Recent work suggests that language production exhibits a biastowards efficient information transmission. Speakers tendto provide more linguistic signal for meaning elements thatare difficult to recover while reducing contextually inferrable(more frequent, probable, or expected) elements. This trade-off has been hypothesized to shape grammatical systems overgenerations, contributing to cross-linguistic patterns. We putthis idea to an empirical test using miniature artificial languagelearning over variable input. Two experiments were conductedto demonstrate that the inferrability of plurality informationinversely predicts the likelihood of overt plural marking, aswould be expected if learners prefer communicatively efficientsystems. The results were obtained even with input frequencycounts of the plural marker counteracting the bias, and thusprovide strong support for a critical role of inferrability ofmeaning in language learning, production, as well as in typo-logically attested variations.

To navigate many-to-many mappings between referents and linguistic expressions, listeners need to calibrate likelihood estimates for different referential expressions taking into account both the context and speaker-specific variation. Focusing on speaker variation, we present three experiments. Experiment 1 establishes that listeners generalize speakerspecific patterns of pre-nominal modification use across different adjective types. Experiment 2 examined a) the dimension of generalization (form-based or informativitybased); b) effects of the strength of the evidence (implicit or explicit); and c) individual differences in dimensions of generalization. Experiment 3 asked parallel questions for exposure to over-specified utterances; we predicted more conservative generalizations because in spontaneous utterances, speakers are more likely to over-modify than under-modify.