Resultatives in word embeddings

Abstract

Verbs can be categorised into classes based on the syntactic frames and alternations they can and cannot participate in (Levin, 1993). This study investigates whether pre-trained language model (PLM) word embeddings can be used to determine if a verb participates in the resultative construction. This construction describes that a change of state has taken place as the result of an action (Goldberg and Jackendoff, 2004; Levin, 1993). In this thesis, we extend the lexical dataset LaVA presented in Kann et al. (2019), to include the resultative construction, adding 70 new verbs, and extending the annotations of the existing verbs in the dataset. We further train a logistic regression classifier on the verb embeddings from BERT (Devlin et al., 2019) to determine which verbs participate in which frames. Our analysis shows that the performance on resultative frames using static pre-trained embeddings is consistent with similar works.