Elsevier

Cortex

Volume 68, July 2015, Pages 155-168
Cortex

Special issue: Review
A predictive coding framework for rapid neural dynamics during sentence-level language comprehension

https://doi.org/10.1016/j.cortex.2015.02.014Get rights and content

Abstract

There is a growing literature investigating the relationship between oscillatory neural dynamics measured using electroencephalography (EEG) and/or magnetoencephalography (MEG), and sentence-level language comprehension. Recent proposals have suggested a strong link between predictive coding accounts of the hierarchical flow of information in the brain, and oscillatory neural dynamics in the beta and gamma frequency ranges. We propose that findings relating beta and gamma oscillations to sentence-level language comprehension might be unified under such a predictive coding account. Our suggestion is that oscillatory activity in the beta frequency range may reflect both the active maintenance of the current network configuration responsible for representing the sentence-level meaning under construction, and the top-down propagation of predictions to hierarchically lower processing levels based on that representation. In addition, we suggest that oscillatory activity in the low and middle gamma range reflect the matching of top-down predictions with bottom-up linguistic input, while evoked high gamma might reflect the propagation of bottom-up prediction errors to higher levels of the processing hierarchy. We also discuss some of the implications of this predictive coding framework, and we outline ideas for how these might be tested experimentally.

Introduction

Reading, or listening to someone speaking, are the simple kinds of tasks that most people engage in every day of their lives without much difficulty. Yet if one considers that the average reader can easily manage between 250 and 300 words per minute (e.g., Rayner, Pollatsek, Ashby, & Clifton, 2012), it becomes clear that the processing carried out by the language comprehension system must be extremely fast and dynamic. One possible explanation for this speed (to be sure, one among many) is that the system may make predictions about upcoming linguistic input. From such a perspective it is surprising that models of language comprehension based on the passive building up of semantic and syntactic structures (from the lexical building blocks activated upon perception of linguistic input) dominated the psycholinguistics literature for so long (e.g., Forster, 1981, Seidenberg et al., 1982, Zwitserlood, 1989). Arguments that prediction was not likely to be involved in language comprehension were generally made based on the observation that at any point while reading or listening there are a large number of possible continuations. Processing costs involved in making incorrect predictions, along with the presumed low percentage of benefits accrued (predictions would not often be correct) made predictive processing accounts unappealing (see van Petten & Luka, 2012 for discussion).

On the other hand, a large number of studies began to show that the processing of a word in a sentence can be facilitated by the constraining sentence context (sometimes even before the word can be uniquely identified; e.g., Altmann and Kamide, 1999, Altmann and Kamide, 2007, Balota et al., 1985, van den Brink et al., 2001, Ehrlich and Rayner, 1981, Federmeier and Kutas, 1999, Kamide, 2008, Kamide et al., 2003, Kamide et al., 2003, Knoeferle et al., 2005, MacDonald et al., 1994, McRae et al., 2005, van Petten et al., 1999, Schwanenflugel and Lacount, 1988, Sussman and Sedivy, 2003). The idea that predictive processing could, at least in some circumstances, be beneficial for language comprehension has slowly grown in popularity. By now the notion that (at least some of the time) prediction plays an important role in rapid, dynamic, real-time language comprehension is a widely accepted view (Pickering & Garrod, 2007).

However, within this emerging view there are many outstanding questions. For instance, what are the details about exactly when prediction plays a role (is the system always making predictions or only under certain circumstances when this may be a useful strategy?). How do predictions interact with real-time comprehension? What kinds of information might lead to (strong) predictions? And, crucially, how does the brain implement predictive processing? While we briefly discuss each of these questions we acknowledge that it is not possible to do justice to them all in a single review. The main focus of this review is to outline some ways in which we think that the study of electrophysiology, and in particular oscillatory neural dynamics measured using electroencephalography (EEG) and magnetoencephalography (MEG) can contribute to our understanding of predictive processing during language comprehension beyond the level of individual words.

In the last ten to fifteen years a number of ERP studies have investigated the potential role of prediction during sentence-level language comprehension (see e.g., van Berkum et al., 2005, DeLong et al., 2005, Otten et al., 2007, Szewczyk and Schriefers, 2013, Wicha et al., 2004). The common ingredient used in all these studies to investigate predictive processing was agreement relations between a particular noun and some element preceding the noun. If the constraining sentence context allows readers/listeners to make predictions about the following noun, then lexical information associated with that noun should be available to the comprehension system before the noun is presented and should have an influence on the processing of agreement relations between the noun and the element preceding it.

