How Long Does a Quick Kiss Take? Studying Event Duration of Light Verb Constructions Using Explicit Word Embeddings

Abstract

Psycholinguistic research indicates that choosing one syntactic construction over another to describe an event can influence its perceived duration: Light Verb Constructions (LVCs) such as punctive events in count syntax (to give a kiss) and durative events in mass syntax (to do research) are perceived as taking less time than their Full Verb Constructions (FVCs; to kiss and to research). Similar computational results were achieved using BERT embeddings to semantically project events onto a one-dimensional Duration scale. We reproduce and further develop this experiment with explicit word embeddings from our own co-occurrence count-based vector space. By semantically projecting 158 LVC-FVC pairs onto our Duration scale, we find that LVCs are modelled as significantly shorter than FVCs. However, we do not find an overall statistically significant difference in duration between sentences containing the target LVCs and FVCs. We demonstrate that semantic properties observed in human experiments and in BERT embeddings can also be modelled using explicit word embeddings, which have the advantage of being fully transparent. However, using transcripts from spoken conversations can be challenging when studying a specific construction: optimising the extraction of sentences containing the target expressions and composition of their meanings are to be addressed in future work.

Resources

Details