Method for dynamically generating personalized learning paths based on large models

Abstract

The application discloses a personalized learning path dynamic generation method based on a large model, relates to the technical field of education, and adjusts the rhythm density adaptively to obtain rhythm mapping values of nodes, which are used to reflect the relative position deviation of target nodes on the rhythm axis. Finally, the system adjusts the time sequence node spacing of the target nodes in the knowledge path according to the rhythm mapping values, generates a dynamic personalized learning path with variable rhythm density without changing the original logical relationship of the knowledge structure, realizes the transformation from a static linear learning sequence to an adaptive time layer scheduling path, and makes the learning process more in line with the individual rhythm law and cognitive evolution characteristics.

Description

Technical Field

[0001] This invention relates to the field of educational technology, specifically to a method for dynamically generating personalized learning paths based on a large model. Background Technology

[0002] In intelligent education systems, personalized learning path generation technology is a crucial means of achieving differentiated instruction for learners. Its goal is to dynamically adjust the order and time intervals of learning tasks based on learners' behavioral characteristics and knowledge acquisition status, while ensuring the consistency of knowledge logic. Currently, intelligent teaching systems based on large models possess semantic understanding, behavior recognition, and content generation capabilities. However, they still lack a unified mapping mechanism between the temporal dimension and semantic level of the learning process, making it difficult to simultaneously consider the impact of learning pace and knowledge absorption delays on the path structure. This limits the adaptive depth and dynamic adjustment precision of the learning path.

[0003] Most current personalized learning path generation methods are based on static features or single-dimensional statistical data, such as total learning time, accuracy rate, or tiered question difficulty. These methods generally ignore the dynamic temporal and semantic evolution characteristics of the learning process, making it difficult for the system to accurately capture changes in learner pace. For example, when learners answer questions correctly consecutively in a short period, traditional systems usually directly increase the difficulty of the questions without considering that their reaction time may be significantly shortened or that semantic absorption delays may not be fully stabilized. At this time, although the learner's accuracy rate increases, their knowledge structure has not yet been consolidated, making them prone to comprehension gaps in subsequent abstract questions. Furthermore, some systems only rely on answer scores to adjust paths without integrating semantic feedback content and changes in learners' language representation at different stages, resulting in a disconnect between the recommended path and the learning pace. Taking existing adaptive assessment systems as an example, although they can allocate questions based on answer speed and accuracy, they lack coupled judgment of fluctuations in learning pace and changes in semantic mastery, making it difficult to dynamically adjust the time density and knowledge sequence structure between paths. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides a method for dynamically generating personalized learning paths based on large models, thus solving the problems mentioned in the background technology.

[0005] To achieve the above objectives, the present invention provides the following technical solution: a method for dynamically generating personalized learning paths based on a large model, the method comprising:

[0006] The learner's interactive events are timestamped on the learning interface. After data preprocessing, behavioral data is generated, which includes reaction time series and knowledge absorption delay. Based on the behavioral data, the original rhythm quantity is obtained.

[0007] Based on the original rhythm quantity, the rhythm quantity is obtained by analyzing the degree of constraint of the semantic state of the previous time window on the current learning rhythm. After differential calculation, the rhythm change stability index is obtained to identify whether the rhythm fluctuation between windows is stable and generate a stability indicator function.

[0008] By using a large model, cross-window semantic comparison is performed on the stable window, the semantic change rate is calculated, and a compensating semantic change rate is generated in the unstable window through interpolation compensation, so as to form semantic mastery in the semantic vector space.

[0009] The knowledge node set is constructed and traversed. By analyzing the local error state of each knowledge node, the degree of local and global error clustering in the knowledge node set is analyzed to obtain the local error clustering strength and the global error clustering strength. Combined with the semantic mastery, the window to be reconstructed is locked, and the personalized path reconstruction program is started to generate a dynamic personalized learning path.

[0010] The present invention has the following beneficial effects:

[0011] By using a behavioral dataset and incorporating the semantic mastery level of the previous window, a rhythm quantity is formed to dynamically reflect the constraint effect of semantic state on learning rhythm. For example, when a learner's semantic mastery level is low in the previous window, the system will detect a decrease in the rhythm quantity, thereby inferring a slowdown in their learning speed and automatically extending the task push interval. Subsequently, the system identifies whether the rhythm fluctuations between windows are stable by calculating the difference and stability index of the rhythm quantity, and uses a stability indicator function to filter out stable intervals that can be used for cross-window comparison. With the help of a large model semantic embedding layer, a semantic vector comparison and compensation mechanism is performed on the stable window, calculating the semantic change rate and dependency difference integral quantity, and finally generating the semantic mastery level, realizing a continuous mapping and accurate recognition from temporal rhythm to semantic mastery.

[0012] The local error clustering strength is calculated based on the local error variation and centrality parameter, and then the global error clustering strength is obtained by averaging the results globally, thus realizing a multi-level representation of error distribution within the structural domain. When the strength exceeds a preset clustering threshold, a personalized path reconstruction program is automatically triggered, realizing a dynamic anomaly detection and intervention mechanism for the learning state.

