download datasets

$ cd data/
$ ./download_datasets.sh
$ cd ../

Download models

$ cd model/
$ ./download_models.sh
$ cd ../

After the datasets and models are downloaded, the data/ and model/ directories are as follows:

├── data
│   ├── AllNLI.tsv
│   ├── download_datasets.sh
│   └── downstream
│       ├── COCO
│       ├── CR
│       ├── get_transfer_data.bash
│       ├── MPQA
│       ├── MR
│       ├── MRPC
│       ├── SICK
│       ├── SNLI
│       ├── SST
│       ├── STS
│       ├── SUBJ
│       ├── tokenizer.sed
│       └── TREC
├── model
│   ├── bert-base-nli-mean-tokens
│   ├── bert-base-uncased
│   ├── bert-large-nli-mean-tokens
│   ├── bert-large-uncased
│   └── download_models.sh

BERT without whitening

$ python3 ./eval_without_whitening.py

Results:

BERT with whitening(target)

$ python3 ./eval_with_whitening\(target\).py

Results:

BERT with whitening(NLI)

$ python3 ./eval_with_whitening\(nli\).py

Results:

Semantic retrieve with FAISS

An important function of BERT-whitening is that it can not only improve the effect of semantic similarity retrieval, but also reduce memory usage and increase retrieval speed. In this experiment, we use Quora Duplicate Questions Dataset and FAISS, a vector retrieval engine, to measure the retrieval effect and efficiency of different models. The dataset contains more than 400,000 pairs of question1-question2, and it is marked whether they are similar. We extract all the semantic vectors of question2 and store them in FAISS (299,364 vectors in total), and then use the semantic vectors of question1 to retrieve them in FAISS (290,654 vectors in total). MRR@10 is used to measure the effect of retrieval, Average Retrieve Time (ms) is used to measure retrieval efficiency, and Memory Usage (GB) is used to measure memory usage. FAISS is configured in CPU mode, nlist = 1024'' and nprobe = 5'', and the CPU is Intel(R) Xeon(R) CPU E5-2699 v4 @ 2.20GHz.

Modify model_name'' in qqp_search_with_faiss.py'', and then execute:

$ python3 qqp_search_with_faiss.py

The experimental results of different models are as follows:

From the experimental results, the use of whitening to reduce the vector sizes of BERTbase and BERTlarge to 256 and 384, respectively, can significantly reduce memory usage and retrieval time, while improving retrieval results. The memory usage is strictly proportional to the vector dimension, while the average retrieval time is not strictly proportional to the vector dimension. This is because FAISS has a difference in clustering question2, which will cause some fluctuations in retrieval efficiency, but in general, the lower its dimensionality, the higher the retrieval efficiency.

References

[1] 苏剑林, 你可能不需要BERT-flow：一个线性变换媲美BERT-flow, 2020.

[2] 苏剑林, Keras版本BERT-whitening, 2020.