An effect of prediction on ERP responses has been shown in the context of both gender-marked determiners (Wicha et al., 2004), and adjectives (van Berkum et al., 2005, Otten et al., 2007) preceding some highly expected noun in strongly constraining sentence contexts. These congruity (congruous or incongruous gender agreement) effects prior to the presentation of the word eliciting them are not the result of simple word-priming (Otten et al., 2007) and can occur more than a single word in advance of the target noun (van Berkum et al., 2005). Along similar lines, the effects of prediction on ERP responses have been shown to be graded in nature (DeLong et al., 2005), dependent on the target noun's cloze probability (a normative measure that in most circumstances can be taken as a proxy for how predicted a particular word is in a given sentence context; cf., Kutas & Federmeier, 2011). In addition to grammatical (van Berkum et al., 2005, Otten et al., 2007, Wicha et al., 2004) and phonological (DeLong et al., 2005) information, it has recently been shown that semantic information (in this case the semantic class of animacy) about an upcoming noun may also be predicted, and has an effect on ERP responses before the target noun (Szewczyk & Schriefers, 2013).

Taken together these studies make a strong case for (graded) predictions during sentence-level language comprehension, and not simply predictions about particular words but also about (at least some) semantic categories of words. They also show that electrophysiological brain responses (in this case ERPs) are sensitive to (at least some of) the processing consequences of these predictions.

In addition to syntactic features associated with specific lexical items (e.g., gender or number marking), other non-local syntactic dependencies may also lead to predictive processing, and the prediction of particular syntactic structures. For example, one prominent account of the P600 ERP component is as a reflection of processes of reanalysis and repair (e.g., Friederici, 2002, Friederici and Mecklinger, 1996). A P600 effect has been reported in the case of syntactic garden path sentences (e.g., Osterhout, Holcomb, & Swinney, 1994), for syntactic ambiguity resolution with object-compared to subject-relative clauses (Mecklinger, Schriefers, Steinhauer, & Friederici, 1995), and for syntactic violations (Hagoort, Brown, & Groothusen, 1993). All these cases have in common that they involve a preferred syntactic structure that is constructed and needs to be revised or repaired at a point where the input indicates that it is not correct (Friederici & Mecklinger, 1996). Although they have not traditionally be interpreted in this way, it is possible to argue that all these cases involve a prediction (by the language comprehension system) that a particular syntactic construction will accurately account for the linguistic input. At some point during the sentence, the input provides evidence that this prediction was incorrect, and the P600 may be thought of as the brain's response to such a failed prediction. Reframing some of the classical P600 findings in this way may provide a hint that readers/listeners make predictions about likely syntactic structures or structural dependencies while reading/listening.

Before moving on to discuss oscillatory neural dynamics during sentence-level language comprehension, we first outline a framework within which we can describe the relationship between predictive processing, the language comprehension system, and their associated functional brain network dynamics. There are a number of models relating the cognitive architecture of sentence-level language comprehension to its underlying neural infrastructure (e.g., Friederici, 2002, Jung-Beeman, 2005, Lau et al., 2008), but we adopt the framework and terminology used by Hagoort and co-workers (Baggio and Hagoort, 2011, Hagoort, 2005, Hagoort, 2013, Hagoort et al., 2009). A memory component, implemented by left temporal cortical areas, is responsible for the retrieval of lexical building blocks containing phonological, syntactic, and semantic properties of individual words. A unification component on the other hand is responsible for combining these building blocks to form a meaningful interpretation of the linguistic input. The unification component is implemented by left inferior frontal cortical regions, and their dynamic, coordinated interaction with left temporal and left inferior parietal cortex (Hagoort, 2014). A third component, the control component, completes the memory, unification, and control (MUC) framework (Hagoort, 2005, Hagoort, 2013) but for our purposes we will focus mainly on the memory and unification components.

The ERP studies discussed in Section 1.2 suggest that the system is likely engaged in ongoing predictive processing whenever possible during sentence-level language comprehension. Typically, while one reads or listens to a sentence, the predictability of upcoming linguistic information increases from beginning to end (because there are more plausible possibilities for continuing the sentence at the beginning than near the end of a sentence). In terms of the MUC framework pre-activation of specific lexical items or semantic categories, as well as biases towards particular syntactic structures due to prediction are the result of the dynamic interaction between the unification component and the memory component. The unification component sends feedback to the memory component during each word processing cycle (Hagoort, 2013), and in the case of highly constraining contexts (or in other cases where the system might use predictive processing) this information may prompt the memory component to pre-activate highly predicted lexical items (or at least some of the information associated with those items, e.g., their semantic category). Similarly, predictions about particular syntactic structures could bias the weighting of connections between nodes in the syntactic representation being built. The unification component would be responsible for this weighting, while the memory component would activate the syntactic treelets containing relevant syntactic nodes (Hagoort, 2005).

