Awesome

xTune

Code for ACL2021 paper Consistency Regularization for Cross-Lingual Fine-Tuning.

Environment

DockerFile: dancingsoul/pytorch:xTune

Install the fine-tuning code: pip install --user .

Data & Model Preparation

XTREME Datasets

Create a download folder with mkdir -p download in the root of this project.
manually download panx_dataset (for NER) [here][2], (note that it will download as AmazonPhotos.zip) to the download directory.
run the following command to download the remaining datasets: bash scripts/download_data.sh The code of downloading dataset from XTREME is from [xtreme offical repo][1].

Note that we keep the labels in test set for easier evaluation. To prevent accidental evaluation on the test sets while running experiments, the code of [xtreme offical repo][1] removes labels of the test data during pre-processing and changes the order of the test sentences for cross-lingual sentence retrieval. Replace csv.writer(fout, delimiter='\t') with csv.writer(fout, delimiter='\t', quoting=csv.QUOTE_NONE, quotechar='') in utils_process.py if using XTREME official repo.

Translations

XTREME provides translations for SQuAD v1.1 (only train and dev), MLQA, PAWS-X, TyDiQA-GoldP, XNLI, and XQuAD, which can be downloaded from [here][3]. The xtreme_translations folder should be moved to the download directory.

The target language translations for panx and udpos are obtained with Google Translate, since they are not provided. Our processed version can be downloaded from [here][4]. It should be merged with the above xtreme_translations folder.

Bi-lingual dictionaries

We obtain the bi-lingual dictionaries from the [MUSE][6] repo. For convenience, you can download them from [here][7] and move it to the download directory, i.e., ./download/dicts.

Models

XLM-Roberta is supported. We utilize the [huggingface][5] format, which can be downloaded with bash scripts/download_model.sh.

Fine-tuning Usage

Our default settings were using Nvidia V100-32GB GPU cards. If there were out-of-memory errors, you can reduce per_gpu_train_batch_size while increasing gradient_accumulation_steps, or use multi-GPU training.

xTune consists of a two-stage training process.

Stage 1: fine-tuning with example consistency on the English training set.
Stage 2: fine-tuning with example consistency on the augmented training set and regularize model consistency with the model from Stage 1.

It's recommended to use both Stage 1 and Stage 2 for token-level tasks, such as sequential labeling, and question answering. For text classification, you can only use Stage 1 if the computation budget was limited.

bash ./scripts/train.sh [setting] [dataset] [model] [stage] [gpu] [data_dir] [output_dir]

where the options are described as follows:

[setting]: translate-train-all (using input translation for the languages other than English) or cross-lingual-transfer (only using English for zero-shot cross-lingual transfer)
[dataset]: dataset names in XTREME, i.e., xnli, panx, pawsx, udpos, mlqa, tydiqa, xquad
[model]: xlm-roberta-base, xlm-roberta-large
[stage]: 1 (first stage), 2 (second stage)
[gpu]: used to set environment variable CUDA_VISIBLE_DEVICES
[data_dir]: folder of training data
[output_dir]: folder of fine-tuning output

Examples: XTREME Tasks

XNLI fine-tuning on English training set and translated training sets (`translate-train-all`)

# run stage 1 of xTune
bash ./scripts/train.sh translate-train-all xnli xlm-roberta-base 1
# run stage 2 of xTune (optional)
bash ./scripts/train.sh translate-train-all xnli xlm-roberta-base 2

XNLI fine-tuning on English training set (`cross-lingual-transfer`)

# run stage 1 of xTune
bash ./scripts/train.sh cross-lingual-transfer xnli xlm-roberta-base 1
# run stage 2 of xTune (optional)
bash ./scripts/train.sh cross-lingual-transfer xnli xlm-roberta-base 2

Paper

Please cite our paper \cite{bo2021xtune} if you found the resources in the repository useful.

@inproceedings{bo2021xtune,
author = {Bo Zheng, Li Dong, Shaohan Huang, Wenhui Wang, Zewen Chi, Saksham Singhal, Wanxiang Che, Ting Liu, Xia Song, Furu Wei},
booktitle = {Proceedings of ACL 2021},
title = {{Consistency Regularization for Cross-Lingual Fine-Tuning}},
year = {2021}
}