A learner cognitive state determination method and system based on human-computer collaborative interaction process data

Abstract

This invention belongs to the field of artificial intelligence and educational data mining technology, and provides a method and system for determining learner cognitive state based on human-computer collaborative interaction process data. By collecting front-end interaction timestamps, text, reading pauses, and copy-paste data, response latency, reading rate, word frequency density, semantic similarity, and text similarity are extracted. After being aggregated into a three-dimensional temporal feature tensor through sliding window slicing, it is input into a long short-term memory neural network for decoding, which determines whether the learner is in a shallow cognitive outsourcing state corresponding to content reuse or a deep active construction state corresponding to content reconstruction, thus achieving real-time and accurate assessment of the learner's cognitive state.

Description

Technical Field

[0001] This application relates to the fields of artificial intelligence and educational data mining technology, and more specifically, to a method and system for determining learner cognitive state based on human-computer collaborative interaction process data. Background Technology

[0002] With the widespread application of generative artificial intelligence technology, human-machine collaborative learning (such as AI-assisted writing and AI-assisted programming) has become the main form of intelligent education. In this specific scenario, intelligent systems improve the efficiency of learning tasks by generating content in real time, but this also raises the hidden danger of cognitive outsourcing for learners—learners relinquish higher-order thinking processes to AI, resulting in low cognitive load and mechanically copying generated content rather than actively reconstructing knowledge.

[0003] To address the aforementioned issues, existing technologies face the following challenges: First, current assessment methods heavily rely on static outcome data (such as post-task tests, subjective questionnaires, and the quality of the final work), which is lagging and prone to creating the illusion of high scores but low ability. They also fail to provide real-time status monitoring during the learning process, missing optimal opportunities for instructional intervention. Second, current process data mining in human-computer collaborative learning primarily relies on macroscopic, surface-level behavioral indicators (such as total page dwell time and total click frequency), resulting in significant loss of underlying, subtle features such as millisecond-level interaction rhythms and reading speeds. This makes it impossible to distinguish between deep task-oriented thinking and cognitive overload or operational stagnation. Third, existing technologies lack methods to quantify cognitive depth from a semantic perspective. The static text of the final product alone cannot reconstruct the dynamic trajectory of learners' thought processes and knowledge reconstruction, and cannot effectively distinguish between superficial vocabulary substitution in AI content and deep reconstruction through multiple rounds of questioning and exploration.

[0004] To address the aforementioned issues, existing technologies urgently need improvement. Summary of the Invention

[0005] The purpose of this application is to provide a method and system for determining learner cognitive state based on human-computer collaborative interaction process data. It has the advantages of being able to achieve non-intrusive and accurate assessment of learners' true cognitive state, effectively distinguishing between content reuse behavior and content reconstruction behavior, and improving the real-time performance and accuracy of cognitive engagement state recognition.

[0006] Firstly, this application provides a method for determining learner cognitive states based on human-computer collaborative interaction process data, including: At the human-computer collaborative interaction front end, timestamp data, text content data, learner reading dwell time, and copy-paste operation data of multiple rounds of interaction are collected. Based on the timestamp data and reading dwell time, the response delay time and reading time index are determined. Based on the text content data, the word frequency density of preset feature words and the semantic similarity of input text in adjacent interaction rounds are extracted. Based on the copy and paste operation data and text content data, the copy and paste frequency, fragment length and text similarity between the final input text and the AI-generated text are extracted. After aligning the response delay time, reading time index, word frequency density, semantic similarity, copy and paste frequency, fragment length, and text similarity according to the timestamp, a sliding time window is used to perform slice aggregation to generate a three-dimensional temporal feature tensor. The three-dimensional temporal feature tensor is input into a long short-term memory neural network for processing, and the learner's current cognitive state is determined to be either a content reuse state or a content reconstruction state based on the output probability distribution.

[0007] Furthermore, the collection of learners' reading dwell time includes: Monitor the proportion of the visible area of ​​AI-generated content in the browser viewport. When the proportion of the visible area exceeds a preset proportion threshold, it is determined that the content is visible in the viewport. Listen for page switching events and listen for mouse movement, scroll wheel or keyboard input events at preset time intervals. When the mouse, scroll wheel or keyboard input events are detected within a preset time threshold, the system is determined to be in an interactive active state. The reading dwell time is accumulated when both the viewport visibility state and the interactive active state are satisfied; the timer is paused when either the viewport visibility state or the interactive active state is not satisfied.

[0008] Furthermore, determining the response latency and reading time index based on the timestamp data and reading dwell time includes: The response delay time is obtained by subtracting the start timestamp of the learner's input from the end timestamp of the AI-generated content in the corresponding round. Divide the reading dwell time by the number of words in the AI-generated content of the corresponding round to obtain the reading time per unit word, which is used as the reading time index.

[0009] Furthermore, before dividing the reading dwell time by the number of words in the AI-generated content of the corresponding round, the method further includes: The reading dwell time is cleaned of outliers based on Z-score statistics to remove outliers that deviate from the normal reading time range. The cleaned reading dwell time is then used to calculate the reading time index.