A large amount of evidence has accumulated over the last two or more decades suggesting that the coupling and uncoupling of functional networks in the brain is related to patterns of neural synchronization and desynchronization (Bastiaansen and Hagoort, 2006, Bastiaansen et al., 2012, Pfurtscheller and Lopes da Silva, 1999b, Singer, 1993, Singer, 2011, Varela et al., 2001, Womelsdorf et al., 2007). One instance of this occurs when areas that are part of the same functional network are linked by synchronous oscillatory firing in the same frequency range. Conceptually, synchronous repetitive firing of neurons increases the probability that they entrain one another and thereby activates participating functional networks at particular frequencies (König & Schillen, 1991). In this way the brain achieves frequency-specific segregation of information being processed by different functional networks. On the other hand, frequency-specific oscillatory neural synchrony also binds together information represented in different elements or subcomponents of the same functional network (Gray, König, Engel, & Singer, 1989). Another instance that has recently received a large amount of interest (and which is beyond the scope of this review) is cross-frequency coupling, where the phase of low frequency oscillations modulate the amplitude of oscillations in a higher frequency range (e.g., Lisman & Jensen, 2013). Such oscillatory neural phenomena typically have similar functions across multiple spatial and temporal scales. Modulations of frequency-specific power are often associated with synchrony within local neural populations, while modulations of frequency-specific phase coupling measures (e.g. coherence or phase-locking value) are most often associated with synchrony between more distant neural populations (inter-areal synchrony). There is however no clear distinction between local and inter-areal synchrony, and hence no guarantee that power always measures local synchrony and coherence always measures inter-areal communication (Varela et al., 2001).

We should note that the alpha frequency band (around 8–12 Hz) and perhaps, to some extent also the beta frequency band (around 13–30 Hz), do not straightforwardly fit in this framework. It has been observed that desynchronization in the alpha frequency range in a specific brain area sometimes entails the engagement/activation of that brain area, especially when related to motor (Bastiaansen and Brunia, 2001, Pfurtscheller and Lopes da Silva, 1999a, Pfurtscheller et al., 1997) and sensory (Bastiaansen & Brunia, 2001) processing. Similarly, beta-band desynchronization has often been related to motor cortex activation (see e.g., Parkes et al., 2006, Pfurtscheller et al., 1996, Pfurtscheller et al., 1998). The exact relationship between alpha/beta desynchronization and the process of functional network recruitment (in the sense of König & Schillen, 1991) and binding (in the sense of Gray et al., 1989) is yet to be established.

A growing body of literature has accumulated relating sentence-level language comprehension to event-related changes in EEG and MEG oscillations (e.g., Bastiaansen and Hagoort, 2010, Bastiaansen et al., 2010, Peña and Melloni, 2012; for review see Bastiaansen et al., 2012, Lewis et al., 2015). Such studies typically investigate patterns of fast temporal dynamics associated with the coupling and uncoupling of nodes in the brain's language network. Effects have been found in all the classical frequency ranges, with for example theta oscillations being linked to lexical retrieval operations and semantic working memory, and alpha being linked to task-specific working memory load (Bastiaansen and Hagoort, 2006, Bastiaansen et al., 2012, Meyer et al., 2013, Weiss et al., 2005). Bastiaansen and Hagoort (2010) proposed the ‘frequency segregation of unification’ hypothesis, which integrates a substantial body of empirical data into a framework in which oscillatory activity in the beta and gamma frequency bands reflect syntactic and semantic unification operations respectively. We have recently reviewed the evidence both for and against that hypothesis, and suggested that beta and gamma oscillations during sentence-level language comprehension may be at least equally well, or perhaps even better explained in relation to predictive processing, and maintenance/change of the current cognitive set respectively (Lewis et al., 2015).

In Section 2.1 we will outline how we think (at least some aspects of) oscillatory neural dynamics during sentence-level language comprehension might be explained in a predictive coding framework. Here we provide a brief outline of the particular flavor of predictive coding we will adopt (Friston, 2005). In this framework, brain systems are considered to be hierarchically structured, with higher hierarchical levels creating probabilistic (or forward) models designed to explain cortical (or sub-cortical) activity at lower levels (Clark, 2013, Friston, 2005).

