Computational neuroanalysis methods and related equipment based on hierarchical Bayesian models

Abstract

This application provides a computational neural analysis method and related equipment based on a hierarchical Bayesian model, belonging to the field of computational neuroscience technology. The application includes: constructing a dynamic language stimulus material library; collecting electroencephalogram (EEG) signals during a language comprehension task based on dynamic language stimulus sequences from the library; preprocessing and performing time-frequency analysis on the EEG signals to obtain time-frequency analysis data; using the time-frequency analysis data, employing a cluster basis permutation test to detect significant neural activity clusters that meet preset significance conditions; extracting trial-level neural features from the significant neural activity clusters; constructing a hierarchical Gaussian filter model based on the dynamic language stimulus sequence, and inferring trial-level computational variables through variational Bayesian inference; and quantifying the contribution relationship between trial-level computational variables and trial-level neural features based on the trial-level computational variables and trial-level neural features. The embodiments of this application can jointly model the "computational-neural" process of semantic expectation.

Description

Technical Field

[0001] This application relates to the field of computational neuroscience, and in particular to computational neuroanalysis methods and related devices based on hierarchical Bayesian models. Background Technology

[0002] In related technologies, traditional EEG ERPs or time-frequency EEG analyses largely rely on conditional averaging, a static method of "averaging trials." Mainstream semantic EEG research typically infers semantic processing by comparing the average differences between conditions of N400 or neural oscillations. This carries risks: first, missing the time or space of actual effects; and second, incorrectly attributing multiple overlapping neural processes to a single component.

[0003] In summary, the technical problems existing in the relevant technologies need to be improved. Summary of the Invention

[0004] The main objective of this application is to propose a computational neural analysis method and related equipment based on a hierarchical Bayesian model, which can achieve precise alignment of computational variables with EEG signals on a trial-by-trial basis, and realize joint computational-neural modeling of semantic expectation processes.

[0005] To achieve the above objectives, one aspect of this application proposes a computational neural analysis method based on a hierarchical Bayesian model, the method comprising the following steps: Construct a dynamic language stimulus material library; the dynamic language stimulus material library includes several dynamic language stimulus sequences consisting of trials organized alternately according to different expected effectiveness blocks; Collecting EEG signals during the execution of a language comprehension task based on the dynamic language stimulus sequence in the dynamic language stimulus material library; The EEG signals are preprocessed and subjected to time-frequency analysis to obtain time-frequency analysis data; Based on the time-frequency analysis data, the cluster basis permutation test method is used to detect significant neural activity clusters that meet the preset significance conditions; Extract trial-level neural features from the significant neural activity clusters; Based on the dynamic language stimulus sequence, a hierarchical Gaussian filter model is constructed, and the trial secondary computational variables are inferred and deduced through variational Bayesian inference. Based on the trial-level computational variables and the trial-level neural features, the contribution relationship between the trial-level computational variables and the trial-level neural features is quantified.

[0006] In some embodiments, constructing the dynamic language stimulus material library includes: Several dynamic language stimulus sequences are set up; the dynamic language stimulus sequences consist of trials organized alternately according to different expected validity blocks; the trials present language units word by word and include target word presentation events; the expected validity blocks include high expected validity blocks and low expected validity blocks; the proportion of trials in the high expected validity blocks that meet the preset conditions of the expected category is higher than the proportion of trials in the low expected validity blocks that meet the preset conditions of the expected category; Based on the dynamic language stimulus sequence, a dynamic language stimulus material library is obtained.

[0007] In some embodiments, the acquisition of EEG signals during the execution of a language comprehension task based on the dynamic language stimulus sequences in the dynamic language stimulus material library includes: Perform a language comprehension task based on the dynamic language stimulus sequence in the dynamic language stimulus material library, and simultaneously collect the subject's raw EEG signals during the task execution; In each trial of the dynamic language stimulus sequence, a hardware trigger signal is sent to the EEG acquisition device at the moment when the target word is presented, so that the moment when the target word is presented is marked as the time lock point of the EEG signal of the corresponding trial. The original EEG signal is time-aligned based on the hardware trigger signal to obtain an EEG signal aligned with the target word event of each trial.

[0008] In some embodiments, the preprocessing and time-frequency analysis of the EEG signal to obtain time-frequency analysis data includes: The EEG signal is subjected to bandpass filtering, rereference, and artifact correction to obtain the first data. The first data is segmented with the time when the target word is presented as the zero point to obtain the second data; The second data for each trial is subjected to a continuous wavelet transform using Morlet wavelets to obtain time-frequency analysis data.

[0009] In some embodiments, the step of using a cluster basis permutation test method to detect significant neural activity clusters that meet preset significance conditions based on the time-frequency analysis data includes: The time-frequency analysis data is statistically tested point by point in the electrode-time-frequency space, and adjacent points with statistical significance exceeding a preset threshold are aggregated to form candidate clusters. The null distribution of cluster statistics was constructed by repeatedly permuting experimental condition labels; The candidate clusters are compared with the zero distribution to obtain clusters of significant neural activity.

[0010] In some embodiments, extracting secondary neural features from the significant neural activity clusters includes: The average event-related potential amplitude is extracted from the significant neural activity clusters, or the average oscillation power is extracted from the preset frequency range of the significant neural activity clusters, as a secondary neural feature.

[0011] In some embodiments, the step of constructing a hierarchical Gaussian filter model based on the dynamic language stimulus sequence and inferring the trial secondary computational variables through variational Bayesian inference includes: Based on the dynamic language stimulus sequence, a hierarchical Gaussian filter model is constructed. The hierarchical Gaussian filter model includes a stimulus category layer and an environmental statistics layer. The stimulus category layer is used to represent the predictability state of the target word in a single trial. The environmental statistics layer is used to represent the individual's continuous belief in the overall expected tendency of the language environment. Variational Bayesian inference is used to invert the hierarchical Gaussian filter model to estimate the hidden state of each trial. During the inversion, trial-level computational variables representing the individual's trial-by-trial update of their belief in the statistical structure of the linguistic environment are extracted. These trial-level computational variables include prior belief, posterior belief, and precision-weighted expectation error. The prior belief represents the individual's expectation of the contextual statistical structure based on past sequences before observing the current trial input. The posterior belief represents the updated belief in the contextual statistical structure after observing the current trial input. The precision-weighted expectation error represents an uncertainty-weighted error signal calculated based on the deviation between the current trial input and the prior belief.

[0012] In some embodiments, quantifying the contribution relationship between the trial-order calculated variables and the trial-order neural features, based on the trial-order calculated variables and the trial-order neural features, includes: A single-trial generalized linear model is established, with the trial-level calculated variables as predictors and the trial-level neural features as response variables. Trial-by-trial regression analysis is performed to quantify the contribution relationship between the trial-level calculated variables and the trial-level neural features.

[0013] To achieve the above objectives, another aspect of this application proposes a computational neural analysis system based on a hierarchical Bayesian model to implement the method described above. The system includes: The stimulus construction and data acquisition preprocessing module is used to construct a dynamic language stimulus material library; the dynamic language stimulus material library includes several dynamic language stimulus sequences consisting of trials organized alternately according to different expected effectiveness blocks; the module acquires EEG signals when performing a language comprehension task based on the dynamic language stimulus sequences in the dynamic language stimulus material library; the module preprocesses and performs time-frequency analysis on the EEG signals to obtain time-frequency analysis data; The significant neural cluster detection module based on cluster basis permutation test is used to detect significant neural activity clusters that meet preset significance conditions based on the time-frequency analysis data and using the cluster basis permutation test method; and to extract secondary neural features from the significant neural activity clusters. An interpretable hierarchical Bayesian model construction and inversion module is used to construct a hierarchical Gaussian filter model based on the dynamic language stimulus sequence and invert the trial secondary computational variables through variational Bayesian inference. The computation-neural mapping module based on the single-trial generalized linear model is used to quantify the contribution relationship between the trial-trial computation variables and the trial-trial neural features.

[0014] To achieve the above objectives, another aspect of this application provides an electronic device, which includes a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the method described above.

[0015] To achieve the above objectives, another aspect of the embodiments of this application proposes a computer-readable storage medium storing a computer program that, when executed by a processor, implements the methods described above.

[0016] To achieve the above objectives, another aspect of the embodiments of this application proposes a computer program product, including a computer program that, when executed by a processor, implements the aforementioned method.

[0017] The embodiments of this application include at least the following beneficial effects: This application provides a computational neural analysis method, system, electronic device, storage medium, and program product based on a hierarchical Bayesian model. This application includes: constructing a dynamic language stimulus material library; collecting EEG signals during a language comprehension task based on dynamic language stimulus sequences from the dynamic language stimulus material library; preprocessing and performing time-frequency analysis on the EEG signals to obtain time-frequency analysis data; based on the time-frequency analysis data, using a cluster basis permutation test to detect significant neural activity clusters that meet preset significance conditions; extracting trial-level neural features from the significant neural activity clusters; constructing a hierarchical Gaussian filter model based on the dynamic language stimulus sequence, and inferring trial-level computational variables through variational Bayesian inference; quantifying the contribution relationship between trial-level computational variables and trial-level neural features based on the trial-level computational variables and trial-level neural features. The embodiments of this application can jointly model the "computational-neural" process of semantic expectation. Attached Figure Description

[0018] Figure 1 This is a flowchart of the computational neural analysis method based on a hierarchical Bayesian model provided in the embodiments of this application; Figure 2 This is a schematic diagram of the structure of the hierarchical Bayesian generative model in language understanding provided in the embodiments of this application; Figure 3 This is a schematic diagram of the variational inference process in the hierarchical Gaussian filter (HGF) provided in the embodiments of this application; Figure 4 This is a schematic diagram of the interpretable hierarchical Bayesian computation-neural framework provided in the embodiments of this application. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of this application and are not intended to limit it. In the following description, when referring to the accompanying drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with those of this application; they are merely examples of systems and methods consistent with some aspects of the embodiments of this application as detailed in the appended claims.