[0010] Furthermore, the step of extracting the word frequency density of preset feature words and the semantic similarity of input text in adjacent interaction rounds based on the text content data includes: The learner's input text is segmented into words, and the ratio of the number of words that match the preset feature word library to the total number of words is used as the word frequency density. The pre-trained sentence encoding model is used to map the input text of the nth round and the input text of the (n+1)th round into high-dimensional sentence vectors, and the cosine similarity between the two sentence vectors is calculated as the semantic similarity.

[0011] Furthermore, the step of extracting the text similarity between the final input text and the AI-generated text based on the text content data includes: The literal overlap rate between the final input text and the AI-generated text is calculated using the edit distance algorithm or the ROUGE algorithm, and the cosine similarity of the sentence vectors between the final input text and the AI-generated text is calculated using the pre-trained sentence encoding model. The literal overlap rate and the sentence vector cosine similarity are used as the two components of the text similarity.

[0012] Furthermore, the step of using a sliding time window to perform slice aggregation and generate a three-dimensional temporal feature tensor includes: Set the window length of the sliding time window to be greater than the sliding step size to create a time overlap area between adjacent time slices; The entire human-computer collaborative learning process is divided into multiple time slices according to the sliding step size. Within each time slice, the response delay time, reading time index, word frequency density, semantic similarity, and text similarity are statistically aggregated to generate a three-dimensional temporal feature tensor with three dimensions: number of samples, time step T, and feature dimension.

[0013] Furthermore, the step of determining whether the learner's current cognitive state is a content reuse state or a content reconstruction state based on the probability distribution of the output includes: The long short-term memory neural network extracts the fluctuation trajectory and cumulative effect of the three-dimensional temporal feature tensor on the time axis through the forget gate, input gate and output gate, and outputs the probability value of the learner being in the content reconstruction state through the sigmoid activation function; When the probability value is greater than or equal to a set threshold, it is determined to be in a content reconstruction state; otherwise, it is determined to be in a content reuse state.

[0014] Secondly, this application also proposes a learner cognitive state determination system based on human-computer collaborative interaction process data, including: The data acquisition module is used to collect timestamp data, text content data, learner reading dwell time, and copy-paste operation data from multiple rounds of interaction. The feature extraction module is used to determine the response delay time and reading time index based on the timestamp data and reading dwell time, extract the word frequency density of preset feature words and the semantic similarity of the input text in adjacent interaction rounds based on the text content data, and extract the copy and paste frequency, fragment length and text similarity between the final input text and the AI-generated text based on the copy and paste operation data and text content data. The generation module is used to align the response delay time, reading time index, word frequency density, semantic similarity, copy and paste frequency, fragment length and text similarity according to the timestamp, and then use a sliding time window to perform slice aggregation to generate a three-dimensional temporal feature tensor with three dimensions: number of samples, time step T and feature dimension. The state determination module is used to input the three-dimensional temporal feature tensor into the long short-term memory neural network for processing, and determine the learner's current cognitive state as either a content reuse state or a content reconstruction state based on the output probability distribution.

[0015] Thirdly, this application also proposes an electronic device comprising: one or more processors, and a memory for storing one or more computer programs; the computer programs are configured to be executed by the one or more processors, and the programs include steps for performing the learner cognitive state determination method based on human-computer collaborative interaction process data as described in the first aspect.

[0016] As can be seen from the above, the learner cognitive state determination method and system provided in this application based on human-computer collaborative interaction process data solves the problems of traditional assessment methods relying on static results, resulting in strong lag, loss of subtle features in macroscopic surface behavioral indicators, and lack of semantic-level cognitive depth quantification by constructing a front-end high-fidelity micro-behavioral log collection mechanism, an automated extraction pipeline for multi-dimensional behavioral features, and a long short-term memory neural network temporal decoding model based on overlapping sliding time windows. It has the advantages of being able to achieve non-intrusive real-time assessment of learners' true cognitive state, accurately distinguishing between content reuse and content reconstruction behaviors, and improving the accuracy of cognitive engagement state identification and the timeliness of teaching intervention. Attached Figure Description

[0017] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation on the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a flowchart illustrating the steps of the learner cognitive state determination method based on human-computer collaborative interaction process data disclosed in an embodiment of the present invention. Figure 2 This is a flowchart illustrating the steps for collecting learner reading dwell time as disclosed in an embodiment of the present invention; Figure 3 This is a schematic diagram of the learner cognitive state determination system based on human-computer collaborative interaction process data disclosed in an embodiment of the present invention. Detailed Implementation

[0019] 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 these embodiments belong; the terminology used herein and in the specification of the application is for the purpose of describing particular embodiments only and is not intended to limit these embodiments; the terms "comprising" and "having," and any variations thereof, in the specification of these embodiments and the foregoing drawings, are intended to cover non-exclusive inclusion. The terms "first," "second," etc., in the specification of these embodiments and the foregoing drawings are used to distinguish different objects, not to describe a particular order.