[0013] The system performs adaptive adjustment of rhythm density to obtain rhythm mapping values ​​for each node, reflecting the relative positional deviation of the target node on the rhythm axis. Finally, based on these rhythm mapping values, the system readjusts the temporal node spacing of each target node in the knowledge path. Without altering the original logical relationships of the knowledge structure, it generates a dynamic, personalized learning path with variable rhythm density, transforming the learning process from a static linear learning sequence to an adaptive time-layer scheduling path. This makes the learning process more aligned with individual rhythmic patterns and cognitive evolution characteristics. Attached Figure Description

[0014] Figure 1 This is a schematic diagram of the method flow of the present invention;

[0015] Figure 2 This is a line graph showing the rhythm mapping of the personalized path reconstruction stage in the method of this invention. Detailed Implementation

[0016] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. 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.

[0017] Please see Figure 1 This invention provides a method for dynamically generating personalized learning paths based on large models, the method comprising:

[0018] The learner's interactive events are timestamped on the learning interface. After data preprocessing, behavioral data is generated, which includes reaction time series and knowledge absorption delay. Based on the behavioral data, the original rhythm quantity is obtained.

[0019] Based on the original rhythm quantity, the rhythm quantity is obtained by analyzing the degree of constraint of the semantic state of the previous time window on the current learning rhythm. After differential calculation, the rhythm change stability index is obtained to identify whether the rhythm fluctuation between windows is stable and generate a stability indicator function.

[0020] By using a large model, cross-window semantic comparison is performed on the stable window, the semantic change rate is calculated, and a compensating semantic change rate is generated in the unstable window through interpolation compensation, so as to form semantic mastery in the semantic vector space.

[0021] The knowledge node set is constructed and traversed. By analyzing the local error state of each knowledge node, the degree of local and global error clustering in the knowledge node set is analyzed to obtain the local error clustering strength and the global error clustering strength. Combined with the semantic mastery, the window to be reconstructed is locked, and the personalized path reconstruction program is started to generate a dynamic personalized learning path.

[0022] In this embodiment of the invention, firstly, by timestamping the start and end times of each learner's interactive events (such as answering questions, reading, modifying answers, and viewing explanations) on the learning interface, the system generates reaction time series. These time series, after data preprocessing, are combined with the time difference between the learner's first exposure to the knowledge point and their first correct answer to form behavioral data. The reaction time series reflects the learner's reaction speed and concentration in the short term, while the knowledge absorption delay reflects the learner's lag in understanding. Through statistical analysis and temporal variance calculation of the behavioral data, the system obtains the raw rhythm quantity, achieving a quantitative characterization of the learner's temporal rhythm state. This process enables the subsequent system to identify the rhythmic change patterns at different learning stages, providing a basis for dynamic path adjustment.

[0023] Secondly, rhythm quantity analysis based on the original rhythm quantity achieves a two-way coupling between learning rhythm and semantic mastery status. The system introduces the semantic mastery level of the previous time window into the original rhythm quantity, calculates the rhythm quantity through a semantic adjustment function, and obtains a rhythm change stability index through differential calculation. The rhythm change stability index is used to determine whether the learner's rhythm is stable, thereby generating a stability indicator function. When the index is below a threshold, the window is marked as a stable window, representing a smooth learning rhythm; when it exceeds the threshold, it is marked as an unstable window, representing significant rhythm fluctuations. This mechanism is used to prevent learning rhythm fluctuations from interfering with semantic evaluation. The logic of this unit is to first detect whether the rhythm is stable, and then decide whether to enter the semantic layer analysis stage, thereby achieving a sequential connection between the time domain and the semantic domain.

[0024] Then, the large-scale model semantic comparison achieves stable generation and anomaly compensation for semantic mastery. Within a window where the stability indicator function is 1, the system extracts the question stem, answer text, and feedback explanation text, generates a semantic vector set through the large-scale model semantic embedding layer, and performs cross-window semantic comparison to calculate the semantic change rate. The semantic change rate represents the learner's speed of semantic comprehension transfer within adjacent time windows and is a key parameter for assessing their knowledge absorption and semantic transfer abilities. When the stability indicator function is 0 (i.e., an unstable window), the system generates a compensating semantic change rate through an interpolation compensation mechanism to maintain the continuity of the semantic vector space, thereby preserving the integrity of the learning trajectory at the semantic level. Through this mechanism, the system can maintain continuous tracking of the semantic evolution of the learning content even under fluctuating rhythm conditions.

[0025] Next, knowledge node clustering analysis enables the dynamic quantification and localization of error distribution patterns. Throughout the learning process, the system constructs and traverses a set of knowledge nodes, each representing a knowledge point. By recording the number of errors, the time to the first correct answer, the corresponding rhythm quantity, and semantic mastery, the system calculates the local error variation and centrality parameter. The node's centrality parameter reflects its importance in the knowledge network, while the local error variation measures the fluctuation in its knowledge mastery. For example, in a geometric knowledge network, nodes representing parallel lines have high centrality. When their error variation increases, the system considers that this node affects multiple subsequent knowledge points, thus increasing its weight in error clustering analysis. The global error clustering strength, obtained after weighted calculation, reflects the degree of error aggregation in the knowledge structure, providing a basis for the system to subsequently identify abnormal learning stages or knowledge structure gaps.

[0026] Finally, path reconstruction and rhythm mapping enable adaptive reconstruction of learning paths and dynamic allocation of rhythm density. The path reconstruction program determines the basic rhythm allocation for each knowledge node based on its local error clustering strength and rhythm quantity within the window to be reconstructed. It then adaptively adjusts the rhythm density and temporal node spacing, taking into account the knowledge absorption delay, to generate a dynamic, personalized learning path. For example, if a learner has a high local error clustering strength and a long absorption delay in the probability and statistics module, the system will automatically increase the rhythm density of that knowledge segment (i.e., extend the learning time interval) and prioritize pushing relevant concept review content. This process achieves dynamic adjustment based on rhythm density without altering the logical knowledge structure, ultimately forming a path sequence that matches the individual's learning pace.

