Exploring politeness control in NMT: fine-tuned vs. multi-register models in Castilian Spanish

Resumen

Nowadays neural machine translation can generate high quality translations with regard to grammatical accuracy and fluency. Therefore, it is time to broaden research efforts to consider aspects of language that go beyond the mentioned attributes to keep pushing the limits of the technology. In this work, we focus on politeness. Specifically, we adapt and explore, for Castilian Spanish, two different domain-adaptation approaches: fine-tuning and multilingual models. Results from automatic and manual evaluations seem to indicate that the latter might be a better solution to strike a quality balance between all registers (formal, informal, and neutral). Fine-tuning a baseline system for each specific register seems to suffer from a degree of catastrophic forgetting, which leads to a worse overall performance of the engines.