[0020] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used herein is for the purpose of describing embodiments of this application only and is not intended to limit this application.

[0021] Before providing a detailed description of the embodiments of this application, some of the nouns and terms involved in the embodiments of this application will be explained first. The nouns and terms involved in the embodiments of this application are subject to the following interpretations.

[0022] 1) EEG: electroencephalogram.

[0023] 2) ERP: event-related potential.

[0024] 3) TF: time–frequency, time frequency.

[0025] 4) CBPT: cluster-based permutation test (a nonparametric multiple comparison control method).

[0026] 5) HGF: Hierarchical Gaussian Filter (Hierarchical Bayesian Perception Model).

[0027] 6) pwPE: precision-weighted prediction error.

[0028] 7) VB: Variational Bayes.

[0029] 8) Prior: a priori.

[0030] 9) Posterior: posterior perspective.

[0031] 10) GLM: general linear model (trial-by-trial regression).

[0032] 11) FFT: Fast Fourier Transform.

[0033] 12) IFFT: inverse FFT, inverse fast Fourier transform.

[0034] 13) RSVP: rapid serial visual presentation.

[0035] 14) N400: Negativity 400 component.

[0036] 15) LAP: Late Anterior Positivity.

[0037] 16) P600: Late posterior positivity, the positive component of the late posterior part.

[0038] 17) Theta-band: 4–8 Hz.

[0039] 18) Beta-band: 13–30 Hz.

[0040] 19) Trial: refers to a complete and independent basic observation unit in an experiment. In this application, a trial is a complete sentence comprehension process: starting from the fixation point, a sentence is presented word by word until the target word at the end of the sentence is presented and the induced EEG response is recorded.

[0041] 20) Morlet wavelet: A specific mathematical function (wavelet) used for time-frequency analysis. Its waveform is composed of a complex sine wave modulated by a Gaussian envelope, resembling a local vibrational wave packet in time.

[0042] 21) Zero distribution: In hypothesis testing, it refers to the theoretical or empirical probability distribution of a test statistic (such as t-value, cluster statistic) under the condition that the null hypothesis (usually "no effect") is true.

[0043] 22) E-Prime, a software for generating and running psychology experiments.

[0044] 23) ICA, Independent Component Analysis, is a statistical and computational technique for blind source separation used to decompose statistically independent source signals from multi-channel mixed signals.

[0045] 24) FieldTrip, an open-source EEG / MEG data analysis toolkit based on MATLAB, is a key computational platform for cluster basis permutation testing. It provides the infrastructure and functions for performing nonparametric statistical tests in the high-dimensional space of "electrode-time-frequency", supporting the core step of data-driven automatic detection of significant neural clusters in this application.

[0046] In related technologies, traditional probabilistic statistical language models and deep neural network models cannot perform hierarchical inference of semantic expectations in constantly changing natural language environments. Traditional language models typically suffer from the following core drawbacks: 1. Lack of explainability.

[0047] Whether it's statistical co-occurrence-based models like N-grams or deep neural network models such as RNNs and Transformers, their internal representations typically exhibit highly distributed and high-dimensional feature space structures. While they can achieve effective language prediction at the behavioral level, they struggle to establish verifiable one-to-one correspondences with key cognitive variables in language understanding (such as prior beliefs, prediction errors, or belief update magnitudes). Therefore, these models tend to focus on functional fitting rather than mechanistic explanation, making it difficult to reveal the underlying hierarchical computational structures and dynamic inference mechanisms in the language prediction process.

[0048] 2. Unable to model higher-order hidden states.

[0049] These traditional statistical language models and deep neural network language models mainly achieve prediction by learning statistical patterns of surface word order and local context. Their internal representations are difficult to explicitly encode high-level latent environmental variables such as "context stability" and "context reliability". Therefore, they are difficult to characterize the dynamic characteristics of the statistical structure of the language environment changing over time, and they are also difficult to simulate the cognitive process of human adaptive prediction and adjustment in uncertain contexts.

[0050] 3. Cannot be matched with neurological indicators.

[0051] Because these models lack explicit representations of belief updating, precision adjustment, and hierarchical uncertainty estimation processes, they cannot use these computational resources to explain neural activity patterns such as N400, LAP, or θ / β oscillations in the brain. Therefore, there is currently a lack of mature and effective technical pathways to jointly model hierarchical cognitive variables and brain electrodynamic processes at the trial level. This results in a significant gap between the computational layer and the neural implementation layer of semantic processing. Consequently, the complete computational neural mechanism of semantic processing has not yet been theoretically closed and cannot be systematically verified in experimental data. 4. Lack of hierarchical structure.

[0052] These models cannot naturally represent the multi-level cognitive processing in language understanding, from the lexical level to the discourse level. Therefore, traditional language models cannot provide an interpretable and verifiable neural computational framework for dynamic semantic prediction processes. Secondly, existing models generally lack interpretable hierarchical Bayesian computational processes, failing to systematically characterize crucial computational quantities in natural language understanding such as prior, posterior, prediction errors, and precision-weighted prediction error (pwPE), and even more so, failing to establish the correspondence between these computational variables and different EEG indicators. For example, existing frameworks cannot explain which neural components reflect word-level belief updates, which reflect re-estimation of contextual statistical structures, or reveal the neural representations of higher-level states such as environmental stability.

[0053] 5. Methodological constraints of fragmented research processes.

[0054] Research workflows often present fragmented toolchains: steps such as stimulus material processing, EEG data preprocessing, statistical modeling, and cognitive modeling are typically scattered across different platforms and methods, lacking complete integration. This not only reduces research efficiency but also results in poor reproducibility of findings across studies. Therefore, a unified methodological framework is needed that can operate end-to-end, tightly couple cognitive models with EEG neural indicators, and possess a hierarchical structure, interpretability, and generalizability.

[0055] It is known that existing technologies cannot simultaneously characterize the multi-layered semantic expectation update mechanism in dynamic language environments, and lack a unified framework for accurately mapping the computational variables of hierarchical Bayesian models to trial-order EEG neural signals. In view of this, embodiments of this application provide a computational neural analysis method and related equipment based on hierarchical Bayesian models. This scheme constructs a comprehensive framework combining a dynamic semantic expectation processing model based on hierarchical Bayesian models with trial-order EEG neural analysis, achieving accurate modeling of multi-layered belief state updates in dynamic language environments, and establishing a systematic correspondence between key EEG indicators such as N400, LAP, and θ / β and computational variables at the neural level. Through this unified cognitive computation-neural framework, this application can deeply reveal the complete computational neural mechanism of multi-layered expectation mechanisms in language understanding, providing a new, interpretable, scalable, and reproducible methodological system for language cognitive neuroscience.

[0056] The computational neuroanalysis method based on a hierarchical Bayesian model provided in this application relates to the field of computational neuroscience. The computational neuroscience method provided in this application can be applied to a terminal, a server, or software running on a terminal or server. In some embodiments, the terminal can be a smartphone, tablet, laptop, or desktop computer, but is not limited to these. The server can be configured as an independent physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server providing basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, CDN, and big data and artificial intelligence platforms. The server can also be a node server in a blockchain network. The software can be an application implementing the computational neuroanalysis method based on a hierarchical Bayesian model, but is not limited to the above forms.

[0057] This application can be used in a wide variety of general-purpose or special-purpose computer system environments or configurations. Examples include: personal computers, server computers, handheld or portable devices, tablet devices, multiprocessor systems, microprocessor-based systems, network PCs, minicomputers, mainframe computers, and distributed computing environments including any of the above systems or devices. This application can be described in the general context of computer-executable instructions executed by a computer, such as program modules. Generally, program modules include routines, programs, objects, components, data structures, etc., that perform specific tasks or implement specific abstract data types. This application can also be practiced in distributed computing environments where tasks are performed by remote processing devices connected via a communication network. In distributed computing environments, program modules can reside in local and remote computer storage media, including storage devices.

[0058] It should be noted that in all specific embodiments of this application, when processing data related to user identity or characteristics, such as user information, user behavior data, user historical data, and user location information, user permission or consent is obtained first. Furthermore, the collection, use, and processing of this data comply with relevant laws, regulations, and standards. In addition, when embodiments of this application require access to sensitive personal information of users, separate permission or consent from the user is obtained through pop-ups or redirection to confirmation pages. Only after obtaining the user's separate permission or consent is the necessary user-related data required for the proper functioning of these embodiments acquired.

[0059] Figure 1 This is an optional flowchart of a computational neural analysis method based on a hierarchical Bayesian model provided in the embodiments of this application. Figure 1 The method may include, but is not limited to, steps S101 to S107.

