Previous studies using artificial languages suggest that sentence production can be guided by verb-specific as well as verb-general statistics present in the language input. Here we investigated whether the statistical properties of ongoing input in the speakers' native language systematically affected their sentence production. Three experiments used a training-plus production test paradigm to examine whether sentence production comes to rely more on verb-specific biases or on verb-general semantic mappings depending on the relative predictive validities of those cues in the input. We found different patterns in speakers' structural choices and utterance characteristics depending on the relative validity of different cues. These results show that native language production can be updated by new input and suggest that the nature of the updating can be explained at least in part by adopting a cue-validity approach.