Children acquire mental state verbs (MSVs) much later than other, lower-frequency, words. One factor proposed to contribute to this delay is that children must learn various semantic and syntactic cues that draw attention to the difficult-to-observe mental content of a scene. A novel computational approach is developed to explore the role of such cues, and it is shown that the model can replicate aspects of the developmental trajectory of MSV acquisition. MSVs, such as think, know, and want, are very frequent in child-directed language, yet children use them productively much later than lower-frequency action verbs. Psycholinguistic theories have suggested that there is a delay in the acquisition of MSVs because they require certain cognitive and/or linguistic skills that are not available during the early stages of language development. For example, MSVs typically occur with a sentential complement (SC) that refers to the propositional content of the mental state, as in He thinks Mom went home. Children have to reach a stage of syntactic development that includes some facility with SCs in order to fully acquire MSVs. However, even at 3–5 years old, children are able to process SCs only imperfectly. Even when children are able to produce SCs with other verbs, there is a lag before they productively use MSVs referring to actual mental content. Psycholinguists have suggested that young children lack the conceptual ability to conceive that others have mental states separate from their own, further delaying the acquisition of MSVs. Another factor suggested to contribute to the difficulty of acquiring MSVs is their informational requirements. Children learn word meanings by figuring out which aspects of an observed scene are referred to by a particular word. MSVs often refer to aspects of the world that are not directly observable, such as the beliefs and desires of another entity. Thus, in addition to the above-mentioned challenges posed by children’s developing linguistic/conceptual abilities, children may simply have difficulty in identifying the relevant mental content necessary to learning MSVs. In particular, Papafragou et al. (2007) have shown that even given adequate conceptual and linguistic abilities, the mental events in a scene are not attended to as much as the actions, unless there are cues that heighten the salience of the mental content. PCG further demonstrate that children’s sensitivity to such cues lags behind that of adults, suggesting an additional factor in the acquisition of MSVs which is the developmental change in how strongly such cues are associated with the relevant mental content. A computational model of MSV acquisition is developed to further illuminate these issues. The model extends an existing model of verb argument structure acquisition to enable the representation and processing of mental state semantics and syntax. The model simulates the developmental change proposed by PCG through a gradually increasing ability in the model to appropriately attend to the mental content of a scene. In addition, the model suggests that even when the learner’s semantic representation is biased towards the action content, the learner attends to the observed SC syntax in an MSV utterance. This is especially important to account for the pattern of errors in child data. The model thus extends the account of PCG to show that a probabilistic interplay of the semantic and syntactic features of a partial and somewhat erroneous perception of the input, combined with a growing ability to attend to cues indicative of mental content, can help to account for children’s developmental trajectory in learning MSVs. To investigate the linguistic and contextual cues that could help in learning MSVs, PCG use a procedure called the Human Simulation Paradigm. In this paradigm, subjects are put in situations intended to simulate various word learning conditions of young children. For example, in one condition, adults watch silent videos of caregivers interacting with children, and are asked to predict the verb uttered by the caregiver. In another condition, subjects hear a sentence containing a nonce verb, and are asked what gorp might mean. The model focuses on two factors investigated by PCG in the performance of adults and children in identifying MSVs. The first factor they investigated involved the syntactic frame used when subjects were given a sentence with a nonce verb. PCG hypothesized that an SC frame would be a cue to mental content (and an MSV), since the SC refers to propositional content. The second factor PCG examined was whether the video described a “true belief” or a “false belief” scene. A true belief scene shows an ordinary situation which unfolds as the character in the scene expects, while a false belief scene has an unexpected outcome for the character. PCG’s results showed that both adults and children were sensitive to both the scene and syntax cues, but children’s ability to draw on such cues was inferior to that of adults. They thus propose that the difference between children and adults is that children have not yet formed as strong an association as adults between the cues and the mental content of a scene as required to match the performance of adults. Nonetheless, their results suggest that the participating children had the conceptual and linguistic abilities required for MSVs, since they were able to produce them under conditions with sufficiently strong cues. The model simulates PCG’s experiments using a novel computational approach. Following PCG, it is assumed that even when a learner is able to perceive the general semantic and syntactic properties of a belief scene and associated utterance, they may not attend to the mental content in every situation, and that this ability improves over time. The model models a developmental change in a learner’s attention to mental content: At early stages, corresponding to the state of young children, the learner largely focuses on the action aspects of a belief scene, even in the presence of an utterance using an MSV. Over time, the learner gradually increases in the ability to attend appropriately to the mental aspects of such a scene and utterance, until adult-like competence is achieved.
