The Impact of Romanian Diacritics on Intent Detection and Slot Filling

This paper aims to provide an empirical analysis of the effect of Romanian diacritics on the performance of two well known frameworks for intent detection and slot filling: the RASA framework and Facebook's Wit.ai API. We use a previously defined home assistant scenario and datasets, which we modify to include diacritics. The results of several alternative processing pipelines in the two frameworks suggest that diacritics introduce a new level on learning complexity, in addition to the ones considered initially - synonymy, missing slot values and class imbalance - and they contribute to augmenting the effect of those complexities on the performance of the learning models. Moreover, the slot filling task seems to be the most affected by this aspect, with a drop in performance as high as ~30% for RASA and more that 40% for Wit.ai.