[0020] The implementation details of the technical solution in this embodiment are described in detail below: This application proposes a method for determining learner cognitive states based on human-computer collaborative interaction process data, such as... Figure 1 As shown, the method includes: S101 collects timestamp data, text content data, learner reading dwell time, and copy-paste operation data from multiple rounds of interaction at the human-computer collaborative interaction front end.

[0021] In this embodiment, at the human-computer collaborative interaction front end, the system first establishes a seamless and high-fidelity silent data pipeline to accurately capture micro-behavioral sequences reflecting the learner's cognitive input and processing depth. For macro-level interaction metadata, the system uniformly uses the browser's underlying `performance.now()` API to obtain sub-millisecond-level high-precision timestamps. On the AI ​​side, the system is deeply integrated into the streaming interface for communication between the front end and the large model. Upon receiving the first data packet, it records the AI ​​generation start timestamp; upon receiving the end marker, it records the AI ​​generation end timestamp, and simultaneously extracts the complete payload text stream for word count. On the learner side, the system uses bidirectional data binding in the front-end input box to trigger the learner's input timing using focus or keyboard events, and intercepts submission events to record the final submission timestamp and complete text content. Finally, these four types of time points and text entities are encapsulated into structured JSON data for a single interaction cycle. In addition, the system also establishes a silent clipboard listening mechanism in the text editing area of ​​the interactive interface to record the learner's copy and paste actions. The system intercepts the underlying paste action using paste event modifiers and forcibly extracts plain text content through native objects, thereby stripping away potential external rich text styles and ensuring the purity of the collected corpus. Whenever a paste action is triggered, the system's global paste frequency state immediately increments. Simultaneously, it uses a high-precision API to obtain an absolute timestamp and combines it with the text's length attribute to construct a snapshot of behavioral data containing time, length, and specific content, which is then pushed into a local reactive upload queue. To ensure front-end rendering performance and a non-blocking interactive experience for learners, the collected micro-process data is not reported in real-time using high-frequency network requests. By combining component uninstallation lifecycle hooks or page visibility changes, the system calls the `navigator.sendBeacon()` interface just before the user leaves the page to asynchronously and seamlessly report a structured JSON sequence to the server, providing underlying native material for subsequent natural language processing pipelines and deep feature extraction.

[0022] Furthermore, in S101, the collection of the learner's reading dwell time, such as... Figure 2 As shown, it includes: S1001, Monitor the proportion of the visible area of ​​AI-generated content in the browser viewport. When the proportion of the visible area exceeds a preset proportion threshold, determine that the content is visible in the viewport. In this embodiment, effective reading time is captured through a custom Vue directive, which encapsulates the IntersectionObserver API in its lifecycle hooks. For dynamically rendered AI text sub-components, an intersection threshold is set. When a node is mounted and more than 50% of its visible area is exposed, the directive immediately triggers a callback, updating the component's reactive state, declaratively binding the timing trigger point, and monitoring the proportion of the AI-generated content visible in the browser viewport. When this proportion exceeds a preset threshold, the content is determined to be visible in the viewport.

[0023] S1002, listen to page switching events, and listen to mouse movement, scroll wheel or keyboard input events at preset time intervals. When the mouse, scroll wheel or keyboard input events are detected within the preset time threshold, it is determined that the system is in an interactive active state. In this embodiment, the system utilizes a front-end state management library to construct an independent heartbeat state machine. In the `onMounted` hook of the application root component, screen-switching listeners based on the Page Visibility API and micro-interaction event listeners for the global Window object are centrally mounted. These micro-interaction events include mouse movement, scroll wheel, and keyboard key events, and their frequency is strictly limited according to preset time intervals to ensure system performance. These events update the global active flag state in the heartbeat state machine in real time and are automatically unbound when the component is unmounted to prevent memory leaks. Thus, the system listens for page-switching events and monitors mouse movement, scroll wheel, or keyboard input events at preset time intervals. When a mouse, scroll wheel, or keyboard input event is detected within a preset time threshold, the system is determined to be in an interactively active state. S1003, the reading dwell time is accumulated when and only when the viewport visible state and the interactive active state are simultaneously satisfied; the timer is paused when either the viewport visible state or the interactive active state is not satisfied.

[0024] In this embodiment, an effective timing logic gate is constructed within the AI ​​text component. Vue's listener dynamically tracks the intersection of the local visibility state and the active states in the global heartbeat state machine. The millisecond-level timer within the component accumulates to obtain the reading dwell time only when both the viewport visibility state and the interactive active state are simultaneously satisfied. Once any reactive state becomes false (indicating screen switching, hanging up, or moving out of the viewport), the listener immediately suspends the timing. The logic for accumulating the effective reading time can be represented by the following Boolean condition: Only if and only if Condition valid When true, the current time step Δt is included in the total effective reading time T of the AI ​​text. reading In the middle. Among them, Conditionvalid The condition is true if the component is visible in the viewport and has received a global action response within the past 5 seconds. This logic gate mechanism largely isolates environmental noise, ensuring extremely high purity of the data collection time. The data will ultimately be reported non-blockingly and asynchronously by calling navigator.sendBeacon() before the component triggers its onBeforeUnmount destruction action, or when intercepted by Vue Router's global route guard.

