Politeness control as a domain adaptation problem in NMT: fine-tuning vs. multi-register models for Castilian Spanish

Abstract

(EN) Given that Neural Machine Translation (NMT) systems have been proven to generate translations with a quality that is being regarded as close to that of a human, we believe it is the time to start paying attention to aspects of language that go beyond grammatical accuracy. In our research, we explore different approaches towards training a NMT system from English to Castilian Spanish with politeness control of the output and deal with the task as a domain-adaptation problem. Our results show that training a multi-register model to deal with different registers following Sennrich et al.’s approach (2016) is the best option when trying to find a balance between overall performance across different registers and types of segments as well as accuracy for producing the right honorific, while fine-tuning a baseline towards each specific register suffers from catastrophic forgetting, thus leading to a worse overall performance of such engines.