Unsupervised Sentence Embedding Benchmark (USEB)
This repository hosts the data and the evaluation script for reproducing the results reported in the paper: “TSDAE: Using Transformer-based Sequential Denoising Auto-Encoder for Unsupervised Sentence Embedding Learning”. This benchmark (USEB) contains four heterogeous, task- and domain-specific datasets: AskUbuntu, CQADupStack, TwitterPara and SciDocs. It directly works with SBERT. For details, pleasae refer to the paper.
Install
pip install useb # Or git clone and pip install .
python -m useb.downloading all # Download both training and evaluation data
Usage & Example
After data downloading, one can either run (it needs ~8min on a GPU)
python -m useb.examples.eval_sbert
to evaluate an SBERT model (really an awesome repository for sentence embeddings, and the lastest model there is much better) on all the datasets; or run this same code below:
from useb import run
from sentence_transformers import SentenceTransformer # SentenceTransformer is an awesome library for providing SOTA sentence embedding methods. TSDAE is also integrated into it.
import torch
sbert = SentenceTransformer('bert-base-nli-mean-tokens') # Build an SBERT model
# The only thing needed for the evaluation: a function mapping a list of sentences into a batch of vectors (torch.Tensor)
@torch.no_grad()
def semb_fn(sentences) -> torch.Tensor:
return torch.Tensor(sbert.encode(sentences, show_progress_bar=False))
results, results_main_metric = run(
semb_fn_askubuntu=semb_fn,
semb_fn_cqadupstack=semb_fn,
semb_fn_twitterpara=semb_fn,
semb_fn_scidocs=semb_fn,
eval_type='test',
data_eval_path='data-eval' # This should be the path to the folder of data-eval
)
assert round(results_main_metric['avg'], 1) == 47.6
It is also supported to evaluate on a single dataset (please see useb/examples/eval_sbert_askubuntu.py):
python -m useb.examples.eval_sbert_askubuntu
Data Organization
.
├── data-eval # For evaluation usage. One can refer to ./unsupse_benchmark/evaluators to learn about how to loading these data.
│ ├── askubuntu
│ │ ├── dev.txt
│ │ ├── test.txt
│ │ └── text_tokenized.txt
│ ├── cqadupstack
│ │ ├── corpus.json
│ │ └── retrieval_split.json
│ ├── scidocs
│ │ ├── cite
│ │ │ ├── test.qrel
│ │ │ └── val.qrel
│ │ ├── cocite
│ │ │ ├── test.qrel
│ │ │ └── val.qrel
│ │ ├── coread
│ │ │ ├── test.qrel
│ │ │ └── val.qrel
│ │ ├── coview
│ │ │ ├── test.qrel
│ │ │ └── val.qrel
│ │ └── data.json
│ └── twitterpara
│ ├── Twitter_URL_Corpus_test.txt
│ ├── test.data
│ └── test.label
├── data-train # For training usage.
│ ├── askubuntu
│ │ ├── supervised # For supervised training. *.org and *.para are parallel files, each line are aligned and compose a gold relevant sentence pair (to work with MultipleNegativeRankingLoss in the SBERT repo).
│ │ │ ├── train.org
│ │ │ └── train.para
│ │ └── unsupervised # For unsupervised training. Each line is a sentence.
│ │ └── train.txt
│ ├── cqadupstack
│ │ ├── supervised
│ │ │ ├── train.org
│ │ │ └── train.para
│ │ └── unsupervised
│ │ └── train.txt
│ ├── scidocs
│ │ ├── supervised
│ │ │ ├── train.org
│ │ │ └── train.para
│ │ └── unsupervised
│ │ └── train.txt
│ └── twitter # For supervised training on TwitterPara, the float labels are also available (to work with CosineSimilarityLoss in the SBERT repo). As reported in the paper, using the float labels can achieve higher performance.
│ ├── supervised
│ │ ├── train.lbl
│ │ ├── train.org
│ │ ├── train.para
│ │ ├── train.s1
│ │ └── train.s2
│ └── unsupervised
│ └── train.txt
└── tree.txt
Citation
If you use the code for evaluation, feel free to cite our publication TSDAE: Using Transformer-based Sequential Denoising Auto-Encoderfor Unsupervised Sentence Embedding Learning:
@article{wang-2021-TSDAE,
title = "TSDAE: Using Transformer-based Sequential Denoising Auto-Encoderfor Unsupervised Sentence Embedding Learning",
author = "Wang, Kexin and Reimers, Nils and Gurevych, Iryna",
journal= "arXiv preprint arXiv:2104.06979",
month = "4",
year = "2021",
url = "https://arxiv.org/abs/2104.06979",
}