Carolin Dudschig, Universität Tübingen

The N400 is a well-established event-related potential (ERP) component  
that is widely studied in the field of cognitive neuroscience. It is  
typically larger following the detection of a semantic violation or  
other incongruities within language processing tasks. The N400  
provides valuable insights into the cognitive processes underlying  
language comprehension. For example, it has been used to investigate  
whether linguistic and world-knowledge violations are integrated in  
parallel during comprehension (e.g., Dudschig, Maienborn & Kaup, 2016;  
Hagoort, Bastiaansen & Petersson, 2004). Nevertheless, to date, it is  
still under debate what processes or information are reflected in the  
N400 and whether integration of basic operators such as negation are  
reflected in the N400 (e.g., Dudschig et al., 2019). The accounts  
range from integration views - suggesting the N400 reflects  
integration processes - to the lexical view - suggesting that the N400  
is non-combinatorial in nature and dominant prediction-based accounts  
that focus on the predictability of the critical word (for a review,  
see Kutas & Federmeier, 2011). Recent developments in large language  
models (LLMs) have opened new avenues to investigate what processes  
are reflected and what insights can be gained from examining the N400.  
This presentation aims to investigate the extent to which the N400  
amplitude can be better explained by pre-determined discrete condition  
labels (e.g., correct vs. world-knowledge vs. semantic violation)  
versus continuous word-pair embedding measures derived from multiple  
LLMs. Overall, this presentation aims to bridge the gap between  
traditional N400 research and the emerging field of natural language  
modeling.