[0027] In summary, this invention employs a cross-dimensional fusion mechanism that uses rhythm stability to filter semantic analysis windows, semantic change rate to form semantic mastery, error clustering to reflect knowledge mastery distribution, and rhythm mapping to dynamically reconstruct learning paths. This enables the system to achieve truly adaptive and personalized learning path generation through the linkage of time, semantics, and structure, overcoming the problems of learning rhythm imbalance, semantic gaps, and path reconstruction delays in existing technologies. It also improves the accuracy of learning rhythm perception and the dynamic rationality of path generation.

[0028] In a preferred embodiment of the present invention, the learner's various interaction events are timestamped through the learning platform logs on the learning interaction interface. After data preprocessing, behavioral data is generated, including reaction time series and knowledge absorption delay. Based on the behavioral data, the original rhythm quantity is obtained, including:

[0029] The learning interface records the timestamps of learners' responses to questions, including the start and end timestamps of each interaction event, to form a reaction time series. Where R is the reaction time series, , and These represent the learner's reaction time for the first interaction event (first response submission), the second interaction event, and the nth interaction event within the corresponding learning cycle; the length n of the entire sequence corresponds to the number of events recorded by the system within that cycle.

[0030] Converting time information into a computable input signal sequence involves transforming discrete timestamp information into a continuous time interval sequence, which in turn transforms log text or timestamps into vectorized data that can be numerically computed.

[0031] Reaction time series represents the set of learners' reaction times for multiple interactive events (such as answering questions, reading, clicking to confirm, etc.) within a learning cycle. The reaction time series is timestamped to fill in missing values ​​(by interpolating between adjacent time intervals). The reaction time series is divided into several time windows. Difference operations are performed on adjacent reaction times within each time window to obtain the variance of the difference result over the entire time window, representing the temporal fluctuation of the current window. The intensity of the temporal fluctuation is obtained. The time difference between the learner's first exposure to a corresponding knowledge point and their first correct answer is statistically analyzed; this time difference is defined as the knowledge absorption delay, representing the learner's comprehension time span at the knowledge point level. Behavioral data within each time window is constructed, including the reaction time series within that time window and the knowledge absorption delay of the corresponding knowledge point.

[0032] Among them, the variance value is a window-level time fluctuation scalar, used to simultaneously measure the dispersion and consistency of time changes, i.e., the intensity of time-series fluctuations;

[0033] The width of each time window is determined by a fixed number of interaction events (e.g., 10-20 response records) to ensure the stability of the time series density.

[0034] Based on behavioral data, we conduct a preliminary analysis of learners’ time rhythm status within each time window to obtain the original rhythm quantity, which is the original intensity index of learning rhythm calculated based on pure time characteristics.

[0035] The method for calculating the original rhythm quantity is as follows: ,in, Let be the original rhythm quantity within the t-th time window. Within each time window, the system calculates intermediate parameters by coupling reaction time fluctuations with knowledge absorption delays. These parameters are used to quantify the original intensity of the learner's temporal rhythm state, representing the numerical intensity of the current temporal change. This provides standardized input for subsequent dynamic rhythm adjustment and personalized path generation. Let be the intensity of the temporal fluctuation within the t-th time window (i.e., the variance of the reaction time difference). The global average response time represents the average response time across all windows and is used as a global benchmark for time standardization. Let t be the average knowledge absorption delay within the t-th time window. It provides a benchmark across the window for normalization, serving as the global average absorbed delay; t is the time window index. It reflects the relative relationship between learners' knowledge absorption rate and the group benchmark. It reflects the intensity of fluctuations in learners' temporal behavior relative to a benchmark;

[0036] In this embodiment of the invention, a reaction time series is a set representing the reaction time of various behaviors of a learner within a learning cycle, reflecting the rhythmic characteristics between the learner's thinking and operation. For example, when a learner's answer time for the same type of question gradually decreases, the system can detect by recording timestamps that the reaction time in the latter half of the sequence continuously decreases, indicating that the learner's proficiency in that knowledge point has improved. By establishing this time series, learning behavior is transformed from a static performance indicator into a quantifiable temporal signal, providing a basic input for rhythm analysis.

[0037] During the recording process, the reaction time series is divided into windows based on a fixed number of interaction events. The width of each time window is determined by the number of events rather than the duration, ensuring consistent sample density across different windows and avoiding variance distortion due to varying time spans. The system performs difference operations on adjacent reaction times within each time window and calculates their variance to obtain the temporal fluctuation intensity. Temporal fluctuation intensity is a scalar measure representing the dispersion of learner behavior rhythm within the same window, reflecting the stability of their thinking and operations. For example, when reaction times in a window are successively 8 seconds, 7 seconds, 6 seconds, 7 seconds, and 8 seconds, the variance is small, indicating a stable rhythm. By calculating the variance, the system can quantitatively identify periods of rhythm stability and periods of rhythm abrupt change, laying the foundation for rhythm change analysis.

