xmod-base

X-MOD is a multilingual masked language model trained on filtered CommonCrawl data containing 81 languages. It was introduced in the paper Lifting the Curse of Multilinguality by Pre-training Modular Transformers (Pfeiffer et al., NAACL 2022) and first released in this repository.

Because it has been pre-trained with language-specific modular components (language adapters), X-MOD differs from previous multilingual models like XLM-R. For fine-tuning, the language adapters in each transformer layer are frozen.

Usage

Tokenizer

This model reuses the tokenizer of XLM-R.

Input Language

Because this model uses language adapters, you need to specify the language of your input so that the correct adapter can be activated:

from transformers import XmodModel

model = XmodModel.from_pretrained("facebook/xmod-base")
model.set_default_language("en_XX")

A directory of the language adapters in this model is found at the bottom of this model card.

Fine-tuning

In the experiments in the original paper, the embedding layer and the language adapters are frozen during fine-tuning. A method for doing this is provided in the code:

model.freeze_embeddings_and_language_adapters()
# Fine-tune the model ...

Cross-lingual Transfer

After fine-tuning, zero-shot cross-lingual transfer can be tested by activating the language adapter of the target language:

model.set_default_language("de_DE")
# Evaluate the model on German examples ...

Bias, Risks, and Limitations

Please refer to the model card of XLM-R, because X-MOD has a similar architecture and has been trained on similar training data.

Citation

BibTeX:

@inproceedings{pfeiffer-etal-2022-lifting,
    title = "Lifting the Curse of Multilinguality by Pre-training Modular Transformers",
    author = "Pfeiffer, Jonas  and
      Goyal, Naman  and
      Lin, Xi  and
      Li, Xian  and
      Cross, James  and
      Riedel, Sebastian  and
      Artetxe, Mikel",
    booktitle = "Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies",
    month = jul,
    year = "2022",
    address = "Seattle, United States",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2022.naacl-main.255",
    doi = "10.18653/v1/2022.naacl-main.255",
    pages = "3479--3495"
}

Languages

This model contains the following language adapters: