Evaluating the performance of multilingual models in answer extraction and question generation

Abstract

Abstract Multiple-choice test generation is one of the most complex NLP problems, especially in languages other than English, where there is a lack of prior research. After a review of the literature, it has been verified that some methods like the usage of rule-based systems or primitive neural networks have led to the application of a recent architecture, the Transformer architecture, in the tasks of Answer Extraction (AE) and Question Generation (QG). Thereby, this study is centred in searching and developing better models for the AE and QG tasks in Spanish, using an answer-aware methodology. For this purpose, three multilingual models (mT5-base, mT0-base and BLOOMZ-560 M) have been fine-tuned using three different datasets: a translation to Spanish of the SQuAD dataset; SQAC, which is a dataset in Spanish; and their union (SQuAD + SQAC), which shows slightly better results. Regarding the models, the performance of mT5-base has been compared with that found in two newer models, mT0-base and BLOOMZ-560 M. These models were fine-tuned for multiple tasks in literature, including AE and QG, but, in general, the best results are obtained from the mT5 models trained in our study with the SQuAD + SQAC dataset. Nonetheless, some other good results are obtained from mT5 models trained only with the SQAC dataset. For their evaluation, the widely used BLEU1-4, METEOR and ROUGE-L metrics have been obtained, where mT5 outperforms some similar research works. Besides, CIDEr, SARI, GLEU, WER and the cosine similarity metrics have been calculated to present a benchmark within the AE and QG problems for future work.