[0038] The system uses temporal fluctuation intensity and knowledge absorption delay as inputs, combined with global average reaction time and global average absorption delay, to calculate the original rhythm quantity. The original rhythm quantity is an intermediate parameter characterizing the overall rhythm intensity of the learner within a given time window, obtained through the coupling of reaction time fluctuation and knowledge absorption delay. For example, within a certain window, if the learner's reaction time variance is large but the knowledge absorption delay is short, it indicates that although the rhythm fluctuates, comprehension is rapid, and the original rhythm quantity is in the medium-to-high range. If the reaction time fluctuation is small but the absorption delay is long, the original rhythm quantity is low, indicating stable rhythm but slow learning speed. Through this comprehensive indicator, the system can quantitatively describe the intensity of the learning rhythm within each time window, forming a standardized input for subsequent rhythm adjustment and path optimization. Through this link, the system directly constructs a learning rhythm cognitive model based on behavioral time characteristics without relying on question difficulty or scoring. This allows subsequent rhythm adjustment, semantic mastery calculation, and path reconstruction to be dynamically driven based on the quantified rhythm input, thus providing a stable and adaptive temporal rhythm foundation for the generation of personalized learning paths driven by the entire large model.

[0039] In a preferred embodiment of the present invention, the rhythm quantity is obtained by analyzing the degree to which the semantic state of the previous time window restricts the current learning rhythm. After differential calculation, a rhythm change stability index is obtained as a criterion for whether the rhythm fluctuation between windows is stable, including:

[0040] Based on the original rhythm quantity, by introducing the semantic mastery degree within the previous time window after standardization and segmentation mapping, the degree of constraint of the semantic state of the previous time window on the current learning rhythm is analyzed to obtain the rhythm quantity. The rhythm quantity is a dynamic parameter reflecting the learner's current learning rhythm state and is used to control the task push interval in the task scheduling process. By indexing and storing the rhythm quantity with the time window, a rhythm parameter sequence is formed so as to track and compare the rhythm change trend hourly in subsequent stages.

[0041] The formula for calculating rhythm quantity is: ,in, Let be the rhythm quantity within the t-th time window; it is a dynamic parameter reflecting the learner's current learning rhythm state. The semantic mastery level within the (t-1)th time window. The semantic mastery adjustment function is defined for the (t-1)th time window. This function is based on semantic mastery and is a functional mapping. The input is the semantic mastery of the previous time window, and the output is a rhythm adjustment coefficient (dimensionless). This coefficient is used to semantically correct the original rhythm value. Specifically, within each time window, the semantic mastery is obtained from the previous time window, standardized using the min-max normalization method, and a corresponding piecewise function is selected based on a preset interval threshold to obtain the corresponding... ;

[0042] By dynamically adjusting the value of the rhythm quantity, the time, content, and path structure of the learning process can be adjusted to avoid the learning pace being too fast or too slow. At the same time, the path structure can be personalized and self-corrected. An increase in the value indicates that the task push is faster and the time interval is shorter; a decrease in the value indicates that the task push is slower and the learning pace is slower.

[0043] The rhythm parameter sequence is subtracted from adjacent rhythm quantities to obtain the rhythm change rate Δβ(t) = β(t) - β(t-1). The rhythm change stability index S(t) = |Δβ(t) - Δβ(t-1)| is calculated based on the rhythm change rate, where Δβ(t) and Δβ(t-1)| are the rhythm change rates between different time windows, β(t) and β(t-1) are the rhythm quantities in different time windows, and S(t) is the rhythm change stability index. The degree of rhythm oscillation is judged by the rhythm change stability index.

[0044] Identify whether the rhythm fluctuations between windows are stable and generate a stability indicator function, including:

[0045] Within a time window, learners and their corresponding question stems, answers, and feedback explanations are extracted and semantically encoded through the semantic embedding layer of the large model to generate a set of semantic vectors. At the same time, the set of semantic vectors from the previous time window is read, and a time alignment table is established based on the time index and rhythm quantity, so that the sets of semantic vectors in adjacent time windows have a mapping relationship in the time series.

[0046] Based on the comparison between the rhythm change stability index and the preset stability threshold, a stability indicator function is generated. If the rhythm change stability index does not exceed the stability threshold, the stability indicator function is 1, and the corresponding time window is included in the stable set as a stable window. At this time, cross-window semantic pairing calculation is performed. Otherwise, the stability indicator function is 0, and the corresponding time window is not included in the stable set as an unstable window. At this time, the semantic comparison of the window is paused, and its time interval gap is recorded for subsequent backtracking.

[0047] In this embodiment of the invention, the temporal rhythm and semantic mastery are first coupled into a schedulable dynamic parameter: based on the original rhythm quantity, the semantic mastery level of the previous time window is introduced and standardized and segmented to obtain the rhythm quantity. The rhythm quantity serves as a direct control quantity for the task push interval, enabling scheduling to not only consider time fluctuations but also perceive the constraints of semantic state on the rhythm. For example, if the semantic mastery level of the previous window is low and easily confused concepts are present, the rhythm quantity will be suppressed after mapping, and the system will automatically increase the interval of the next batch of tasks, prioritizing the insertion of reinforcement questions and brief explanations. Conversely, when the semantic mastery level steadily increases, the rhythm quantity increases, allowing for shorter intervals and faster progress.

[0048] The rhythm quantity is a dynamic variable used to characterize the current learning rhythm state and directly drive the adjustment of task intervals and rhythm density.

[0049] Semantic mastery is measured by the level of cross-window semantic understanding, which is formed by semantic vector comparison and dependency difference integration.

