We present a neural network model of referent identification in a visual world task. Inputs are visual representations of item pairs unfolding with sequences of phonemes identifying the target item. The model is trained to output the semantic representation of the target and to suppress the distractor. The training set uses a 200-word lexicon typically known by toddlers. The phonological, visual, and semantic representations are derived from real corpora. Successful performance requires correct association between labels and visual and semantic representations, as well as correct location identification. The model reproduces experimental evidence that phonological, perceptual, and categorical relationships modulate item preferences. The model provides an account of how language can drive visual attention in the inter-modal preferential looking task.