A contextualized adaptive interaction evaluation method and system

Abstract

The application discloses a kind of contextual adaptive interaction evaluation method and system, it is related to human-computer interaction evaluation, data processing and artificial intelligence technical field, in acquisition end gathers region information, main use language, education years, historical evaluation record and compliance record to generate cultural context embedding;In side presentation contextual interaction topic and interaction task and gather answer duration, prompt number, error type, touch or mouse track, voice response and image or image data;According to perplexity and answer behavior trigger alternative question state machine and calculate rectification cost;Based on external objective calibration index, scale or task score to consistency score monotone mapping is established, and the consistency score is used as fusion anchor quantity;For each mode, programmable expert output layered judgment result and uncertainty are set, to improve the consistency and robustness of evaluation in cross-region, cross-language and low-education population scenarios.

Description

Technical Field

[0001] This invention relates to the fields of human-computer interaction evaluation, data processing and artificial intelligence technology, specifically a contextualized adaptive interaction evaluation method and system. Background Technology

[0002] In environments with differences in educational background, language, dialect, or cross-cultural contexts, interactive assessment processes based on fixed questions or expressions are easily affected by non-target factors such as misunderstandings of questions, differences in expression abilities, and differences in operating habits. This leads to insufficient comparability of assessment scores among different groups, resulting in biases in stratified results, trend judgments, or anomaly alerts. While existing computerized assessment systems can automate the collection of information such as response time, number of prompts, error types, voice responses, or interaction trajectories to some extent, they still face the following problems in cross-regional, cross-language, and low-education scenarios.

[0003] The questions and presentation methods often lack contextualized adaptation mechanisms related to region, language, and years of education, making it difficult to distinguish between errors caused by comprehension difficulties and errors caused by genuine changes in ability or state, thus affecting the robustness of stratification conclusions. Multimodal fusion often employs simple feature concatenation, fixed weighting, or unconstrained model integration methods, making it difficult to maintain a stable output scale and interpretability across different populations, and it is insufficiently adaptable to missing modalities and scenarios with limited edge resources. There is a lack of mechanisms to use external objective calibration indicators to calibrate the consistency of interaction evaluation scores, making it difficult to establish a stable mapping relationship between interaction scores and objective reference scales, thereby affecting the setting of anomaly thresholds, trend comparisons across time windows, and consistency judgments across populations. Some systems rely on stable networks and cloud computing power; when faced with network instability or interruptions, it is difficult to continuously record evaluation logs and ensure idempotent submissions and conflict merging, resulting in insufficient continuity in closed-loop management.

[0004] Therefore, there is an urgent need for a contextualized adaptive interactive assessment method and system for multi-regional, multilingual, and low-education populations. This system should be able to adaptively collect and integrate multi-source data, such as scales or tasks, behavioral trajectories, speech and language, and images or videos, under collaborative conditions between the edge and cloud. It should combine external objective calibration for consistency, programmable expert integration, and cultural context gating mechanisms to output hierarchical conclusions and deviation or anomaly scores. Furthermore, it should maintain closed-loop continuity under unstable network conditions through local encrypted caching, version identification, idempotent submission, and conflict merging. This would overcome the shortcomings of existing technologies in terms of cultural adaptability, consistency calibration, interpretable integration, and application in resource-constrained environments. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of the prior art and propose a contextualized adaptive interactive evaluation method and system to solve the above-mentioned problems.

[0006] The objective of this invention is achieved through the following technical solution: a context-based adaptive interaction evaluation method, comprising: S1 record building and context collection: Collect regional information, main language, years of education, historical assessment records and compliance records at the collection end to generate cultural context embedding; initialize the question bank index and voice pack on the client side, and establish local encrypted cache and version identifier; S2 Contextualized Adaptive Interaction Evaluation: Contextualized interactive questions and tasks are presented on the device side, and the answering time, number of prompts, error type, touch or mouse trajectory, voice response and image or video data are collected simultaneously. S3 Understanding Difficulty Judgment and Alternative Question State Machine: Based on the confusion level, answering time, number of prompts, and consecutive errors of similar questions output by the edge interaction engine, the alternative question state machine is activated when the trigger conditions are met, and it is executed in the order of rewriting prompts, downgrading the alternative question, and changing the modality; the correction cost is calculated for each trigger and written to the evaluation log for subsequent scoring correction and weight adjustment; S4 External Objective Calibration Consistency Calibration: Based on external objective calibration indicators, establish a monotonic mapping from scale or task scores to consistency scores, use the monotonic mapping to calibrate scale or task scores to obtain consistency scores, and use the consistency scores as anchor quantities for subsequent fusion and anomaly alerts. S5 Expert Integration Preparation: Set up programmable experts for scale or task modalities, behavioral trajectory modalities, external objective calibration index modalities, image or video modalities, and speech language modalities respectively, output their respective hierarchical judgment results and uncertainties, and record interpretable basis; S6 Multiplicative Expert Integration and Cultural Context Gating: Based on the uncertainty of each expert, the consistency measure between each modality and the consistency score, and the gating coefficient of the cultural context embedding output, the weight of each modality is determined, missing modalities are automatically ignored, multiplicative multimodal integration is completed, hierarchical conclusions and deviation or anomaly scores are generated, and key evidence and weight sources are output. S7 suggestion generation and distribution: Based on the stratification conclusions, deviation or abnormal scores and compliance history, automatically generate a suggestion set containing interactive training tasks, learning tasks or service process items, and adapt the difficulty according to the performance on the client side; the suggestion set is presented in a contextualized form through the client side interaction engine; S8 Continuous Monitoring and Local Anomaly Alerts: Records task completion rate, evaluation results, and compliance curves on the device at set intervals; when the consistency score of continuous evaluations drops below the first threshold or the deviation or anomaly score rises above the second threshold within a preset time window, a local anomaly alert is triggered, and event notifications are pushed to the management terminal and authorized receiving terminal. S9 Synchronization and Consistency Merging: When the network is unstable or interrupted, the data and event logs collected on the edge are written to the local encrypted cache; when the network recovers, idempotent commit and conflict merging are performed according to the version identifier, and the cloud completes deduplication, verification and write-back to ensure the continuity of the evaluation loop. The linkage between S3 to S6 is as follows: the cost of correction caused by difficulty in understanding reduces the contribution of language-related items to the stratification and increases the weight of non-linguistic modalities; consistency scores serve as anchor quantitative constraints on the weight allocation of multimodal integration; cultural context gating is used to simultaneously adjust the question type selection and modal weights, thereby reducing misjudgments and omissions across regions, languages, and low-education populations, and maintaining consistent decision-making logic on the edge and in the cloud.

