Applying "Load What You Need: Smaller Versions of Multilingual BERT" to LaBSE

Last update: Sep 02, 2022

Overview

smaller-LaBSE

LaBSE(Language-agnostic BERT Sentence Embedding) is a very good method to get sentence embeddings across languages. But it is hard to fine-tune due to the parameter size(~=471M) of this model. For instance, if I fine-tune this model with Adam optimizer, I need the GPU that has VRAM at least 7.5GB = 471M * (parameters 4 bytes + gradients 4 bytes + momentums 4 bytes + variances 4 bytes). So I applied "Load What You Need: Smaller Multilingual Transformers" method to LaBSE to reduce parameter size since most of this model's parameter is the word embedding table(~=385M).

The smaller version of LaBSE is evaluated for 14 languages using tatoeba dataset. It shows we can reduce LaBSE's parameters to 47% without a big performance drop.

If you need the PyTorch version, see https://github.com/Geotrend-research/smaller-transformers. I followed most of the steps in the paper.

Model	#param(transformer)	#param(word embedding)	#param(model)	vocab size
tfhub_LaBSE	85.1M	384.9M	470.9M	501,153
15lang_LaBSE	85.1M	133.1M	219.2M	173,347

Used Languages

English (en or eng)
French (fr or fra)
Spanish (es or spa)
German (de or deu)
Chinese (zh, zh_classical or cmn)
Arabic (ar or ara)
Italian (it or ita)
Japanese (ja or jpn)
Korean (ko or kor)
Dutch (nl or nld)
Polish (pl or pol)
Portuguese (pt or por)
Thai (th or tha)
Turkish (tr or tur)
Russian (ru or rus)

I selected the languages multilingual-USE supports.

Scripts

A smaller version of the vocab was constructed based on the frequency of tokens using Wikipedia dump data. I followed most of the algorithms in the paper to extract proper vocab for each language and rewrite it for TensorFlow.

Convert weight

mkdir -p downloads/labse-2
curl -L https://tfhub.dev/google/LaBSE/2?tf-hub-format=compressed -o downloads/labse-2.tar.gz
tar -xf downloads/labse-2.tar.gz -C downloads/labse-2/
python save_as_weight_from_saved_model.py

Select vocabs

./download_dataset.sh
python select_vocab.py

Make smaller LaBSE

./make_smaller_labse.py

Evaluate tatoeba

./download_tatoeba_dataset.sh
# evaluate TFHub LaBSE
./evaluate_tatoeba.sh
# evaluate the smaller LaBSE
./evaluate_tatoeba.sh \
    --model models/LaBSE_en-fr-es-de-zh-ar-zh_classical-it-ja-ko-nl-pl-pt-th-tr-ru/1/ \
    --preprocess models/LaBSE_en-fr-es-de-zh-ar-zh_classical-it-ja-ko-nl-pl-pt-th-tr-ru_preprocess/1/

Results

Tatoeba

Model	fr	es	de	zh	ar	it	ja	ko	nl	pl	pt	th	tr	ru	avg
tfHub_LaBSE(en→xx)	95.90	98.10	99.30	96.10	90.70	95.30	96.40	94.10	97.50	97.90	95.70	82.85	98.30	95.30	95.25
tfHub_LaBSE(xx→en)	96.00	98.80	99.40	96.30	91.20	94.00	96.50	92.90	97.00	97.80	95.40	83.58	98.50	95.30	95.19
15lang_LaBSE(en→xx)	95.20	98.00	99.20	96.10	90.50	95.20	96.30	93.50	97.50	97.90	95.80	82.85	98.30	95.40	95.13
15lang_LaBSE(xx→en)	95.40	98.70	99.40	96.30	91.10	94.00	96.30	92.70	96.70	97.80	95.40	83.58	98.50	95.20	95.08

Accuracy(%) of the Tatoeba datasets.
If the strategy to select vocabs is changed or the corpus used in the selection step is changed to the corpus similar to the evaluation dataset, it is expected to reduce the performance drop.

References

Load What You Need: Smaller Versions of Multilingual BERT (Paper: https://arxiv.org/abs/2010.05609, GitHub: https://github.com/Geotrend-research/smaller-transformers)
Language-agnostic BERT Sentence Embedding: https://arxiv.org/abs/2007.01852
TFHub - LaBSE: https://tfhub.dev/google/LaBSE/2
LaBSE blog post: https://ai.googleblog.com/2020/08/language-agnostic-bert-sentence.html
Massively Multilingual Sentence Embeddings for Zero-Shot Cross-Lingual Transfer and Beyond: https://arxiv.org/abs/1812.10464

Comments

Training time and Machine configuration

Hi, thanks for your sharing model. I want to make a smaller model, just contains two languages(en, zh). And I want to know the kind of machine GPU and how long does it need to cost?

opened by QzzIsCoding 2
Publish model to HuggingFace Model Hub?

I migrated the full LaBSE model from TF to PyTorch and uploaded them to the HuggingFace model hub. I saw this model on the TF hub and started migrating it for uploading to the HF Hub. I realized then that this wasn't published by Google but by @jeongukjae, so wanted to check with you before uploading it.

I have exported the model locally. I'm happy to check the changes in and upload the exported model if that's fine for you :).

opened by setu4993 2

Applying "Load What You Need: Smaller Versions of Multilingual BERT" to LaBSE

Related tags

Overview

smaller-LaBSE

Used Languages

Scripts

Convert weight

Select vocabs

Make smaller LaBSE

Evaluate tatoeba

Results

Tatoeba

References

You might also like...

Comments

Training time and Machine configuration

Publish model to HuggingFace Model Hub?

Releases(15lang-1)

15lang-1(Sep 19, 2021)

Owner

Jeong Ukjae

Just Another Telegram Ai Chat Bot Written In Python With Pyrogram.

Natural Language Processing library built with AllenNLP 🌲🌱

PIZZA - a task-oriented semantic parsing dataset

Text Normalization（文本正则化）

Large-scale Self-supervised Pre-training Across Tasks, Languages, and Modalities

Blackstone is a spaCy model and library for processing long-form, unstructured legal text

This repository details the steps in creating a Part of Speech tagger using Trigram Hidden Markov Models and the Viterbi Algorithm without using external libraries.

Finally, some decent sample sentences

PyTorch Implementation of Meta-StyleSpeech : Multi-Speaker Adaptive Text-to-Speech Generation

⚡ Automatically decrypt encryptions without knowing the key or cipher, decode encodings, and crack hashes ⚡

📜 GPT-2 Rhyming Limerick and Haiku models using data augmentation

IndoBERTweet is the first large-scale pretrained model for Indonesian Twitter. Published at EMNLP 2021 (main conference)

NLP codes implemented with Pytorch (w/o library such as huggingface)

华为商城抢购手机的Python脚本 Python script of Huawei Store snapping up mobile phones

Use fastai-v2 with HuggingFace's pretrained transformers

Code for EMNLP20 paper: "ProphetNet: Predicting Future N-gram for Sequence-to-Sequence Pre-training"

A tool helps build a talk preview image by combining the given background image and talk event description

A crowdsourced dataset of dialogues grounded in social contexts involving utilization of commonsense.

🤕 spelling exceptions builder for lazy people

This is the code for the EMNLP 2021 paper AEDA: An Easier Data Augmentation Technique for Text Classification