[0025] S102, based on the timestamp data and reading dwell time, determine the response delay time and reading time index; based on the text content data, extract the word frequency density of preset feature words and the semantic similarity of input text in adjacent interaction rounds; based on the copy and paste operation data and text content data, extract the copy and paste frequency, fragment length, and text similarity between the final input text and the AI-generated text.

[0026] Furthermore, in S102, determining the response delay time and reading time index based on the timestamp data and reading dwell time includes: subtracting the generation end timetamp of the AI-generated content in the corresponding round from the input start timestamp of the learner's input content to obtain the response delay time; dividing the reading dwell time by the number of words in the text of the AI-generated content in the corresponding round to obtain the reading time index per unit word.

[0027] In this embodiment, the backend system receives and parses the JSON serialized time-series data from the frontend, performing deep time-series processing and feature quantization. The system reconstructs the complete lifecycle of a single interaction on the timeline and calculates the time difference between adjacent action nodes. The system primarily utilizes the time-series processing capabilities of the Pandas library, calling the `shift` and `diff` methods to calculate the difference values ​​of key time nodes within an interaction cycle. During this process, the core feature metric extracted by the system is the learner response latency, calculated using the following formula:

[0028] By calculating the absolute timestamp of when the learner started input (Timestamp) student_start (and the absolute timestamp of the end of the previous AI generation) ai_complete The difference between the two values ​​represents the implicit cognitive load period during which learners process information from AI, perform internal cognitive processing, schema assimilation, and conceive the next exploration strategy.

[0029] Subsequently, the system further extracts the word count attribute of the AI-generated text entities and, combined with the effective reading time, calculates a standardized dynamic reading time index. The system leverages the underlying vectorized computation capabilities of the NumPy library to perform efficient batch calculations of this feature; the specific calculation formula is as follows:

[0030] Among them, T reading Represents effective reading time, converted to a uniform number of seconds, Count. ai_words This represents the total number of words in the AI-generated text within the same data row. By calculating this reading time index and its dynamic fluctuation trajectory across multiple rounds of interaction, the system can objectively and instantly reveal the direct impact of the redundancy or complexity of the generated information on the learner's processing pace at the behavioral level, thus serving as a key feature for quantifying their cognitive engagement.

[0031] Furthermore, in S102, the extraction of copy-paste frequency, fragment length, and text similarity between the final input text and the AI-generated text based on copy-paste operation data and text content data includes: the system summarizing the paste logs reported from the front end, aggregating and grouping the copy-paste operation data, counting the frequency of paste actions, and extracting the fragment length from the logs as a quantitative feature of the copy operation behavior. It should be noted that this feature is not used alone as a criterion for determining the 'intensity of direct adoption behavior,' but is jointly analyzed with the subsequently calculated text similarity between the final input text and the AI-generated text (including literal overlap rate and sentence vector cosine similarity): if the pasted fragment length is long and the text similarity is high, it indicates that the learner may be engaging in content reuse behavior; if the text similarity is low due to significant modifications after pasting, it may indicate content reconstruction behavior. Therefore, the copy-paste behavior feature and the text similarity feature together constitute the basis for determining the learner's cognitive state, rather than a single feature directly representing the intensity of adoption.

[0032] Further, in S102, the extraction of text similarity between the final input text and the AI-generated text includes: calculating the literal overlap rate between the final input text and the AI-generated text using an edit distance algorithm or a ROUGE algorithm; calculating the sentence vector cosine similarity between the final input text and the AI-generated text using the pre-trained sentence encoding model; and using the literal overlap rate and the sentence vector cosine similarity as two components of the text similarity.

[0033] In this embodiment, the system extracts text adoption and knowledge reconstruction features to reveal the learner's direct adoption and deep internalization of AI-generated knowledge. The system uses AI-generated text as the baseline text and the learner's final submitted text as the target text for comparison. The comparison not only measures the superficial literal overlap rate but also delves into the semantic similarity between the two. The system uses the edit distance algorithm or the ROUGE algorithm in a Python environment to compare the overlap rate between the AI-generated text and the learner's final submitted text. The logic for calculating the superficial literal overlap rate using standardized edit distance can be expressed as follows:

[0034] Among them, T ai For the original text generated by AI, T final The system ultimately submits the text to the learner. Beyond quantifying the degree of literal appropriation, it further utilizes a deep learning model to map the two texts into high-dimensional dense sentence vectors and uses cosine similarity to calculate the overlap of their deep semantics, thereby accurately distinguishing whether the learner has mechanically copied or deeply reconstructed the content. The literal overlap rate and the cosine similarity of the sentence vectors are used as two components of text similarity to quantify the learner's direct adoption and deep reconstruction of the AI-generated content.

[0035] Furthermore, before dividing the reading dwell time by the number of words in the AI-generated content of the corresponding round, the method further includes: performing outlier cleaning on the reading dwell time based on Z-score statistics, removing outliers that deviate from the normal reading time range, and using the cleaned reading dwell time to calculate the reading time consumption index.

