Skip to main page content
U.S. flag

An official website of the United States government

Dot gov

The .gov means it’s official.
Federal government websites often end in .gov or .mil. Before sharing sensitive information, make sure you’re on a federal government site.

Https

The site is secure.
The https:// ensures that you are connecting to the official website and that any information you provide is encrypted and transmitted securely.

Access keys NCBI Homepage MyNCBI Homepage Main Content Main Navigation
. 2019 Jan 16;14(1):e0207741.
doi: 10.1371/journal.pone.0207741. eCollection 2019.

Hierarchical structure guides rapid linguistic predictions during naturalistic listening

Affiliations

Hierarchical structure guides rapid linguistic predictions during naturalistic listening

Jonathan R Brennan et al. PLoS One. .

Abstract

The grammar, or syntax, of human language is typically understood in terms of abstract hierarchical structures. However, theories of language processing that emphasize sequential information, not hierarchy, successfully model diverse phenomena. Recent work probing brain signals has shown mixed evidence for hierarchical information in some tasks. We ask whether sequential or hierarchical information guides the expectations that a human listener forms about a word's part-of-speech when simply listening to every-day language. We compare the predictions of three computational models against electroencephalography signals recorded from human participants who listen passively to an audiobook story. We find that predictions based on hierarchical structure correlate with the human brain response above-and-beyond predictions based only on sequential information. This establishes a link between hierarchical linguistic structure and neural signals that generalizes across the range of syntactic structures found in every-day language.

PubMed Disclaimer

Conflict of interest statement

The authors have declared that no competing interests exist.

Figures

Fig 1
Fig 1. Surprisal distributions from each of three models along with mean surprisal ±1 standard error of the mean (black bars).
Surprisal from a model where each POS tag is uniformly probably is indicated with the dashed line.
Fig 2
Fig 2. Language models and data analysis.
Word-by-word surprisal values (top) estimated from one hierarchy-based and two sequential language models are time-aligned to the audiobook stimulus (middle). Epochs aligned with the onset of each word are extracted from filtered EEG data and amplitudes from each time-point and electrode serve as the dependent measure of a regression model (right) that includes surprisal values and low-level covariates as predictors.
Fig 3
Fig 3. Whole-head regression results.
β coefficients (M ± CI95) and model-reconstructed ERPs. (A) Regression time-series for log-transformed word frequency fit against content-word measurements. Dark grey shading indicates significant time-points and the inset shows significant channels and coefficient averages across significant time-points. (B) Estimated content-word ERP for three word frequency values ranging from low (200 corpus counts) to high (2,000,000 corpus counts) from a represenative central channel (inset) shows a classic N400 effect. (C) Regression time-series for hierarchical CFG surprisal fit against content-word data (inset and shading as in (A)). (D) Estimated content-word ERP for three CFG surprisal values from a representative channel (inset). (E) Regression time-series for sequential Ngram surprisal fit against function-word data (inset and shading as in (A)). (F) Estimated function-word ERP for three Ngram surprisal values from a representative channel (inset).
Fig 4
Fig 4. Model comparison results.
WAIC difference scores (± standard error) indicate changes in model fit across six ROIs (columns). Each set of rows tests a different statistical question using step-wise model comparison. Terms that are being evaluated are indicated to the left of “>”; interactions with word-class are indicated with “:WC”. For each row-set, the baseline model includes all control covariates along with the indicated surprisal term(s) and interactions between word-class and those surprisal terms. The WAIC values are scaled so that positive numbers represent improvements for the larger model, while negative numbers indicate that the added complexity of the larger model is not matched by a better fit. Bold-face indicates WAIC improvements that are more than two standard errors from zero.

Similar articles

Cited by

References

    1. Hauser MD, Chomsky N, Fitch WT. The Faculty of Language: What is it, Who has it, and How did it Evolve? Science. 2002;298:1569–1579. 10.1126/science.298.5598.1569 - DOI - PubMed
    1. Frank SL, Bod R, Christiansen MH. How hierarchical is language use? Proceedings of the Royal Society B: Biological Sciences. 2012;. 10.1098/rspb.2012.1741 - DOI - PMC - PubMed
    1. Ding N, Melloni L, Tian X, Poeppel D. Rule-based and word-level statistics-based processing of language: insights from neuroscience. Language, Cognition and Neuroscience. 2017;32(5):570–575. 10.1080/23273798.2016.1215477 - DOI - PMC - PubMed
    1. Frank SL, Christiansen MH. Hierarchical and sequential processing of language. Language, Cognition and Neuroscience. 2018; p. 1–6.
    1. Ding N, Melloni L, Zhang H, Tian X, Poeppel D. Cortical tracking of hierarchical linguistic structures in connected speech. Nat Neurosci. 2016;19(1):158–64. 10.1038/nn.4186 - DOI - PMC - PubMed

Publication types