[0007] The external objective calibration consistency calibration adopts monotonic mapping, with the optimization objective being to minimize the consistency error between the scale or task score and the external objective calibration index. The mapping parameters are obtained by supervised training with regularization. The monotonic mapping maintains a stable ranking in different language and education groups. The obtained consistency score is used for anomaly threshold determination and prompt triggering. The external objective calibration index includes laboratory indicators or biomarkers.

[0008] The triggering conditions for the difficulty discrimination and substitution question state machine include at least two of the following: the answering time exceeds the threshold, the number of prompts exceeds the threshold, and consecutive errors of the same type of question satisfy the threshold. The substitution question state machine is executed in the order of rewriting prompts, downgrading substitution questions, and changing modalities. The correction cost is included in the scale subdomain score correction and used to reduce the weight of language modality in expert integration.

[0009] The programmable expert for behavioral trajectory modal uses trajectory heatmaps, pause and hesitation ratios, path complexity, and micro-motion rhythms as features to output hierarchical judgment results and uncertainties. When the confusion level increases, the weight of the behavioral trajectory modal is increased, making the hierarchical results robust to language comprehension biases.

[0010] Cultural context gating uses an embedding of regional information, primary language, years of education, and time series of confusion levels as input, and is also used for the selection of question types and presentation methods as well as the weight adjustment of multimodal integration, thereby establishing a causal closed loop between the interactive evaluation process and the integrated decision-making.

[0011] The weights of multiplicative expert integration are determined by uncertainty, consistency measures between consistency scores and cultural context gating coefficients. When there are missing modalities, the output of missing modalities is automatically ignored, maintaining consistent integration rules between the end side and the cloud. The weights and interpretable basis of each modality are displayed on the management end.

[0012] The edge interaction engine adopts a three-branch sparse attention structure, including a compressed retrieval branch for historical context, a global branch for selecting continuous blocks based on importance, and a local branch with a fixed window. The results of the three branches are fused on the edge in a gating manner for confusion discrimination, alternative question triggering, and prompt generation under long-term, multi-turn interactions. It is also combined with fixed-point quantization and intra-group shared indexes to reduce edge latency and storage consumption.

[0013] Synchronization and consistency merging includes local encrypted caching, version identification, idempotent commits and conflict merging. Conflict merging follows the time priority principle and the latest version priority principle. After deduplication, integrity verification and write-back are performed. The device side performs feature encoding on voice responses, image or video data and behavioral trajectory data. The feature encoding results and encrypted digests are written to the local encrypted cache and used for synchronization. The original data is deleted or encrypted and stored on the device side according to the preset retention strategy.

[0014] Deviation or outlier scores are determined based on a weighted combination of at least the following: differences from individual baseline scores, correction costs, individual expert uncertainties, and consistency measures.

[0015] A contextualized adaptive interactive evaluation system includes an edge-side interactive terminal, a cloud service, and a data management layer. The edge-side interactive terminal is equipped with a contextualized presentation module, a confusion discrimination and substitution question state machine module, a behavior trajectory collection and encoding module, a three-branch sparse attention reasoning module, and a local anomaly prompting module. The cloud service is equipped with an external objective calibration consistency calibration module, a programmable expert management and integration module, a cultural context gating module, a suggestion generation and task orchestration engine, and a continuous monitoring and anomaly prompting module. The data management layer is equipped with encrypted storage, version identification, idempotent submission, and conflict merging mechanisms. The system is configured to execute a contextualized adaptive interactive evaluation method and output hierarchical conclusions, deviation or anomaly scores, and interpretable evidence. Under network instability conditions, it maintains closed-loop management through edge-side local anomaly prompts and data caching.

[0016] The beneficial effects of this invention are: By combining data collection with contextualized adaptive interactive assessment, the system achieves adaptation for different regions, language habits, and years of education. In step S1, the system incorporates regional information, primary language, years of education, and historical compliance records to create a cultural context embedding. In step S2, it selects questions and presentation methods closely aligned with local life scenarios. In step S3, it dynamically adjusts the prompts, question difficulty, and modal format through confusion detection and a substitute question state machine, ensuring users complete the interaction under relatively consistent comprehension conditions. Compared to solutions with fixed questions and single-language expressions, this reduces interference caused by comprehension biases and improves the comparability of cross-population assessment results.

[0017] By introducing external objective calibration consistency calibration and multimodal programmable expert integration, the consistency between interactive scores and external objective reference scales is improved. In step S4, the system establishes a monotonic mapping from scale or task scores to consistency scores, and optimizes the consistency error between interactive evaluation results and external objective calibration indicators to calibrate scores affected by cultural and educational differences to a unified consistency scale. In steps S5 and S6, programmable experts are configured for scale or task modalities, behavioral trajectory modalities, external objective calibration indicator modalities, image or video modalities, and speech language modalities, and integrated using a multiplicative approach. At the same time, the consistency score is used as an anchor quantifier in modal weight allocation, thereby maintaining a more stable output scale in missing modalities and cross-group scenarios, and providing interpretable evidence.

[0018] By deeply integrating behavioral trajectory modalities with cultural context gating, the system's robustness is enhanced in scenarios where language comprehension is limited. In step S2, the behavioral trajectory acquisition and encoding module continuously records touch or mouse trajectories, pauses, hesitations, and operational rhythms. In step S5, a programmable expert for the behavioral trajectory modality analyzes these features. When language confusion increases, cultural context gating automatically increases the weight of the behavioral modality in the integration process in step S6. This allows for a relatively accurate assessment of a subject's cognitive state even when their language expression ability is limited or they are unfamiliar with Mandarin. This complementary relationship between language and behavioral modalities effectively compensates for the limitations of purely language-based assessments in low-education populations and improves the robustness of the evaluation results.

[0019] By employing a three-branch sparse attention inference module and an edge-side local anomaly alerting mechanism, this invention ensures long-duration, multi-turn interaction capabilities while also considering availability under edge-side resource constraints and network instability conditions. The edge-side interaction engine performs compressed retrieval of historical context, global block selection, and local window modeling during interaction evaluation and confusion resolution. It reduces edge-side latency and storage consumption through a sparse attention structure and fixed-point quantization strategy. In step S8, the local anomaly alerting module, based on the consistency score and deviation or anomaly score obtained from continuous evaluation, can independently trigger anomaly alerts and record events even when the cloud is unavailable or the network is unstable, thereby reducing the risk of closed-loop interruption caused by network outages.