[0036] Specifically, in this embodiment, the system uses Python's Pandas library to read log content and construct a multidimensional data frame. For the reading time field, a Z-score-based statistical algorithm is used for outlier detection. This algorithm accurately identifies and smooths or directly removes noisy data that significantly deviates from the normal physiological limits of human reading. The cleaned reading dwell time is used to calculate the reading time index, thereby ensuring the statistical validity and reliability of subsequent feature calculations. For example, when the AI-generated text contains 100 characters, the time required for a normal human to read this content is typically within a reasonable range of 10 to 30 seconds, corresponding to a reading time index of approximately 0.10 to 0.30 seconds per character. If a log record shows a learner's reading dwell time of 300 seconds, the Z-score calculation shows that this value significantly deviates from the mean, with a standardized score absolute value greater than 3. The system determines this data as an outlier caused by extreme idle time or system underreporting and removes it. Conversely, if a record shows a reading dwell time of only 0.5 seconds, it is also identified as an outlier deviating from the normal physiological limits. After this cleaning, the remaining valid data is used to calculate the reading time index, ensuring that the feature values ​​truly reflect the learner's cognitive processing pace.

[0037] Further, in S102, the step of extracting the word frequency density of preset feature words and the semantic similarity of input texts in adjacent interaction rounds based on the text content data includes: performing word segmentation on the learner's input text, counting the ratio of the number of words matching the preset feature word library to the total number of words, and using it as the word frequency density; using a pre-trained sentence encoding model to map the input text of the nth round and the input text of the (n+1)th round into high-dimensional sentence vectors respectively, and calculating the cosine similarity of the two sentence vectors as the semantic similarity.

[0038] Specifically, in this embodiment, the backend natural language processing pipeline receives the cleaned interactive text sequence from the frontend and extracts the learner's plain text input log. First, the system calls Python's jieba natural language processing library for accurate word segmentation and part-of-speech tagging. Combining a pre-built dictionary of verbs in the educational cognitive domain, the system automatically calculates the word frequency density of specific cognitive verbs in the text using regular expression matching and word frequency statistics algorithms, thereby quantifying the learner's surface-level expressive complexity. The formula for calculating the word frequency density can be expressed as:

[0039] Where, Count cognitive_verbs Total represents the number of cognitive verbs contained in the input text of this round. words This represents the total number of words in the text for this round.

[0040] Secondly, the system aims to reconstruct the evolutionary trajectory of learners' thinking. The system leverages a pre-trained Chinese large language model within a deep learning framework, mapping the interactive texts of adjacent rounds into high-dimensional, dense semantic vectors of a fixed dimension. Subsequently, it calls the scikit-learn library to calculate the cosine similarity between the two sets of vectors in multidimensional space, thus characterizing the divergence and convergence trajectory of learners' questions at the semantic level. The core formula for calculating semantic similarity is as follows:

[0041] Among them, V n With V n+1 These represent the sentence vectors of the input text in rounds n and n+1, respectively. The continuous fluctuation line of similarity values ​​will serve as the core high-order feature for quantifying students' coherence of thought and focus of inquiry. Simultaneously, the system constructs an intent encoding table based on Bloom's taxonomy of cognitive objectives, and uses support vector machines or prompting engineering to call the large language model interface to automatically classify each learner's input instruction. Using zero-shot or few-shot methods, the interactive text is labeled with cognitive level tags indicating memorization, understanding, application, analysis, evaluation, or creation, and these tags are appended as enumerated feature metadata to the JSON feature matrix of this round of interaction.

[0042] S103, after aligning the response delay time, reading time index, word frequency density, semantic similarity, copy and paste frequency, fragment length and text similarity according to the timestamp, the sliding time window is used to perform slice aggregation to generate a three-dimensional temporal feature tensor.

[0043] In this embodiment, after automatically extracting micro-behavioral and semantic features, the system proceeds to the behavioral sequence feature overlapping sliding window slicing step. First, strict spatiotemporal alignment is performed on the discrete, multi-source heterogeneous data. The multi-dimensional feature vectors extracted in the previous steps, including time rhythm quantification features, semantic evolution features, copy-paste behavior features, and text adoption features, are uniformly mapped onto the same global time axis according to the absolute timestamp of the action trigger. This alignment eliminates temporal misalignment between heterogeneous data, ensuring accurate correspondence of multimodal features on the physical time scale. Subsequently, an overlapping sliding time window mechanism with set window length and step size is used to resample and slice the data stream of the entire human-machine collaborative learning process. Response latency, reading time index, word frequency density, semantic similarity, copy-paste frequency, segment length, and text similarity are transformed into high-dimensional feature tensors containing time dependencies.

[0044] Furthermore, in S103, the step of using a sliding time window to aggregate slices and generate a three-dimensional temporal feature tensor includes: setting the window length of the sliding time window to be greater than the sliding step size, so that there is a time overlap area between adjacent time slices; dividing the entire human-computer collaborative learning process into multiple time slices according to the sliding step size; statistically aggregating the response delay time, reading time index, word frequency density, semantic similarity, and text similarity in each time slice to generate a three-dimensional temporal feature tensor with three dimensions: number of samples, time step T, and feature dimension.

