Neural evidence for lexical parafoveal processing

The study addresses whether lexical information from upcoming words is extracted parafoveally during natural reading—a key controversy between serial attention shift models (e.g., E‑Z Reader) and parallel graded processing models (e.g., SWIFT, OB1-reader). While eye-tracking studies typically do not show parafoveal-on-foveal lexical effects on fixation durations, they only indirectly index parafoveal processing. The authors propose using rapid invisible frequency tagging (RIFT) with MEG during natural sentence reading to directly measure neural excitability linked to covert attention to upcoming (parafoveal) words. They hypothesize that if lexical information is previewed parafoveally, the 60 Hz tagging response during fixations on the pre-target word will be modulated by the lexical frequency of the target word.

Prior work shows parafoveal information is crucial for fluent reading despite reduced acuity. Serial models posit lexical processing of one word at a time with possible attentional shifts to the next word, explaining effects like word skipping. Parallel graded models propose simultaneous, graded processing of multiple words within the perceptual span. Eye-tracking research often reports no parafoveal-on-foveal lexical frequency effect, supporting serial models; some boundary paradigm studies report delayed effects (n+2). RIFT has previously indexed covert spatial attention with stronger tagging for attended stimuli. EEG/MEG studies show lexical frequency effects for fixated words typically arise at or after ~100 ms, informing temporal constraints on models. These literatures set the stage for directly probing parafoveal lexical processing neurally during natural reading.

Participants: Forty-three recruited; 39 included (25 females; mean age 22±2.6), right-handed, normal/corrected vision; ethics approved; informed consent; compensated. Stimuli: 228 plausible sentences (two sets). Set 1: 142 sentences embedding 71 low/high-frequency target pairs with two sentence frames; structure pre-target, target, post-target as adjective+noun+verb; counterbalanced versions. Set 2: 86 sentences adapted from Degno et al., each containing two targets of the same frequency; participants read either low- or high-frequency version. Pre-tests ensured plausibility and low predictability. Word characteristics matched (e.g., target length) and frequencies defined via CELEX. Procedure: Participants read silently in MEG, 145 cm from screen; Courier New monospace font (size 22); each character 0.35°; sentences up to 27° width. Trials began with fixation cross (1.2–1.6 s), then a gaze-triggered onset; comprehension questions on 25% of trials (accuracy 95.4±4.7%). Eye movements recorded continuously (EyeLink 1000 Plus, 1000 Hz). RIFT: A Gaussian-masked rectangular patch beneath the target word flickered at 60 Hz (100% sinusoidal luminance modulation) using a PROPixx projector at 1440 Hz refresh (12 grayscale frames via RGB×quadrants multiplexing). Patch width covered target plus surrounding spaces (2–3°), height 1.5× word height (~1°). Patch was largely invisible; only 3/39 noticed. A photodiode recorded the flicker timing. Data acquisition: MEG (306-channel Elekta TRIUX; 204 planar gradiometers, 102 magnetometers), 0.1–330 Hz online filters, 1000 Hz sampling. Head position digitization and MRI (3T PRISMA, T1) for source modeling; 3 participants lacked MRI (one with robust tagging used MNI template). EOG recorded. Eye-tracking calibration with drift correction throughout. Preprocessing: MEG band-pass 0.5–100 Hz; epochs −0.5 to 0.5 s time-locked to first fixation onset on pre-target, target, post-target words; baseline epochs 1 s from pre-sentence fixation cross. Fixations 0.08–1 s retained. Demeaning/linear detrend; bad sensors removed (0–2 per participant); ICA to remove blinks, eye movements, heartbeat; manual artefact rejection. Coherence analysis: Time-resolved coherence between MEG sensors and photodiode at 40–80 Hz (2 Hz steps; 10 Hz smoothing) using Hilbert analytic signals. Sensor selection: Compared 60 Hz coherence during pre-target epochs (flicker present) vs baseline (no flicker) using Monte Carlo permutation (10,000 shuffles) over 52 visual cortex planar sensors; sensors exceeding p<0.01 considered tagging-responsive. Twenty-six of 39 participants had robust tagging responses (mean 5.4±4.0 sensors). Source analysis: DICS beamforming at 60 Hz using single-shell head models and common spatial filters to localize sources coherent with the photodiode; relative change ((pre-target−baseline)/baseline) computed and group-averaged. FRFs: For all 39, 1 s pre-target epochs low-pass 35 Hz, equalized trial counts, baseline −0.2–0 s; planar gradiometers combined via RMS; cluster-based permutation (0–0.5 s, two-tailed, 1000 permutations, p<0.05) for lexical effects. Eye movement analyses: First fixation durations and gaze durations compared across conditions via two-sided paired t-tests. Statistical approach: Two-sided paired Student’s t-tests (R); Spearman correlation for coherence difference vs reading speed; subsampling to equate trial numbers across lexical conditions; averaging windows for coherence set individually to the minimum fixation duration per participant to avoid contamination from subsequent fixations.

