Search Abstracts | Symposia | Slide Sessions | Poster Sessions
Backward surprisal in serial and parallel reading
Poster Session B, Friday, September 12, 4:30 - 6:00 pm, Field House
Amilleah Rodriguez1, Nigel Flower1, Liina Pylkkänen1; 1New York University
Predictive processing is central to our neurobiological models of language, but recent work across domains has questioned the traditional focus on forward prediction—the likelihood of the present given the past—by emphasizing backward prediction, or the likelihood of the past given the present. Backward prediction has been shown to influence linguistic processing (Onnis et al., 2022), statistical learning and perception (de Lange & Press, 2025), and decision making (Sharp & Eldar, 2024). Yet its role in language neuroscience remains largely unexamined. For example, the extent to which the N400 reflects backward rather than forward prediction is unclear, as the two are naturally correlated. We used MEG to disentangle forward and backward prediction, quantified using surprisal, during the reading of adjective-noun phrases. To test whether any observed effects are dependent on the adjective temporally preceding the noun, we presented stimuli both serially using RSVP and in parallel using RPVP. Prior MEG work identifies the left anterior temporal lobe (LATL) as a key site for combinatory processing (Pylkkänen, 2019), making it a plausible locus for backward surprisal effects. In contrast, classic N400 signals localize to the left posterior temporal lobe and the LIFG (Lau et al., 2013; Halgren et al., 2002). While high forward surprisal is typically associated with increased neural responses, backward surprisal may have the opposite effect. Prior studies have found that strongly associated bigrams evoke larger LATL responses (Li et al., 2021), and it has been hypothesized that high backward probabilities may lead to stronger bidirectional feedback between the current item and its context, enhancing activation for sequences with stronger backward associations (Onnis & Huettig, 2021). Participants read adjective-noun phrases that varied in forward and backward surprisal—such as ‘romantic comedy’ (LowForw-LowBack), ‘lunar calendar’ (LowForw-HighBack), ‘live coverage’ (HighForw-LowBack), or ‘untitled piece’ (HighForw-HighBack)—and were then presented with a semantic task probe. Analyses were time-locked to the noun’s onset in RSVP and to the whole phrase onset in RPVP, enabling us to directly compare the neural signals elicited by the noun in its temporally preceding context and that same noun when its context could be read in parallel with the noun. Cluster-based permutation tests revealed a robust main effect of backward surprisal in the LATL at 207–300ms, with higher amplitudes for less surprising contexts—an anti-surprisal effect. This effect was consistent across presentation modes. Forward surprisal also showed main effects, though only marginally after correction, emerging in the LIFG (203–306ms), LATL (474–580ms), and left posterior temporal lobe (522–584 ms), without interactions with presentation mode. Overall, backward surprisal was a more reliable and earlier modulator of neural responses than forward surprisal, with effects beginning before the typical N400 window. The anti-surprisal direction of the backward effect also differentiates it from the typical forward surprisal response. Notably, presenting the two words in parallel did not eliminate the distinction between forward and backward surprisal. This suggests that both predictive (forward) and retrodictive (backward) processes are active during phrase comprehension, even when the stimulus does not unfold over time.
Topic Areas: Meaning: Lexical Semantics, Reading