Slide Slam C13 Sandbox Series
Do sentences modulate the low-frequency neural response to words?
Sophie Slaats1, Hugo Weissbart2, Jan-Mathijs Schoffelen2, Andrea E. Martin1,2; 1Max Planck Institute for Psycholinguistics, 2Donders Institute for Brain, Cognition and Behaviour
Listeners have the remarkable ability to combine acoustic information from speech with abstract linguistic knowledge, resulting in a structured representation of intended meaning. Recent work in psycho- and neurolinguistics has revealed signatures of this process in the brain: in the delta band (≤4Hz) – the timescale of occurrence of words/phrases – speech tracking is affected if the linguistic structure or content is manipulated (Blanco-Elorrieta et al., 2020; Molinaro & Lizarazu, 2018; Kaufeld et al., 2020). In the theta band (4-8 Hz) – the timescale of syllables – speech tracking is affected by modifications of the acoustic signal (Etard & Reichenbach, 2019; Peelle, Gross & Davis, 2012; Doelling et al., 2014). Furthermore, lexical features also drive low-frequency neural responses (e.g., Weissbart et al., 2019; Brodbeck et al., 2018; Broderick et al., 2018), but it is not yet clear how linguistic structure affects these. Here we ask therefore whether and how the neural response to spoken words in the delta and theta bands changes when words appear in sentences or word lists. We analyze published MEG data from 102 participants listening to Dutch sentences (9-15 words, various syntactic structures) and word lists: scrambled sentences (Schoffelen et al., 2019). We model the source-localized neural response to words using temporal response functions (TRFs) and compare these between the conditions. This approach allows for the estimation of effects above and beyond differences in the acoustic signal. To this end, we add speech envelope- and word onset-features to all models. The feature we use as a proxy for lexical information is word frequency, a crucially unigram feature. To isolate the sentence effect from sequential predictability effects, we include estimates of surprisal and entropy. Because these estimates contain frequency information, including them in the models offers the strongest test of sentence effects on the neural response to word frequency. Using cross-validation, we estimate TRFs at lags from -200ms to 800ms, and quantify how well they reconstruct the neural signal. On the basis of analysis-by-synthesis accounts of language comprehension (Martin, 2016; 2020; Poeppel & Monahan, 2011), we expect the neural response to word frequency to be modulated by the words’ involvement in sentences. Specifically, we hypothesize that word frequency modulates the neural response in both conditions in temporal and left inferior parietal areas (Hagoort, 2013). In the sentences, however, this modulation should be more pronounced and spatially more widespread as a result of propagation of lexical information when this becomes included in larger structures (Martin, 2020). The spread is likely visible in the left inferior frontal gyrus (Hagoort, 2013). We expect this pattern in the delta band. In theta, we might see a neural response to words due to supply of syllabic information (Brown, Tanenhaus & Dilley, 2021), but the response may not differ between conditions. The results of this study will speak to how the neural representation of words is affected by structural embedding, and as such provides insight into how the brain instantiates compositionality in language (Martin, 2016; 2020).