[0020] By designing a synchronization and consistency merging mechanism, this invention achieves seamless integration between edge-side data caching and cloud-based data management. In step S9, the system assigns a version identifier to each record generated on the edge, employs idempotent commit and conflict merging strategies, automatically performs data deduplication and merging upon network recovery, and reconciles with local anomaly alert records on the edge to ensure that continuous monitoring trajectories and evaluation logs do not experience breakpoints or duplications during long-term operation. This data management approach satisfies both data integrity and traceability requirements while also accommodating real-world scenarios involving offline or low-bandwidth terminals, reducing the workload of manual processing and verification.

[0021] By creating a closed loop of suggestion generation and distribution, along with continuous monitoring, this invention integrates assessment, suggestion, monitoring, and anomaly alerts. In step S7, the system automatically generates a set of suggestions based on hierarchical conclusions, deviation or anomaly scores, and compliance history, presenting them in a contextualized manner on the client-side. In step S8, the system continuously monitors task completion rate, assessment results, and compliance curves. When a decrease in consistency score or an increase in deviation or anomaly score is detected, an anomaly alert is triggered, supporting management review and process adjustments. Compared to solutions that only output results once or provide only simple reminders, this improves the continuity and traceability of closed-loop management. Attached Figure Description

[0022] Figure 1 The process of this invention Figure 1 ; Figure 2 The process of this invention Figure 2 ; Figure 3 The process of this invention Figure 3 . Detailed Implementation

[0023] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention. Example

[0024] like Figure 1 As shown, this embodiment provides a contextualized adaptive interactive assessment method for multi-regional, multilingual, and low-education populations. Through the collaborative work of the terminal-side interactive terminal, cloud services, and data management layer, it realizes a closed-loop process of file creation and context collection, contextualized interactive assessment, comprehension difficulty identification and alternative question state machine, external objective calibration consistency calibration, multimodal programmable experts and multiplicative integration, suggestion generation and distribution, continuous monitoring and local anomaly prompts, and synchronization and consistency merging.

[0025] In step S1, the system creates user profiles and collects contextual information at the data acquisition terminal. The acquisition terminal inputs or collects the user's regional information, primary language, years of education, historical assessment records, and compliance records, and assigns or binds a unique identifier field to the user for linking data across different sessions and terminals. Based on the discrete and continuous characteristics of the regional information, primary language, and years of education, the system converts this information into a fixed-length numerical vector using a combination of a pre-trained encoding network and lookup table mapping, forming a cultural context embedding for subsequent gating adjustments. As a reproducible example, the cultural context embedding dimension is... Region and language are mapped to 16-dimensional vectors through an embedding table. Years of education, after Min-Max normalization, are input into a two-layer fully connected network (32 neurons in the first layer, ReLU activation; 16 neurons in the second layer, ReLU activation) to obtain a 16-dimensional vector. The three vectors are concatenated and projected onto a linear layer to a 32-dimensional vector as the cultural context embedding. Upon first use, the terminal synchronizes a question bank index and voice pack related to the region and language from the cloud and establishes a local encrypted cache file and version identifier field in local storage. The version identifier is obtained by combining at least a unique user identifier, an event timestamp, and a locally incrementing counter on the terminal.

[0026] In step S2, the system performs contextualized adaptive interactive assessment on the edge terminal. Based on cultural context embedding and a question bank index, the edge terminal selects a set of questions relevant to daily life for the current round of assessment, presenting them with animations, background images, and voice prompts to guide the user through the interactive task. During the user's response, the edge terminal collects the response time for each question, the number of times the prompt button is clicked, the error type code, touch or mouse trajectory data, and the audio of the voice response; when the device is equipped with a camera or can access image / video input, it collects image or video data or its characteristic results.

[0027] To reduce privacy risks and bandwidth consumption, this embodiment preferably performs feature encoding on the speech, image, or video data, and behavioral trajectory data at the device side, and generates an encrypted digest. Feature encoding may include: Mel spectrum or self-supervised speech embedding for speech, visual embedding or keypoint vector for image or video, and heatmap and statistics for trajectory. The event log includes at least the question number, presentation timestamp, submission timestamp, answer result, number of prompts, error type encoding, and trajectory compression representation, and together with the feature encoding result, forms a round of evaluation log.

[0028] In step S3, the system executes the logic of the difficulty-to-understand judgment and substitution question state machine on the edge. The interaction engine calculates the perplexity value during the presentation and answering of each question. The perplexity value can be estimated based on the difference between the user's voice response or text input and the target answer / target operation using a language model or multimodal alignment model. For each question, the system records the answering time, the number of prompts, and consecutive errors, respectively denoted as... , and And preset a threshold for the response time. The threshold for the number of prompts and continuous error threshold Calculate the triggering factor: when If it is determined that the user has difficulty understanding the question, the alternative question state machine is activated, executing in the order of rewriting the hint, downgrading the question, and changing the modality. The cost of correction is calculated for each state transition. ; Correction costs The score is written to the evaluation log and stored along with the scale subdomain or task subdomain to which the question belongs, for subsequent score correction and adjustment of language-related modality weights. As an example, this embodiment takes... , , , , , .

[0029] In step S4, the system performs external objective calibration consistency calibration on the cloud or local server. External objective calibration indicators are used to provide an objective reference scale related to the evaluation object. These external objective calibration indicators can be one or a combination of laboratory indicators, biomarkers, imaging quantitative indicators, physiological signal indicators, or third-party measurement indicators obtained through standardized procedures, and are denoted as... The system summarizes the raw scores of the scales or tasks to obtain the raw total score. And introduce a mapping function whose parameters are trainable and maintain monotonicity. Map the original total score to a consistency score: The system minimizes the consistency loss on the sample set: The consistency score is designed to maintain a stable and consistent relationship with external objective indicators across different language and education groups; in this embodiment, the consistency score serves as an anchor for subsequent fusion and anomaly alerts.

[0030] To ensure right To achieve a monotonically non-decreasing and reproducible implementation, this embodiment preferably employs a piecewise linear monotonically mapping: within the interval Top settings Nodes (For example ),make in Let be a cap function or a ReLU basis function on the interval of nodes, and for all constraint Thus ensuring Monotonic and does not decrease. The Adam optimizer (learning rate) is used during training. , , Batch size 64, training epochs 100, regularization coefficient Pick to This embodiment can also incorporate an intra-group order consistency constraint: for any group Internal sample pairs like Then punish and take it as Additional items.

