Search Abstracts | Symposia | Slide Sessions | Poster Sessions
Information-theoretic storage cost in the brain: fMRI and EEG evidence from naturalistic sentence comprehension
Poster Session D, Thursday, October 1, 4:30 - 6:30 pm, Wangari Maathai
This poster is part of the Sandbox Series.
Yushi Sugimoto1, Kohei Kajikawa2, Shinnosuke Isono3; 1The University of Osaka, 2Georgetown University, 3National Institute for Japanese Language and Linguistics
Sentence comprehension requires maintaining contextual information to anticipate upcoming input, imposing a load on working memory. This storage cost has traditionally been formalized using symbolic grammars, as in Dependency Locality Theory (DLT; Gibson, 2000), which defines storage cost as the number of predicted syntactic heads at each word. While grammar-based storage cost has received empirical support, it relies on specific syntactic theories and assigns uniform, discrete costs to all predictions. A recent work proposed an information-theoretic reformulation of storage cost (InfoStor) that is continuous, grammar-independent, probabilistic, and estimable from neural language models (Kajikawa et al., 2026). This measure quantifies, for each preceding word, how much information it carries about future sequence, formalized as the (contextualized half-pointwise) mutual information between that word and future words, estimated using BERT. Summed over all prior words, this yields a word-by-word index of memory load. Prior reading-time analyses demonstrate that InfoStor outperforms DLT storage cost in predicting processing difficulty, suggesting it captures a more precise index of memory load during comprehension. Building on this, the present study examines whether InfoStor outperforms DLT storage cost as a predictor of neural activity during naturalistic sentence comprehension. We report preliminary fMRI and EEG analyses from the open-source dataset The Alice Datasets (Bhattasali et al., 2020), examining the unique neural contributions of DLT storage cost and information-theoretic storage cost. For fMRI, we analyzed six left-hemisphere language ROIs (LH_IFGorb, LH_IFG, LH_MFG, LH_AntTemp, LH_PostTemp, LH_AngG), defined by functional localizer parcellation (Fedorenko et al., 2010). The baseline model included word rate, sentence boundaries, prosodic breaks, and unigram surprisal, with by-subject random slopes for word rate. For EEG, we examined three ERP components (LAN, N400, P600) with analogous baseline controls including speech envelope and unigram surprisal. We assessed unique predictor contributions via likelihood ratio tests, conducted sequential model comparisons. Preliminary results suggest that InfoStor explains neural variance above and beyond DLT storage cost. In fMRI, the strongest incremental effects of InfoStor were observed in temporal (LH_AntTemp and LH_PostTemp) and angular gyrus (LH_AngG) regions, with InfoStor adding significant fit over DLT storage across multiple ROIs. In EEG, InfoStor alone did not reach significance in the P600 window, yet showed reliable additional predictive power when added over DLT storage cost. While these findings are preliminary, they suggest that information-theoretic storage cost captures neural variance that grammar-based storage cost alone cannot account for, pointing toward InfoStor as a more neurobiologically plausible index of the memory demands underlying real-time sentence comprehension. To more precisely characterize the temporal dynamics of these effects, we are currently conducting multivariate temporal response function (mTRF) analyses on the EEG data, which will allow us to model the continuous neural response to each predictor over time; these results will be presented at the conference.
Topic Areas: Computational Approaches,