[0060] Step S101: Construct a dynamic language stimulus material library; the dynamic language stimulus material library includes several dynamic language stimulus sequences consisting of trials organized alternately according to different expected effectiveness blocks; Step S102: Collect EEG signals when performing a language comprehension task based on dynamic language stimulus sequences from a dynamic language stimulus material library; Step S103: Preprocess and perform time-frequency analysis on the EEG signal to obtain time-frequency analysis data; Step S104: Based on time-frequency analysis data, the cluster basis permutation test method is used to detect significant neural activity clusters that meet the preset significance conditions; Step S105: Extract trial secondary neural features from significant neural activity clusters; Step S106: Based on the dynamic language stimulus sequence, construct a hierarchical Gaussian filter model, and infer the trial secondary computational variables through variational Bayesian inference; Step S107: Based on the trial-level computational variables and trial-level neural features, quantify the contribution relationship between the trial-level computational variables and trial-level neural features.

[0061] Steps S101 to S107 as illustrated in this embodiment construct a dynamic language stimulus sequence with alternating high and low expected effectiveness blocks in step S101, simulating statistical fluctuations in the natural language environment. This provides a quantifiable framework to address the problem that traditional research relies on static sentence filling and cannot characterize dynamic learning. Furthermore, steps S102-S105 preprocess the millisecond-level hardware-triggered synchronously acquired EEG signals and use a data-driven cluster-basis permutation test to automatically identify significant neural activity clusters and extract secondary features, thereby overcoming the reliance on traditional EEG analysis. Subjective bias and multiple comparison issues in the prior time window / electrode region; finally, by constructing a hierarchical Gaussian filter model in step S106, the trial-level computational variables (prior, posterior, and precision-weighted expected error) are inverted, and combined with the single-trial generalized linear model established in step S107, the interpretable cognitive computational variables and brain neural signals are accurately aligned and quantitatively correlated at the trial-by-trial level. This systematically reveals how the updating of high-level beliefs in semantic expectation processing drives neural dynamics, providing an end-to-end, engineerable analytical solution for the computational neural mechanisms of language understanding.

[0062] In some embodiments, step S101 may include, but is not limited to, steps S111 to S112: Step S111: Set up several dynamic language stimulus sequences; the dynamic language stimulus sequences consist of trials organized alternately according to different expected validity blocks; each trial presents language units word by word and includes a target word presentation event; the expected validity blocks include high expected validity blocks and low expected validity blocks; the proportion of trials in the high expected validity blocks that meet the preset conditions of the expected category is higher than the proportion of trials in the low expected validity blocks that meet the preset conditions of the expected category. Step S112: Based on the dynamic language stimulus sequence, obtain the dynamic language stimulus material library.

[0063] In steps S111 to S112 of some embodiments, step S111 sets up a dynamic language stimulus sequence organized by alternating high and low predictability validity blocks to construct a controllable experimental paradigm that can accurately simulate the statistical fluctuations of the natural language environment. Specifically, "high predictability validity blocks" refer to experimental paragraphs with a significantly high probability of predictable target words appearing (e.g., 80% of sentences end with predictable words), while "low predictability validity blocks" refer to paragraphs with a low probability of predictable words appearing (e.g., only 20%), and the two are presented alternately in a specific order. This design, by manipulating the "trial ratio" (i.e., the ratio of predictable to unpredictable trials in different blocks), enables subjects to implicitly learn and continuously update their internal estimates of the statistical structure of the environment during the experiment, thereby solving the problem that traditional language understanding research relies on static context and cannot induce and quantify hierarchical expectation mechanisms. Based on this, step S112 integrates the above sequence into a structured dynamic language stimulus material library, providing stable and repeatable input conditions for subsequent computational modeling, and realizing high-fidelity simulation and manipulation of real language learning dynamics in a laboratory environment.

[0064] In some embodiments, step S102 may include, but is not limited to, steps S201 to S203: Step S201: Perform a language comprehension task based on dynamic language stimulus sequences in a dynamic language stimulus material library. During the task execution, the subject's raw EEG signals are collected simultaneously. Step S202: In each trial of the dynamic language stimulus sequence, a hardware trigger signal is sent to the EEG acquisition device at the time of the target word presentation, so that the time of the target word presentation is marked as the time lock point of the corresponding trial's EEG signal. Step S203: Perform time alignment processing on the original EEG signal based on the hardware trigger signal to obtain EEG signals aligned with the target word events of each trial.

[0065] In steps S201 to S203 of some embodiments, step S201 performs a language comprehension task based on dynamic language stimulus sequences and simultaneously acquires raw EEG signals, allowing neural data and cognitive tasks to be acquired synchronously in time. Based on this, in step S202, at the precise millisecond moment when the target word (i.e., the key expected word at the end of the sentence, such as the expected "lawsuit" or the unexpected "football match") is presented in each trial, a specific hardware trigger signal is sent to the EEG device, marking this moment as the time lock point of the neural data for that trial, thereby establishing a precise time anchor in the continuously recorded EEG signal stream. Finally, in step S203, these hardware trigger signals are used to perform time alignment processing on the raw EEG signals, cutting and aligning all EEG data segments of all trials based on the presentation time of their respective target words. This series of operations achieves millisecond-level precise alignment between language input events and high temporal resolution EEG signals, laying an irreplaceable foundation of temporal precision for subsequent trial-by-trial analysis. It overcomes the problems of signal average distortion and loss of variation information between trials caused by temporal ambiguity in traditional EEG analysis, enabling the precise capture and modeling of the independent neural dynamics of each trial.

[0066] In some embodiments, step S103 may include, but is not limited to, steps S301 to S303: Step S301: Bandpass filtering, rereference, and artifact correction are performed on the EEG signal to obtain the first data; Step S302: Divide the first data into segments with the time when the target word is presented as the zero point of time to obtain the second data; Step S303: Use Morlet wavelet to perform continuous wavelet transform on the second data of each trial to obtain time-frequency analysis data.

[0067] In steps S301 to S303 of some embodiments, step S301 performs bandpass filtering, rereference, and artifact correction on the raw EEG signal to systematically eliminate environmental noise, individual differences, and non-brain-derived interference in the signal, thereby extracting pure and reliable first data. Based on this, step S302 segments the first data with the presentation time of the target word in each trial as the time zero point to obtain second data that is precisely aligned in time, establishing a unified neural time reference for subsequent analysis. Finally, step S303 uses Morlet wavelets to perform continuous wavelet transform on the second data in each trial, decomposing the time-domain signal into high-resolution time-frequency analysis data, which can simultaneously reveal the temporal process and frequency characteristics of neural activity. This series of standardized preprocessing procedures not only improves the signal-to-noise ratio and comparability of EEG signals, but also stably extracts neural oscillation features in the θ (4-8Hz) and β (13-30Hz) frequency bands, which are closely related to semantic expectation processing, through time-frequency decomposition technology. This provides high-quality, multi-dimensional, trial-aligned neural feature inputs for subsequent data-driven statistical testing and computational modeling, overcoming the shortcomings of traditional methods, such as unstable feature extraction and single information dimension caused by inconsistent preprocessing or reliance on time-domain analysis.

[0068] In some embodiments, step S104 may include, but is not limited to, steps S401 to S403: Step S401: Perform point-by-point statistical testing of time-frequency analysis data in the electrode-time-frequency space, and aggregate adjacent points with statistical significance exceeding a preset threshold to form candidate clusters; Step S402: Construct the null distribution of cluster statistics by repeatedly permuting the experimental condition labels; Step S403: Compare the candidate clusters with the zero distribution to obtain the clusters of significant neural activity.

[0069] In steps S401 to S403 of some embodiments, step S401 performs point-by-point statistical testing in the three-dimensional data space of "electrode-time-frequency" and aggregates adjacent points with statistical significance exceeding a preset threshold (e.g., p<0.05) to form candidate clusters, thereby achieving automated and unbiased preliminary detection of spatiotemporal-frequency multidimensional features in EEG signals. Based on this, step S402 uses multiple random permutations of experimental condition labels (e.g., "predictable" and "unpredictable" labels) to construct an empirical null distribution of cluster statistics, thereby establishing a statistical benchmark for judging cluster significance while maintaining data structure and correlation. Finally, step S403 compares the statistics of each candidate cluster with the null distribution to screen out significant neural activity clusters that exceed the level of random fluctuation. This complete process based on cluster-based permutation testing enables fully data-driven detection of significant effects in high-dimensional EEG data analysis without the need for pre-specifying regions of interest or time windows. It overcomes the shortcomings of traditional EEG research, such as reliance on subjective prior assumptions, serious multiple comparison correction problems, and poor result stability, ensuring the objectivity, robustness, and reproducibility of the identified neural features.

[0070] In some embodiments, step S105 may include, but is not limited to, step S501: Step S501: Extract the average event-related potential amplitude from the significant neural activity clusters, or extract the average oscillation power from the preset frequency range of the significant neural activity clusters, as the secondary neural feature.

[0071] In step S501 of some embodiments, the average event-related potential amplitude (e.g., amplitude of N400 and P600 components) corresponding to each trial is extracted from the significant neural activity clusters objectively identified by the CBPT method, or the average oscillation power is extracted from a preset frequency range (e.g., 4-8Hz in the theta band and 13-30Hz in the beta band). These indicators are used as key trial-level neural features. This operation accurately converts neural indicators (e.g., average ERP amplitude and average energy of frequency bands) that are traditionally only used for group-level comparisons into quantitative features that correspond one-to-one with each independent language stimulus trial, thereby providing directly mappable, stable, and reliable neural observation variables for subsequent trial-by-trial computational modeling. This not only overcomes the problem of losing dynamic information between trials due to conditional averaging in traditional methods, but also realizes the transformation of neural signals from "group-level descriptive statistics" to "individual trial-based modelable variables," providing a data foundation for revealing how the brain processes expectations and errors in real time during a single language event.

