The Belebele Benchmark for Massively Multilingual NLU Evaluation

Belebele is a multiple-choice machine reading comprehension (MRC) dataset spanning 122 language variants. This dataset enables the evaluation of mono- and multi-lingual models in high-, medium-, and low-resource languages. Each question has four multiple-choice answers and is linked to a short passage from the FLORES-200 dataset. The human annotation procedure was carefully curated to create questions that discriminate between different levels of generalizable language comprehension and is reinforced by extensive quality checks. While all questions directly relate to the passage, the English dataset on its own proves difficult enough to challenge state-of-the-art language models. Being fully parallel, this dataset enables direct comparison of model performance across all languages. Belebele opens up new avenues for evaluating and analyzing the multilingual abilities of language models and NLP systems.

Please refer to our paper for more details, presented at ACL 2024: The Belebele Benchmark: a Parallel Reading Comprehension Dataset in 122 Language Variants. Or get more details at https://github.com/facebookresearch/belebele

Citation

If you use this data in your work, please cite:

@inproceedings{bandarkar-etal-2024-belebele,
    title = "The Belebele Benchmark: a Parallel Reading Comprehension Dataset in 122 Language Variants",
    author = "Bandarkar, Lucas  and
      Liang, Davis  and
      Muller, Benjamin  and
      Artetxe, Mikel  and
      Shukla, Satya Narayan  and
      Husa, Donald  and
      Goyal, Naman  and
      Krishnan, Abhinandan  and
      Zettlemoyer, Luke  and
      Khabsa, Madian",
    booktitle = "Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)",
    month = aug,
    year = "2024",
    address = "Bangkok, Thailand and virtual meeting",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2024.acl-long.44",
    pages = "749--775",
}

Composition

• 900 questions per language variant
• 488 distinct passages, there are 1-2 associated questions for each.
• For each question, there is 4 multiple-choice answers, exactly 1 of which is correct.
• 122 language/language variants (including English).
• 900 x 122 = 109,800 total questions.

Further Stats

• 122 language variants, but 115 distinct languages (ignoring scripts)
• 27 language families
• 29 scripts
• Avg. words per passage = 79.1 (std = 26.2)
• Avg. sentences per passage = 4.1 (std = 1.4)
• Avg. words per question = 12.9(std = 4.0)
• Avg. words per answer = 4.2 (std = 2.9)

Pausible Evaluation Settings

Thanks to the parallel nature of the dataset and the simplicity of the task, there are many possible settings in which we can evaluate language models. In all evaluation settings, the metric of interest is simple accuracy (# correct / total).

Evaluating models on Belebele in English can be done via finetuning, few-shot, or zero-shot. For other target languages, we propose the incomprehensive list of evaluation settings below. Settings that are compatible with evaluating non-English models (monolingual or cross-lingual) are denoted with ^.

No finetuning

• Zero-shot with natural language instructions (English instructions)
  • For chat-finetuned models, we give it English instructions for the task and the sample in the target language in the same input.
  • For our experiments, we instruct the model to provide the letter A, B, C, or D. We perform post-processing steps and accept answers predicted as e.g. (A) instead of A. We sometimes additionally remove the prefix The correct answer is for predictions that do not start with one of the four accepted answers.
  • Sample instructions can be found at the dataset github repo.
• Zero-shot with natural language instructions (translated instructions) ^
  • Same as above, except the instructions are translated to the target language so that the instructions and samples are in the same language. The instructions can be human or machine-translated.
• Few-shot in-context learning (English examples)
  • A few samples (e.g. 5) are taken from the English training set (see below) and prompted to the model. Then, the model is evaluated with the same template but with the passages, questions, and answers in the target language.
  • For our experiments, we use the template: P: <passage> \n Q: <question> \n A: <mc answer 1> \n B: <mc answer 2> \n C: <mc answer 3> \n D: <mc answer 4> \n Answer: <Correct answer letter>. We perform prediction by picking the answer within [A, B, C, D] that has the highest probability relatively to the others.
• Few-shot in-context learning (translated examples) ^
  • Same as above, except the samples from the training set are translated to the target language so that the examples and evaluation data are in the same language. The training samples can be human or machine-translated.

With finetuning

• English finetune & multilingual evaluation
  • The model is finetuned to the task using the English training set, probably with a sequence classification head. Then the model is evaluated in all the target languages individually. For results presented in the paper we used the HuggingFace library.
• English finetune & cross-lingual evaluation
  • Same as above, except the model is evaluated in a cross-lingual setting, where for each question, the passage & answers could be provided in a different language. For example, passage could be in language x, question in language y, and answers in language z.
• Translate-train ^
  • For each target language, the model is individually finetuned on training samples that have been machine-translated from English to that language. Each model is then evaluated in the respective target language.
• Translate-train-all
  • Similar to above, except here the model is trained on translated samples from all target languages at once. The single finetuned model is then evaluated on all target languages.
• Translate-train-all & cross-lingual evaluation
  • Same as above, except the single finetuned model is evaluated in a cross-lingual setting, where for each question, the passage & answers could be provided in a different language.
• Translate-test
  • The model is finetuned using the English training data and then the evaluation dataset is machine-translated to English and evaluated on the English.
  • This setting is primarily a reflection of the quality of the machine translation system, but is useful for comparison to multilingual models.

In addition, there are 83 additional languages in FLORES-200 for which questions were not translated for Belebele. Since the passages exist in those target languages, machine-translating the questions & answers may enable decent evaluation of machine reading comprehension in those languages.

Training Set

As discussed in the paper, we also provide an assembled training set consisting of samples at the github repo.

The Belebele dataset is intended to be used only as a test set, and not for training or validation. Therefore, for models that require additional task-specific training, we instead propose using an assembled training set consisting of samples from pre-existing multiple-choice QA datasets in English. We considered diverse datasets, and determine the most compatible to be RACE, SciQ, MultiRC, MCTest, MCScript2.0, and ReClor.

For each of the six datasets, we unpack and restructure the passages and questions from their respective formats. We then filter out less suitable samples (e.g. questions with multiple correct answers). In the end, the dataset comprises 67.5k training samples and 3.7k development samples, more than half of which are from RACE. We provide a script (assemble_training_set.py) to reconstruct this dataset for anyone to perform task finetuning.

Since the training set is a joint sample of other datasets, it is governed by a different license. We do not claim any of that work or datasets to be our own. See the Licenses section in the README of https://github.com/facebookresearch/belebele .

Languages in Belebele