Preface to special issue “prediction in language comprehension and production”

Humans use prediction in many cognitive domains, for example, to avoid getting hit by a bus, to putt a golf ball, or to find the right moment to interrupt a chatty interlocutor. It is experimentally well established that we make predictions during language comprehension and production, whether during reading, listening or conversation. It is less clear, however, how prediction functionally contributes to different aspects of language processing, from speech perception to language acquisition, and by which cognitive and neural mechanisms this is accomplished. Pre-activation, priming, predictive coding, Bayesian inference – are there fundamental principles of prediction that apply to all aspects of language processing, or do different phenomena require different approaches?

For this special issue “Prediction in language comprehension and production” in Language, Cognition and Neuroscience we invited opinion papers, theory-guided reviews and empirical studies that offer new insights into the phenomenon of prediction. They do so by defining or differentiating the concept of prediction itself, by providing behavioural or neuroscientific evidence for the functional role of prediction in specific domains of language processing, or by suggesting computational or neural mechanisms by which prediction may be implemented.

Kuperberg and Jaeger (2016) unpack the “loaded” concept of prediction at several levels. They address questions about the levels of representation at which prediction operates, how predictions are used and updated in context, their implementation at the algorithmic and neural levels, and whether there are predictive mechanisms that are generally involved in language comprehension. Models based on surprisal and prediction error, pre-activation and integration are discussed. The authors argue that prediction can be best understood within a probabilistic framework, taking into account the goals of the comprehender and the utility or prediction, and involving generative models, Bayesian inference and principles of predictive coding.

Huettig and Mani (2016) argue that prediction is not necessary for many aspects of language comprehension. They discuss five arguments that have been used to argue for a general role or prediction in language comprehension, and conclude for example that there is currently little empirical evidence that prediction is necessary for language learning, that the detection of statistical regularities is based on prediction, or that predictive coding is a fundamental neural principle for language processing. Rather, the literature shows that prediction varies considerably with respect to subject population and is highly context-dependent, and much of the evidence in favour of predictive mechanisms stems from “prediction-encouraging” paradigms. They doubt that prediction provides a unified framework for language science, and instead suggest that prediction may lend a “helping hand” in tasks where it can be beneficial for performance.

Norris, McQueen, and Cutler (2016) challenge the commonly held view that prediction implies feedback from higher to lower levels of the processing hierarchy. They argue that at least for speech recognition, simple activation feedback from lexical to pre-lexical levels is not beneficial. They analyse the possible roles of feedback within a Bayesian inference framework, and conclude that feedback can be beneficial for the adaptation to changing inputs and to distinguish competing sounds.

Gambi and Pickering (2016) investigate the similarities between predicting and imagining language. They suggest imagination is a form of “offline prediction”. Different kinds of prediction, whether they are concerned with oneself or others, are mirrored by different kinds of imagination. For example, predicting or imagining other people's utterances may occur via corresponding simulation or association routes, depending on the comprehender's knowledge about the speaker's intentions, and the similarity between speaker and comprehender. The authors then discuss how these concepts can be used to analyse the role of prediction in conversation.

Lowder and Ferreira (2016) point out that the role of prediction in the repair of disfluent speech is still under-researched. They review existing findings that demonstrate that listeners use semantic information about the to-be-repaired part of speech as well as contextual information to predict likely repairs. They reconcile two theoretical approaches that differ with respect to their reliance on statistical and linguistic information required for repair mechanisms, namely the Overlay and the Noisy-Channel model. While the latter can be used to describe how our comprehension system detects disfluencies using probabilistic analysis of the input, the former may explain how such disfluencies are repaired. They propose future experiments to investigate the mechanisms of disfluency repairs in speech comprehension in more detail, highlighting the usefulness of visual world paradigms.

Huettig and Janse (2016) emphasise the importance of individual differences in predictive strategies. They present results from a novel study using an individual differences approach, showing that working memory and processing speed explain most of the variance (compared to a non-verbal intelligence measure and age) in language-mediated anticipatory eye movements in a visual world paradigm. They argue that these factors should be taken into account by models of predictive language processing, and that working memory poses a crucial link between linguistic and visual–spatial representations.

