cisnlp / mPLM-Sim

mPLM-Sim: Better Cross-Lingual Similarity and Transfer in Multilingual Pretrained Language Models

mPLM-Sim: Better Cross-Lingual Similarity and Transfer in Multilingual Pretrained Language Models

mplm-sim is a language similarity tool providing:

• Loader: Accessing high-quality language similarity results directly.
• Executor: Obtaining similarity results from scratch.

Quickstart

Download the repo for use or alternatively install with PyPi

pip install mplm_sim

or directly with pip from GitHub

pip install --upgrade git+https://github.com/cisnlp/mPLM-Sim.git#egg=mplm_sim

Loader

from mplm_sim import Loader

# loading existing results given model_name and corpus_name
loader = Loader.from_pretrained(model_name='cis-lmu/glot500-base', corpus_name='flores200')
# Or loading results given similarity file
# loader = Loader.from_tsv('your_similarity_file.tsv')

# Getting similarity given language pairs
# iso3_script
sim = loader.get_sim('eng_Latn', 'cmn_Hani')
# or language name
sim = loader.get_sim('English', 'Chinese')

Executor

from mplm_sim import Loader

# model_name: any text/speech language model support by Huggingface
# corpus_name: specific corpus name for saving
# corpus_path: path for multi-parallel corpora, see corpora_demo for file formatting
# corpus_type: text or speech
executor = Executor(model_name='cis-lmu/glot500-base', corpus_name='own',
                    corpus_path='corpora/own', corpus_type='text')

# Run
executor.run()

Citation

@article{DBLP:journals/corr/abs-2305-13684,
  author       = {Peiqin Lin and
                  Chengzhi Hu and
                  Zheyu Zhang and
                  Andr{\'{e}} F. T. Martins and
                  Hinrich Sch{\"{u}}tze},
  title        = {mPLM-Sim: Unveiling Better Cross-Lingual Similarity and Transfer in
                  Multilingual Pretrained Language Models},
  journal      = {CoRR},
  volume       = {abs/2305.13684},
  year         = {2023},
  url          = {https://doi.org/10.48550/arXiv.2305.13684},
  doi          = {10.48550/ARXIV.2305.13684},
  eprinttype    = {arXiv},
  eprint       = {2305.13684},
  timestamp    = {Mon, 05 Jun 2023 15:42:15 +0200},
  biburl       = {https://dblp.org/rec/journals/corr/abs-2305-13684.bib},
  bibsource    = {dblp computer science bibliography, https://dblp.org}
}