Stimulus dependencies---rather than next-word prediction---can explain pre-onset brain encoding in naturalistic listening designs

Technical Summary

Problem Statement
The prevailing view in language neuroscience posits that the human brain continuously predicts upcoming words during naturalistic listening. Recently, a novel analytical approach has emerged to directly capture this neural prediction by using vector representations of words (embeddings) to model brain activity prior to word onset. Two specific findings have been proposed as hallmarks of genuine neural next-word prediction: (i) the presence of significant encoding of future words before they occur, and (ii) the modulation of this pre-onset encoding by the predictability of the upcoming word. However, natural language is characterized by dense temporal correlations, where statistical information in preceding words inherently shares information with upcoming words. This raises a critical methodological question: do the observed hallmarks reflect the brain's active predictive computation, or are they artifacts arising because the regression models exploit these inherent stimulus dependencies?

Methodology
To distinguish between active prediction and stimulus-driven dependencies, the authors applied the same encoding analysis used in previous prediction studies to "passive control systems." These are representational systems that encode the stimulus but, by definition, cannot predict upcoming content. Specifically, the study tested two control systems:

1. Word Embeddings: The vector representations themselves, which contain the statistical structure of language but lack a temporal predictive mechanism.
2. Speech Acoustics: The raw acoustic features of the speech signal.

The authors examined whether the two proposed hallmarks of prediction (pre-onset encoding and its modulation by predictability) emerged in these passive systems. Furthermore, they evaluated whether proposed methodological corrections designed to account for stimulus dependencies were sufficient to eliminate these effects.

Key Results
The analysis revealed that both hallmarks of neural prediction emerged robustly in the passive control systems:

• Pre-onset Encoding: Significant encoding of future words was observed in both word embeddings and speech acoustics prior to word onset.
• Modulation by Predictability: The strength of this pre-onset encoding was modulated by word predictability in these control systems, mirroring the patterns previously attributed to neural prediction.

Crucially, the authors demonstrated that existing methods intended to correct for stimulus dependencies were insufficient; the effects persisted even after applying these corrections.

Key Contributions

• Identification of a Confound: The paper identifies that temporal correlations in natural language stimuli can generate spurious pre-onset encoding effects in regression models, independent of any neural predictive mechanism.
• Validation via Control Systems: By demonstrating that these effects appear in systems incapable of prediction (embeddings and acoustics), the study provides strong evidence that the observed "hallmarks" may be statistical artifacts rather than neural phenomena.
• Evaluation of Corrections: The study shows that current correction techniques do not fully resolve the issue of stimulus dependency in these encoding models.

Significance and Claims
The paper concludes that pre-onset prediction of brain activity in naturalistic listening designs may reflect the statistical dependencies inherent in natural language rather than predictive computations performed by the brain. Consequently, the authors argue that this challenges the extent to which the current encoding-based method can be reliably used to study neural prediction. The findings suggest that previous interpretations of pre-onset encoding as direct evidence of brain prediction may need to be re-evaluated in light of these stimulus-driven confounds.