[0072] In some embodiments, step S106 may include, but is not limited to, steps S601 to S602: Step S601: Based on the dynamic language stimulus sequence, construct a hierarchical Gaussian filter model; the hierarchical Gaussian filter model includes a stimulus category layer and an environmental statistics layer; the stimulus category layer is used to represent the predictability state of the target word in a single trial; the environmental statistics layer is used to represent the individual's continuous belief in the overall expected tendency of the language environment; Step S602: The hierarchical Gaussian filter model is inverted using variational Bayesian inference to estimate the hidden state of each trial. In the inversion, trial-level computational variables representing the individual's trial-by-trial belief update of the statistical structure of the linguistic environment are extracted. The trial-level computational variables include prior belief, posterior belief, and precision-weighted expected error. Prior belief represents the individual's expectation of the contextual statistical structure based on past sequences before observing the current trial input. Posterior belief represents the updated belief in the contextual statistical structure after observing the current trial input. Precision-weighted expected error represents the error signal calculated based on the deviation between the current trial input and the prior belief, weighted by uncertainty.

[0073] In steps S601 to S602 of some embodiments, a hierarchical Gaussian filter model including a stimulus category layer and an environmental statistics layer is constructed in step S601 to establish a precise computational framework for the dynamic language understanding process. The stimulus category layer models the predictability of the target word in each trial (e.g., "predictable" or "unpredictable") as a discrete state; the environmental statistics layer models the expected tendency of the entire language block (e.g., "is the current environment highly predictable or unpredictable") as a continuously evolving belief. Based on this, variational Bayesian inference is used to invert the model in step S602, estimating the hidden cognitive state trial by trial and extracting three key interpretable trial-level computational variables: prior belief (the overall predictive strength of whether the current sentence is predictable based on history before seeing specific words), posterior belief (the updated predictive strength after seeing specific words, combined with new evidence), and precision-weighted expectation error (the magnitude of the deviation between the actual input and the expectation, weighted considering the uncertainty of the current environment). This series of operations enables millisecond-level, quantifiable mathematical characterization of the implicit, continuous belief update trajectory in natural reading tasks. It transforms the internal cognitive processes that could only be inferred through behavioral questionnaires or subjective reports into a series of time-series data that can be precisely aligned with EEG signals. This provides a directly verifiable method for the question of "how computational variables drive neural responses," overcoming the limitations of traditional cognitive models lacking neural predictive capabilities or deep learning models lacking interpretability.

[0074] In some embodiments, step S107 may include, but is not limited to, step S701: Step S701: Establish a single-trial generalized linear model, using trial-level calculated variables as predictors and trial-level neural features as response variables, and perform trial-by-trial regression analysis to quantify the contribution relationship between trial-level calculated variables and trial-level neural features.

[0075] In step S701 of some embodiments, a single-trial generalized linear model is established. Trial-level computational variables (such as prior beliefs, posterior beliefs, and precision-weighted expectation errors) obtained from the HGF model inversion are used as predictors, and trial-level neural features extracted from EEG data (such as N400 amplitude and theta band power) are used as response variables. Trial-by-trial regression analysis is then performed, thereby achieving a one-to-one, interpretable quantitative integration of cognitive computation theory and neural observation data based on time alignment. This crucial step can precisely quantify specific relationships such as "how many microvolts the N400 amplitude decreases for every unit increase in prior belief" or "how the magnitude of the precision-weighted expectation error specifically modulates the energy of the prefrontal theta oscillations." This method surpasses traditional correlation analysis based on conditional averaging. By revealing the covariation patterns between computational variables and neural signals in trial-level dynamics, it provides computational neural mechanism models with causal explanatory potential for higher cognitive functions such as semantic expectation, error processing, and belief updating, advancing brain science research from phenomenological description to a stage of mechanistic explanation based on computational principles.

[0076] This application also provides a computational neural analysis system based on a hierarchical Bayesian model to implement the method described above. The system includes: The stimulus construction and data acquisition preprocessing module is used to construct a dynamic language stimulus material library. The dynamic language stimulus material library includes several dynamic language stimulus sequences consisting of trials organized alternately according to different expected effectiveness blocks. The module collects EEG signals when performing a language comprehension task based on the dynamic language stimulus sequences in the dynamic language stimulus material library. The EEG signals are preprocessed and subjected to time-frequency analysis to obtain time-frequency analysis data. The significant neural cluster detection module based on cluster basis permutation test is used to detect significant neural activity clusters that meet the preset significance conditions based on time-frequency analysis data and the cluster basis permutation test method; and extract trial secondary neural features from the significant neural activity clusters. An interpretable hierarchical Bayesian model construction and inversion module is used to construct a hierarchical Gaussian filter model based on dynamic language stimulus sequences and invert the trial secondary computational variables through variational Bayesian inference. The computation-neural mapping module based on the single-trial generalized linear model is used to quantify the contribution relationship between trial-trial computation variables and trial-trial neural features based on trial-trial computation variables and trial-trial neural features.

[0077] As an optional implementation, this application proposes an end-to-end computational-neural analysis method based on a hierarchical Bayesian model (computational neuroanalysis method based on a hierarchical Bayesian model) and related equipment for parsing the hierarchical computational mechanism of semantic expectation processing in a dynamic language environment and achieving trial-by-trial precise alignment of computational variables with electroencephalogram (EEG) signals. This application integrates four modules—(1) language stimulus library construction, EEG data acquisition and preprocessing, (2) automatic identification of salient neural clusters, (3) interpretable hierarchical Bayesian model construction and inversion, and (4) neural-computational modeling based on trial-by-trial GLM—into a continuous analysis link, thereby realizing the joint computational-neural modeling of the semantic expectation process.

[0078] The core technical solution of this application includes the following four parts: (1) Stimulus construction and data acquisition preprocessing module: This application proposes a method for constructing and presenting a language stimulus library for dynamic semantic anticipation processing experiments. By combining a structured language material library with an executable experimental program, it automatically generates dynamic contexts and achieves fine-grained control over experimental trials and precise synchronization with EEG signals. During the experiment, the system presents language stimuli word-by-word while simultaneously recording the participants' behavioral responses in real time. A hardware triggering mechanism enables high-precision time stamping of stimulus events and EEG data, ensuring millisecond-level time alignment between language input and neural signals. For EEG data processing, this application designs an automated preprocessing workflow, including bandpass filtering, reference reconstruction, and automatic correction of artifacts such as electrooculography (EOG) and electromyography (EMG). Continuous EEG data is segmented based on key time-locking points to ensure precise matching between neural signals and stimulus events. Furthermore, this application employs a continuous wavelet transform method based on complex Morlet wavelets to perform time-frequency decomposition of EEG signals in the 1–30 Hz frequency band, thereby stably extracting neural oscillation features related to language processing and providing reliable feature indicators for subsequent data-driven statistical analysis.

[0079] (2) A significant neural cluster detection module based on cluster basis substitution test (CBPT): This application proposes a nonparametric statistical method for automatically identifying significant neural effects under dynamic language environment modulation. Based on point-by-point statistics and cluster-level permutation inference mechanisms in a three-dimensional "electrode-time-frequency" space, this method can automatically detect significant neural activity clusters in ERP signals and time-frequency power data, achieving objective identification of key neural effects. Unlike traditional methods that rely on preset electrode positions or time windows, this method does not require manual specification of the region of interest and can automatically complete multiple comparison corrections in a multidimensional data space, thereby reducing subjective bias and improving the robustness and reliability of the results. Simultaneously, this method can output cluster-level feature indices that are stable among subjects. Building upon this, this application further extracts single-trial neural feature indices from significant neural clusters for subsequent secondary statistical analysis and computational model fitting, thus providing data support for revealing the computational mechanism of language prediction processing.

[0080] (3) Explainable hierarchical Bayesian model construction and inversion: This application proposes a hierarchical Gaussian filter (HGF) modeling method suitable for simulating semantic expectation processing in dynamic language environments. This method characterizes the trial-by-trial belief update mechanism of an individual regarding the statistical structure of the language environment during language comprehension. The method employs a multi-level generative structure, models changes in the environmental state through a Gaussian random walk process, and dynamically adjusts the belief update amplitude using a precision-weighted expectation error signal. This allows for the inversion of the individual's prior beliefs about the high-level contextual structure and their time-varying updated posterior states (prior beliefs). Post-update verification Considering that natural reading tasks typically do not involve explicit behaviors such as key presses, this application employs a model containing only a perceptual layer to achieve trial-by-trial latent state estimation of the statistical structure of the linguistic environment without relying on continuous behavioral output. This method can stably extract key computational variables (such as precision-weighted expected error). This provides interpretable computational metrics for subsequent coupling analysis with EEG signals.

[0081] (4) Computational Neural Mapping Module Based on Trial-based Generalized Linear Model (GLM): This application proposes a trial-by-trial generalized linear model (GLM) method that maps the hidden states of a hierarchical Bayesian model to EEG neurodynamics. This method uses prior beliefs, posterior beliefs, and precision-weighted expectation errors output by a hierarchical Gaussian filter (HGF) as predictors, and ERP amplitude and oscillation power extracted through salient clusters as neural response variables to construct a statistical model of the contribution of computational variables to neural indices. This mapping framework can reveal the representation of different levels of beliefs in brain ERP dynamics and θ / β oscillations during semantic expectation at the trial-by-trial level, providing an interpretable quantitative approach to the correspondence between cognitive computational variables and neurodynamics. Furthermore, this method is engineerable, forming a unified workflow from computational modeling to neural data analysis, providing an operable technical solution for brain-cognitive mechanism research.

