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

Area of Science:

• Cognitive Neuroscience
• Computational Linguistics
• Neuroimaging

Background:

• The human brain is believed to predict upcoming words during language comprehension.
• Recent methods use word embeddings to predict neural activity before word onset, identifying hallmarks like pre-onset encoding and modulation by predictability.
• However, natural language's inherent temporal correlations complicate the interpretation of these findings.

Purpose of the Study:

• To investigate whether observed hallmarks of neural prediction truly reflect active brain prediction or merely statistical dependencies within language stimuli.
• To determine if current encoding-based methods can reliably distinguish genuine predictive processing from stimulus-driven confounds.

Main Methods:

• Applied encoding analysis, typically used for neural prediction, to passive control systems (word embeddings, speech acoustics).
• These control systems process stimuli but lack predictive capabilities.
• Evaluated if the hallmarks of prediction persist in these non-predictive systems and after applying standard correction methods.

Main Results:

• Both hallmarks of neural prediction—pre-onset encoding and modulation by predictability—were observed in word embeddings and speech acoustics.
• These effects persisted even after applying proposed correction methods for stimulus dependencies.
• Control systems demonstrated that these hallmarks can arise from inherent language structure, not just active prediction.

Conclusions:

• The findings challenge the interpretation of pre-onset neural activity as direct evidence of next-word prediction.
• Current encoding-based methods may be confounded by statistical dependencies in natural language, potentially overestimating predictive processing.
• Further research is needed to refine methods for accurately measuring predictive computations in the brain.