[0031] In step S5, the system assigns programmable experts to the scale or task modality, behavioral trajectory modality, external objective calibration index modality, image or video modality, and speech / language modality, respectively, outputs their respective hierarchical judgment results and uncertainties, and records the interpretable basis. For missing modalities, the system does not force their acquisition and automatically ignores them during subsequent fusion.

[0032] In step S6, the system performs multiplicative expert integration and cultural context gating. For the currently existing modality set... The system integrates the output probabilities of each mode using a multiplicative approach: Modal weights Determined jointly by uncertainty, consistency measure, and cultural context gating coefficient: in The result of the uncertainty transformation, This is a measure of consistency between modality and consistency score. This is the gating coefficient. The system outputs the overall hierarchical conclusion and continuous deviation or anomaly scores, and records the modal weights and key evidence in the management interface.

[0033] This embodiment provides a set of definitions for uncertainty and consistency measures: Uncertainty It can be obtained from the predicted distribution entropy output by the modal expert. and take

[0034] 2) Consistency Measurement Used to depict the first Modal output and anchoring quantity The degree of consistency. This embodiment preferably uses the most recent... Secondary evaluation window (e.g.) to Intra-calculation consistency: Let the first... The continuous score output by the modal expert is For example, taking the probability of the modality with respect to the anomaly category or its monotonic transformation), the consistency metric is taken as... in For the reason The distribution obtained by normalization, For the reason The distribution obtained by normalization, The divergence is Kullback-Leibler; when no historical window is available, it degenerates into distance consistency of a single evaluation. in To be Mapped to It is a monotonic function.

[0035] The above definition makes It is also influenced by whether the modality is confident, whether the modality is consistent with external objective anchors, and whether the user's cultural context needs to improve the modality.

[0036] In step S7, the system generates and distributes a set of suggestions based on the stratification conclusions, deviation or anomaly scores, and compliance history. The suggestion set includes interactive training tasks, learning tasks, or service process items, and its difficulty can be adaptively adjusted according to the client-side performance. As an example, the suggestion set is represented in a structured form, with each item including task type, frequency, duration, and priority; the frequency is 3 to 7 times per week, the duration is 5 to 20 minutes per session, and the priority is 1 to 5 levels.

[0037] In step S8, the system performs continuous monitoring and local anomaly alerts. The terminal initiates new assessments at set intervals and records task completion rate, assessment results, and compliance curves. When the consistency score drops below a first threshold or the deviation or anomaly score rises above a second threshold within a preset time window, the terminal triggers a local anomaly alert and pushes an event notification to the management terminal and authorized receivers. For example, the preset time window is 7 or 14 days, the first threshold is a 10% to 20% decrease in consistency score, and the second threshold is a 15% to 30% increase in deviation or anomaly score.

[0038] This embodiment provides a deviation or anomaly score. One method of composition. Let the individual baseline consistency score be... (For example, taking the average of the first three assessments), the current consistency score is The cost of corrective action is summarized as follows: The uncertainty of each expert is Consistency measure is Then the abnormality score can be: in to Non-negative weights and satisfy As an example, take , , , This anomaly score takes into account baseline differences, correction costs, expert uncertainty, and consistency measures.

[0039] In step S9, the system ensures the continuity of the evaluation loop through a synchronization and consistency merging mechanism. The endpoint writes the evaluation log, suggested execution records, and exception notification events to a local encrypted cache and assigns a version identifier. After network recovery, the data is submitted to the cloud with idempotency based on the version identifier. The cloud deduplicates data based on the version identifier and performs conflict merging, following the principles of time priority and latest version priority. After integrity verification, a confirmation identifier is written back.

[0040] To further reduce the risk of personal information exposure, this embodiment preferably only uploads the characteristic encoding results and encrypted digests, and deletes or seals the original voice, image or video and behavioral trajectory data on the device side according to a preset retention strategy; as an example, the original data is sealed on the device side with AES-256-GCM, and the retention period is 24 hours to 7 days, and it is automatically deleted after the retention period.

[0041] Example 2 like Figure 1 and Figure 2 As shown, this embodiment, based on the contextualized adaptive interaction evaluation method described in Embodiment 1, further refines the processing procedures of the edge-side interaction engine and the behavior trajectory modality, and improves the synchronization and consistency merging mechanism. This enables the system to maintain stable, low-latency, and robust hierarchical conclusions and deviation or abnormal score outputs even in scenarios with unstable networks, difficult language understanding, and long-term multi-turn interactions. In this embodiment, the overall process of S1 to S9, the archiving and context acquisition, external objective calibration consistency, programmable expert integration, and suggestion set generation can be referred to in Embodiment 1. This embodiment focuses on explaining the specific implementation of the programmable expert of the behavior trajectory modality, the three-branch sparse attention structure of the edge-side interaction engine, and the synchronization and consistency merging mechanism.

[0042] In this embodiment, step S2 presents contextualized questions and interactive tasks on the device side while simultaneously sampling touch or mouse operations at high frequency to form raw behavioral trajectory data. The terminal records the location coordinates, event type, and timestamp each time an interactive event occurs, obtaining a behavioral trajectory sequence. in and For the first Screen coordinates of the next interactive event For timestamps, This represents the total number of events related to this round of tasks. The trajectory sequence is indexed in the local cache by question number and task number, along with the answer result and prompt information.

[0043] In programmable experts focusing on behavioral trajectory modalities, the system constructs a trajectory heatmap based on trajectory sequences to characterize concentrated areas of user interaction on the interface. The trajectory heatmap is defined on a screen coordinate grid, and the heatmap values... The following formula is given: in This is a radius threshold for statistically analyzing nearby events; as an example, the screen resolution is... Time to take Pixel.

[0044] The system analyzes pauses and hesitations in behavioral sequences and defines the pause / hesitation ratio. This is the ratio of pause duration during effective interaction to the total interaction duration. in This refers to the sum of interaction event intervals that are greater than a preset threshold; as an example, the interval threshold is set to... To reflect the complexity of the behavioral trajectory, the system calculates the path complexity. ; in The total length of the trajectory. Path complexity. A larger value indicates that the trajectory direction changes frequently, which can be used to characterize behavioral patterns such as target search—getting lost—backtracking.

[0045] In terms of frequency domain characteristics, the system constructs rhythm signals from the time series of behavioral events. The frequency is obtained through discrete Fourier transform. Power spectral density at : Through analysis By analyzing energy distribution across different frequency bands, behavioral modal experts can detect whether users exhibit patterns such as high-frequency jitter, low-frequency sluggishness, or unstable rhythms.