For any two hierarchical levels, predictions are made at the higher level and propagated to the lower level in a top-down fashion. Here, such predictions are matched with bottom-up information (the actual activity in that lower level based on its inputs) to compute a prediction error (the difference between the prediction and the actual activity). Bottom-up information takes the form of the prediction errors themselves, so that this information contains only the amount of surprisal (an information theoretic measure of how far off the predictions actually were) based on the mismatch between the prediction and the actual activity in the lower hierarchical level. Those prediction errors are then used by the higher hierarchical level to update the predictions being made about the activity at the level below (Clark, 2013, Friston, 2005).

In the framework of Friston (2005), each hierarchical level contains both representational units and error units. Representational units represent activity at the current hierarchical level, provide top-down predictions to lower hierarchical levels, and receive inputs from error units at lower hierarchical levels with which they update predictions (or forward models) at the current level. Error units on the other hand receive input from representational units at the current and at higher hierarchical levels. They compute prediction errors based on the mismatch between top-down predictions and bottom-up information (activity at the current level) and send those prediction errors to higher hierarchical levels. Error units at the same hierarchical level also interact in order to decorrelate and laterally inhibit one another. The system attempts to minimize prediction error, and in this way achieve an optimal model of the events or causes of activity at different hierarchical levels (for more details see Friston, 2005).

The actual organization of the brain is of course far more complicated than the simple picture outlined above, with multiple hierarchical layers (often embedded within sub-layers) and fast, dynamic interactions between layers resulting in the constant updating and refinement of myriad forward-models at different hierarchical levels. Nonetheless, the static view outlined above provides a useful descriptive tool for probing the relationship between predictive coding, oscillatory neural dynamics, and various cognitive phenomena (for a formal description of a more dynamic implementation see e.g., Friston, 2005). This framework has been highly successful in accounting for a wide range of phenomena with a relatively simple mechanistic explanation of how information flows between and within cortical (and sub-cortical) hierarchies (e.g., spike-time dependent plasticity (STDP) during learning, classical and extra-classical receptive field properties in vision, repetition suppression, priming effects, ERP responses to learned sequences; see Friston, 2005 for details). In the next section we will see whether applying some of these principles to sentence-level language comprehension might prove useful in better understanding its neural implementation.

Section snippets

Beta and gamma oscillatory dynamics during language comprehension

Thus far we have briefly discussed prediction during sentence-level language comprehension, oscillatory neural dynamics in relation to the formation of functional brain networks, and a predictive coding framework for understanding the flow of information between hierarchical levels in the brain. In Section 2.1 we attempt to unify these three areas of investigation, and in 2.2 Beta oscillations and NeuroCognitive Networks, 2.3 Gamma oscillations and predictive processing we review the available

Some suggestions for future research

To recap briefly, we have suggested a role for oscillatory activity in the beta frequency range during sentence-level language comprehension in the active maintenance or change of an NCN responsible for representing the current sentence-level meaning under construction. We have also implicated beta in providing top-down predictions to lower hierarchical levels on the basis of that sentence-level meaning. Our proposal links oscillatory activity in the low and middle gamma frequency range to a

Conclusions

In this paper we have suggested how the extant findings relating oscillatory neural dynamics in the beta and gamma frequency ranges to sentence-level language comprehension may be given a unified explanation under a predictive coding framework. We have proposed that beta activity reflects both the active maintenance of the current NCN responsible for the construction and representation of a sentence-level meaning, and the top-down propagation of predictions based on that meaning to lower levels

Acknowledgments

We would like to thank two anonymous reviewers for their very helpful comments on an earlier version of the manuscript, and for many excellent suggestions that have improved the manuscript considerably. This work is partly supported by an IMPRS PhD fellowship from the Max Planck Society to A.G.L.

References (99)

  • A.K. Engel et al.

    Beta-band oscillations–signalling the status quo?

    Current Opinion in Neurobiology

    (2010)
  • K. Federmeier et al.

    A rose by any other name: long-term memory structure and sentence processing

    Journal of Memory and Language

    (1999)
  • A.D. Friederici

    Towards a neural basis of auditory sentence processing

    Trends in Cognitive Sciences

    (2002)
  • K.J. Friston et al.

    LFP and oscillations-what do they tell us?

    Current Opinion in Neurobiology

    (2015)
  • P. Hagoort

    On Broca, brain, and binding: a new framework

    Trends in Cognitive Sciences

    (2005)
  • P. Hagoort

    Nodes and networks in the neural architecture for language: Broca's region and beyond

    Current Opinion in Neurobiology

    (2014)
  • L.A. Hald et al.

    EEG theta and gamma responses to semantic violations in online sentence processing

    Brain and Language

    (2006)
  • C.S. Herrmann et al.

