Different Reading Processing Stages or Different Brain Areas? A Computational Cognitive Investigation on N400, P600, and PNP

Lavinia Salicchi, Yu-Yin Hsu

Keywords: Cognitive modeling, Surprisal, Entropy, Cosine Similarity, ERPs

TL;DR: A computational investigation of the cognitive dynamics taking place during reading in the N400 and P600 time windows, with a differentiation between frontal and centro-parietal P600

Abstract:

The classical distinction between N400 as an index of semantic processing and P600 as a marker of syntactic
processing has been challenged by studies reporting P600 effects in response to semantic violations. This has
led to debates about the functional roles of these event-related potentials (ERPs), particularly the frontal P600
(PNP) and its relationship with N400 and posterior P600 [1]. Computational metrics like surprisal, entropy, and
semantic similarity, mathematically representing cognitive dynamics, have been employed to model these ERPs
[2, 3], to directly test which mechanisms take place at different reading stages. However, little computational
research has been conducted on P600 and PNP, especially in non-alphabetic languages like Mandarin Chinese.

We analyzed EEG data from 38 participants reading 280 grammatical Mandarin Chinese sentences without
semantic violations. This type of data allows us to extend previous research to Sinitic languages and create
a general baseline for future investigations. We extracted N400, P600, and PNP, employing the channels
selected in [4]. Using a Chinese GPT-2 model for conditional probabilities and word embeddings, we computed
surprisal, entropy, entropy variation, and three semantic similarity metrics: sentword, a context-word similarity
employed in [5], the semantic similarity between the upcoming word and the most expected word (cosk1) or a
general concept based on the five most expected words (cosk5). We created 10 linear mixed-effect models:
a baseline model, including word-level features only, 6 models employing the word-level regressors and one
computational metric, and three general models, including all the features. As in [6], the baseline signal was
as a covariate of no interest, and word ID and participant ID were random intercepts. To assess each metric’s
predictive power, we computed the target model - baseline model log likelihood difference (ΔLL).

Surprisal was the strongest predictor of N400 amplitude (ΔLL = 6.94, significantly different from zero - p <
0.001), suggesting that in early processing stages, readers are sensitive to the absence of expected lexical
items. Entropy variation and expectation-driven semantic similarity (cosk5) predicted PNP (ΔLL = 4.43, p
= 0.001 and ΔLL = 3.30, p = 0.004), suggesting that in later stages, readers perform a higher-level semantic
evaluation and suppress previous expectations. The context-word semantic similarity predicted both P600
and PNP, indicating a semantic integration happening in later stages and involving a wide network. Entropy
significantly modulated all ERPs.

Our findings support a multi-stage model: In the early stages of word processing, a centro-parietal network
assesses whether $w_n$ matches the predictions generated by the preceding context ($C_{n−1}$), with unexpected
words requiring greater cognitive resources. Simultaneously, the number of possible continuations maintained
in working memory increases cognitive effort. In later stages, if $w_n$ introduces new sentence constraints, the
resulting cognitive demand can be traced as frontal brain activity. Meanwhile, the reader evaluates the degree
to which wn fits $C_{n−1}$, with poorer matches inducing a higher cognitive load across frontal and posterior
areas. Finally, a frontal network compares wn’s semantics to the predicted general concept, with conceptual
mismatches being cognitively more expensive.

Submission Number: 11