Implicit indefinite objects at the syntax-semantics-pragmatics interface: a probabilistic model of acceptability judgments

Optionally transitive verbs, whose Patient participant is semantically obligatory but syntactically optional (e.g., to eat, to drink, to write), deviate from the transitive prototype defined by Hopper and Thompson (1980). Following Fillmore (1986), unexpressed objects may be either indefinite (referring to prototypical Patients of a verb, whose actual entity is unknown or irrelevant) or definite (with a referent available in the immediate intra- or extra-linguistic context). This thesis centered on indefinite null objects, which the literature argues to be a gradient, non-categorical phenomenon possible with virtually any transitive verb (in different degrees depending on the verb semantics), favored or hindered by several semantic, aspectual, pragmatic, and discourse factors. In particular, the probabilistic model of the grammaticality of indefinite null objects hereby discussed takes into account a continuous factor (semantic selectivity, as a proxy to object recoverability) and four binary factors (telicity, perfectivity, iterativity, and manner specification). This work was inspired by Medina (2007), who modeled the effect of three predictors (semantic selectivity, telicity, and perfectivity) on the grammaticality of indefinite null objects (as gauged via Likert-scale acceptability judgments elicited from native speakers of English) within the framework of Stochastic Optimality Theory. In her variant of the framework, the constraints get floating rankings based on the input verb’s semantic selectivity, which she modeled via the Selectional Preference Strength measure by Resnik (1993, 1996). I expanded Medina’s model by modeling implicit indefinite objects in two languages (English and Italian), by using three different measures of semantic selectivity (Resnik’s SPS; Behavioral PISA, inspired by Medina’s Object Similarity measure; and Computational PISA, a novel similarity-based measure by Cappelli and Lenci (2020) based on distributional semantics), and by adding iterativity and manner specification as new predictors in the model. Both the English and the Italian five-predictor models based on Behavioral PISA explain almost half of the variance in the data, improving on the Medina-like three-predictor models based on Resnik’s SPS. Moreover, they have a comparable range of predicted object-dropping probabilities (30-100% in English, 30-90% in Italian), and the predictors perform consistently with theoretical literature on object drop. Indeed, in both models, atelic imperfective iterative manner-specified inputs are the most likely to drop their object (between 80% and 90%), while telic perfective non-iterative manner-unspecified inputs are the least likely (between 30% and 40%). The constraint re-ranking probabilities are always directly proportional to semantic selectivity, with the exception of Telic End in Italian. Both models show a main effect of telicity, but the second most relevant factor in the model is perfectivity in English and manner specification in Italian.