Are Word Embeddings Really a Bad Fit for the Estimation of Thematic Fit?

While neural embeddings represent a popular choice for word representation in a wide variety of NLP tasks, their usage forthematic fitmodelinghas been limited, as they have been reported to lag behind syntax-based count models. In this paper, we propose a completeevaluation of count models and word embeddings on thematic fit estimation, by taking into account a larger number of parametersand verb roles and introducing also dependency-based embeddings in the comparison. Our results show a complex scenario, wherea determinant factor for the performance seems to be the availability to the model of reliable syntactic information for building thedistributional representations of the roles.