• Eye movements: No effect of target lexical frequency on pre-target first fixation durations: t(38)=0.17, p=0.86, d=0.03. Target first fixation durations were longer for low- vs high-frequency targets: t(38)=6.94, p=3×10^−8, d=1.11. Similar pattern for gaze durations (supplementary).
• RIFT/MEG tagging: Robust 60 Hz coherence during pre-target fixations over left visual cortex sensors in 26/39 participants; sources localized to early visual cortex (BA 17/18), peak MNI [−4, −97, 3].
• Lexical parafoveal effect: During pre-target fixations, 60 Hz coherence was stronger when the upcoming target was low vs high lexical frequency. Paired t-test on averaged 60 Hz coherence (trial counts equated; window set by each participant’s shortest pre-target fixation): t(25)=2.20, p=0.037, d=0.43. The effect emerged around ~100 ms after pre-target fixation onset.
• No contamination from foveal processing: During target fixations, no coherence difference between low vs high lexical frequency targets (t(25)=0.01, p=0.992, d=0.002).
• Orthographic confounds: Bi/trigram token frequency and neighborhood size co-varied with word frequency, but splitting trials by these variables did not yield significant pre-target coherence differences (p=0.69, 0.06, 0.44; Bonferroni-corrected), suggesting the effect is lexical rather than orthographic.
• FRFs: Cluster-based test revealed higher pre-target FRFs for trials followed by low- vs high-frequency targets, peaking around ~0.4 s over left posterior sensors (p_cluster<0.05). No significant FRF differences during target fixations.
• Behavior–neural link: Individual difference in pre-target coherence (low−high) positively correlated with reading speed (words/s): Spearman r=0.449, p=0.022 (n=26). Participants with stronger lexical parafoveal tagging tended to read faster.
• Additional: Pre-target coherence onset latency modulated when both pre-target and target words were short (supplementary).

The results provide neural evidence that lexical information from upcoming words is processed in the parafovea during natural reading, supporting parallel graded processing models over strictly serial attention models. A pre-target lexical frequency effect in tagging coherence emerged around ~100 ms, earlier than typical foveal lexical effects reported in EEG/MEG, and difficult to reconcile with serial models that restrict lexical processing to one word at a time during saccade programming. Although pre-target fixation durations did not reflect lexical parafoveal effects, the magnitude of the neural parafoveal effect predicted individual reading speed, suggesting that covert allocation of attention to less frequent upcoming words facilitates fluent reading without prolonging current fixations. The tagging effect localized to early visual cortex, consistent with interactive processing and top-down feedback whereby lexical difficulty increases covert spatial attention, boosting sensory tagging responses. Lack of lexical effects in foveal tagging likely reflects lower flicker sensitivity in the fovea (fewer rods), smaller/brief flicker stimuli here compared to prior RIFT studies, and short fixation durations. Together, RIFT and FRFs provide complementary temporal windows: early covert attentional modulation (~100 ms) and later contextual integration (~400 ms). The findings indicate that natural reading involves simultaneous, graded processing of foveal and parafoveal words at the lexical level.

This study introduces rapid invisible frequency tagging combined with MEG and eye-tracking to directly measure parafoveal processing during natural reading. It demonstrates stronger pre-target 60 Hz coherence when the upcoming target word is low vs high frequency, with sources in early visual cortex and an onset around ~100 ms, providing neural evidence for parallel lexical processing of parafoveal words. The neural effect predicts individual reading speed, underscoring its functional relevance. Future work using RIFT could probe parafoveal processing at higher linguistic levels (semantic, syntactic), elucidate determinants of reading proficiency, and investigate reading disorders such as dyslexia, potentially linking magnocellular pathway function and covert spatial attention to reading outcomes.

• Only 26 of 39 participants exhibited robust tagging responses at visual sensors, constraining some MEG analyses to this subset.
• Source-level differences between low- and high-frequency targets were not strong enough to yield reliable contrasts, limiting inferences about differential source localization of lexical effects.
• No lexical effect was observed in foveal tagging during target fixations, likely due to smaller/brief flicker stimuli and reduced foveal flicker sensitivity; this limits direct comparison of foveal vs parafoveal tagging within the same paradigm.
• Eye movement measures did not show parafoveal-on-foveal lexical effects, so behavioral corroboration of the neural effect is indirect (though linked to reading speed).
• Orthographic variables co-varied with frequency (e.g., trigram token frequency), necessitating corrections; although analyses suggest minimal confounding, residual covariance cannot be entirely excluded.