[0046] The aforementioned trajectory heatmap, pause / hesitation ratio, path complexity, and rhythmic features constitute the input feature set of the behavioral trajectory modality. The behavioral trajectory modality programmable expert model outputs the hierarchical judgment results and uncertainty estimates for that modality. During system operation, when language perplexity significantly increases due to multiple substitution questions, the cultural context gating module automatically increases the weight of the behavioral trajectory modality in the multiplicative expert ensemble based on the perplexity time series. This makes the final conclusion rely more on behavioral patterns than on language understanding, thereby improving robustness in populations with strong dialects or lower levels of education.

[0047] Regarding the client-side interaction engine, this embodiment introduces a three-branch sparse attention structure to handle historical context information in long-term, multi-turn interactions. For each turn of interaction, the system maintains a query vector. Key vector sequence Sum value vector sequence To reduce edge storage and computational overhead, the system divides historical time steps into several segments of length [missing information]. The block, record the first The block contains a set of time steps. Define the compressed key vector of this block. :

[0048] The system according to and Importance score of inner product calculation block : Select the highest-scoring consecutive blocks to form the global branch attention key set. , At the same time, the recent history with a fixed window length is retained as a local branch. , Compressed retrieval branches are formed using compressed blocks. , .

[0049] This embodiment provides a set of example parameters: block length Choose from 16 to 32, preferably. Number of consecutive blocks selected in the global branch Choose from 2 to 6, preferably. (i.e., selection) (Key values ​​corresponding to the four largest consecutive blocks); fixed window length for local branches. Choose from 16 to 64, preferably. (That is, retain the key values ​​of the most recent 32 time steps). The compressed block of the compressed retrieval branch represents either the mean within the block (as shown in the above formula) or the attention-weighted mean within the block.

[0050] For the current time step The output vector The system performs gated fusion of the attention results from compressed branches, global branches, and local branches:

[0051] Gating coefficient After Softmax normalization: Through this gating mechanism, the system dynamically adjusts the relative contributions of the three attention branches based on the current interactive content.

[0052] When implementing the three-branch sparse attention structure on the edge, the system introduces a fixed-point quantization strategy and an intra-group shared index mechanism to reduce storage and computational overhead: fixed-point quantization will... , , Convert to a low-bit integer representation using a scaling factor; as an example, use 8-bit specific-point quantization (int8) for any vector. Get scaling factor And quantified as During reasoning Dequantization or maintaining fixed-point domain computation during the accumulation phase. In a multi-head attention structure, a group-shared index mechanism allows heads within the same group to share the block index list selected by the global branch, avoiding redundant sorting and collection; as an example, attention heads are grouped by size. Grouping, sharing within the same group A global block index.

[0053] Regarding the merging of synchronization and consistency, this embodiment refines the version identification and conflict handling strategies. When the client writes the evaluation results, behavioral trajectory characteristics, and local anomaly alert events to the local encrypted cache, it assigns a version identifier to each record. : in A unique identifier for the user. For event timestamps, This is the locally incrementing counter value on the endpoint; as an example, the...

[0054] HMAC-SHA256 is used with the built-in key as the HMAC key, and a 32-byte digest is output. After receiving the records, the cloud first deduplicates them based on the version identifier to ensure that the same version identifier is processed only once; then it processes the record set that may belong to the same logical event. The final records to be retained are selected based on the principles of time priority and latest version priority. : After deduplication and conflict merging are completed, the cloud performs integrity verification and sends back a list of confirmed version identifiers, which the client uses to clear the synchronized cache. The cloud also reconciles local anomaly alert records with the final merging results, verifying whether the anomaly alert triggering conditions match the final stratification conclusions, deviations, or anomaly score changes; if inconsistencies are found, they are marked as events requiring manual review.

[0055] Example 3 like Figures 1 to 3As shown, based on Embodiments 1 and 2, this embodiment provides a contextualized adaptive interactive evaluation system. Through the collaboration of the edge interactive terminal, cloud services, and data management layer, the method steps are implemented in a specific software and hardware architecture, realizing a closed-loop process from interactive evaluation and fusion decision-making to suggestion issuance and continuous monitoring. This embodiment focuses on the system-level structural design and data flow between modules. The algorithms disclosed in the previous embodiments are only referenced as necessary, and the description focuses on the deployment method of each module, the calling relationship, and the working mechanism under unstable network conditions.

[0056] The system in this embodiment includes a client-side interactive terminal, a cloud service, and a data management layer. The client-side interactive terminal is an interactive device with client-side applications installed. The cloud service provides functions such as external objective calibration consistency, programmable expert management and integration, cultural context gating, suggestion generation, and continuous monitoring and anomaly alerts. The data management layer is configured with encrypted storage, version identification, idempotent commit, conflict merging, and reconciliation mechanisms.

[0057] The terminal-side interactive system includes a contextualized presentation module. This module renders interactive assessment tasks and training / learning tasks related to local life scenarios on the terminal screen based on the assessment schemes and suggestion sets distributed from the cloud. After retrieving the question bank index, scene images, and voice packets from the cloud, the contextualized presentation module creates an index table in the local cache and presents the questions and tasks sequentially on the display layer using the user ID and session ID as keys.

[0058] The terminal also includes a state machine module for judging comprehension difficulties and alternative questions. It receives user answers, prompts, voice input summaries, and contextual representations output by the terminal inference module in real time. It judges the comprehension difficulties in the answering process and calls up alternative questions or changes the presentation mode when necessary.

[0059] The behavior trajectory acquisition and encoding module is attached to the terminal input event distribution layer. It listens for touch coordinates, mouse movement, click actions, and gesture trajectories at a fixed sampling frequency, and encodes them as behavior trajectory features. These features are then sent as behavior trajectory modal input to the cloud or on-device expert module. Simultaneously, the on-device can perform feature encoding on voice, image, or video data and generate encrypted digests to reduce bandwidth and privacy risks.

[0060] The edge-side inference module is used to maintain the dialogue context and evaluate the sub-states in long-term, multi-turn interaction scenarios, and participates in perplexity estimation, prompt generation, and alternative question triggering decisions. This embodiment preferably adopts the three-branch sparse attention structure described in Embodiment 2, and compresses the attention calculation to an acceptable latency range through fixed-point quantization and intra-group shared index compression key-value caching, enabling users to obtain near real-time feedback.

[0061] The local anomaly alert module shares the consistency scores and deviation or anomaly scores from the most recent evaluations with the edge-side inference module. Even when the network is disconnected or the cloud is unavailable, it can still determine whether to trigger an anomaly alert event based on changes in the local cache. When triggered, the local anomaly alert module calls the system notification interface to issue a notification and writes the event to the local encrypted cache, waiting for the network to be restored before reporting it to the cloud.