    Cognitive functions of gamma-band activity: memory match and utilization

    Trends in Cognitive Sciences

    (2004)
  • M. Jung-Beeman

    Bilateral brain processes for comprehending natural language

    Trends in Cognitive Sciences

    (2005)
  • Y. Kamide et al.

    The time-course of prediction in incremental sentence processing: evidence from anticipatory eye movements

    Journal of Memory and Language

    (2003)
  • P. Knoeferle et al.

    The influence of the immediate visual context on incremental thematic role-assignment: evidence from eye-movements in depicted events

    Cognition

    (2005)
  • R. Levy

    Expectation-based syntactic comprehension

    Cognition

    (2008)
  • J.E. Lisman et al.

    The theta-gamma neural code

    Neuron

    (2013)
  • Y. Luo et al.

    Electroencephalogram oscillations differentiate semantic and prosodic processes during sentence reading

    Neuroscience

    (2010)
  • L. Meyer et al.

    Left parietal alpha enhancement during working memory-intensive sentence processing

    Cortex

    (2013)
  • N. Molinaro et al.

    Long-range neural synchronization supports fast and efficient reading: EEG correlates of processing expected words in sentences

    NeuroImage

    (2013)
  • L.M. Parkes et al.

    Combining EEG and fMRI to investigate the post-movement beta rebound

    NeuroImage

    (2006)
  • B. Penolazzi et al.

    Gamma EEG activity induced by semantic violation during sentence reading

    Neuroscience Letters

    (2009)
  • A. Pérez et al.

    Oscillatory dynamics related to the unagreement pattern in Spanish

    Neuropsychologia

    (2012)
  • G. Pfurtscheller et al.

    Event-related EEG/MEG synchronization and desynchronization: basic principles

    Clinical Neurophysiology

    (1999)
  • G. Pfurtscheller et al.

    Foot and hand area mu rhythms

    International Journal of Psychophysiology

    (1997)
  • G. Pfurtscheller et al.

    Post-movement beta synchronization. A correlate of an idling motor area?

    Electroencephalography and Clinical Neurophysiology

    (1996)
  • G. Pfurtscheller et al.

    Event-related beta synchronization after wrist, finger and thumb movement

    Electroencephalography and Clinical Neurophysiology

    (1998)
  • M.J. Pickering et al.

    Do people use language production to make predictions during comprehension?

    Trends in Cognitive Sciences

    (2007)
  • M.S. Seidenberg et al.

    Automatic access of the meanings of ambiguous words in context: some limitations of knowledge-based processing

    Cognitive Psychology

    (1982)
  • W. Singer

    Dynamic formation of functional networks by synchronization

    Neuron

    (2011)
  • N.J. Smith et al.

    The effect of word predictability on reading time is logarithmic

    Cognition

    (2013)
  • J.M. Szewczyk et al.

    Prediction in language comprehension beyond specific words: an ERP study on sentence comprehension in Polish

    Journal of Memory and Language

    (2013)
  • C. van Petten et al.

    Prediction during language comprehension: benefits, costs, and ERP components

    International Journal of Psychophysiology

    (2012)
  • S. Weiss et al.

    The contribution of EEG coherence to the investigation of language

    Brain and Language

    (2003)
  • S. Weiss et al.

    Increased neuronal communication accompanying sentence comprehension

    International Journal of Psychophysiology

    (2005)
  • P. Zwitserlood

    The locus of the effects of sentential-semantic context in spoken-word processing

    Cognition

    (1989)
  • L.H. Arnal et al.

    Transitions in neural oscillations reflect prediction errors generated in audiovisual speech

    Nature Neuroscience

    (2011)
  • G. Baggio et al.

    The balance between memory and unification in semantics: a dynamic account of the N400

    Language and Cognitive Processes

    (2011)
  • M. Bastiaansen et al.

    Frequency-based segregation of syntactic and semantic unification

  • M.C.M. Bastiaansen et al.

    Syntactic unification operations are reflected in oscillatory dynamics during on-line sentence comprehension

    Journal of Cognitive Neuroscience

    (2010)
  • M.C.M. Bastiaansen et al.

    Beyond ERPs: oscillatory neuronal dynamics

  • A. Clark

    Whatever next? Predictive brains, situated agents, and the future of cognitive science

    The Behavioral and Brain Sciences

    (2013)
  • K.A. DeLong et al.

    Probabilistic word pre-activation during language comprehension inferred from electrical brain activity

    Nature Neuroscience

    (2005)
  • Cited by (0)

    View full text