The Perfect Recipe: Add SUGAR, Add Data

We present the FBK participation at the EVALITA 2018 Shared Task ``SUGAR -- Spoken Utterances Guiding Chef's Assistant Robots''. There are two peculiar, and challenging, characteristics of the task: first, the amount of available training data is very limited; second, training consists of pairs \texttt{[audio-utterance, system-action]}, without any intermediate representation. Given the characteristics of the task, we experimented two different approaches: (i) design and implement a neural architecture that can use as less training data as possible, and (ii) use a state of art tagging system, and then augment the initial training set with synthetically generated data. In the paper we present the two approaches, and show the results obtained by their respective runs.