Abstract: Beyond individual languages, multilingual natural language processing (NLP) research increasingly aims to develop models that perform well across languages generally. However, evaluating these systems on all the world’s languages is practically infeasible. To attain generalizability, representative language sampling is essential. Previous work argues that generalizable multilingual evaluation sets should contain languages with diverse typological properties. However, ‘typologically diverse’ language samples have been found to vary considerably in this regard, and popular sampling methods are flawed and inconsistent. We present a language sampling framework for selecting highly typologically diverse languages given a sampling frame, informed by language typology. We compare sampling methods with a range of metrics and find that our systematic methods consistently retrieve more typologically diverse language selections than previous methods in NLP. Moreover, we provide evidence that this affects generalizability in multilingual model evaluation, emphasizing the importance of diverse language sampling in NLP evaluation.
This repository contains the implementations, results and visualizations for the paper A Principled Framework for Evaluating on Typologically Diverse Languages. If you use any contents from this repository for your work, we kindly ask you to cite our paper:
@misc{ploeger2024principledframework,
title={A Principled Framework for Evaluating on Typologically Diverse Languages},
author={Esther Ploeger and Wessel Poelman and Andreas Holck Høeg-Petersen and Anders Schlichtkrull and Miryam de Lhoneux and Johannes Bjerva},
year={2024},
eprint={2407.05022},
archivePrefix={arXiv},
primaryClass={cs.CL},
url={https://arxiv.org/abs/2407.05022},
}
For the specific version related to the initial preprint, please check the preprint-v1 branch.
The package is tested with Python 3.10.
git clone https://github.com/esther2000/typdiv-sampling
cd typdiv-sampling
pip install .Alternatively, with visualization support:
pip install ".[vis]"Or for development:
pip install ".[dev]"
pre-commit install
pre-commit run --all-filesThe data can be downloaded and prepared using the following script:
./prepare-data.shfrom typdiv_sampling import Sampler
from pathlib import Path
# A list of glottocodes to sample from.
frame = ['stan1293', 'russ1263', 'finn1318', 'nucl1301', 'stan1290', 'kore1280']
k = 3 # The number of languages to sample.
seed = 1 # A random seed for the non-deterministic methods.
# Initialize with default setup.
sampler = Sampler()
sampler.sample_maxsum(frame, k)
> ['kore1280', 'russ1263', 'stan1290']Sampling methods include: sample_maxsum() (MaxSum), sample_maxmin() (MaxMin) and several baselines: sample_random(), sample_convenience(), sample_random_family(), sample_random_genus().
Most options are also available from the cli of the package:
sample --helpAn example of sampling usage in practice is found in: evaluation/experiment.py.
from typdiv_sampling.evaluation import Evaluator
# With default settings.
evaluator = Evaluator()
sample = ['kore1280', 'russ1263', 'stan1290']
evaluator.evaluate_sample(sample)
> Result(
run=None, # Optional result to keep track of averages across runs, unused here.
ent_score_with=0.5374,
ent_score_without=0.4954,
fvi_score=0.7686,
mpd_score=0.7836,
fvo_score=0.6302,
sample={'russ1263', 'kore1280', 'stan1290'},
)An example of evaluation usage in practice is found in use_cases/next-best.ipynb.
The results and visualizations from the paper can be reproduced with the following scripts or notebooks:
- Language families and typological similarities (Figure 1, page 5):
analysis/typ-data-analysis.ipynb - Sampling algorithm visualization (Figure 2, page 10):
analysis/algo-vis.ipynb - Intrinsic evaluation graph (Figure 5, page 14):
evaluation/intrinsic-eval.sh - Intrinsic evaluation table (Table 1, page 15):
evaluation/tables/table.ipynb - Tokenization boxplots (Figure 6, page 16):
use_cases/tokenization/visualizations.ipynb - Dataset expansion results (Table 2, page 18):
use_cases/dataset_expansion/next-best.ipynb - UD expansion case study (Table 3, page 19):
use_cases/dataset_expansion/next-best.ipynb - Geographical distance plot (Figure 7, page 20):
use_cases/geo_dist/visualize-dist.ipynb
- Grambank is licensed under a Creative Commons 4.0 BY International License.
- WALS is licensed under a Creative Commons 4.0 BY International License.
- Glottolog is licensed under a Creative Commons 4.0 BY International License.
The code and data in this repo are licensed under a Creative Commons 4.0 BY International License.