Rabagliati, Gambi, and Pickering (2016) critically review the evidence for the view that children learn language by testing their predictions on speakers’ behaviour. Language learning has traditionally been described as an off-line process, and the role of online mechanisms such as prediction is not yet clear. It is possible that children use prediction to aid learning, but also that as a consequence of language development they learn to predict. The authors describe several theoretical perspectives, including network models and models based on prediction error. Empirically distinguishing between these different theories is not easy, in particular with respect to the role of predictive learning. The authors lay out necessary criteria for predictive learning, and argue on the basis of their literature review that the current evidence for this view is suggestive but not conclusive.

Baayen, Shaoul, Willits, and Ramscar (2016) argue that learning the skill of speech comprehension does not rely on the isolation of word forms, and take a non-decompositional computational perspective. In their view, the relationship between form and meaning results from discriminative learning that directly links forms and meanings without intermediate abstract representations such as phonemes or word forms. They use a two-layer Rescorla-Wagner network to demonstrate that their approach can explain a range of phenomena including categorical perception and speech acquisition.

Johnson, Turk-Browne, and Goldberg (2016) investigated in two functional magnetic resonance imaging (fMRI) studies whether abstract phrasal patterns can be used to generate predictions during sentence comprehension. Their results suggest that adult comprehenders can use newly learned knowledge of grammatical constructions to predict the content of visual scenes. They identified ventral striatum and occipital cortex as brain structures that are implicated in the corresponding prediction processes. The ventral striatum in particular appears to be sensitive to prediction error.

Molinaro, Monsalve, and Lizarazu (2016) discuss the role of oscillatory brain dynamics for prediction mechanism in language comprehension and production. They review the electroencephalography and magnetoencephalography (EEG and MEG) literature on language-related brain oscillations. In one of their own MEG studies, they found differences in beta band desynchronisation prior to target words embedded in sentences, depending on prediction demands. These results resembled those found in previous production studies. They argue that beta band desynchronisation may reflect mechanisms of predictive coding during language processing, but note that we need more fine-grained paradigms to characterise the neurophysiological link between language comprehension and production in more detail.

Simanova, Francken, De Lange, and Bekkering (2016) discuss the interaction between top-down and bottom-up information processing during conceptual processing in the brain, with a focus on the role of predictive coding. They argue that predictive mechanisms would require early modulation of sensory brain responses in sensory areas, as well as modulation of brain activity before stimulus presentation. They provide a brief review of fMRI and electrophysiological studies, and conclude that the empirical evidence for these requirements is still inconclusive. In particular, temporal aspects of language–perception interaction and the possible predictive mechanisms require further investigation.

All authors agree that prediction is not a unitary phenomenon, and that it can appear in many different colours, depending on whether we look, for example, at speech perception, sentence comprehension, discourse, language acquisition, etc. “Prediction” cannot be considered a magic wand that elucidates the mysteries of language with a single touch. First of all, we need explicit computational models that specify how prediction might contribute to specific aspects of language processing. Our special issue discusses several theoretical approaches, such as neural network models, Bayesian decision-making and predictive coding. Second, we need experimental approaches and methodology that can test specific predictions of these models. Our paradigms should not be biased towards prediction strategies, and should take the influence of tasks, context and individual strategies into account. Our methodology should be able to dissociate effects prior to stimulus presentation from those during stimulus processing, and for the latter should be able to distinguish early perceptual biases from later selection or decision processes. It will be exciting to watch the development of methodology that combines the high temporal resolution of EEG and MEG with online behavioural measures, such as eye movements in visual world paradigms or text reading. Our issue provides the research community with an overview over the different aspects of prediction, and presents novel experimental and methodological approaches to characterise those in more detail.

Disclosure statement

No potential conflict of interest was reported by the author.