Trainable NLG for Data to Portuguese - With application to a Medication Assistant

Abstract

New equipments, such as smartphones and tablets, are changing human computer interaction. These devices present several challenges, especially due to their small screen and keyboard. In order to use text and voice in multimodal interaction, it is essential to deploy modules to translate the internal information of the applications into sentences or texts, in order to display it on screen or synthesize it. Also, these modules must generate phrases and texts in the user's native language; the development should not require considerable resources; and the outcome of the generation should achieve a good degree of variability.

Our main objective is to propose, implement and evaluate a method of data conversion to Portuguese which can be developed with a minimum of time and knowledge, but without compromising the necessary variability and quality of what is generated. The developed system, for a Medication Assistant, is intended to create descriptions, in natural language, of medication to be taken. Motivated by recent results, we opted for an approach based on machine translation, with models trained on a small parallel corpus.

For that, a new corpus was created. With it, two variants of the system were trained: phrase-based translation and syntax-based translation. The two variants were evaluated by automatic measurements -- BLEU and Meteor -- and by humans. The results showed that a phrase-based approach produced better results than a syntax-based one: human evaluators evaluated 60% of phrase-based responses as good, or very good, compared to only 46% of syntax-based responses. Considering the corpus size, we judge this value (60%) as good.