Pickering & Garrod (P&G) have done the field a huge favour by conceptualising language processing in terms of an integrated action-perception system, which highlights the centrality of prediction. It seems to us that taking a similar approach to the field of language acquisition would be equally productive. After all, the task of learning a language requires a close integration of oral motor action and sound perception. P&G's model adopts a theory of acquisition from motor learning, in which forward models learn to bridge between action and perception. In this commentary, we explore the implications of P&G's forward model for our understanding of psycholinguistic processes, with a particular focus on language acquisition.

There is now substantial evidence that children use prediction in comprehension in much the same way as adults. For example, Lew-Williams and Fernald (Reference Lew-Williams and Fernald2007) reported that Spanish-learning 3-year-old children can use the grammatical gender of the article (la/el) to predict the referent of the next word (la pelota vs. el zapato) in a looking-while-listening task. Mani and Huettig (Reference Mani and Huettig2012) have shown that 2-year-olds can use a verb's semantic affordances in a similar manner to adults (e.g., eat predicts cake), and Borovsky et al. (Reference Borovsky, Elman and Fernald2012) have reported similar findings in older children and adults. Importantly, the acquisition studies have demonstrated that children's predictive abilities correlate with their knowledge of language. In all of these studies, children (and adults) with bigger productive (and/or receptive) vocabularies tend to be faster and more accurate at prediction during online sentence comprehension. Faster prediction correlates with positive language outcomes longitudinally: Marchman and Fernald (Reference Marchman and Fernald2008) reported that the speed at which 25-month-olds process lexical information correlates with linguistic knowledge 6 years later. These findings suggest that prediction is not only part of the language processing system, but is tightly linked to language acquisition mechanisms.

Scrutinizing the mechanism underlying prediction therefore appears an important priority. Here we compare P&G's model to an alternative language production model, Chang's (Reference Chang2009) Dual-path model, concentrating on how each model explains prediction. By way of example, we consider verb-object affordances (e.g., eat-cake), where both children and adults have been found to predict a verb's object before the object is encountered in speech (Altmann & Kamide Reference Altmann and Kamide1999). These results are particularly interesting, because Mani and Huettig (Reference Mani and Huettig2012) found a significant relationship between expressive vocabulary and prediction behaviour in children, which supports both P&G's and the Dual-path model's hypothesis that production representations support prediction.

P&G's model places prediction within a forward model that implements constraints on word order (prediction-as-processing). The model explains verb-object affordances in the following manner: a production command EAT(CAKE) can be derived from hearing the word eat and seeing the cake in the visual scene. The command passes through the production forward system to activate the triplet p[cake, NP, /cake/]. Crucially, the model predicts that only syntactically and semantically appropriate predictions will be generated. In the eat-cake example, the prediction is correct, but is not borne out in every instance. For instance, Kamide et al. (Reference Kamide, Scheepers and Altmann2003) reported that participants made predictive looks to a cabbage after the processing the fragment The hare will be eaten …, where the passivized verb should have restricted looks to an animate agent (e.g., a fox).

In contrast to P&G's model, the Dual-path model acquires syntactic representations within a network that contains separate meaning and sequencing pathways (Elman Reference Elman1990). Its learning algorithm compares the predicted next word with the actual comprehended next word, and the mismatch or error is used to adjust the model's internal representations (error-based learning, Rumelhart et al. Reference Rumelhart, Hinton and Williams1986). The Dual-path model is able to explain the phenomena of structural priming as error-based learning (Chang et al. Reference Chang, Dell and Bock2006), and this ability requires that prediction-for-learning is constantly taking place during language comprehension. For the eat-cake prediction, the input word eat activates the concept CAKE because the model's meaning pathway learns associations between words in utterances and elements in messages. Critically, the same word-concept mechanism learns to associate eaten with cabbage. This associative word-concept prediction mechanism is different from the structure-sensitive prediction mechanism in the sequencing system, which can explain why eaten increases predictive looks to likely agents like the fox in Kamide et al. (Reference Kamide, Scheepers and Altmann2003). Thus, the pathways in the model can explain the different types of prediction. Crucially, the similarity in prediction in children and adults can be explained by the idea that humans are constantly doing prediction-for-learning to adjust their language representations to the input (Kidd Reference Kidd2012; Rowland et al. Reference Rowland, Chang, Ambridge, Pine and Lieven2012).

Learning processes can explain prediction in processing, but language acquisition constraints are also critical for learning the syntactic and semantic representations that support prediction in P&G's model. P&G's theory is based on Wolpert et al.'s (Reference Wolpert, Diedrichsen and Flanagan2011) theory of motor planning and perception, which uses error-based learning for motor and forward model learning (Jordan & Rumelhart Reference Jordan and Rumelhart1992; Plaut & Kello Reference Plaut, Kello and MacWhinney1999). According to Wolpert et al.'s theory, the forward model is learned by mapping from muscle commands to the perception of one's arm in three-dimensional space. These algorithms work because humans can directly perceive their arm's position. In P&G's theory, the forward model maps from a message-like production command to a triplet including syntax and semantics. This is problematic: We cannot directly perceive syntax and semantics, and hence the learning mechanism in the motor theory cannot explain how P&G's forward model learns to make these language predictions. When error-based learning is used to map from production commands to sentences, Chang (Reference Chang2002) demonstrated that abstract syntax was not always learned unless the model had language acquisition constraints like those in the Dual-path architecture. Therefore, P&G's forward model may need a similar architecture to yield the appropriate predictions.

Language processing theories like P&G's account treat language learning as a peripheral process. We argue that prediction in processing is actually a by-product of learning. Prediction is a critical component of error-based learning, which is one of the most successful accounts of both motor and language learning.