Differentiating valence patterns

A quantitative analysis based on formal and semantic attributes

Abstract

A verb licenses multiple constructions. For instance, the verb replace licences (at least) three constructions: (i) Alice replaced Bill with Charlotte, (ii) Alice replaced Bill, (iii) Bill replaced Charlotte. Each construction realises different arrays of semantic roles (i.e., Alice as the responsible entity for the transition of Bill and/or Charlotte, Bill as the ``new'' entity, and Charlotte as the ``old'' entity). Since each construction represent different meanings, a hearer (or reader) must identify the appropriate construction to understand a given utterance. Despite its theoretical importance, the cues used in identifying each construction have yet to be explored. This paper aims to reveal the features used in identifying each construction by annotating data with semantic features (i.e.,  Animacy of subjects and objects) and a formal feature (i.e., the presence of with-phrase). The analysis based on the conditional inference tree showed that active sentences with the verb replace can be classified by nearly 80% accuracy, which was fairly ``good'' (baseline: 35.4%). Moreover, the analysis of misclassifications revealed that a finer-grained characterisation of semantic types is needed for a more accurate specification of a construction, which suggests the necessity of event ontologies in semantic role labelling.