Measuring Transitivity Using Untrained Annotators

摘要: Hopper and Thompson (1980) defined a multi-axis theory of transitivity that goes beyond simple syntactic captures how much "action" takes place in sentence. Detecting these features requires deep understanding lexical semantics real-world pragmatics. We propose two general approaches for creating corpus sentences labeled with respect to the Hopper-Thompson schema using Amazon Mechanical Turk. Both assume no existing resources incorporate all necessary annotation into single system; this is done allow future generalization other languages. The first task attempts use language-neutral videos elicit human-composed specified attributes. second uses an iterative process label actors objects then annotate sentences' transitivity. examine success techniques perform preliminary classification held-out data.