The relative strength of hierarchical structure and statistics differs across the measures in naturalistic reading

Technical Summary: The Relative Strength of Hierarchical Structure and Statistics in Naturalistic Reading

Problem Statement
The paper addresses a long-standing debate in psycholinguistics regarding the mechanisms driving online language comprehension. Two primary, often rival, accounts exist: the hierarchical account, which posits that comprehension relies on the construction of abstract phrase structures and syntactic dependencies independent of linear word order; and the non-hierarchical account, which argues that comprehension is driven by local statistics, sequential prediction, and linear adjacency without strong sensitivity to hierarchical organization. While evidence exists for both, the field lacks a consensus on when and how strongly hierarchical structure exerts its influence relative to statistical factors. The authors argue that framing this as a "winner-take-all" contest is unproductive. Instead, they propose investigating how these factors jointly organize comprehension across different phenomena (word-level vs. sentence-level) and timescales.

Methodology
The study utilizes co-registered EEG and eye-tracking data from the ZuCo corpus (Hollenstein et al., 2018), comprising 12 native English speakers reading 400 movie review sentences. The authors operationalize hierarchical structure using syntactic depth (the number of dependency links from a word to the sentence root), a metric that varies independently of linear position and lexical predictability.

The study employs three converging analyses to compare hierarchical depth against non-hierarchical alternatives (linear sequence and statistical surprisal):

1. Transitional Probability Analysis (Scanpath Geometry):

   • Goal: To test if eye movements (scanpaths) are organized by linear adjacency (sequence-based) or syntactic centrality (hierarchical).
   • Method: Words were categorized as "heads" (syntactic depth 0 or 1) or "non-heads" (depth ≥ 2). The authors computed transitional probabilities for four transition types (head-to-head, head-to-non-head, etc.) in two sequences: the original linear text order and the actual fixation-derived order.
   • Comparison: They compared the frequency of "head-to-head" transitions in the fixation sequence against the linear baseline. This analysis specifically tested for anticipatory behavior (pre-fixation) and ruled out artifacts caused by regressions (backward eye movements).

2. Bayesian Network Analysis (Sentence-Level Strength):

   • Goal: To quantify the relative predictive weight of hierarchical structure versus statistical factors in explaining deviations from linear reading order.
   • Method: The outcome variable was edit distance (the minimal operations required to transform the linear text sequence into the fixation-derived sequence). A Gaussian Bayesian network was learned with five predictors: two structural (maximum parsing depth, number of clauses) and two lexical/statistical (sentence familiarity, sentence surprisal derived from GPT-2).
   • Analysis: Structure learning identified the dependency graph, and local linear-Gaussian regressions at the outcome node provided standardized coefficients to determine the relative strength of each predictor.

3. Time-Resolved Regression of Fixation-Related Potentials (Word-Level Neural Alignment):

   • Goal: To determine the timing and strength of hierarchical processing at the neural level, distinguishing it from statistical processing.
   • Method: Fixation-Related Potentials (FRPs) were extracted from EEG data time-locked to fixation onset. Ridge regression was performed on trial-wise EEG epochs using five standardized predictors: syntactic depth, syntactic surprisal (computed via a CCG-supertagging LSTM), word length, lexical frequency, and lexical surprisal.
   • Inference: Group-level inference used cluster-based permutation tests across six Regions of Interest (ROIs) to identify significant time windows and topographic dissociations between predictors.

Key Results

• Scanpath Organization (Analysis 1): Readers exhibited a significant bias toward transitioning between syntactic heads (low-depth words) in the fixation-derived sequence compared to the linear text baseline. This "head-to-head" preference was robust even after removing regressions, indicating that forward saccades are guided by hierarchical structure rather than just linear adjacency.
• Sentence-Level Strength (Analysis 2): In the Bayesian network, syntactic depth emerged as the strongest direct predictor of scanpath deviation (edit distance), with a standardized coefficient ($\beta \approx 0.40$) exceeding that of sentence familiarity ($\beta \approx 0.37$) and sentence surprisal ($\beta \approx 0.19$). This suggests that at the sentence level, hierarchical structure drives deviations from linear reading more strongly than statistical predictability.
• Neural Timing and Strength (Analysis 3):
  • Timing: Hierarchical processing (syntactic depth) modulated neural activity in the right-frontal region starting 108 ms before fixation onset, indicating an anticipatory role. Statistical predictors (lexical frequency and surprisal) modulated activity in the left-posterior region, also beginning pre-fixation (approx. 90–130 ms).
  • Strength Dissociation: While both effects were reliable, the magnitude differed by scale. At the word-level neural response, lexical surprisal showed a much larger effect size than syntactic depth. Conversely, at the sentence-level scanpath, syntactic depth was the dominant predictor.
  • Syntactic Surprisal vs. Depth: Crucially, syntactic depth produced a robust neural cluster, whereas syntactic surprisal (a statistical measure of syntactic uncertainty) produced no significant cluster. This dissociation suggests that the neural signature is specific to hierarchical representations rather than flat statistical probabilities of syntactic categories.

Significance and Claims
The paper claims that the hierarchical vs. non-hierarchical debate should not be viewed as a binary opposition where one factor universally dominates. Instead, the strength of hierarchical structure is dependent on the phenomenon being explained (the explanandum):

1. Scale Dependence: Hierarchical structure is the primary driver of sentence-level phenomena (specifically the geometry of scanpaths and deviations from linear order), whereas statistical factors (surprisal) are stronger drivers of word-level phenomena (fixation-locked neural activity and local processing costs).
2. Anticipatory Processing: The study provides evidence that hierarchical structure guides comprehension anticipatorily. The pre-fixation onset of the syntactic depth signal (108 ms prior) and the bias toward syntactic heads in scanpaths suggest that readers utilize an internalized hierarchical scaffold to guide attention before foveal uptake.
3. Beyond Linear Statistics: The finding that syntactic depth predicts neural activity while syntactic surprisal does not offers positive evidence that online comprehension involves genuine hierarchical representations, not merely the accumulation of sequential statistical probabilities.

The authors conclude that hierarchical structure is an active dimension that shapes where eyes go, how scanpaths deviate from linear order, and the nature of pre-foveal neural signals, with its influence varying in strength across different facets of reading behavior.