LangRank is a program for Choosing Transfer Languages for Cross-lingual Transfer Learning, described by our paper on the topic at ACL 2019. Cross-lingual transfer, where a high-resource transfer language is used to improve the accuracy of a low-resource task language, is now an invaluable tool for improving performance of natural language processing (NLP) on low-resource languages. However, given a particular task language, it is not clear which language to transfer from, and the standard strategy is to select languages based on ad hoc criteria, usually the intuition of the experimenter. Since a large number of features contribute to the success of cross-lingual transfer (including phylogenetic similarity, typological properties, lexical overlap, or size of available data), even the most enlightened experimenter rarely considers all these factors for the particular task at hand.
LangRank is a program to solve this task of automatically selecting optimal transfer languages, treating it as a ranking problem and building models that consider the aforementioned features to perform this prediction.
For example, let's say you have a machine translation (MT) task, and you want to know which languages/datasets you should use to build a system for the low-resource language Azerbaijani.
In this case, you would prepare an example of the type of data you want to translate (in word and sub-word format, details below), and the language code "aze", then run a command like the following (where -n 3 are the top 3 languages):
python3 langrank_predict.py -o word-data.aze -s subword-data.aze -l aze -n 3 -t MT
which would give you results like the following, showing which related datasets to transfer from, and why the model chose those datasets:
Ranking (top 3):
1. ted_tur : score=-0.01
1. Transfer over target size ratio : score=0.40;
2. Transfer lang dataset size : score=0.40;
3. Overlap word-level : score=0.18
2. ted_ara : score=-0.63
1. Transfer over target size ratio : score=0.25;
2. Transfer target TTR distance : score=0.17;
3. Overlap word-level : score=0.10
3. ted_aze : score=-1.67
1. Transfer target TTR distance : score=0.12;
2. Overlap word-level : score=0.06;
3. SYNTACTIC : score=0.06
In experiments on representative NLP tasks such as machine translation, entity linking, POS tagging, and dependency parsing, we have found that LangRank predicts good transfer languages much better than ad hoc baselines considering single features in isolation. Try it out below if you want to figure out which language you should be using to solve your low-resource NLP task!
Steps to install:
First install the latest version (1.1.6) of lang2vec:
git clone https://github.com/antonisa/lang2vec.git
cd lang2vec
wget http://www.cs.cmu.edu/~aanastas/files/distances.zip .
mv distances.zip lang2vec/data/
python3 setup.py install
or use the test PyPI distribution as follows:
pip install -i https://test.pypi.org/simple/ lang2vec
Now clone and install langrank (future: install it as module)
cd ../
git clone https://github.com/neulab/langrank.git
cd langrank
pip install -r requirements.txt
wget http://phontron.com/data/langrank/indexed.tar.gz .
tar -xzvf indexed.tar.gz
rm indexed.tar.gz
You can run langrank_predict.py to predict transfer languages by providing an unsegmented dataset, a segmented dataset
(using sentencepiece), the language code of your datasets, and the NLP task on hand.
python3 langrank_predict.py -o sample-data/ted-train.orig.aze -s sample-data/ted-train.orig.spm8000.aze -l aze -n 3 -t MT
The supported NLP tasks include: machine translation (MT), entity linking (EL), part-of-speech tagging (POS), and dependency parsing (DEP).
A detailed walk-through of langrank_predict.py is provided below. (ran in the langrank directory):
python3
>>> import langrank as lr
>>>
>>> # Load some data
>>> with open("sample-data/ell.tok") as inp:
... lines = inp.readlines()
...
>>> # Just to show that we loaded something
>>> len(lines)
10133
>>> lines[0]
"o milissi c' e cikala isan adèrfia .\n"
>>>
>>> # Now prepare the dataset
>>> prepared = lr.prepare_new_dataset('ell', task="MT", dataset_source=lines)
NOTE: no subword-level dataset provided, will only extract word-level features.
>>>
>>> # And rank the candidates (this could be set to 'all' so that it would rank all available datasets)
>>> lr.rank(prepared, task='MT' candidates=['tur','ara', 'aze'])
Ranking (top 3):
1. ted_tur : score=-0.01
1. Transfer over target size ratio : score=0.40;
2. Transfer lang dataset size : score=0.40;
3. Overlap word-level : score=0.18
2. ted_ara : score=-0.63
1. Transfer over target size ratio : score=0.25;
2. Transfer target TTR distance : score=0.17;
3. Overlap word-level : score=0.10
3. ted_aze : score=-1.67
1. Transfer target TTR distance : score=0.12;
2. Overlap word-level : score=0.06;
3. SYNTACTIC : score=0.06
We have also made public in an easy-to-browse format the raw data that went into training the model. We have released data for MT, EL, POS tagging, and dependency parsing as described in the paper, and also some additional data for zero-shot MT. This may be easy to browse if you'd like to see, for example, which languages we empirically found the most useful for transferring into other languages.
If you use LangRank, we'd appreciate if you cite the paper about it!
@inproceedings{lin19acl,
title = {Choosing Transfer Languages for Cross-Lingual Learning},
author = {Yu-Hsiang Lin and Chian-Yu Chen and Jean Lee and Zirui Li and Yuyan Zhang and Mengzhou Xia and Shruti Rijhwani and Junxian He and Zhisong Zhang and Xuezhe Ma and Antonios Anastasopoulos and Patrick Littell and Graham Neubig},
booktitle = {The 57th Annual Meeting of the Association for Computational Linguistics (ACL)},
address = {Florence, Italy},
month = {July},
year = {2019}
}
LangRank was supported by NSF Award #1761548 "Discovering and Demonstrating Linguistic Features for Language Documentation", and the DARPA Low Resource Languages for Emergent Incidents (LORELEI) program under Contract No. HR0011-15-C0114.