[0082] The solutions of the embodiments of this application will be described in detail and explained with reference to specific application examples: This application proposes an end-to-end computational-neural analysis method and framework based on the Hierarchical Bayesian Model (HMBM) for parsing the computational neural mechanisms of semantic expectation processing in dynamic language environments and aligning them with EEG signals trial by trial. This application integrates four modules into a unified analysis chain: (1) language stimulus construction, experimental procedure design, behavior acquisition, EEG neural signal acquisition and preprocessing; (2) automatic detection of significant neural clusters based on cluster basis permutation test (CBPT); (3) interpretable HMBM construction and inversion; and (4) HMBM hidden state-neural signal mapping of a trial-by-trial generalized linear model (GLM). The key implementation process is as follows: 1. Stimulus & Behavioral Data Acquisition and Preprocessing Module: This module aims to construct a structured stimulus library for dynamic language prediction experiments and generate an executable E-Prime experimental program based on this library, enabling synchronized language stimulus presentation, trial control, and EEG event triggering. This module includes: stimulus library construction, dynamic context block design, E-Prime program design, stimulus presentation and behavior, EEG data acquisition and preprocessing, and an EEG time-frequency analysis method based on complex Morlet continuous wavelet transform.

[0083] (1) Construction of the Stimulus Database: This application constructs a structured Chinese stimulus library for dynamic semantic prediction experiments, containing 320 sets of highly constrained sentence stimuli. The stimulus library is based on 160 highly constrained contextual sentences with a cloze probability greater than 0.65. Each contextual sentence is configured with one predictable target word and one unpredictable but semantically acceptable target word, resulting in 160 predictable sentences and 160 unpredictable sentences. To ensure strict control over predictability, this application uses an independent cloze probability test for quantitative evaluation (30 native Chinese speakers participated). The average cloze probability of predictable target words is 0.82, while the cloze probability of unpredictable target words is close to zero. Simultaneously, to ensure semantic reasonableness, this application uses an independent semantic acceptability scoring task for verification (27 native Chinese speakers participated). Both types of sentences received high acceptability scores, ensuring that the experimental materials are semantically natural and reasonable. Furthermore, lexical variables such as word frequency and stroke count of target words are strictly matched to eliminate interference from lexical factors. To simulate statistical fluctuations in natural language environments, this application further constructs a dynamic context structure, dividing 320 trials into four blocks, two of which are of high predictability (80% predictable) and the other two are of low predictability (only 20% predictable), presented in an alternating order of high and low effectiveness. This forms a dynamic language environment that can be used to trigger the brain's hierarchical expectation mechanism, providing stable and reliable input conditions for the subsequent hierarchical Bayesian model (HGF) to invert and model high-level latent states.

[0084] (2) E-Prime experimental procedure design: This application constructs an executable E-Prime experimental program based on a structured stimulus library to achieve language stimulus presentation, behavioral response acquisition, and EEG time synchronization. The experimental program uses a rapid sequence presentation (RSVP) method to display sentence stimuli and controls the timing of all events with millisecond-level precision. In each trial, a 300ms fixation point and a 100ms blank screen are presented sequentially to stabilize attention, followed by a word-by-word presentation phase. Each word is displayed for 400ms with a 200ms interval between words, all presented in black, fixed-width font and automatically recorded with precise timestamps. The key event in each trial is the target word; the system sends hardware trigger signals (predictable: 11; unpredictable: 12) to the EEG device upon its presentation, serving as the time lock point for ERP and time-frequency analysis. A 1200ms blank screen window is presented after the target word to record neural components such as N400, LAP, P600, and θ / β oscillations. To monitor comprehension levels, the program presented comprehension questions after approximately 20% of the trials and recorded key presses, reaction times, accuracy rates, and related EEG triggers (question code 99, key press codes 201–202). Furthermore, the experiment employed a dynamic contextual structure, dividing the 320 trials into blocks of high predictability (80% predictable) and low predictability (20% predictable), which were presented alternately in sequence. Block switching was invisible to the participants, allowing them to implicitly learn and form internal estimates of the environmental statistical structure.

[0085] (3) Stimulus presentation & behavior and EEG data acquisition and preprocessing: EEG data were acquired using a 64-channel Brain Products actiCHamp system, with electrodes arranged according to the international 10–20 system. The left mastoid process served as an online reference, AFz as the ground electrode, and an additional electrooculography (EOG) electrode was used to monitor eye movement artifacts. EEG was sampled at 1000 Hz, with electrode impedance kept below 5 kΩ. All trigger signals (target word trigger codes 11 / 12, question trigger codes 99, key trigger codes 201–202, and block trigger codes 31 / 32) were recorded synchronously with continuous electroencephalography (EEG). Offline preprocessing was performed using the EEGLAB and ERPLAB toolboxes in MATLAB, including bandpass filtering (0.02–30 Hz), rereference to mean values ​​from the left and right mastoid processes, electrophysiological artifact correction, ICA separation, and removal of eye movement / blink components. For ROI ERP analysis, trials exceeding ±50 μV were automatically excluded (average rejection rate approximately 14%); however, all trials were retained for trial-wise modeling (GLM / HGF) to maximize trial-wise signal variability. The continuous data was then segmented according to the trigger points of the key target words, from The analysis interval is defined as a time window from 200ms to 1000ms. The baseline is 200–0ms. All EEG epochs use the target word as the time zero point to achieve unified neural temporal alignment for semantically expected operations.

[0086] (4) EEG time-frequency analysis method based on complex Morlet continuous wavelet transform: This application provides a time-frequency feature extraction method for EEG based on continuous wavelet transform (CWT) for identifying neural oscillation patterns related to semantic expectation in dynamic language understanding tasks. This method employs complex Morlet wavelet (CMW) to perform high-resolution time-frequency decomposition of trial-level EEG data in the 1–30 Hz frequency band, and improves computational efficiency through fast frequency domain convolution. The steps are as follows: First, the continuous EEG signal is segmented according to the time of target word presentation. Analysis window from 200ms to 1000ms, and using The 200-0ms timeframe was used as the baseline. Subsequently, a wavelet family consisting of frequencies from 1 to 30 Hz was constructed, with higher frequencies corresponding to larger wavelet periods to ensure a dynamic balance between time and frequency resolution. To achieve efficient convolution, a Fast Fourier Transform (FFT) was used to convert the EEG signal and wavelet kernel to the frequency domain and multiply them point-by-point, followed by an Inverse Fourier Transform (IFFT) to obtain the complex form of the time-frequency signal. By squaring the complex coefficients modulowise, the oscillation energy of each trial at different time and frequency points was obtained, and further normalization in decibels (dB) was performed using the baseline window power to eliminate the influence of background noise, electrode differences, and individual differences. This method can stably extract the dynamic oscillation characteristics of semantically expected frequency bands such as θ (4–7 Hz) and β (13–30 Hz), and is suitable for subsequent cluster basis permutation test (CBPT), hierarchical Bayesian model (HGF) inversion, and single-trial GLM modeling analysis.

[0087] 2. Automatic detection of significant neural clusters based on Cluster Basis Permutation Test (CBPT): This application employs a cluster-based permutation test (CBPT) implemented using the FieldTrip toolkit to perform data-driven salient cluster detection on ERP and time-frequency data. This method is used to identify spatiotemporal (-frequency) neural effects modulated by the linguistic environment without prior region or window assumptions. The method automatically aggregates spatially or temporally adjacent salient samples by performing point-by-point nonparametric statistics on the electrode-time-frequency three-dimensional space of the time-domain (ERP) and time-frequency domain (power spectrum) data, thereby constructing candidate spatiotemporal (-frequency) clusters. The permutation test controls for multiple comparison errors. The specific process is as follows: First, perform within-subject t-tests at each electrode, time point, and frequency bin (for oscillation analysis); then, cluster adjacent samples with p < 0.05 according to the spatial neighborhood rule of the 64-channel 10–20 template, and sum the t-values ​​belonging to the same cluster as the cluster statistic; next, generate the null distribution of the maximum cluster statistic by randomly permuting the experimental condition labels (2,000 times); finally, compare the observed cluster statistic with the null distribution, and clusters exceeding the threshold (p < 0.05, two-tailed) are judged as significant clusters. ERP analysis in The analysis was conducted within a time window of 200 to 1,000 ms, and the data were downsampled at 500 Hz and baseline corrected. The time-frequency analysis used the same permutation framework in the time-frequency-electrode space and identified significant oscillation clusters based on the max-sum statistic.

[0088] To ensure the cross-subject stability of significant cluster correspondence effects, this application extracts the mean signal (ERP amplitude or oscillation power) of significant cluster regions identified by CBPT from each subject, and averages the results across all time, electrode (and frequency) points within the cluster to obtain trial-by-trial values. This cluster-derived index is then used for secondary statistical analysis, constructing a 2×2 repeated measures ANOVA with expected environment (high vs. low validity) and predictability (predictable vs. unpredictable) as within-subjects factors, and employing the R-package ezANOVA for statistical testing. Simple effects analysis is further performed when interactions are significant. The resulting trial-level cluster mean index is then used as a neural feature input to a single-trial GLM, and aligned trial-by-trial with the computational variables (prior, posterior, pwPE, etc.) obtained from the hierarchical Bayesian model (HGF) inversion, thereby establishing a mapping relationship between computational variables and neural signals. This CBPT-driven ROI extraction strategy ensures that statistical analysis is entirely based on data-driven salient clusters, avoiding statistical dependence on the cluster formation process, and providing reliable feature inputs for subsequent hierarchical modeling and neural mechanism analysis.