[0050] The semantic adjustment function is used to transform the semantic mastery of the previous window into a coefficient that promotes or inhibits rhythm.

[0051] Next, the stability of the rhythm is transformed into a quantifiable gating signal. Adjacent differences in the rhythm parameter sequence are used to obtain the rhythm change rate, which is further used to form a rhythm stability index. This index serves as a unified criterion for the stability of rhythm fluctuations between windows. For example, when a learner exhibits rhythmic jitter of "fast, slow, fast," the difference in adjacent change rates increases, the stability index exceeds a threshold, and the window is judged as unstable. Based on this, the system temporarily suspends semantic computation to avoid accumulating errors in semantic mastery during noisy periods. The rhythm parameter sequence is a rhythm quantity sequence indexed and stored over time windows to track rhythm evolution trends. The stability indicator function is a binary gating signal for stability and instability, determining whether semantic computation and integration are allowed. The time interval gap is a void segment formed by an unstable window on the time axis, serving as positioning information for subsequent compensation.

[0052] Within each time window, the question stem, answer, and feedback explanation text are extracted and embedded into a large model to obtain a semantic vector set. A time alignment table is then established based on the time index and rhythm quantity to ensure a one-to-one mapping relationship between semantic items in adjacent windows, avoiding mismatches and drift, and facilitating accurate calculation of semantic differences and change rates between two windows. If the rhythm quantity is low, the alignment interval is widened to avoid forcibly pairing semantics that have not yet been consolidated. The time alignment table integrates the time index and rhythm quantity to constrain the cross-window correspondence of semantic items and control the granularity and stride of pairing. The semantic vector set is a multi-source text semantic representation from the question stem, answer, and feedback, used to calculate semantic differences and change rates.

[0053] In summary, the rhythm quantity is regulated by the semantic state of the previous window, avoiding blindly accelerating the process based solely on time acceleration; semantic accumulation is paused during unstable windows to reduce the contamination of the mastery curve during noisy periods. The time alignment table introduces the rhythm quantity into the semantic mapping stride control, avoiding both noise amplification caused by overly dense alignment and information loss caused by overly sparse alignment. For example, pairing is encrypted during high-rhythm phases to capture rapidly progressing fine-grained semantic transfer; pairing is relaxed during low-rhythm phases to provide more robust cross-window comparisons. The rhythm quantity serves as both a control quantity for task intervals and a scale quantity for semantic computation (implemented through stability criteria and the time alignment table).

[0054] In a preferred embodiment of the present invention, cross-window semantic comparison is performed on a stable window, the semantic change rate is calculated, and a compensated semantic change rate is generated in an unstable window through interpolation compensation to form a semantic mastery in the semantic vector space, including:

[0055] When the stability indicator function is 1, the average semantic difference of semantic vector pairs with the same number in the semantic vector sets of adjacent stable windows is calculated, and the average semantic difference of adjacent stable windows is differentially divided to obtain the change in semantic difference.

[0056] Using rhythm quantity as the time scale benchmark, the semantic change rate is calculated as follows: ;in, For semantic change rate, This represents the amount of semantic difference change;

[0057] When the stability indicator function is 0, the above comparison is not performed. Instead, the compensation mechanism is executed only after a stable window appears. If the compensation mechanism is executed, the corresponding unstable window is recorded as the compensation window. Specifically, if there are adjacent stable windows before and after the unstable window and the time interval (time interval gap) between the two stable windows does not exceed a preset time interval threshold, interpolation compensation is used to obtain the compensation semantic change rate. Specifically, if interpolation compensation is used, the following calculations are performed:

[0058] ;

[0059] ;

[0060] ;

[0061] in, To compensate for the average semantic difference, The rhythm amount after compensation. and These represent the rhythm quantities within the (t-1)th and (t+1)th time windows, respectively. To compensate for the rate of semantic change, This is to compensate for the change in semantic differences. and The average semantic difference within different time windows, i.e., the average semantic difference within the (t-1)th and (t+1)th time windows;

[0062] Calculate the semantic vector pairs with the same index in the semantic vector set of adjacent stable windows. The average semantic difference is as follows: ,in, The average semantic difference represents the degree of structural difference in semantic expression, used for subsequent transfer evaluation. K is the number of knowledge points in the current window, and k is the knowledge point number. Cosine similarity of semantic vectors corresponding to the same knowledge point within adjacent windows; and These are the semantic vectors corresponding to the k-th knowledge point within the t-th time window and the k-th knowledge point within the (t-1)-th time window, respectively.

[0063] Within the time window sequence, the system performs semantic encoding on all learning tasks or knowledge units within each window. These tasks include multiple text items such as question stems, options, answer texts, and feedback explanations. After processing by the semantic encoding unit (the semantic embedding module of the large model), the system generates a semantic vector for each knowledge unit. A semantic vector pair with the same number refers to a pair of vectors formed by the system taking the corresponding semantic vectors in two adjacent time windows for the same number (i.e., questions representing the same knowledge unit or the same logical position). Simply put, a semantic vector pair with the same number describes the change in the semantic expression of the same knowledge point in adjacent learning cycles; it is a one-to-one mapping across windows.

[0064] Then, by differentiating adjacent windows, the semantic difference change is obtained: ,in, and The average semantic difference within the (t-1)th and (t+1)th time windows. This represents the amount of semantic difference change;

[0065] If there is no need to perform a compensation mechanism, then only the semantic change rate of the stable window is integrated over time.