[0062] The cloud service is configured with an external objective calibration consistency calibration module, which reads external objective calibration indicators that match the user's identity from the data management layer. and the assessment of the original total score The original total score is converted into a consistency score using a monotonic mapping function. The consistency score is written to the user status table and pushed to the programmable expert management and integration module as a fusion anchor.

[0063] The programmable expert management and integration module maintains a list of expert models for scales or task modalities, behavioral trajectory modalities, external objective calibration index modalities, image or video modalities, and speech and language modalities. After normalizing and formatting the modal features from the edge and external systems, it calls each expert model according to a preset interface, collects their respective hierarchical judgment results and uncertainty estimates, embeds them with consistency scores and cultural context, and inputs them into a multiplicative integration algorithm to generate an overall hierarchical conclusion, deviation or anomaly score, and the weights and interpretable basis of each modality. The results are then sent to the continuous monitoring module and the suggestion generation module.

[0064] The cultural context gating module resides in the cloud service and provides gating coefficients to the contextual presentation module and the expert integration module through a unified context embedding service. This module reads user's regional information, primary language, years of education, and perplexity time series reported from the edge from the data management layer. It inputs these features into the gating network to generate question type selection gating vectors and modality weight gating vectors, which are then fed back to the contextual presentation module and the expert integration module. This allows the system to dynamically adjust the proportion of language questions, diagram questions, and operational questions, as well as the weight of each modality in the integration process, in subsequent rounds of evaluation and task execution, forming a causal closed loop from the interaction process to the fusion decision.

[0065] The suggestion generation module receives hierarchical conclusions and deviation or anomaly scores from the expert integration module, and reads compliance history and past suggestion execution records from the data management layer. Under the combined effect of built-in rules and learning strategies, it generates a set of suggestions targeting the current state and recent trends. This set of suggestions is represented in a structured form, with each item including task type, frequency, duration, and priority. It is pushed from the cloud to the on-device interactive terminal and presented in a contextualized manner using a localized real-life scenario. The suggestion generation module also writes key data points from the generation process into an interpretability log, including the most important modal weights, trigger records, and alternative question chains, for management to view and review.

[0066] The continuous monitoring and anomaly alert module uses a timeline as its main thread, aggregating evaluation results, consistency scores, deviation or anomaly scores, and task completion rates reported from the client side in the cloud, and constructing a trajectory curve for each user. Based on preset time windows and threshold rules, this module triggers cloud-based anomaly events when the consistency score decreases or the deviation or anomaly score increases. Notifications are then sent to the management and authorized receiving ends via push notifications, and the anomaly events are reconciled with local anomaly alert records on the client side.

[0067] The data management layer includes an encrypted storage unit, a version identifier and idempotent commit unit, and a conflict merging and reconciliation unit. The encrypted storage unit encrypts user profiles, evaluation records, external objective metrics, and operational logs before writing them to the database and performs access control. The version identifier and idempotent commit unit generates a version identifier for each record uploaded from the endpoint and performs idempotent processing based on the version identifier when reports are duplicated, preventing data duplication due to network jitter. Version identifier... satisfy: in A unique identifier for the user. For event timestamps, This is the locally incrementing counter value on the endpoint.

[0068] After the conflict merging and reconciliation unit receives the uploaded records, it first deduplicates them based on the version identifier; for sets of records that may belong to the same logical event but have different version identifiers... The final records to be retained are selected based on the principles of time priority and latest version priority. : After the conflict is merged, an integrity check is performed, and the list of accepted version identifiers is written back to the client to clear the cache.

[0069] The conflict merging and reconciliation unit periodically compares cloud-based anomaly events with local anomaly alerts on the endpoint to verify whether the triggering conditions on both ends are consistent within the same time period. If an anomaly alert is found on the endpoint but there is no corresponding record in the cloud, it is marked and the operations and maintenance personnel are notified to check the synchronization link.

[0070] Through the collaboration of the aforementioned end-side interactive terminal, cloud service, and data management layer, the system described in this embodiment can still output hierarchical conclusions, deviation or anomaly scores, and interpretable evidence even in complex scenarios such as network instability, modality loss, and language comprehension difficulties. Furthermore, it maintains closed-loop continuity through end-side anomaly alerts and data caching. All data collection in this invention complies with the requirements of relevant regulations.

[0071] Example 4 Based on the methods and systems described in Examples 1, 2, and 3, Example 4 designs comparative and ablation experiments. For ease of understanding, this example provides a set of comparative and ablation data to illustrate the changes in the effectiveness of the invention in terms of the stability of stratification conclusions, consistency with external objective calibration indicators, compliance, and applicability in low-bandwidth environments. Furthermore, Example 4 provides cognitive impairment screening and individualized intervention as application examples of the invention's methods, demonstrating the invention's scope of application and engineering feasibility.

[0072] In the comparative experiment, the control group used a conventional approach in this field: fixed-item interaction + manual or simple rule explanation + no consistency calibration process. This involved fixed items (or simplified electronic items in a fixed order), without the introduction of a contextualized adaptive question bank, a state machine for identifying and replacing difficult questions, multimodal programmable experts and multiplicative integration, or external objective calibration for interaction scores. Data synchronization was primarily simple uploading, lacking on-device local anomaly alerts and conflict merging mechanisms. The experimental group used the contextualized adaptive interaction assessment system described in Examples 1 to 3, which included a state machine for identifying and replacing difficult questions, external objective calibration, multimodal programmable experts and multiplicative integration, cultural context gating, behavioral trajectory experts, a three-branch sparse attention on-device interaction engine, and local anomaly alerts and synchronization and consistency merging mechanisms. Both groups were included in the same region and under the same inclusion criteria, and baseline assessments and one-year continuous monitoring were completed.

[0073] In this comparative experiment, 600 users were included, with 300 in the control group and 300 in the experimental group. External objective calibration indicators... The comprehensive objective index obtained through the standardized process was selected. This comprehensive objective index is obtained by standardized scaling and weighted fusion of multiple objective measurements (at least one of laboratory indicators, imaging quantitative indicators, and physiological signal indicators), and is used as a reference for consistency evaluation. The evaluation indicators include: (1) abnormal event detection rate (abnormal events judged by the third-party standardized process are considered positive); (2) comprehensive AUC; (3) sensitivity and specificity; (4) deviation or abnormality score and (5) Consistency (expressed as correlation coefficient or consistency measure); (6) Average time spent on a single interaction assessment; (7) One-year continuous monitoring completion rate; (8) User subjective satisfaction. The data shown in Table 1 were obtained.