[0045] Specifically, in this embodiment, the window length of the sliding time window is set to be greater than the sliding step size, so that there is a time overlap region between adjacent time slices. This design can effectively avoid the truncation of boundary behavior information caused by non-overlapping physical slices, ensuring the continuity of the thought evolution sequence. Let the total interaction time of human-computer collaborative learning be L. total The inherent duration of the sliding window is W, and the step size of the window sliding forward each time is S. Since S is less than W, the formula for calculating the total number of continuous time steps T is as follows:

[0046] Taking an implementation scenario with a total duration of 1200 seconds, a window size of 120 seconds, and a step size of 30 seconds as an example, the system can calculate and generate T=37 consecutive feature slices with temporal overlap. The entire human-machine collaborative learning process is divided into T time slices according to the sliding step size. Within each time slice, the response latency, reading time index, word frequency density, semantic similarity, and text similarity are statistically aggregated, making each window a fixed-length multi-dimensional feature snapshot. The system calls the underlying memory view operation interface to complete the data serialization conversion with extremely low memory overhead, ultimately constructing a three-dimensional temporal feature tensor of shape (Samples, T, Features_Dim). This high-order tensor completely encapsulates the fluctuation trajectory and cumulative effect of the behavioral sequence during human-machine collaborative learning, where Samples represents the number of analysis objects, T represents the serialized time step size, and Features_Dim represents the number of feature dimensions contained in each feature snapshot, serving as the standard input format for the next stage of temporal dynamic model training.

[0047] At the technical implementation level, the system calls low-level memory view operation APIs in the NumPy library (such as numpy.lib.stride_tricks.as_strided) to complete the data serialization conversion with extremely low memory overhead. Ultimately, the system transforms the originally discrete planar behavioral features into a high-order input feature tensor X containing explicit temporal dependencies. The mathematical space representation of this tensor is as follows:

[0048] Where N represents the number of analysis objects (learner samples), T represents the time step after serialization (37 in this case), and D represents the number of feature dimensions contained in each feature snapshot. This three-dimensional high-order tensor fully encapsulates the fluctuation trajectory and cumulative effect of the behavioral sequence during human-machine collaborative learning, and will serve as the standard input format for the next stage of temporal dynamic model training.

[0049] S104, the three-dimensional temporal feature tensor is input into the long short-term memory neural network for processing, and the learner's current cognitive state is determined to be either a content reuse state or a content reconstruction state based on the output probability distribution.

[0050] Furthermore, in S104, the step of determining whether the learner's current cognitive state is a content reuse state or a content reconstruction state based on the output probability distribution includes: the long short-term memory neural network extracts the fluctuation trajectory and cumulative effect of the three-dimensional temporal feature tensor on the time axis through the forget gate, input gate and output gate, and outputs the probability value of the learner being in the content reconstruction state through the Sigmoid activation function; when the probability value is greater than or equal to a set threshold, it is determined to be in the content reconstruction state, otherwise it is determined to be in the content reuse state.

[0051] Specifically, in this embodiment, after completing the overlapping sliding window slicing of behavioral sequence features, the system enters the cognitive state decoding and output stage based on temporal mapping. This stage aims to construct a deep temporal neural network, specifically employing a long short-term memory neural network to receive the high-dimensional feature tensor generated in the aforementioned steps as standard input. This model, with its internal memory mechanism, can effectively overcome the gradient vanishing or gradient exploding problems in long-sequence training, making it highly suitable for handling continuous multimodal behavioral nodes generated throughout the human-machine collaborative learning process. This establishes a nonlinear mapping path from microscopic feature sequences to macroscopic cognitive states, ultimately determining the learner's current cognitive state as either a content reuse state or a content reconstruction state based on the output probability distribution.

[0052] Furthermore, in terms of the internal computation mechanism of the model, the system utilizes the gating units unique to long short-term memory neural networks to process slice features step by step. For each time step t, the network unit reads the multidimensional feature vector x of the current time window. t and the hidden state h of the previous time step t-1 The retention, forgetting, and updating of current state information are dynamically determined through the nonlinear transformation of the gating function. This process accurately captures the fluctuation trajectory, decay law, and cumulative effect of learners' behavioral features such as frequent copy-pasting in the early stage and deep semantic reconstruction in the later stage on the time axis. Its core hidden state update logic can be highly abstracted into the following calculation process:

[0053] Among them, c t-1 This refers to the memory state at time t-1 in the previous step. After iterative evolution through all T time steps, the model extracts the final hidden layer state representation h that incorporates global interaction context temporal information. t , which serves as the top-level high-order feature of the learner's cognitive engagement state during decoding.