[0066] The compensation window and the stability window are combined to generate an analysis window set;

[0067] The time interval gap is used to characterize the traceable distance and rhythmic fault length between the current time window and the most recent stable window. It is calculated by the window timestamp difference and serves as the basis for subsequent realignment and interpolation compensation. It ensures that semantic computation is performed only within the time base recoverable window interval. It plays a role in calibrating the interval buffer length in the time series.

[0068] In the semantic vector space, the contextual dependency of any two semantic items within the semantic vector set of the analysis window set is calculated to obtain the contextual dependency matrix of the corresponding time window. By calculating the difference in contextual dependency matrices between adjacent time windows, the stability of the learner's semantic structure at different stages is evaluated to obtain the magnitude of change in semantic dependency structure between adjacent time windows. The magnitude of change in semantic dependency structure between all adjacent windows in the analysis window set is integrated over time according to the Euclidean norm to obtain the dependency difference integral. At the same time, the semantic change rate of all stable windows and the compensating semantic change rate of the compensating windows are included in the time integration to obtain the semantic change integral. Based on the semantic change integral and the dependency difference integral, the semantic mastery is obtained.

[0069] The context dependency matrix is ​​a matrix structure used by a system within a time window to describe the strength of dependencies between semantic items (such as stems, options, answers, and explanations). Specifically, it is not a set of original semantic vectors, but a relational matrix that reflects the degree of dependency, connectivity, and mutual influence between these semantic vectors. Mathematically, its element in row a and column b represents the degree of semantic dependency of the a-th semantic item on the b-th semantic item within the window.

[0070] The process of constructing the context dependency matrix:

[0071] S1: Extract the set of semantic vectors within the window;

[0072] S2: For any two semantic terms, calculate the degree of contextual dependency between them using the cosine similarity algorithm;

[0073] S3: To prevent uneven numerical distribution, the system performs normalization on the context dependency matrix using the Frobenius norm;

[0074] The magnitude of change in semantic dependency structure is the L2 difference of the context dependency matrix within adjacent time windows, which is used to assess the stability of the learner's semantic structure at different stages. The larger the value, the stronger the learner's semantic structure is reorganized (such as semantic confusion or reorganization) in that window.

[0075] The semantic change integral represents the cumulative incremental integral of semantic changes, reflecting the positive growth of learners' semantic mastery; the dependency difference integral represents the complexity integral of semantic context dependency changes, reflecting the consumption or internal friction of semantic relations.

[0076] The semantic mastery is obtained by subtracting the dependency difference integral from the semantic change integral. This subtraction relationship represents an energy conservation-like abstraction in the learning process: the increase in semantic mastery is constrained by semantic complexity. Even if the learner updates the semantic representation rapidly on the surface (high semantic change rate), if the semantic dependencies change drastically (large change in the semantic dependency structure), the net increase in mastery will be offset. Therefore, semantic mastery is equivalent to the cumulative net effective rate of change in semantic learning. It is not a direct semantic similarity, but a dynamic equilibrium indicator.

[0077] In this invention, all parameters are dimensionless by using dimensionless processing technology to remove their dimensions.

[0078] In this embodiment of the invention, by performing cross-window semantic comparison of semantic vector sets from adjacent time windows within a stable window, calculating the average semantic difference, and further determining the change in semantic difference, a semantic change rate model is established using the rhythm quantity as a time scale benchmark. This dynamically quantifies the changing trend of learners' semantic understanding over time. This mechanism controls the time scale through the rhythm quantity, ensuring that the semantic change rate not only reflects textual differences but also embodies the intrinsic influence of learning rhythm, thus forming a continuous and quantifiable trajectory of semantic mastery in the semantic vector space. This enables simultaneous assessment of learners' depth of understanding and speed of mastery. Compared to traditional static assessment methods based on accuracy or scores, it can reveal the magnitude and stability of changes in learners' semantic transfer process in real time, avoiding lag or distortion in semantic mastery status.

[0079] Meanwhile, in situations where the rhythm is unstable or there are time interval gaps, this invention uses a compensation mechanism to mark unstable windows as compensation windows. When stable windows exist before and after the current window and the time interval does not exceed a threshold, interpolation is used to calculate the compensation semantic change rate, ensuring the semantic trajectory remains continuous on the time axis. For example, if a learner fails to complete a task in window t due to a temporary interruption, but the semantic states of the two windows before and after are similar, the system automatically interpolates to generate the compensation semantic change rate for that period, avoiding gaps or abrupt changes in the semantic mastery curve.

[0080] Building upon this foundation, the system combines all stable and compensating windows into an analysis window set and constructs a context dependency matrix in the semantic vector space. It then calculates the context dependency differences between adjacent time windows to quantify the magnitude of changes in the learner's semantic structure. By integrating the semantic change rate and dependency differences over time, the system obtains the semantic change integral and dependency difference integral, ultimately generating a comprehensive index reflecting the depth of semantic understanding and the stability of semantic structure—the semantic mastery. This mechanism ensures that even during unstable learning cycles, the system can continuously and robustly evaluate the learner's semantic state, avoiding interruptions in mastery calculation due to short-term rhythm fluctuations or data gaps, thereby improving the robustness and temporal integrity of semantic understanding modeling.