[0074] Table 1 Comparative experimental results of the traditional process and the system of this invention

[0075] As shown in Table 1, compared with the control group, the experimental group showed improvements in abnormal event detection rate, overall AUC, sensitivity, and specificity. More importantly, the deviation or abnormality score of the experimental group was significantly better than that of the external objective calibration indicators. The consistency between the two groups was significantly improved (correlation coefficient 0.87 vs. 0.52), indicating that the output after external objective calibration and multimodal expert integration is closer to the objective reference scale. Simultaneously, the false positive and false negative rates in the experimental group decreased, demonstrating that distinguishing between comprehension difficulties and genuine changes in task ability, introducing behavioral trajectory modalities, and cultural context gating can effectively reduce false positives caused by language differences and differences in years of education. The average time per assessment was shortened, and the completion rate and satisfaction of the one-year continuous monitoring improved, reflecting the improvement in compliance and experience brought about by contextualized interaction and suggestion sets.

[0076] To verify the contribution of each key technical feature of this invention to the overall performance, this embodiment designed an ablation experiment within the experimental group, comparing the operational indicators of different schemes with and without the full configuration system and the simplified scheme with key modules removed. Scheme A is defined as the full configuration system; Scheme B is defined as removing the external objective calibration consistency module; Scheme C is defined as removing the behavioral trajectory modality expert; Scheme D is defined as removing the cultural context gating; and Scheme E is defined as replacing the three-branch sparse attention with conventional single-branch attention. Using the same evaluation metrics as described above, the data shown in Table 2 were obtained.

[0077] Table 2 Ablation Experiment Results of Key Modules of the System of the Invention

[0078] As can be seen from Table 2: when removing external objective calibration consistency calibration (Scheme B), compared with... The consistency decreased significantly, and the overall AUC and false negative rate worsened, indicating that consistency calibration, as a fusion anchor, is necessary to improve cross-population stability. When the behavioral trajectory expert was removed (Option C), the overall AUC decreased and the continuous monitoring completion rate decreased, indicating that the behavioral trajectory modality has a compensatory effect on language comprehension difficulties. When the cultural context gating was removed (Option D), the performance further decreased, indicating that dynamic question types and weight adjustment are key to adaptability for multi-regional and multilingual populations. When single-branch attention was used to replace three-branch sparse attention (Option E), the performance decreased slightly but the evaluation time increased significantly, indicating that the three-branch structure reduces edge latency while ensuring performance.

[0079] Application Example: Cognitive Impairment Screening and Personalized Intervention The following are application examples of the method of the present invention in the scenario of cognitive impairment screening and personalized intervention, illustrating an optional implementation method of the present invention in the medical and health management scenario. It should be understood that these application examples are not intended to limit the scope of protection of the claims, and the external objective calibration indicators described in the claims are not intended to limit the scope of protection of the claims. This is presented as an optional example only.

[0080] In this application example, the edge interaction evaluation uses a contextualized adaptive set of items and collects speech, behavioral trajectories, and optional image or video modalities. External objective calibration indicators are also used. Select a combination of laboratory indicators or biomarkers and imaging quantification indicators; as an example, it may include amyloid-related biomarkers (e.g., Aβ42 and Aβ40 and their ratio) and structural or functional imaging quantification indicators.

[0081] The system outputs stratification conclusions, deviation or anomaly scores, and interpretable evidence through multimodal programmable experts and multiplicative integration. In this application example, the stratification conclusions can be mapped to stratification results such as recommendations for routine continuous monitoring, further evaluation, or prompt medical consultation and review. The deviation or anomaly score is used to measure the trend of change relative to the baseline. In step S7, the system generates a set of recommendations, which may include cognitive training tasks (belonging to interactive training tasks), lifestyle recommendations (belonging to learning tasks), retest frequency recommendations, and service process items (e.g., appointment / consultation reminders, medical visit reminders). The above content is only an illustration of this application scenario and does not constitute a limitation on the scope of protection of the claims.

[0082] Six hundred subjects were enrolled in a certain region (300 in the control group and 300 in the experimental group). The control group used a traditional scale with manual interpretation and routine continuous monitoring, while the experimental group used the system of this invention. Using specialized neuropsychological assessment and biomarker / image quantification as references, the stratification performance for cognitively normal / mildly cognitively impaired / high risk of dementia was calculated, and the one-year continuous monitoring completion rate was statistically analyzed. Example data are shown in Table 3.

[0083] Table 3 Comparison of application scenarios for cognitive impairment screening index Control group: Traditional process Experimental group: The system of this invention Sample size (example) 300 300 Early detection rate of mild damage (%) 62 82 Comprehensive stratified AUC 0.78 0.90 Consistency with the Aβ42 / 40 ratio (correlation coefficient) 0.49 0.86 Average time per assessment (minutes) 35 28 One-year continuous monitoring completion rate (%) 58 81 As shown in Table 3, the system of the present invention can improve early detection capability and significantly enhance consistency with biomarker reference scale in this application scenario, while improving interaction time and continuous monitoring completion rate.

[0084] The above description is merely a preferred embodiment of the present invention. It should be understood that the present invention is not limited to the forms disclosed herein and should not be construed as excluding other embodiments. It can be used in various other combinations, modifications, and environments, and can be altered within the scope of the concept described herein through the above teachings or related technologies or knowledge. Modifications and variations made by those skilled in the art that do not depart from the spirit and scope of the present invention should be within the protection scope of the appended claims.

