Don’t Invite BERT to Drink a Bottle: Modeling the Interpretation of Metonymies Using BERT and Distributional Representations

In this work, we carry out two experiments in order to assess the ability of BERT to capture themeaning shift associated with metonymic expressions. We test the model on a new dataset that isrepresentative of the most common types of metonymy. We compare BERT with the StructuredDistributional Model (SDM), a model for the representation of words in context which is basedon the notion of Generalized Event Knowledge. The results reveal that, while BERT abilityto deal with metonymy is quite limited, SDM is good at predicting the meaning of metonymicexpressions, providing support for an account of metonymy based on event knowledge