[0089] 3. Interpretable hierarchical Bayesian construction and inversion: Building upon existing research on hierarchical learning in uncertain environments, this study implements and establishes a task-adaptive, interpretable Hierarchical Gaussian Filter (HGF) model for dynamic contextual inference, used to characterize the belief update process at each trial level in dynamic linguistic contexts. HGF is a generative hierarchical Bayesian model that estimates multi-layered hidden states through variational Bayesian inference, thereby simulating adaptive learning in uncertain environments. This model can explain various neurophysiological patterns observed in semantic expectation, and its unique technical characteristics include: multi-layered structure, continuous state evolution, Gaussian random walks, and precision-weighted expectation error (pwPE) updates.

[0090] The Hidden Belief Framework (HGF) comprises two parts: a perceptual model and a response model. The perceptual model describes how the observer updates multiple layers of latent states from the input sequence; the response model generates behavioral responses based on these beliefs. This study employed a natural reading task, where participants only occasionally needed to answer comprehension questions, resulting in very little behavioral data. Based on this characteristic, this study used only the perceptual model to infer the trajectory of internal beliefs within the trial sub-levels, without using the response model.

[0091] In a linguistic context, each layer in the HGF hypothesis hierarchy generates an expectation for the layer below it and updates its own posterior beliefs based on error signals from the layers below. (Second layer) This represents a high-level belief in the statistical structure of the linguistic environment, such as the global probability of a predictable word (category "1") appearing in an experiment. This latent state evolves over time via a Gaussian random walk, reflecting the chronic changes in environmental statistics. (First layer) This represents the stimulus category in a single trial (predictable = 1, unpredictable = 0), and its generation probability is determined by the previous layer. It is obtained by transforming with the sigmoid function. The first input u directly reflects... The realized value is whether the sentence-final word conforms to contextual expectations. Through this structure, HGF can simulate an individual's internal estimation of linguistic environment statistics and how these estimations drive expectations of sentence meaning.

[0092] refer to Figure 2 This demonstrates a two-layer hierarchical Bayesian generative model for modeling semantic expectation processing. The model consists of the following layers from top to bottom: (1) Second layer (environmental validity layer): continuous latent variables This represents an individual's high-level belief about whether sentences in a linguistic environment are predictable, and it is updated with each trial using a Gaussian random walk, with a step size parameter of [parameter value missing]. This layer captures the trend of dynamic environmental structure changing over time; (2) First layer (stimulus category layer): binary latent variables The value is taken between the expected and unexpected categories, and the latent tendency of the second layer is transformed into the expected category of the current trial through the Sigmoid mapping; (3) Input layer (observation layer): The observation variable u represents the actual sentence type (expected or unexpected) presented at the moment, which is generated by the latent variables of the first layer. The right side of this figure shows the probability generation function of each layer and its corresponding interpretation in language understanding. This three-layer structure realizes the probabilistic modeling of the language input at the trial level, so that the system can perform trial-by-trial belief updates and expected error calculations based on each new sentence.

[0093] This study does not include the third-level volatility state in the traditional HGF model. This is because the high / low expectation environment in the experiment is a stable range and does not include dynamic changes in statistical fluctuations.

[0094] The hierarchical dynamics of HGF follow a Gaussian random walk: Generated by Bernoulli distribution: p( (1); The second-level state evolves with the following distribution: p( ) ~ N ( (2); Its update rule follows the general form of expected encoding and reinforcement learning models: (3); At any layer i, the posterior mean is updated as follows: (4); (5); in For the expected error, For accuracy ratio, Precision-weighted expected error (pwPE) is used. GLM analysis primarily focuses on the second layer. This represents the dynamic belief update of the participants regarding the contextual statistical structure.

[0095] The variational inference mechanism in the Hierarchical Gaussian Filter (HGF) employed in this embodiment estimates the hidden states of the model (including prior and posterior states) and model parameters (such as the coupling strength between layers). Since the true posterior distribution typically exhibits high dimensionality and non-Gaussian properties, directly solving for the closed-form analytical solution of the posterior distribution is infeasible. This embodiment introduces Variational Bayes (VB) inference to replace the true posterior distribution with a tractable, simplified distribution, thereby achieving a computationally feasible inference process.

[0096] In this method, the true variational energy function is obtained by calculating the log joint probability of the co-occurrence of the hidden state and the observed input (such as a sentence category sequence) under the generative model (see reference). Figure 3 (Middle blue curve). To obtain a parsely updated form, this embodiment uses the posterior mean μ from the previous time step. (k 1) A second-order Taylor expansion of the energy function is then performed to construct a local quadratic approximation of the energy function, and an approximate Gaussian posterior distribution is obtained (see reference). Figure 3 (Middle red curve). Approximates the traditional Laplace curve (see reference). Figure 3 Compared to the green curve in the middle (which uses the posterior mode as the expansion point), this method uses the posterior expectation of the previous time step as the expansion point, which can significantly reduce computation and storage consumption and is more in line with biological feasibility in online inference scenarios.

[0097] Although this approximation method does not guarantee an accurate true posterior distribution, it can provide a stable and biologically sound belief update mechanism with low computational cost under online conditions, thereby enabling real-time inference of hidden states.

[0098] This application embodiment Figure 3 This demonstrates the computational mechanism by which variational Bayesian (VB) inference approximates the true posterior in the HGF model. The blue curve represents the true variational energy function. I(X) This reflects the joint logarithmic probability of the latent state and observed data under the generative model. Since the true posterior is often high-dimensional and non-Gaussian, direct solution is not analytical. Therefore, HGF uses a local second-order Taylor expansion to make a second approximation of the variational energy: μ = (the posterior mean of the previous trial) (k 1) As the starting point, for I(X) A second-order approximation is performed to obtain a locally Gaussian approximate energy function (red curve), corresponding to the updated posterior for trial k. Unlike the classical Laplace approximation (green curve, with the posterior mode as the expansion point), HGF chooses the posterior mean of the previous trial as the expansion point, making the inference process more computationally efficient and biologically feasible in online update scenarios. This method achieves fast, low-computational-cost updates to hierarchical belief states without significantly sacrificing accuracy.

[0099] 4. Hierarchical Bayesian hidden state-neural signal mapping of a trial-and-error generalized linear model (GLM): This application proposes a computational-neural mapping method based on a trial-by-trial generalized linear model (GLM) for quantitatively correlating the internal latent states obtained from the hierarchical Bayesian model (HGF) with trial-by-trial electroencephalography (EEG) neurodynamic indices. This method can accurately characterize how high-level semantic expectation beliefs modulate neural responses at different time and frequency scales in dynamic language environments, achieving trial-by-trial coupling between language expectation computational variables and ERP / oscillatory features.

[0100] First, neural metrics for each valid trial were extracted using cluster basis permutation test (CBPT) and repeated measures ANOVA. These metrics included trial-level amplitudes for N400, LAP, and P600, and trial power values ​​for the θ (4–8 Hz) and β (18–23 Hz) frequency bands. Subsequently, these neural metrics were mapped one-to-one with the hierarchical hidden states output by the HGF perception model. The hidden states included: the high-level prior belief μ2 before input, the updated posterior belief μ2 after input, and the second-level precision-weighted expectation error ε2 (pwPE). These variables collectively constituted the expectation factors of the single-trial regression model.

[0101] This application uses a generalized linear model to establish the following trial-order mapping relationship: glm(neural_measure~prior+posterior+pwPE,data=m1) (6); Where neural_measure(k) is the ERP amplitude or time-frequency (TF) power of the k-th trial, and prior, posterior, and pwPE represent the trial-level belief variables of the HGF, respectively. The model is implemented using R language... glm() The function executes and outputs the regression coefficients, standard errors, t-values, and significance levels of each expected factor, which are used to determine the contribution of different calculated variables to the neural signal.

[0102] Using this method, this application achieves a fine mapping between hierarchical Bayesian computational variables and trial-level neural dynamics, which can systematically reveal the specific representation of priors, posteriors, and pwPE in semantic expectations in observable ERP components and oscillatory activities in the brain, providing an engineering-reproducible analytical framework for the computational neural mechanisms of dynamic semantic processing.

[0103] refer to Figure 4 Interpretable Hierarchical Bayesian Computation-Neural Framework: Computational Neural Mechanisms for Semantic Expectation Processing in Dynamic Language Environments (A) Sentence Comprehension Task. Participants complete a sentence comprehension task in a language environment with a learnable probabilistic structure. All sentences are presented word-by-word, and participants read highly constrained context sentences (e.g., "Attorney Zhang won this..."). ___(a) Sentence-ending words may be predictable (e.g., "lawsuit") or unpredictable but reasonable (e.g., "football match"). (b) EEG analysis based on cluster basis permutation test (CBPT). Continuous EEG records were analyzed in the time domain (ERP) and time-frequency domain (neural oscillations). The cluster basis permutation test was used to identify robust spatiotemporal-frequency significant clusters from five-dimensional EEG data (subject × frequency × time × number of trials × lead). (c) Environmental probability structure manipulation. The probabilities of predictable and unpredictable sentences alternated over time between high (80% predictability) and low (20% predictability) predictability validity blocks. Participants were not explicitly informed of the environmental statistical structure and had to infer it from the distribution of the input sentences. (d) Hierarchical Gaussian filter model (HGF). To explain the observed To understand the neural dynamics, we compared multiple models and selected the one with the highest fit. Model comparisons showed that the Hierarchical Gaussian Filter (HGF) performed best across multiple evaluation metrics. HGF infers the trial-by-trial dynamic belief trajectory from the input sequence, models the trial-by-trial belief updates at the local semantic and global environmental levels, and obtains the hierarchical hidden states, including the second-layer prior belief (belief2), posterior belief, and precision-weighted expectation error (pwPE2). (E) Individual belief trajectory and expectation error. The trial-by-trial hierarchical latent states generated by HGF serve as the expectation variables of GLM, used to explain the trial-by-trial variation of ERP and neural oscillations in dynamic language environments. The model reveals the complementary computational functions of prior, posterior, and pwPE in tracking the evolution of environmental statistical structure.