[0081] In a preferred embodiment of the present invention, a set of knowledge nodes is constructed and traversed. Based on the local error states of each knowledge node, the degree of local and overall error clustering in the knowledge node set is analyzed to obtain the local error clustering strength and the global error clustering strength, including:

[0082] Throughout the learner's learning and question-answering process, the learning content is analyzed at the knowledge unit level, forming a set of knowledge nodes. Each knowledge node represents an independent concept, and the logical dependencies between nodes are established in the form of directed edges to represent the hierarchical relationship of the knowledge structure in the logical domain. During the learner's task execution, the system records the number of errors for each knowledge node, the index of the first correct answer, the pace of the current time window, and the degree of semantic mastery.

[0083] Iterate through all knowledge nodes in the knowledge node set, calculate the local error change and centrality parameter for each knowledge node, and obtain the local error clustering strength by multiplying the local error change of each knowledge node within each time window by the corresponding centrality parameter. Summate the local error clustering strengths and take their mean to obtain the global error clustering strength, which reflects the clustering strength of errors within the knowledge structure. The specific calculation formula is as follows:

[0084] Where E is the global erroneous clustering strength, K is the number of knowledge nodes within the corresponding time window, and k is the knowledge node index. The centrality parameter of the k-th knowledge node acts as a weighting factor, amplifying the error changes of the core node. This represents the local error change at the k-th knowledge node, reflecting the local shift of the error in the structure.

[0085] After completing the initial calculations, the global error clustering strength is periodically recorded, and a correlation table is established with the semantic mastery level in the same period for subsequent joint analysis.

[0086] The centrality parameter represents the importance of a node's connections in the global knowledge network, and its formula is:

[0087] ;

[0088] Where j and k are distinct knowledge nodes, and p is the number of knowledge nodes in the knowledge node set. The dependency relationship between nodes k and j can be obtained through cosine similarity calculation;

[0089] Based on the global error clustering strength, the window to be reconstructed is locked, including:

[0090] The global error clustering intensity is normalized and compared based on the numerical trend of semantic mastery. The semantic correction factor is calculated, and the corrected global error clustering intensity is obtained by multiplying the semantic correction factor by the global error clustering intensity. This is used to describe the inhibitory effect of semantic mastery level on error clustering intensity.

[0091] The formula for calculating the semantic correction factor is:

[0092] ;

[0093] in, This is a semantic correction factor used to balance the impact of learners' current semantic mastery on incorrect clustering. The semantic mastery level within the t-th time window. To prevent a small constant with a zero denominator when SMI(t) = 0 (generally taken as 10), -3 Up to 10 -5 When semantic mastery is low, the same error variation is more likely to represent a cognitive problem. Therefore, we take its inverse relationship, so that the correction factor is high when mastery is low and low when mastery is high.

[0094] By continuously recording the corrected global erroneous clustering strength over time, the system can identify the range of changes and abnormal upward trends in the global erroneous clustering strength and trigger a path reconstruction command when the threshold is reached.

[0095] If the corrected global erroneous clustering strength exceeds the preset clustering threshold, the personalized path reconstruction program will be started; otherwise, the personalized path reconstruction program will not be started.

[0096] In this embodiment of the invention, by parsing the learning content at the knowledge unit level, a set of knowledge nodes containing multiple knowledge nodes and directed dependencies is constructed, realizing hierarchical modeling of the learner's cognitive structure. During the learning process, the system records dynamic parameters such as the number of errors, the index of the first correct answer, the rhythm, and semantic mastery of each knowledge node in real time, and then calculates the local error change and centrality parameter of the node. By multiplying the local error change of a node by its centrality parameter, the local error clustering strength of that node is obtained. The global error clustering strength is then calculated by averaging across all nodes, comprehensively reflecting the distribution of learning errors in the entire knowledge network. For example, when a learner makes frequent errors on the two high-centrality nodes of conditional probability and Bayesian update, while making fewer errors on other nodes, the system will detect an increase in the global error clustering strength, indicating that the learner's errors have concentrated in key core areas, suggesting a systematic misunderstanding of core knowledge points in the current learning path. This mechanism realizes aggregated analysis from single-question errors to errors at the knowledge structure level, automatically identifying relatively influential error nodes and propagation paths, providing high-resolution error localization basis for subsequent personalized reconstruction, thereby improving the system's accuracy in capturing learning weaknesses.

[0097] This invention, after obtaining the global error clustering strength, introduces a semantic correction factor based on the changing trend of semantic mastery to normalize and correct the error clustering strength, thereby dynamically balancing the weight relationship between semantic understanding depth and error clustering degree. This correction mechanism effectively avoids false anomalies caused by simply using the number of errors as the judgment criterion. For example, when learners transition from basic concepts to comprehensive questions, there may be a short-term increase in errors, but the semantic mastery curve continues to rise. Based on this, the system automatically reduces the corrected error clustering strength, judging that the error belongs to the normal fluctuation during the learning deepening period, and does not trigger path reconstruction. Conversely, if the semantic mastery decreases and the error clustering strength continues to rise, the system judges it as a conceptual deviation and triggers a personalized path reconstruction program. This mechanism introduces the level of semantic understanding into the error judgment logic through the semantic correction factor, enabling the system to continuously track the change range of the corrected global error clustering strength in the time dimension, and automatically judge whether it exceeds the clustering threshold based on its abnormal upward trend, triggering a path reconstruction instruction.

