Skip to content

Latest commit

 

History

History
129 lines (116 loc) · 6.26 KB

README.md

File metadata and controls

129 lines (116 loc) · 6.26 KB
Raw

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" (EMNLP 2021 Findings). 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:

@inproceedings{wang-etal-2021-tsdae-using,
    title = "{TSDAE}: Using Transformer-based Sequential Denoising Auto-Encoderfor Unsupervised Sentence Embedding Learning",
    author = "Wang, Kexin and Reimers, Nils and Gurevych, Iryna",
    booktitle = "Findings of the Association for Computational Linguistics: EMNLP 2021",
    month = nov,
    year = "2021",
    address = "Punta Cana, Dominican Republic",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2021.findings-emnlp.59",
    doi = "10.18653/v1/2021.findings-emnlp.59",
    pages = "671--688",
}

Contact person and main contributor: Kexin Wang, [email protected]

https://www.ukp.tu-darmstadt.de/

https://www.tu-darmstadt.de/

Don't hesitate to send us an e-mail or report an issue, if something is broken (and it shouldn't be) or if you have further questions.

This repository contains experimental software and is published for the sole purpose of giving additional background details on the respective publication.