[0054] Subsequently, the system extracts the top-level temporal feature representation h. t The input is fed into the output layer of a hierarchical fully connected network with a specific activation function for automated decoding of cognitive engagement states. For the two core binary labels defined in this system—content reconstruction and content reuse—the output layer uses the Sigmoid activation function to smoothly compress the feature values ​​after linear mapping of the weight matrix into a probability space of 0 to 1, outputting the predicted probability distribution of the learner's content reconstruction state. The calculation formula for the output decoding is as follows:

[0055] Among them, W out Let b be the weight matrix of the fully connected layer. out This is the bias term. If the predicted probability... If the learner's current cognitive state is greater than or equal to the set classification threshold (0.5), the system decodes and determines it as a content reconstruction state; if it is less than the threshold, the system decodes and determines it as a content reuse state.

[0056] Furthermore, during the supervised learning and training optimization phase of the model, the system weighted and fused the independent transfer test scores obtained by the learners in the post-test with the expert scores of the final text, and then divided them into objective true labels y using a clustering algorithm. (1 represents active construction, 0 represents cognitive outsourcing). The system uses a binary cross-entropy loss function to quantify the model's prediction probability. The deviation between the actual cognitive state label y and the actual cognitive state label y. For a training batch containing N training samples, the loss function is calculated as follows:

[0057] The model relies on the backpropagation algorithm, combined with adaptive optimization algorithms such as Adam, to continuously minimize this loss function. This updates all weight matrices within the network. Once the model training converges and its generalization performance meets the target, the decoding system can perform high-precision forward computation and state monitoring of learners' real-time interactive tensors in actual teaching scenarios, providing data-driven decision-making support for precise teaching interventions.

[0058] Secondly, this embodiment also proposes a learner cognitive state determination system based on human-computer collaborative interaction process data, such as... Figure 3 As shown, it includes: Data acquisition module 301 is used to collect timestamp data, text content data, learner reading dwell time, and copy and paste operation data from multiple rounds of interaction. The feature extraction module 302 is used to determine the response delay time and reading time index based on the timestamp data and reading dwell time, extract the word frequency density of preset feature words and the semantic similarity of the input text in adjacent interaction rounds based on the text content data, and extract the copy and paste frequency, fragment length and text similarity between the final input text and the AI-generated text based on the copy and paste operation data and text content data. The generation module 303 is used to align the response delay time, reading time index, word frequency density, semantic similarity, copy and paste frequency, fragment length and text similarity according to the timestamp, and then use a sliding time window to perform slice aggregation to generate a three-dimensional temporal feature tensor with three dimensions: number of samples, time step T and feature dimension. The state determination module 304 is used to input the three-dimensional temporal feature tensor into the long short-term memory neural network for processing, and determine the learner's current cognitive state as either a content reuse state or a content reconstruction state based on the output probability distribution.

[0059] This system can be used to execute the learner cognitive state determination method based on human-computer collaborative interaction process data as described in the first aspect, which will not be elaborated here.

[0060] Thirdly, this application also proposes an electronic device comprising: one or more processors, and a memory for storing one or more computer programs; the computer programs are configured to be executed by the one or more processors, and the programs include steps for performing the learner cognitive state determination method based on human-computer collaborative interaction process data as described in the first aspect.

[0061] In this embodiment, the electronic device can be a server, desktop computer, laptop computer, or cloud virtual computing node, or any device with data processing and storage capabilities. The processor can be a central processing unit, graphics processing unit, or tensor processor, used to execute front-end data acquisition and control, back-end natural language processing, and deep time-series model training and inference tasks in parallel. The memory includes non-volatile readable storage media and random access memory, used to persistently store the original logs of human-computer collaborative interaction, multi-dimensional behavioral feature vectors extracted through feature engineering, weight parameters of the trained long short-term memory neural network, and a pre-set cognitive feature lexicon.

[0062] The communication interface is used to establish a data transmission channel between the human-computer collaborative interaction front-end and the back-end processing module. It receives structured interaction logs reported by the front-end browser through the beacon sending interface, including timestamp data, text content data, reading dwell time, and copy-paste operation data. After receiving the above data, the processor calls the computer program stored in the memory and executes the learner cognitive state determination method steps based on the human-computer collaborative interaction process data in the first aspect in sequence.

[0063] The above description is merely an embodiment of this application and is not intended to limit the scope of protection of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of protection of this application.

Claims

1. A method for determining learner cognitive state based on human-computer collaborative interaction process data, characterized in that, include: At the human-computer collaborative interaction front end, timestamp data, text content data, learner reading dwell time, and copy-paste operation data of multiple rounds of interaction are collected. Based on the timestamp data and reading dwell time, the response delay time and reading time index are determined. Based on the text content data, the word frequency density of preset feature words and the semantic similarity of input text in adjacent interaction rounds are extracted. Based on the copy and paste operation data and text content data, the copy and paste frequency, fragment length and text similarity between the final input text and the AI-generated text are extracted. After aligning the response delay time, reading time index, word frequency density, semantic similarity, copy and paste frequency, fragment length, and text similarity according to the timestamp, a sliding time window is used to perform slice aggregation to generate a three-dimensional temporal feature tensor. The three-dimensional temporal feature tensor is input into a long short-term memory neural network for processing, and the learner's current cognitive state is determined to be either a content reuse state or a content reconstruction state based on the output probability distribution.

