Humans effortlessly comprehend spoken language despite significant acoustic variability stemming from noise, speaker differences, and accents. This robustness likely arises from predictive processing mechanisms, a neurocomputational framework where predictions actively shape the processing of incoming information streams. In language comprehension, this manifests as the brain anticipating upcoming linguistic units based on prior knowledge. However, the inherent unpredictability of novel sentences presents a challenge. This unpredictability is not absolute; it's influenced by the message's relevance and specificity. Previous linguistic experience enables prediction at various levels, from phonemes to spectro-temporal characteristics of sound segments. A crucial aspect of human language is its nested syntactic structure, upon which meaning is computed. Historically, the generative and unbounded nature of language has been contrasted with distributional and statistical accounts of language processing. This study bridges this dichotomy, proposing that syntactic structure and statistical cues are jointly processed during comprehension. Building on previous research which demonstrated BOLD signal sensitivity to both structure and surprisal, this study leverages magnetoencephalography (MEG) to analyze phase and amplitude dynamics of cortical activity in response to variations in both statistical cues and syntactic features, offering superior temporal resolution than fMRI.

The study builds upon existing research demonstrating the brain's sensitivity to both structural and statistical aspects of language. Brennan et al. (2016) showed that BOLD signal responded to both hierarchical syntactic features and surprisal, but lacked the temporal resolution to analyze high-frequency activity. Previous work has used various linguistic metrics, including features derived from hierarchical trees (e.g., number of brackets, tree depth) reflecting syntactic structure processing, and information-theoretic measures like surprisal and entropy from language models, capturing statistical word predictability. While some studies suggest a dichotomy between rule-based (syntactic) and statistical accounts of language processing, this study synthesizes these views, hypothesizing a joint contribution of both to neural dynamics during comprehension. The authors acknowledge ongoing debates regarding the extent to which statistical models implicitly encode structural information and highlight the improved accuracy of modern language models (like Transformers) in capturing these nuances compared to earlier architectures.

The study used magnetoencephalography (MEG) to measure brain activity during audiobook listening. Participants listened to Dutch and French stories while their brain activity was recorded. Linguistic features reflecting both syntactic complexity and statistical properties were extracted from the stimuli. Rule-based features were derived from constituency tree parses (syntactic depth and number of closing brackets), while statistical features (word surprisal and entropy) were obtained from a large language model (GPT2). Temporal Response Function (TRF) models were constructed using ridge regression to predict MEG signals based on these linguistic features. Model performance was evaluated by correlating predicted and observed neural activities. The study compared the reconstruction accuracy of models using different feature sets (rule-based, statistical, and a joint model) and assessed their temporal dynamics across frequency bands (delta, theta, beta, gamma) by analyzing phase consistency, power modulation, and phase-amplitude coupling (PAC). An innovative approach was employed to estimate PAC for continuous MEG data, using a linear forward model based on the complex analytical signal formed by combining low-frequency phase and high-frequency amplitude. Statistical significance was evaluated through model comparisons, including null models with shuffled feature values, cluster-based permutation tests, and linear mixed-effects models.

The MEG data showed neural oscillations in the alpha and beta bands, with a marginally significant difference in beta power between French and Dutch listening conditions. Cerebro-acoustic coherence analysis revealed phase alignment between sound amplitude and MEG signals in the delta and theta bands, stronger in the French condition. Word-triggered analyses found significant power modulation in the beta band and inter-word phase consistency in the delta and theta bands. TRF models showed that both rule-based syntactic features and statistical features significantly improved MEG signal reconstruction beyond a baseline model. Importantly, the joint model incorporating both feature sets significantly outperformed models using either feature set alone, suggesting a synergistic effect. Time-resolved analysis revealed a temporally broader contribution of syntactic features, potentially reflecting integration of words into larger units. Phase-amplitude coupling (PAC) analysis found significant delta-beta and theta-gamma PAC, particularly in fronto-temporal regions. The novel TRF-based PAC analysis revealed that individual features significantly modulated PAC, with word entropy and the number of closing brackets having a strong influence on theta-gamma PAC. Superior temporal gyri bilaterally showed high PAC coefficients, and the left hemisphere exhibited clusters in the inferior frontal gyrus and anterior temporal lobe, regions associated with syntactic processing and semantic composition.

The findings demonstrate that both syntactic structure and statistical predictability of words jointly contribute to shaping neural dynamics during language comprehension. The superior performance of the joint model highlights the brain's integrated processing of these seemingly distinct aspects of language. The overlapping brain regions involved in processing both feature sets, including the left inferior frontal gyrus and anterior temporal lobe, support the idea of an integrated neural mechanism. The temporal dynamics suggest distinct roles: syntactic features contribute more broadly in time, potentially reflecting the integration of words into larger syntactic structures, whereas statistical features exhibit a more focused, potentially predictive role. The results align with predictive processing frameworks, where both top-down predictions based on linguistic knowledge and bottom-up error signals (represented by surprisal and entropy) are crucial for real-time language understanding. The observed phase-amplitude coupling suggests a mechanism for binding neural representations related to both syntactic and statistical information, enhancing temporal prediction and facilitating the integration of linguistic units.

This study demonstrates the synergistic role of syntactic structure and statistical predictability in shaping neural dynamics during language comprehension. Both rule-based syntactic features and statistically derived word predictability improve MEG signal reconstruction, with a joint model significantly outperforming individual models. The observed phase-amplitude coupling suggests a neural binding mechanism coordinating these information streams. Future research could investigate the role of these mechanisms in other aspects of language processing, such as cross-linguistic comparisons and the influence of individual differences in language proficiency. The use of naturalistic speech paradigms and advanced computational modeling techniques significantly contributes to a deeper understanding of human language comprehension.

While the study uses a novel approach to analyze phase-amplitude coupling in continuous MEG data, it still relies on certain assumptions, such as the linearity of the forward model. The study focuses on word-level features, and finer-grained linguistic units (like phonemes or syllables) could be investigated in future studies to provide a more complete picture of predictive mechanisms. Furthermore, the study did not control for semantic information present in the statistical features, limiting the ability to make strict claims about the independence of syntactic and statistical processing.