Traditional models of language comprehension assume that language processing involves recognizing patterns, for example, words, by mapping the signal onto existing representations, retrieving information associated with these stored representations, and then using rules based on abstract categories (e.g., syntactic rules) to build structured representations. Four aspects of the literature are inconsistent with this framework. First, listeners are exquisitely sensitive to fine-grained, sub-categorical properties of the signal, making use of this information rather than discarding it (McMurray et al. Reference McMurray, Tanenhaus and Aslin2009). Second, comprehenders rapidly integrate constraints at multiple grains. Third, they generate expectations about likely input at multiple levels of representation. Finally, adaptation is ubiquitous in language processing. These results can be unified if we assume that comprehenders use internally generated predictions at multiple levels to explain the source of the input, and that prediction error is used to update the generative models in order to facilitate more accurate predictions in the future.

Extended to the domain of language processing, Clark's framework predicts that expectations at higher levels of representation (e.g., syntactic expectations) should constrain interpretation at lower levels of representation (e.g., speech perception). According to this view, listeners develop fine-grained probabilistic expectations about how lexical alternatives are likely to be realized in context (e.g., net vs. neck) that propagate from top to bottom through the levels of a hierarchically organized system representing progressively more fine-grained perceptual information. Provisional hypotheses compete to explain the data at each level, with the predicted acoustic realization of each alternative being evaluated against the actual form of the input, resulting in a residual feed-forward error signal propagated up the hierarchy. As the signal unfolds, then, the activation of a particular lexical candidate should be inversely proportional to the joint error signal at all levels of the hierarchy (i.e., the degree of divergence between the predicted acoustic realization of that candidate and the actual incoming signal), such that candidate words whose predicted realizations are most congruent with the acoustic signal are favored.

Hierarchical predictive processing therefore provides a potential explanatory framework for understanding a wide variety of context effects and cue integration phenomena in spoken word recognition. Converging evidence suggests that the initial moments of competition between lexical alternatives are constrained by multiple sources of information from different dimensions of the linguistic input (e.g., Dahan & Tanenhaus Reference Dahan and Tanenhaus2004; Kukona et al. Reference Kukona, Fang, Aicher, Chen and Magnuson2011), including information external to the linguistic system, such as visually conveyed social information (Hay & Drager Reference Hay and Drager2010; Staum Casasanto Reference Staum Casasanto, Love, McRae and Sloutsky2008) and high-level information about a speaker's linguistic ability (Arnold et al. Reference Arnold, Hudson and Tanenhaus2007). Crucially, lexical processing is influenced by information preceding the target word by several syllables or clauses (Dilley & McAuley Reference Dilley and McAuley2008; Dilley & Pitt Reference Dilley and Pitt2010) and this information affects listeners' expectations (Brown et al. Reference Brown, Salverda, Dilley and Tanenhaus2011; Reference Brown, Dilley, Tanenhaus, Miyake, Peebles and Cooper2012). The integration of these various constraints, despite their diversity, is consistent with the hypothesis that disparate sources of constraint are integrated within generative models in the language processing system.

Clark's framework also helps explain a recent set of results on context effects in reading that are surprising from the viewpoint of more traditional theories that emphasize the bottom-up, feed-forward flow of information. Farmer et al. (Reference Farmer, Christiansen and Monaghan2006) demonstrated that when a sentential context conferred a strong expectation for a word of a given grammatical category (as in The child saved the…, where a noun is strongly expected), participants were slower to read the incoming noun when the form of it (i.e., its phonological/orthographic properties) was atypical with respect to other words in the expected category. In a subsequent MEG experiment, Dikker et al. (Reference Dikker, Rabagliati, Farmer and Pylkkanen2010) showed that at about 100 msec post-stimulus onset – timing that is unambiguously associated with perceptual processing – a strong neural response was elicited when there was a mismatch between form and syntactic expectation. Moreover, the source of the effect was localized to the occipital lobe, suggesting that the visual system had access to syntactic representations. These results provide support for Clark's hypothesis that “if the predictive processing story is correct, we expect to see powerful context effects propagating quite low down the processing hierarchy” (sect. 3.1, para. 8). Linguistic context is used to generate expectations about form-based properties of upcoming words, and these expectations are propagated to perceptual cortices (Tanenhaus & Hare Reference Tanenhaus and Hare2007).

This framework also serves to specify the functionality of the prediction error that arises when some degree of mismatch between a prediction and the incoming signal occurs. In behavioral and Event-Related Potential (ERP) experiments, prediction-input mismatch frequently results in increased processing difficulty, typically interpreted as evidence that prediction is being made. But, under Clark's framework, the error signal assumes functionality; in part, it serves to adjust higher-level models such that they better approximate future input. The explanatory power of this hypothesis can best be seen when considering the large amount of relatively recent literature on adaptation within linguistic domains. Whether in the domain of speech perception (Kleinschmidt & Jaeger Reference Kleinschmidt and Jaeger2011; Kraljic et al. Reference Kraljic, Samuel and Brennan2008), syntactic processing (Farmer et al. Reference Farmer, Monaghan, Misyak and Christiansen2011; Fine et al. under review; Wells et al. Reference Wells, Christiansen, Race, Acheson and MacDonald2009), prosody (Kurumada et al. Reference Kurumada, Brown, Tanenhaus, Miyake, Peebles and Cooper2012), or pragmatics (Grodner & Sedivy Reference Grodner, Sedivy, Gibson and Pearlmutter2011), it has become increasingly apparent that readers and listeners continually update their expectations about the likelihood of encountering some stimulus based on their exposure to the statistical regularities of a specific experimental context. Adaptation of expectations is predicted by Clark's framework, and it may be taken as evidence that prediction-input mismatch produces an error signal that is fed forward to update the relevant generative models.

In sum, Clark's hierarchical prediction machine hypothesis provides a framework that we believe will unify the literature on prediction in language processing. This unification will necessarily involve systematic examination of what aspects of the stimulus are predicted, when in the chain of processing these predictions are generated and assessed, and the precise form of these generative models. This task will be challenging because it is likely that generative models use signal-relevant properties that do not map to the standard levels of linguistic representation that are incorporated into most models of language processing.