2. The learner cognitive state determination method based on human-computer collaborative interaction process data according to claim 1, characterized in that, The collection of learners' reading dwell time includes: Monitor the proportion of the visible area of ​​AI-generated content in the browser viewport. When the proportion of the visible area exceeds a preset proportion threshold, it is determined that the content is visible in the viewport. Listen for page switching events and listen for mouse movement, scroll wheel or keyboard input events at preset time intervals. When the mouse, scroll wheel or keyboard input events are detected within a preset time threshold, the system is determined to be in an interactive active state. The reading dwell time is accumulated when both the viewport visibility state and the interactive active state are satisfied; the timer is paused when either the viewport visibility state or the interactive active state is not satisfied.

3. The learner cognitive state determination method based on human-computer collaborative interaction process data according to claim 1, characterized in that, The process of determining the response latency and reading time index based on the timestamp data and reading dwell time includes: The response delay time is obtained by subtracting the start timestamp of the learner's input from the end timestamp of the AI-generated content in the corresponding round. Divide the reading dwell time by the number of words in the AI-generated content of the corresponding round to obtain the reading time per unit word, which is used as the reading time index.

4. The learner cognitive state determination method based on human-computer collaborative interaction process data according to claim 3, characterized in that, Before dividing the reading dwell time by the number of words in the AI-generated content of the corresponding round, the method further includes: The reading dwell time is cleaned of outliers based on Z-score statistics to remove outliers that deviate from the normal reading time range. The cleaned reading dwell time is then used to calculate the reading time index.

5. The learner cognitive state determination method based on human-computer collaborative interaction process data according to claim 1, characterized in that, The step of extracting the word frequency density of preset feature words and the semantic similarity of input text in adjacent interaction rounds based on the text content data includes: The learner's input text is segmented into words, and the ratio of the number of words that match the preset feature word library to the total number of words is used as the word frequency density. The pre-trained sentence encoding model is used to map the input text of the nth round and the input text of the (n+1)th round into high-dimensional sentence vectors, and the cosine similarity between the two sentence vectors is calculated as the semantic similarity.

6. The learner cognitive state determination method based on human-computer collaborative interaction process data according to claim 5, characterized in that, The text similarity between the final input text and the AI-generated text includes: calculating the literal overlap rate between the final input text and the AI-generated text using the edit distance algorithm or the ROUGE algorithm, and calculating the sentence vector cosine similarity between the final input text and the AI-generated text using the pre-trained sentence encoding model. The literal overlap rate and the sentence vector cosine similarity are used as the two components of the text similarity.

7. The learner cognitive state determination method based on human-computer collaborative interaction process data according to claim 1, characterized in that, The step of using a sliding time window to perform slice aggregation and generate a three-dimensional temporal feature tensor includes: Set the window length of the sliding time window to be greater than the sliding step size to create a time overlap area between adjacent time slices; The entire human-computer collaborative learning process is divided into multiple time slices according to the sliding step size. Within each time slice, the response delay time, reading time index, word frequency density, semantic similarity, and text similarity are statistically aggregated to generate a three-dimensional temporal feature tensor with three dimensions: number of samples, time step T, and feature dimension.

8. The learner cognitive state determination method based on human-computer collaborative interaction process data according to claim 1, characterized in that, The determination of whether a learner's current cognitive state is a content reuse state or a content reconstruction state based on the probability distribution of the output includes: The long short-term memory neural network extracts the fluctuation trajectory and cumulative effect of the three-dimensional temporal feature tensor on the time axis through the forget gate, input gate and output gate, and outputs the probability value of the learner being in the content reconstruction state through the sigmoid activation function; When the probability value is greater than or equal to a set threshold, it is determined to be in a content reconstruction state; otherwise, it is determined to be in a content reuse state.

9. A learner cognitive state determination system based on human-computer collaborative interaction process data, characterized in that, include: The data acquisition module is used to collect timestamp data, text content data, learner reading dwell time, and copy-paste operation data from multiple rounds of interaction. The feature extraction module is used to determine the response delay time and reading time index based on the timestamp data and reading dwell time, extract the word frequency density of preset feature words and the semantic similarity of the input text in adjacent interaction rounds based on the text content data, and extract the copy and paste frequency, fragment length and text similarity between the final input text and the AI-generated text based on the copy and paste operation data and text content data. The generation module is used to align the response delay time, reading time index, word frequency density, semantic similarity, copy and paste frequency, fragment length and text similarity according to the timestamp, and then use a sliding time window to perform slice aggregation to generate a three-dimensional temporal feature tensor with three dimensions: number of samples, time step T and feature dimension. The state determination module is used to input the three-dimensional temporal feature tensor into the long short-term memory neural network for processing, and determine the learner's current cognitive state as either a content reuse state or a content reconstruction state based on the output probability distribution.

10. An electronic device, the electronic device comprising: One or more processors, a memory for storing one or more computer programs; characterized in that the computer programs are configured to be executed by the one or more processors, the programs including steps for performing the learner cognitive state determination method based on human-computer collaborative interaction process data as described in any one of claims 1-8.

Images