More details will provided soon!
Official repository of the paper Promoting Generalized Cross-lingual Question Answering in Few-resource Scenarios via Self-knowledge Distillation containing the implementation to reproduce it.
Please, refer to the preprint for more details about the fundamental ideas, the method and the evaluation results: URL
First, install Python dependencies and extracting training and evaluation data.
bash ./setup.shThe script creates a Python virtualenv in the venv directory and extract all the necessary data in the corpora directory.
We provide simple script to train the models in the paper. For example, to train the overall best performing model, referred in the paper as mbert-qa-en, skd, mAP@k, run the following steps:
- Standard cross-entropy fine-tuning of the mBERT model for the extractive QA task using the SQuAD v1.1 training dataset in English.
bash train_qa_mbert.sh- Cross-lingual fine-tuning of the previous mBERT model, called mBERT-qa-en, using self-knowledge distillation with mAP@k loss coefficients.
bash train_skd_map.shThe result will be stored in the runs directory along with the tensoboard logs.
All other scripts will fine-tuning the mBERT-qa-en model with different methods. Note that each script is use a configuration of hyperparameters correspoding to the best models. Change the configuration inside the scripts to train different models.
We also provide scripts to evaluate the model after training. For example, to evaluate on the MLQA-test dataset, run:
bash eval_qa.sh <model_path_trained_model> mlqa-testThe evaluation result will be stored inside the trained model directory under the name eval_results_mlqa-test
Is it possible to choose another test set between xquad, mlqa-dev and tydiqa-goldp datasets.
We uploaded the best-performing models on the HuggingFace Models Hub under the names: mBERT-qa-en, skd and mBERT-qa-en, skd, mAP@k
To cite our work use the following BibTex:
@misc{carrino2023promoting,
title={Promoting Generalized Cross-lingual Question Answering in Few-resource Scenarios via Self-knowledge Distillation},
author={Casimiro Pio Carrino and Carlos Escolano and José A. R. Fonollosa},
year={2023},
eprint={2309.17134},
archivePrefix={arXiv},
primaryClass={cs.CL}
}