Improving Translation Capabilities of Pre-Trained Multilingual Sequence-to-Sequence Models for Low-Resource Languages

En-Shiun Annie Lee

Keywords: Pre-Trained Multilingual Sequence-to-Sequence Model, Machine Translation, Low Resource Languages, Empirical Experiments

TL;DR: Empirical experiments on data and techniques for leveraging Pre-Trained Multilingual Sequence-to-Sequence Models for Low-Resource language translation

Abstract: Performance of Pre-trained Multilingual Sequence-to-Sequence (PMSS) models for translation heavily depends on the amount of monolingual data used in model pre-training. Thus these models under-perform for low-resource languages included in the model, even more for the languages unseen by the model. In this paper, we focus on the domain-specific translation of low-resource language (LRL) pairs. For a given domain-specific translation task, we investigate the most effective way of utilizing parallel data from auxiliary domains. the possibility of leveraging the available bitext to improve translation capabilities of PMSS models for low-resource languages. We experiment with several Transfer Learning protocols, considering the domain divergence of the available data.} %Therefore, to determine whether pre-trained multilingual sequence-to-sequence (PMSS) models can be leveraged for low-resource language (LRL) translation, we conducted a large-scale empirical experiment. We systematically studied transfer learning frameworks under a breadth of data scenarios under for LRLs in a non-English-centric manner.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics

Submission Guidelines: Yes

Please Choose The Closest Area That Your Submission Falls Into: Applications (eg, speech processing, computer vision, NLP)