Search Abstracts | Symposia | Slide Sessions | Poster Sessions
Neural activity to unexpected words in the N400 time window preserves both item-specific and condition-general structure
Poster Session D, Thursday, October 1, 4:30 - 6:30 pm, Wangari Maathai
Gina Kuperberg1,2, Aileen Guo1, Thomas Hansen2; 1Massachusetts General Hospital, 2Tufts University
**Introduction: A foundational finding in the neurobiology of language comprehension is that less predictable words elicit larger neural responses than more predictable words between 300–500ms after word onset—the N400 effect. However, the functional significance of this response remains unclear. One possibility is that it only indexes the increased demands of semantic access. This would be consistent with accounts in which the response emerges as a byproduct of inference or as a downstream signal to support longer-term learning. A second possibility is that it also preserves structured, item-specific information about the identity of the unexpected word. This would support predictive coding accounts, which hold that unexpected inputs activate error units that specifically encode representational information about the observed word's form and/or meaning that was not predicted by higher-level states. On this account, instead of functioning as a byproduct of inference, the larger response to unexpected inputs plays a direct functional role in updating higher-level representations so that they better explain the incoming input. **Methods: We used magnetoencephalography (MEG) with univariate and multivariate representational similarity analysis (RSA) to adjudicate between these accounts. Participants read sentences presented word-by-word. Sentence materials were constructed in pairs in which the same critical words appeared as expected completions in two distinct high-constraint contexts or as unexpected completions in two distinct low-constraint contexts. This allowed us to compare the similarity between spatial patterns produced by same-target versus different-target pairs across expected and unexpected conditions, while controlling for order effects, thereby separating target-specific from condition-general structure. To help interpret our findings, we simulated results using an implemented predictive coding model that infers word meaning from orthographic form. **Results: Replicating previous work, unexpected words evoked larger univariate responses than expected words in left ventral fusiform and lateral superior/middle temporal cortices between 300–500ms. RSA revealed that in both regions, within the same time-window, spatial patterns elicited by unexpected same-target pairs were more similar than those elicited by expected same-target pairs. The patterns evoked by same-target unexpected pairs were also more similar than those evoked by different-target unexpected pairs, indicating that unexpected responses preserved target identity rather than only reflecting a non-specific increase in response magnitude. Critically, the two regions dissociated in their condition-general structure: ventral temporal cortex showed no elevated similarity across different-target unexpected (vs. expected) pairs, suggesting its response was dominated by target-specific content. In contrast, lateral temporal cortex showed elevated similarity across different-target unexpected (vs. expected) pairs, revealing a component shared across unexpected inputs. Predictive coding simulations reproduced this dissociation, with orthographic-lexical prediction error capturing the ventral fusiform pattern and lexico-semantic prediction error capturing the lateral temporal pattern. **Conclusions: Neural responses to unexpected input in the N400 window carry structured representational content that is regionally differentiated: Ventral temporal cortex encodes the specific identity of unexpected words, while lateral temporal cortex encodes both item-specific and condition-general prediction error signals. These findings support predictive coding accounts of language comprehension, in which error signals encode the content of unexpected input rather than simply its magnitude.
Topic Areas: Meaning: Lexical Semantics, Computational Approaches