Claims

1. A contextualized adaptive interaction evaluation method, characterized in that, include: S1 Filing and Context Collection: Collect regional information, main language, years of education, historical assessment records and compliance records at the collection end to generate cultural context embedding; Initialize the question bank index and voice package on the client side, and establish a local encrypted cache and version identifier; S2 Contextualized Adaptive Interaction Evaluation: Contextualized interactive questions and tasks are presented on the device side, and the answering time, number of prompts, error type, touch or mouse trajectory, voice response and image or video data are collected simultaneously. S3 Understanding Difficulty Judgment and Alternative Question State Machine: Based on the confusion level, answering time, number of prompts, and consecutive errors of similar questions output by the edge interaction engine, the alternative question state machine is activated when the triggering conditions are met, and it is executed in the order of rewriting prompts, downgrading alternative questions, and changing modal. The cost of corrective action is calculated for each trigger and written to the evaluation log for subsequent score correction and weight adjustment. S4 External Objective Calibration Consistency Calibration: Based on external objective calibration indicators, establish a monotonic mapping from scale or task scores to consistency scores, use the monotonic mapping to calibrate the scale or task scores to obtain consistency scores, and use the consistency scores as anchor quantities for subsequent fusion and anomaly alerts. S5 Expert Integration Preparation: Set up programmable experts for scale or task modalities, behavioral trajectory modalities, external objective calibration index modalities, image or video modalities, and speech language modalities respectively, output their respective hierarchical judgment results and uncertainties, and record interpretable basis; S6 Multiplicative Expert Integration and Cultural Context Gating: Based on the uncertainty of each expert, the consistency measure between each modality and the consistency score, and the gating coefficient of the cultural context embedding output, the weight of each modality is determined, missing modalities are automatically ignored, multiplicative multimodal integration is completed, hierarchical conclusions and deviation or anomaly scores are generated, and key evidence and weight sources are output. S7 Suggestion Generation and Distribution: Based on the hierarchical conclusions, the deviation or anomaly scores, and the compliance history, an automatic suggestion set containing interactive training tasks, learning tasks, or service process items is generated, and the difficulty is adaptively adjusted according to the performance on the client side; the suggestion set is presented in a contextualized form through the client-side interaction engine. S8 Continuous Monitoring and Local Anomaly Alerts: Records task completion rate, evaluation results, and compliance curves on the device at a set period; when the consistency score of the continuous evaluation within a preset time window decreases by more than a first threshold, or the deviation or anomaly score increases by more than a second threshold, a local anomaly alert is triggered, and an event notification is pushed to the management terminal and authorized receiving terminal. S9 Synchronization and Consistency Merging: When the network is unstable or interrupted, the data and event logs collected on the edge are written to the local encrypted cache; when the network recovers, idempotent commit and conflict merging are performed according to the version identifier, and the cloud completes deduplication, verification and write-back to ensure the continuity of the evaluation loop. The linkage between S3 to S6 is as follows: the cost of correction caused by difficulty in understanding reduces the contribution of language-related items to the stratification and increases the weight of non-linguistic modalities. The consistency score serves as the anchor quantitative constraint for the weight allocation of multimodal integration. The cultural context gating is used to simultaneously adjust the question type selection and modal weights, thereby reducing misjudgments and omissions across regions, languages, and low-education populations, and maintaining consistent decision-making logic on the edge and in the cloud.

2. The contextualized adaptive interaction evaluation method as described in claim 1, characterized in that, The external objective calibration consistency calibration adopts a monotonic mapping, and the optimization objective is to minimize the consistency error between the scale or task score and the external objective calibration index. The mapping parameters are obtained by supervised training with regularization. The monotonic mapping maintains a stable ranking in different language and education groups. The obtained consistency score is used for anomaly threshold determination and prompt triggering. The external objective calibration index includes laboratory indicators or biomarkers.

3. The contextualized adaptive interaction evaluation method as described in claim 1, characterized in that, The triggering conditions for the comprehension difficulty discrimination and substitution question state machine include at least two of the following: the answering time exceeds a threshold, the number of prompts exceeds a threshold, and consecutive errors of the same type of question satisfy a threshold. The substitution question state machine is executed in the order of rewriting prompts, downgrading substitution questions, and changing modalities. The correction cost is included in the scale subdomain score correction and used to reduce the weight of language modality in expert integration.

4. The contextualized adaptive interaction evaluation method as described in claim 1, characterized in that, The programmable expert for behavioral trajectory modality uses trajectory heatmaps, pause and hesitation ratios, path complexity, and micro-motion rhythms as features to output hierarchical judgment results and uncertainties. When the confusion level increases, the weight of the behavioral trajectory modality is increased, making the hierarchical results robust to language comprehension biases.

5. The contextualized adaptive interaction evaluation method as described in claim 1, characterized in that, The cultural context gating takes as input an embedding consisting of the regional information, the main language used, the years of education, and the time series of confusion levels. It is also used for the selection of question types and presentation methods, as well as the weight adjustment of multimodal integration, thereby establishing a causal closed loop between the interactive evaluation process and the fusion decision.

6. The contextualized adaptive interaction evaluation method as described in claim 1, characterized in that, The weights of the multiplicative expert integration are jointly determined by the uncertainty, the consistency measure between the consistency score and the cultural context gating coefficient, and automatically ignore the output of missing modalities when there are missing modalities, maintaining consistent integration rules between the end side and the cloud, and displaying the weights and interpretable basis of each modality on the management end.

7. The contextualized adaptive interaction evaluation method as described in claim 1, characterized in that, The edge interaction engine adopts a three-branch sparse attention structure, including a compressed retrieval branch for historical context, a global branch for selecting continuous blocks based on importance, and a local branch with a fixed window. The results of the three branches are fused on the edge in a gating manner for confusion discrimination, alternative question triggering, and prompt generation under long-term, multi-turn interactions. It is also combined with fixed-point quantization and intra-group shared indexes to reduce edge latency and storage consumption.

8. The contextualized adaptive interaction evaluation method as described in claim 1, characterized in that, The synchronization and consistency merging includes local encrypted caching, version identification, idempotent commit, and conflict merging. The conflict merging follows the time priority principle and the latest version priority principle. After deduplication, integrity verification and write-back are performed. The device side performs feature encoding on the voice response, image or video data, and behavior trajectory data. The feature encoding result and encrypted digest are written into the local encrypted cache and used for synchronization. The original data is deleted or encrypted and stored on the device side according to a preset retention strategy.

9. The contextualized adaptive interaction evaluation method as described in claim 1, characterized in that, The deviation or anomaly score is determined based on a weighted combination of at least the following: the difference in score from the individual baseline, the correction cost, the uncertainty of each expert, and the consistency measure.

10. A contextualized adaptive interactive evaluation system, characterized in that, The system includes an edge-side interactive terminal, a cloud service, and a data management layer. The edge-side interactive terminal is equipped with a contextualized presentation module, a confusion discrimination and substitution question state machine module, a behavior trajectory collection and encoding module, a three-branch sparse attention reasoning module, and a local anomaly prompting module. The cloud service is equipped with an external objective calibration consistency calibration module, a programmable expert management and integration module, a cultural context gating module, a suggestion generation and task orchestration engine, and a continuous monitoring and anomaly prompting module. The data management layer is equipped with encrypted storage, version identification, idempotent submission, and conflict merging mechanisms. The system is configured to execute the method described in any one of claims 1 to 9 and output hierarchical conclusions, deviation or anomaly scores, and interpretable evidence. Under network instability conditions, it maintains closed-loop management through edge-side local anomaly prompts and data caching.

Images