[PS-3.79] The relationship between semantics and verb argument structure: A large-scale, crowd-sourced investigation

Hartshorne, J. 1 , Mu, J. 1 , O'Donnell, T. 2 & Palmer, M. 3

1 Boston College
2 Massachusetts Institute of Technology
3 University of Colorado-Boulder

Which verbs are compatible with which syntactic frames? In at least some cases, this seems to be explained by semantics (Levin & Rappaport Hovav, 2005). For instance, the reason ?Agnes broke at the vase? is ungrammatical is likely because the NP V at NP frame is incompatible with verbs entailing a change of state (cf. hit, kick).

However, theories disagree as to how much of argument structure can be explained semantically, how regular the patterns are, and what the specific patterns are (Levin & Rappaport Hovav, 2005). Distinguishing these theories is difficult because most investigations have focused on a small subset of verbs, and the conclusions may not generalize (Hartshorne et al., in press).

We report initial results from a wide-coverage investigation of 6,334 verbs. VerbNet (Kipper, Karin, et al., 2008) classifies these verbs into 487 verb classes based whether they can take each of 286 syntactic frames. We recruited 10,413 annotators who provided 526,670 semantic judgments about the verbs in allowed syntactic frames, focusing on semantic features commonly implicated by theory (see gameswithwords.org/VerbCorner/). The semantic judgments were highly consistent within class (>=95% consistency) while varying across classes, suggesting a highly regular effect of semantics on argument structure.