Predicting the proto-typical "who", "what" and "where" of eating
Being able to learn information about proto-typical event representations, i.e., learning probability distributions over the typical "who"/"whom"/"how"/"where" for specific verbs has many applications in natural language understanding (e.g. inferring a location even if it's not mentioned) as well as cognitive modelling.
Consider for example the verb "to serve": a model also needs to encode which event participants typically occur together: waiters serve food in a restaurant, mum serves food at home, priests serve God in a church, and prisoners serve sentences in jail.
In this talk, I will give an overview of my group's current research in psycholinguistics, cognitive modelling and NLP and will then delve into more depth on recent models we have developed to address the above question. Specifically, I will discuss the use of large vector space models compared to neural networks on the task of predicting typical thematic role fillers based on large amounts of text.