When is BERT Multilingual? Isolating Crucial Ingredients for Cross-lingual Transfer
This repository contains code for our paper titled "When is BERT Multilingual? Isolating Crucial Ingredients for Cross-lingual Transfer". [arXiv]
Table of contents
- Paper in a nutshell
- Installation
- Data and models
- Repository usage
- Links to experiments and results
- Citation
Paper in a nutshell
While recent work on multilingual language models has demonstrated their capacity for cross-lingual zero-shot transfer on downstream tasks, there is a lack of consensus in the community as to what shared properties between languages enable such transfer. Analyses involving pairs of natural languages are often inconclusive and contradictory since languages simultaneously differ in many linguistic aspects. In this paper, we perform a large-scale empirical study to isolate the effects of various linguistic properties by measuring zero-shot transfer between four diverse natural languages and their counterparts constructed by modifying aspects such as the script, word order, and syntax. Among other things, our experiments show that the absence of sub-word overlap significantly affects zero-shot transfer when languages differ in their word order, and there is a strong correlation between transfer performance and word embedding alignment between languages (e.g., Spearman's R=0.94 on the task of NLI). Our results call for focus in multilingual models on explicitly improving word embedding alignment between languages rather than relying on its implicit emergence.
Installation instructions
- Step 1: Install from the conda
.ymlfile.
conda env create -f installation/multilingual.yml
- Step 2: Install
transformersin an editable way.
pip install -e transformers/
pip install -r transformers/examples/language-modeling/requirements.txt
pip install -r transformers/examples/token-classification/requirements.txt
Repository usage
For the commands we used to get the reported numbers in the paper, click here. This file contains common instructions used. This file can automatically generate commands for your use case.
Bilingual pre-training
- For bilingual pre-training on original and derived language pairs, use the flag
--invert_word_orderfor the Inversion transformation,--permute_wordsfor Permutation and--one_to_one_mappingfor Transliteration. Example command for bilingual pre-training for English with Inversion transformation to create the derived language pair.
nohup python transformers/examples/xla_spawn.py --num_cores 8 transformers/examples/language-modeling/run_mlm_synthetic.py --warmup_steps 10000 --learning_rate 1e-4 --save_steps -1 --max_seq_length 512 --logging_steps 50 --overwrite_output_dir --model_type roberta --config_name config/en/roberta_8/config.json --tokenizer_name config/en/roberta_8/ --do_train --do_eval --max_steps 500000 --per_device_train_batch_size 16 --per_device_eval_batch_size 16 --train_file ../../bucket/pretrain_data/en/train.txt --validation_file ../../bucket/pretrain_data/en/valid.txt --output_dir ../../bucket/model_outputs/en/inverted_order_500K/mlm --run_name inverted_en_500K_mlm --invert_word_order --word_modification add &
- For Syntax transformations, the train file used in the following command (
[email protected][email protected]) means that it is the concatenation of French corpus with French modified to English verb and noun order ([email protected][email protected]).
nohup python transformers/examples/xla_spawn.py --num_cores 8 transformers/examples/language-modeling/run_mlm_synthetic.py --warmup_steps 10000 --learning_rate 1e-4 --save_steps -1 --max_seq_length 512 --logging_steps 50 --overwrite_output_dir --model_type roberta --config_name config/fr/roberta_8/config.json --tokenizer_name config/fr/roberta_8/ --do_train --do_eval --max_steps 500000 --per_device_train_batch_size 16 --per_device_eval_batch_size 16 --train_file ../../bucket/pretrain_data/fr/synthetic/[email protected][email protected] --validation_file ../../bucket/pretrain_data/fr/synthetic/[email protected][email protected] --output_dir ../../bucket/model_outputs/fr/syntax_modif_en/mlm --run_name fr_syntax_modif_en_500K_mlm &
- For composed transformations, apply multiple transformations by using multiple flags, e.g.,
--one_to_one_mapping --invert_word_order.
nohup python transformers/examples/xla_spawn.py --num_cores 8 transformers/examples/language-modeling/run_mlm_synthetic.py --warmup_steps 10000 --learning_rate 1e-4 --save_steps -1 --max_seq_length 512 --logging_steps 50 --overwrite_output_dir --model_type roberta --config_name config/en/roberta_8/config.json --tokenizer_name config/en/roberta_8/ --do_train --do_eval --max_steps 500000 --per_device_train_batch_size 16 --per_device_eval_batch_size 16 --train_file ../../bucket/pretrain_data/en/train.txt --validation_file ../../bucket/pretrain_data/en/valid.txt --output_dir ../../bucket/model_outputs/en/one_to_one_inverted/mlm --run_name en_one_to_one_inverted --one_to_one_mapping --invert_word_order --word_modification add &
- Using different domains for the original and derived language.
nohup python transformers/examples/xla_spawn.py --num_cores 8 transformers/examples/language-modeling/run_mlm_synthetic_transitive.py --warmup_steps 10000 --learning_rate 1e-4 --save_steps -1 --max_seq_length 512 --logging_steps 50 --overwrite_output_dir --model_type roberta --config_name config/en/roberta_8/config.json --tokenizer_name config/en/roberta_8/ --do_train --do_eval --max_steps 500000 --per_device_train_batch_size 16 --per_device_eval_batch_size 16 --train_file ../../bucket/pretrain_data/en/train_split_1.txt --transitive_file ../../bucket/pretrain_data/en/train_split_2.txt --validation_file ../../bucket/pretrain_data/en/valid.txt --output_dir ../../bucket/model_outputs/en/one_to_one_diff_source_100_more_steps/mlm --run_name en_one_to_one_diff_source_100_more_steps --one_to_one_mapping --word_modification add &
Fine-tuning and evaluation
This directory contains scripts used for downstream fine-tuning and evaluation.
- Transliteration, Inversion, and Permutation
- Syntax
- Composed transformations
- Using different domains for original and derived languages
Embedding alignment
Use this script to calculate embedding alignment for any model which uses Transliteration as one of the transformations.
Data and models
All the data used for our experiments, hosted on Google Cloud Bucket.
- Pre-training data -
pretrain_data - Downstream data -
supervised_data - Model files -
model_outputs
Links to experiments and results
- Spreadsheets with run descriptions, commands, and weights and biases link
- Spreadsheet with all results
- Links to pre-training runs
- Link to fine-tuning and analysis
Citation
Please consider citing if you used our paper in your work!
To be updated soon!
2 Sep 05, 2022
2.1k Jan 01, 2023
8 Oct 04, 2022
117 Nov 27, 2022
349 Aug 06, 2022
56 Dec 12, 2022
31 Dec 21, 2022
11.3k Dec 30, 2022
103 Nov 22, 2022
153 Jan 04, 2023
14 Nov 13, 2022
2.6k Jan 05, 2023
3 Jan 24, 2022
19 Dec 13, 2022
67 Dec 01, 2022
7.4k Jan 04, 2023
6.2k Jan 01, 2023
77.2k Jan 02, 2023
112 Dec 23, 2022