[0104] The beneficial effects of this application are: The end-to-end hierarchical Bayesian computation-neural analysis method (computational neuroanalysis method based on hierarchical Bayesian model) and framework proposed in this application overcome the shortcomings of existing language cognition and neuroanalysis techniques, such as the inability to achieve cognitive modeling for dynamic semantic expectation processing, the misalignment of EEG neural signals and computational variables, and the reliance on traditional superimposed average statistics. It has the following significant advantages and technical effects: (1) Achieving true trial-by-trial alignment between the hidden states of the interpretable computational model and the neural signals: This application achieves, for the first time, precise trial-by-trial alignment between hierarchical hidden states in an interpretable computational model and EEG neural signals. Existing language understanding and EEG research largely relies on conditional averaging, static contrastive models, or non-hierarchical models, lacking techniques to describe trial-by-trial belief updates simultaneously with neural responses. This application introduces a millisecond-level triggering mechanism from E-Prime, high-resolution time-frequency decomposition of complex Morlet wavelets, and trial-by-trial internal belief trajectories (including prior, posterior, and precision-weighted expected errors) obtained from hierarchical Gaussian filter (HGF) model inversion. Combined with a single-trial generalized linear model (GLM) to establish a one-to-one mapping between neural signals and computational variables, this allows for precise alignment of the language input, internal inference process, and corresponding ERP / oscillatory neural dynamics for each sentence trial on a unified time axis. This enables end-to-end computation-neural synchronic analysis simulating "language input → computational belief update → neural response." This capability is unattainable in traditional ERP averaging methods, static probability models, or deep learning black-box models, and is a key innovation of this application.

[0105] (2) For the first time, the experimental manipulation of a natural and dynamic language environment was realized (solving the problem of filling in sentences in traditional language comprehension research): This application constructs a dynamic context block with alternating high / low expectation validity, addressing the problem of traditional language comprehension research relying on fill-in-the-blank sentences and assuming fixed contexts. This design can characterize statistical fluctuations in real-world language environments, simulate and quantify participants' dynamic context learning, and extract interpretable belief states (prior, posterior, and precision-weighted expectation errors). For the first time, it realizes the long-lacking ability of "computational modeling of hierarchical expectation mechanisms" in language comprehension research. Through this novel and ecologically valid experimental paradigm, each trial can have a real impact on the subsequent language environment, providing a quantifiable framework for trial-by-trial research on expectation adaptation.

[0106] (3) Data-driven, unbiased neural cluster detection (solving the problem of traditional EEG relying on prior windows / electrodes), realizing a fully data-driven, prior-free automatic neural cluster detection method: Traditional ERP and time-frequency studies heavily rely on manually defined time windows and electrode regions, which easily introduces subjective bias and leads to serious multiple comparison problems in high-dimensional EEG data. This application employs FieldTrip's Cluster Basis Permutation Test (CBPT) to automatically identify spatially-temporally-frequency significant clusters without any prior assumptions, strictly controls multiple comparisons, and ensures that cluster detection is completely independent of subsequent statistical analysis (ANOVA / GLM), thereby greatly improving the objectivity, stability, and reproducibility of neural feature extraction.

[0107] (4) This application constructs an interpretable computation-neural mapping system (which traditional deep learning models cannot do): While existing deep learning models can fit neural data, they lack cognitive interpretability and cannot explain the computational origin of ERPs or neural oscillations. This application combines the hierarchical hidden states retrieved by HGF inversion with single-trial GLM modeling to explicitly reveal how N400, LAP, P600, and θ / β oscillations correspond to prior, posterior beliefs, and precision-weighted expectation error signals in semantic expectations. This directly verifies the functional division of neural indicators within the hierarchical expectation framework, thereby establishing a computational-neural mechanism model with causal explanatory power.

[0108] (5) This application establishes a reproducible, transferable, and deployable end-to-end analysis framework, improving the reproducibility of research and the deployability of engineering: This application's workflow encompasses stimulus construction, behavioral / EEG acquisition, preprocessing, CBPT feature extraction, HGF hierarchical modeling, and single-trial GLM mapping, forming a complete engineering chain from data input to model output. This framework is highly standardized, modular, and automated, and can be directly transferred to other language, cognitive tasks, or clinical research. It can also be applied to neural data from other modalities, including fMRI and MEG, significantly improving the reproducibility, scalability, and engineering deployment value of experimental research and computational simulations.

[0109] (6) This application has significant potential for clinical and educational applications: By quantifying an individual's prior beliefs, posterior updates, and precision-weighted expected errors on a trial-by-trial basis, and mapping these computational variables to observable EEG patterns, this application can be used for assessing the mechanisms of language disorders (such as aphasia and dyslexia), evaluating children's and adolescents' language expectation abilities, testing dynamic text comprehension abilities, and quantitatively detecting abnormalities in expectation coding in mental disorders, thus demonstrating broad application prospects in clinical diagnosis, educational assessment, and individual difference research.

[0110] In summary, the core innovations and benefits of this application can be summarized as: "For the first time, end-to-end, trial-by-trial, and interpretable joint analysis of computational variables and EEG signals based on hierarchical Bayesian modeling is achieved in a dynamic language environment." Key points of this application: 1. Trial-by-trial alignment of language input, computational model, and neural signals: This application achieves end-to-end trial-by-trial alignment of "sentence input → hierarchical belief update → EEG neural response," solving the problem that traditional EEG research relies on conditional averaging or non-hierarchical models, which cannot track the expectation and update process of each trial. This method achieves high-precision synchronization between language input and neural activity through millisecond-level E-Prime triggering, complex Morlet wavelet time-frequency decomposition, trial-by-trial belief trajectories generated by the HGF model, and GLM-based neural-computation mapping, providing unprecedented trial-by-trial dynamic tracking capabilities for studying semantic expectation processing.

[0111] 2. Experimental manipulation and hierarchical expectation modeling of natural dynamic language environments: This application constructs a dynamic context block alternating between high and low expectation validity, ensuring that each trial has a real impact on subsequent contexts, thereby simulating the dynamic changes of statistical features in real-world language environments. Combined with the inversion of the HGF model, the prior, posterior, and precision-weighted expectation errors of participants can be quantified trial by trial, characterizing their dynamic learning process under continuous language input. This method, for the first time, achieves "computational modeling of hierarchical expectation mechanisms," a long-standing deficiency in language comprehension research, providing a quantifiable framework for studying how individuals adjust their expectations in constantly changing language environments.

[0112] 3. Automatic detection and computation of neural clusters without prior knowledge – Neural mapping system: This application employs FieldTrip's Cluster Basis Permutation Test (CBPT) to achieve fully data-driven spatial-temporal-frequency significant cluster detection, effectively controlling for multiple comparisons and avoiding subjective bias caused by manually setting time windows or electrode regions. Simultaneously, by combining HGF hierarchical latent states with single-trial GLM modeling, interpretable mappings of N400, LAP, P600, and θ / β oscillations to belief states are achieved, establishing an interpretable, reproducible, and transferable end-to-end analytical framework. This provides a highly standardized and scalable tool for language, cognitive, and clinical research.

[0113] This application constructs a dynamic context block alternating between high and low expectation validity, and combines this with an HGF model to quantify belief states trial by trial, thereby achieving hierarchical expectation modeling and trial-by-trial expectation adaptation research in dynamic language environments. Theoretically, alternative solutions might include: Other dynamic context manipulation ratios: Besides the 80% / 20% alternating ratio of high and low expected validity used in this application, other ratios (such as 70% / 30%, 60% / 40%, etc.) can theoretically be chosen to construct dynamic context environments, thereby introducing different degrees of statistical fluctuations and expected uncertainty. These alternative ratios can still achieve dynamic linguistic environment manipulation, but may affect the significance of expected signals and the quantifiability of trial-by-trial belief states. In contrast, the 80% / 20% alternating ratio, while maintaining environmental controllability, can generate clear and stable statistical signals, enabling the HGF model to more effectively invert the prior, posterior, and precision-weighted expected errors of each trial, thereby optimizing the accuracy of belief quantification and neural-computation mapping.

[0114] Alternatives to other sentence types: Besides the predictable / unpredictable sentence comparison used in this application, theoretically, incongruent sentences could also be used to construct dynamic contexts and manipulate the effectiveness of environmental expectations. Such alternative designs can also introduce contextual statistical fluctuations and trigger expectation error signals; however, compared to the predictable / unpredictable design, incongruent sentences may introduce excessively strong anomalous effects, affecting the statistical separability of consecutive trials and the stability of trial-by-trial belief quantification. Therefore, this application chooses predictable / unpredictable sentence operations, which can realistically simulate the expected characteristics of natural language environments.

[0115] Other hierarchical Bayesian models: In addition to HGF, other hierarchical Bayesian or state-space models (such as Kalman filter, Dynamic Causal Modeling, etc.) can theoretically be used to invert trial-by-trial belief updates. However, when dealing with complex language sequences and precision-weighted expected errors, these models often struggle to simultaneously consider hierarchical structure, trial refinement, and interpretability, especially in multidimensional EEG time-frequency data, where the stability and reproducibility of model mappings may be limited.

