A central tenet of cognitive linguistics is that adults’ knowledge of language consists of a structured inventory of constructions, including various two-argument constructions such as the active (e.g., Lizzy rescued John), the passive (e.g., John was rescued by Lizzy) and “fronting” constructions (e.g., John was the one Lizzy rescued). But how do speakers choose which construction to use for a particular utterance, given constraints such as discourse/information structure and the semantic fit between verb and construction? The goal of the present study was to build a computational model of this phenomenon for two-argument constructions in Mandarin. First, we conducted a grammaticality judgment study with 60 native speakers which demonstrated that, across 57 verbs, semantic affectedness – as determined by further 16 native speakers – predicted each verb’s relative acceptability in the bei-passive and ba-active constructions, but not the Notional Passive and SVO Active constructions. Second, in order to simulate acquisition of these competing constraints, we built a computational model that learns to map from corpus-derived input (information structure + verb semantics + lexical verb identity) to an output representation corresponding to these four constructions (+“other”). The model was able to predict judgments of the relative acceptability of the test verbs in the ba-active and bei-passive constructions obtained in Study 1, with model-human correlations in the region of r = 0.5 and r = 0.3, respectively. Surprisingly, these correlations increased (to r = 0.75 and r = 0.5 respectively) when lexical verb identity was removed; perhaps because this information leads to over-fitting of the training set. These findings suggest the intriguing possibility that acquiring constructions involves forgetting as a mechanism for abstracting across certain fine-grained lexical details and idiosyncrasies.