Disentangling modal meanings with distributional semantics

This paper investigates the collocational behavior of English modal auxiliaries such as may and might with the aim of finding corpus-based measures that distinguish between different modal expressions and that allow insights into why speakers may choose one over another in a given context. The analysis uses token-based semantic vector space modeling (Heylen et al. 2015, Hilpert and Correia Saavedra 2017) in order to determine whether different modal auxiliaries can be distinguished in terms of their collocational profiles. The analysis further examines whether different senses of the same auxiliary exhibit divergent collocational preferences. The results indicate that near-synonymous pairs of modal expressions, such as may and might or must and have to, differ in their distributional characteristics. Also different senses of the same modal expression, such as deontic and epistemic uses of may, can be distinguished on the basis of distributional information. We discuss these results against the background of previous empirical findings (Hilpert 2016, Flach in press) and theoretical issues such as degrees of grammaticalization (Correia Saavedra 2019) and the avoidance of synonymy (Bolinger 1968).