Search Abstracts | Symposia | Slide Sessions | Poster Sessions
Probing LLM-brain alignment when structure dissociates from statistics
Poster Session D, Saturday, September 13, 5:00 - 6:30 pm, Field House
This poster is part of the Sandbox Series.
Junyuan Zhao1, Jonathan Brennan1; 1University of Michigan
[Introduction] Recent proposals consider large language models (LLMs) as compelling computational models of human language processing (Goldstein et al., 2022; Schrimpf et al., 2021; Tuckute, Kanwisher, et al., 2024). Yet, structural representations which are not explicitly encoded in LLMs capture additional variance in behavioral and neural patterns above and beyond LLM-based surprisal (Huang et al., 2024; Stanojević et al., 2023). Current data appear mixed because structural factors can be largely confounded with statistics language data (Slaats & Martin, 2023) and also due to different LLM-derived measures being probed across studies (hidden layers: Schrimpf et al., 2021; surprisal: Stanojević et al., 2023). Here, we test if LLM-brain alignment in structural (syntactic and semantic) computations extends to low predictability sentences where statistics and structural well-formedness are de-confounded. To this end, we manipulate syntactic and semantic composition in English sentences while minimizing predictability. Then, we compare layer-wise LLM activation with EEG recordings as a proxy of LLM-brain alignment. [Design] We recorded 32-channel EEG while native English-speaking participants (N = 35) read English sentences in a rapid serial visual presentation experiment. The materials involve six conditions: (1) normal [low surprisal, +syntax, +compositional semantics, +associative semantics]: grammatical, meaningful sentences (e.g., Bright colorful flowers bloom richly), (2) syntactic prose [high surprisal, +syntax, +compositional semantics, -associative semantics]: sentences that follow English grammatical rules but lack a coherent meaning (e.g., Colorless green ideas sleep furiously), (3) Jabberwocky [+syntax, -compositional semantics, -associative semantics]: sentences that only contain (inflected) English pseudowords (e.g., Rupe sleen althes rore smountly), along with conditions 4-6 that are derived by shuffling words from 1-3. We match conditions 2-6 in terms of GPT2-xl-derived surprisal (Radford et al., 2019); F(5, 395) = 1.60, p = 0.177. Each condition has 80 items, yielding 480 critical items alongside 400 filler sentences to balance acceptability in the stimuli. Our design allows stepwise comparisons of the role of different structural factors in LLM-brain alignment. For example, comparing jabberwocky [+syntax, -compositional semantics, -associative semantics] and its surprisal-matched, shuffled counterpart [-syntax, -compositional semantics, -associative semantics] sheds light on the role of syntax in brain-LLM alignment. [Proposed Analyses] LLM-brain alignment is evaluated with a linear encoding model from LLM activations to EEG signal (Tuckute et al., 2024). The goodness of fit of this model quantifies the similarity between them. We hypothesize that if the brain and LLM similarly represent a certain linguistic factor, then goodness of fit should be similar in both the presence and absence of that factor. We will also examine effects in time-locked EEG signal, in both the time and frequency domains (Hahne & Jescheniak, 2001; Kaufeld et al., 2020). Pilot validation analyses (N = 10) indicate (1) an N400 component for normal sentences vs. syntactic prose, (2) a P600 component for Jabberwocky vs. normal sentences and syntactic prose, and that (3) hidden layers of the GPT-2 model (note it is a smaller version) reliably encode information about syntax (original vs. shuffled) for all conditions, as revealed by LLM-probing (Liu et al., 2019; Marvin & Linzen, 2018).
Topic Areas: Syntax and Combinatorial Semantics, Computational Approaches