You are viewing the SNL 2017 Archive Website. For the latest information, see the Current Website.

Poster A37, Wednesday, November 8, 10:30 – 11:45 am, Harborview and Loch Raven Ballrooms

No evidence for semantic predictions? Inability to decode predictable semantic categories from EEG during silent pauses in spoken language

Edvard Heikel1, Jona Sassenhagen1, Christian J. Fiebach1;1Goethe University Frankfurt

Predictive coding has rapidly become a prevailing theory of neuronal processing. It proposes that the brain continuously predicts incoming sensory information in a probabilistic, Bayesian-like manner, based on mental models of the external (and internal) sources of sensory information. These predictions are utilized to minimize information processing requirements by ‘explaining away’ predictable input, which is highly compatible with the assumption of incremental parsing. If sensory input is incompatible with these internally generated predictions, a prediction error signal is generated to update these models thereby optimizing their predictions. Recently theories of language processing have made steps towards incorporating predictive coding into their frameworks. However, while many neural correlates of language processing seem to be plausibly associated with the prediction error (such as, e.g., the N400 component of the ERP), the neural signatures of linguistic predictions themselves have so far remained elusive. One plausible approach to identifying neural correlates of top-down predictions during perception is to abolish bottom-up sensory information, for example through omissions or the occlusion of sensory stimuli. In a strict interpretation of predictive coding, the resulting prediction error should in this situation represent the prediction – which should thus be decodable from the neural prediction error signal. We aimed to investigate the neurophysiological realization of linguistic predictions during the perception of spoken sentences, by using EEG and multivariate pattern analysis (MVPA) to decode the semantic category of a predictable word from brain activity during a 1 second pause preceding the word itself.. If predictive coding operates in speech processing, we should be able to decode above chance the upcoming word’s semantic category during the brief period of silence preceding the actual perception of the word. In a first study (N=39), pauses were inserted prior to target words occuring in sentence contexts that were constraining towards (and therefore, allowing prediction of) either an animate or an inanimate noun. In the second study (N=39), the predictable target words were either abstract or concrete nouns. The constraining (prediction licensing) nature of the sentences was verified on the basis of cloze ratings. Analysing EEG activity during the pre-target word pause with a multitude of MVPA approaches (Generalization Across Time – King & Dehaene, 2014; XDAWN – Rivet et al, 2009; Common Spatial Patterns – Zoltan et al, 1991) and machine learning algoritms, we found no evidence for above-chance decoding, neither for animacy nor for concreteness (all p > .05; accuracies ~50%). Experiment 2 was designed after obtaining the null-result in Experiment 1, with the hypothesis that concreteness entails a clearer difference in neurophysiological response than animacy, and thus should increase the chances of decoding linguistic expectations prior to word onset. We conclude that either 1. our manipulation was insufficiently strong to induce a semantic category-level prediction, 2. there was such a prediction maintained by the subjects, but (at least our) EEG data does not contain enough information to decode this prediction, or 3. at least some aspects of language processing are not captured by the predictive coding framework.

Topic Area: Meaning: Lexical Semantics

Back to Poster Schedule