This repository contains the code and pre-trained models for our paper Compressing Sentence Representation for Semantic Retrieval via Homomorphic Projective Distillation.

• Thanks for your interest in our repo!

We propose Homomorphic Projective Distillation (HPD) to learn compressed sentence embeddings. Our method augments a small Transformer encoder model with learnable projection layers to produce compact representations while mimicking a large pre-trained language model to retain the sentence representation quality. The following figure is an illustration of our models.

Our implementation is based on Sentence-Transformers. You can get an easy-to-use sentence embedding tool by installing sentence-transformers:

# install with pip
pip install -U sentence-transformers
# or install with conda
conda install -c conda-forge sentence-transformers

Note that if you want to enable GPU encoding, you should install the correct version of PyTorch that supports CUDA. See PyTorch official website for instructions.

After installing the package, you can simply load our model

from sentence_transformers import SentenceTransformer
model = SentenceTransformer('Xuandong/HPD-MiniLM-F128')

There are two available models: HPD-MiniLM-F128 and HPD-TinyBERT-F128.

Then you can use our model for encoding sentences into embeddings

sentences = ['He plays guitar.', 'A street vendor is outside.', 'A woman is reading.']
sentence_embeddings = model.encode(sentences)

for sentence, embedding in zip(sentences, sentence_embeddings):
    print("Sentence:", sentence)
    print("Embedding:", embedding)
    print("")

Our released models are listed as following. You can import these models by using the sentence-transformers package or using HuggingFace's Transformers.

You should first get the augmented data and put it in the datasets folder. Or you can just use ALLNLI dataset (worse performance).

mkdir datasets embs output

By running the script, we store all teacher's embeddings to embs/simcse-train-F128.pt and embs/simcse-valid-F128.pt.

bash run_emb.sh

Then we can train the model. Example in script,

bash run.sh

Our evaluation code for sentence embeddings is based on a modified version of SentEval and SimCSE. It evaluates sentence embeddings on semantic textual similarity (STS) tasks.

Before evaluation, please download the evaluation datasets by running

cd SentEval/data/downstream/
bash download_dataset.sh

Then come back to the root directory, you can evaluate any transformers-based pre-trained models using our evaluation code. For example,

python senteval.py \
    --pooler avg 
    --task_set sts 
    --mode test 
    --cuda cuda:0 
    --model_type st 
    --model_name_or_path Xuandong/HPD-MiniLM-F128

which is expected to output the results in a tabular format:

------ test ------
+-------+-------+-------+-------+-------+--------------+-----------------+-------+
| STS12 | STS13 | STS14 | STS15 | STS16 | STSBenchmark | SICKRelatedness |  Avg. |
+-------+-------+-------+-------+-------+--------------+-----------------+-------+
| 74.94 | 84.52 | 80.25 | 84.87 | 81.90 |    84.98     |      81.15      | 81.80 |
+-------+-------+-------+-------+-------+--------------+-----------------+-------+

Please see Evaluation of SimCSE for more details about the arguments. We add one argument to choose sentence-transformers model or simcse model.

• --model_type: Which model type to use. Now we support
  • st (default): sentence-transformers based models
  • simcse: SimCSE based models

Please cite our paper if you use HPD in your work:

@inproceedings{zhao-etal-2022-compressing,
    title = "Compressing Sentence Representation for Semantic Retrieval via Homomorphic Projective Distillation",
    author = "Zhao, Xuandong  and
      Yu, Zhiguo  and
      Wu, Ming  and
      Li, Lei",
    booktitle = "Findings of the Association for Computational Linguistics: ACL 2022",
    month = may,
    year = "2022",
    address = "Dublin, Ireland",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2022.findings-acl.64",
    pages = "774--781",
}