[0116] Conclusion: Although several potential alternatives exist, the block+HGF scheme proposed in this application has significant advantages in terms of controllability, ecological validity, trial-by-trial belief quantification capability, and interpretability of neural-computational mapping. It is currently the optimal scheme for realizing hierarchical expectation modeling and trial-by-trial expectation adaptation research in dynamic language environments.

[0117] This application also provides an electronic device, which includes a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the above-described method.

[0118] It is understood that the content of the above method embodiments is applicable to this device embodiment. The specific functions implemented by this device embodiment are the same as those of the above method embodiments, and the beneficial effects achieved are also the same as those achieved by the above method embodiments.

[0119] This application also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described method.

[0120] It is understood that the content of the above method embodiments is applicable to this storage medium embodiment. The specific functions implemented in this storage medium embodiment are the same as those in the above method embodiments, and the beneficial effects achieved are also the same as those achieved in the above method embodiments.

[0121] This application also provides a computer program product, including a computer program that, when executed by a processor, implements the above-described method.

[0122] It is understood that the content of the above method embodiments is applicable to the embodiments of this program product. The specific functions implemented by the embodiments of this program product are the same as those of the above method embodiments, and the beneficial effects achieved are also the same as those achieved by the above method embodiments.

[0123] The computational neural analysis method, system, electronic device, storage medium, and program product based on hierarchical Bayesian models provided in this application, through the construction of an end-to-end analysis framework of "dynamic stimulus generation - synchronous EEG acquisition - data-driven feature extraction - interpretable computational modeling - trial-by-trial statistical mapping," achieves for the first time, in a dynamic language environment, the precise alignment and quantitative correlation between cognitive computational models (such as prior beliefs, posterior beliefs, and precision-weighted expectation errors derived from hierarchical Gaussian filter inversion) and brain neural signals (such as ERP amplitude and neural oscillation power) at the single trial level. This reveals the computational mechanism by which "internal belief updates" drive "external neural responses" in the semantic expectation processing process, overcoming the shortcomings of traditional methods that rely on static conditional averaging and cannot analyze trial-by-trial dynamic variations and computational roots.

[0124] The embodiments described in this application are for the purpose of more clearly illustrating the technical solutions of the embodiments of this application, and do not constitute a limitation on the technical solutions provided by the embodiments of this application. As those skilled in the art will know, with the evolution of technology and the emergence of new application scenarios, the technical solutions provided by the embodiments of this application are also applicable to similar technical problems.

[0125] Those skilled in the art will understand that the technical solutions shown in the figures do not constitute a limitation on the embodiments of this application, and may include more or fewer steps than shown, or combine certain steps, or different steps.

[0126] Those skilled in the art will understand that all or some of the steps in the methods disclosed above, as well as the functional modules / units in the systems and devices, can be implemented as software, firmware, hardware, or suitable combinations thereof.

[0127] The preferred embodiments of the present application have been described above with reference to the accompanying drawings, but this does not limit the scope of the claims of the present application. Any modifications, equivalent substitutions, and improvements made by those skilled in the art without departing from the scope and substance of the embodiments of the present application shall be within the scope of the claims of the present application.

Claims

1. A computational neural analysis method based on a hierarchical Bayesian model, characterized in that, The method includes the following steps: Construct a dynamic language stimulus material library; the dynamic language stimulus material library includes several dynamic language stimulus sequences consisting of trials organized alternately according to different expected effectiveness blocks; Collecting EEG signals during the execution of a language comprehension task based on the dynamic language stimulus sequence in the dynamic language stimulus material library; The EEG signals are preprocessed and subjected to time-frequency analysis to obtain time-frequency analysis data; Based on the time-frequency analysis data, the cluster basis permutation test method is used to detect significant neural activity clusters that meet the preset significance conditions; Extract trial-level neural features from the significant neural activity clusters; Based on the dynamic language stimulus sequence, a hierarchical Gaussian filter model is constructed, and the trial secondary computational variables are inferred and deduced through variational Bayesian inference. Based on the trial-level computational variables and the trial-level neural features, the contribution relationship between the trial-level computational variables and the trial-level neural features is quantified.

2. The method according to claim 1, characterized in that, The construction of the dynamic language stimulus material library includes: Several dynamic language stimulus sequences are set up; the dynamic language stimulus sequences consist of trials organized alternately according to different expected validity blocks; the trials present language units word by word and include target word presentation events; the expected validity blocks include high expected validity blocks and low expected validity blocks; the proportion of trials in the high expected validity blocks that meet the preset conditions of the expected category is higher than the proportion of trials in the low expected validity blocks that meet the preset conditions of the expected category; Based on the dynamic language stimulus sequence, a dynamic language stimulus material library is obtained.

3. The method according to claim 1, characterized in that, The acquisition of EEG signals during the execution of a language comprehension task based on the dynamic language stimulus sequences in the dynamic language stimulus material library includes: Perform a language comprehension task based on the dynamic language stimulus sequence in the dynamic language stimulus material library, and simultaneously collect the subject's raw EEG signals during the task execution. In each trial of the dynamic language stimulus sequence, a hardware trigger signal is sent to the EEG acquisition device at the moment when the target word is presented, so that the moment when the target word is presented is marked as the time lock point of the EEG signal of the corresponding trial. The original EEG signal is time-aligned based on the hardware trigger signal to obtain an EEG signal aligned with the target word event of each trial.

4. The method according to claim 1, characterized in that, The preprocessing and time-frequency analysis of the EEG signals to obtain time-frequency analysis data includes: The EEG signal is subjected to bandpass filtering, rereference, and artifact correction to obtain the first data. The first data is segmented with the time when the target word is presented as the zero point to obtain the second data; The second data for each trial is subjected to a continuous wavelet transform using Morlet wavelets to obtain time-frequency analysis data.

5. The method according to claim 1, characterized in that, Based on the time-frequency analysis data, the cluster basis permutation test method is used to detect significant neural activity clusters that meet preset significance conditions, including: The time-frequency analysis data is statistically tested point by point in the electrode-time-frequency space, and adjacent points with statistical significance exceeding a preset threshold are aggregated to form candidate clusters. The null distribution of cluster statistics was constructed by repeatedly permuting experimental condition labels; The candidate clusters are compared with the zero distribution to obtain clusters of significant neural activity.

6. The method according to claim 1, characterized in that, The extraction of secondary neural features from the significant neural activity clusters includes: The average event-related potential amplitude is extracted from the significant neural activity clusters, or the average oscillation power is extracted from the preset frequency range of the significant neural activity clusters, as a secondary neural feature.

7. The method according to claim 1, characterized in that, The process of constructing a hierarchical Gaussian filter model based on the dynamic language stimulus sequence and inferring the trial secondary computational variables through variational Bayesian inference includes: Based on the dynamic language stimulus sequence, a hierarchical Gaussian filter model is constructed. The hierarchical Gaussian filter model includes a stimulus category layer and an environmental statistics layer. The stimulus category layer is used to represent the predictability state of the target word in a single trial. The environmental statistics layer is used to represent the individual's continuous belief in the overall expected tendency of the language environment. Variational Bayesian inference is used to invert the hierarchical Gaussian filter model to estimate the hidden state of each trial. During the inversion, trial-level computational variables representing the individual's trial-by-trial update of their belief in the statistical structure of the linguistic environment are extracted. These trial-level computational variables include prior belief, posterior belief, and precision-weighted expectation error. The prior belief represents the individual's expectation of the contextual statistical structure based on past sequences before observing the current trial input. The posterior belief represents the updated belief in the contextual statistical structure after observing the current trial input. The precision-weighted expectation error represents an uncertainty-weighted error signal calculated based on the deviation between the current trial input and the prior belief.

8. The method according to claim 1, characterized in that, The step of quantifying the contribution relationship between the trial-level calculated variables and the trial-level neural features, based on the trial-level calculated variables and the trial-level neural features, includes: A single-trial generalized linear model is established, with the trial-level calculated variables as predictors and the trial-level neural features as response variables. Trial-by-trial regression analysis is performed to quantify the contribution relationship between the trial-level calculated variables and the trial-level neural features.

9. A computational neural analysis system based on a hierarchical Bayesian model, used to implement the method as described in any one of claims 1 to 8, characterized in that, The system includes: The stimulus construction and data acquisition preprocessing module is used to construct a dynamic language stimulus material library; the dynamic language stimulus material library includes several dynamic language stimulus sequences consisting of trials organized alternately according to different expected effectiveness blocks; the module acquires EEG signals when performing a language comprehension task based on the dynamic language stimulus sequences in the dynamic language stimulus material library; the module preprocesses and performs time-frequency analysis on the EEG signals to obtain time-frequency analysis data; The significant neural cluster detection module based on cluster basis permutation test is used to detect significant neural activity clusters that meet preset significance conditions based on the time-frequency analysis data and using the cluster basis permutation test method; and to extract secondary neural features from the significant neural activity clusters. An interpretable hierarchical Bayesian model construction and inversion module is used to construct a hierarchical Gaussian filter model based on the dynamic language stimulus sequence and invert the trial secondary computational variables through variational Bayesian inference. The computation-neural mapping module based on the single-trial generalized linear model is used to quantify the contribution relationship between the trial-trial computation variables and the trial-trial neural features.

10. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the method of any one of claims 1 to 8.

Images