Interpersonal verbs often implicitly attribute the cause for the action or change of state they express to either their subject or their object. Knowledge of this semantic property of verbs is key to recognising the correct causal links in text. This implicit information can, however, pose a processing challenge when it contradicts explicit signals from the context of the verb. We investigate implicit causality bias and its interference effects in large pre-trained language models with methods drawn from psycholinguistic studies of this phenomenon.
Yova is a PhD fellow, soon to be PostDoc, at the CoAStaL NLP group (University of Copenhagen). Interested in multilinguality, low-resource training, language modeling and language itself.
Related work from the talk: link