[0098] Therefore, this invention realizes intelligent detection and adaptive control of abnormal learning states in dynamic learning scenarios, which not only prevents excessive intervention, but also enables timely path reconstruction when a real structural error occurs, ensuring the stability and personalized optimization of the learning process.

[0099] Please see Figure 2 In a preferred embodiment of the present invention, initiating a personalized path reconstruction procedure includes:

[0100] The time window corresponding to the initiation of the personalized path reconstruction program is determined as the reconstruction window, and the knowledge nodes involved in the reconstruction window are designated as target nodes. Target nodes are then selected based on constraints, which include two sets of constraints:

[0101] Focus on the absolute level of the current cluster strength: If the local erroneous cluster strength of the target node in the current reconstruction window exceeds the preset strength threshold, it is used to identify nodes whose erroneous cluster strength itself exceeds the average level;

[0102] Focus on the upward trend between adjacent time windows: If the difference between the cluster strength of the target node in the current reconstruction window and the local erroneous cluster in the previous detection exceeds the preset difference threshold, it is used to identify nodes whose cluster strength shows a significant upward trend over time.

[0103] Together, these two factors constitute the dual conditions for identifying abnormal node regions, enabling the system to detect strong temporally local clustering features while maintaining structural consistency.

[0104] Take all target nodes that satisfy at least one constraint condition and find their union to obtain the target set. Based on the rhythm amount of the window to be reconstructed, determine the basic rhythm allocation amount of each target node. Combined with the knowledge absorption delay of each target node, adaptively adjust the rhythm density of each target node in the target set to obtain the rhythm mapping value. The rhythm mapping value is used to reflect the relative position deviation of the target node on the rhythm axis. The rhythm mapping value is used as the time layer input signal to readjust the temporal node spacing of each target node in the knowledge path, so that the system can achieve a dynamic learning path with "variable rhythm density between different nodes" without changing the knowledge structure logic.

[0105] The rhythm density of each target node in the target set is adaptively adjusted, and the specific acquisition method is as follows:

[0106] ;

[0107] in, Assign a rhythm value to the i-th target node in the target set within the t-th time window. Let be the average knowledge absorption delay within the t-th time window; The number of knowledge nodes within the t-th time window (referring to the corresponding window to be reconstructed); This is an adjustment coefficient used to control the magnitude of the impact of the delay correction term on the basic rhythm; The knowledge absorption delay deviation of the i-th target node in the target set within the t-th time window is specifically the knowledge absorption delay of the i-th target node in the target set within the t-th time window minus the average knowledge absorption delay within the corresponding time window. It is the base rhythm allocation for each node, representing the average progress of each node within the current time window under the default rhythm. Increasing the spacing between nodes reduces their basic advancing distance, and vice versa. It is a correction term used to ensure individual differences in rhythm between nodes. The delay deviation of the node within the corresponding window, when When the distance between nodes increases, the spacing between them increases; conversely, it decreases.

[0108] The rhythm allocation value is then scaled linearly to obtain the rhythm mapping value of the corresponding target node in the target set. The specific process is as follows: First, the rhythm allocation value of the corresponding target node in the target set and the basic rhythm allocation of other target nodes in the corresponding window to be reconstructed are statistically analyzed and summed to obtain the total rhythm allocation value. The rhythm value of the window to be reconstructed is divided by the total rhythm allocation value to obtain the mapping coefficient. The rhythm allocation value or basic rhythm allocation of each target node is linearly scaled using the mapping coefficient to obtain the rhythm mapping value. The rhythm mapping value is the intermediate medium for converting individual learning speed into time structure scheduling signal. Its role is equivalent to labeling each knowledge node with a rhythm weight at the time level, thereby determining which knowledge nodes should be pushed in advance and which should be reproduced later.

[0109] At the same time, a unit time period (e.g., 0.5 days or 2 days) is defined as the basic time scale. By multiplying the unit time period by the rhythm mapping value, the temporal node spacing of the corresponding node is obtained, which is used to form the temporal distribution of rhythm density changes.

[0110] The goal of the screening is to identify nodes that exhibit significant local clustering levels within a certain time window or time period;

[0111] The target set contains multiple target nodes that satisfy at least one constraint condition;

[0112] Among them, by Figure 2 As can be seen, the horizontal and vertical axes represent rhythm mapping values, which reflect the rhythm density of nodes on the rhythm axis after reconstruction. The horizontal axis represents the knowledge node index, the horizontal dashed line represents the baseline before reconstruction, and the vertical shaded area represents the window to be reconstructed. Target nodes in the target set are marked with an upper triangle to indicate the specific upward adjustment value, while non-target sets (all nodes are lowered with the same magnitude) are marked with a lower triangle.

[0113] In this embodiment of the invention, by setting dual constraints, high-risk knowledge nodes are accurately screened within the reconstruction time window. This screening process can simultaneously capture nodes with consistently high error rates and nodes with sudden increases in errors, thereby dynamically locking in key knowledge points that need adjustment at different learning stages. For example, when a learner makes multiple consecutive errors on a conditional probability node and the error clustering strength exceeds a threshold, the system directly identifies it as an anomalous node. Conversely, if a learner previously had a high accuracy rate on a Bayesian update node but recently experienced a sharp increase in errors, the system identifies it as a sudden risk node based on a difference threshold. This dual-condition union screening mechanism ensures that the system can capture both long-term accumulated problems and respond to short-term sudden deviations, avoiding the drawback of traditional systems that only focus on the total number of errors while ignoring the trend of error changes. Therefore, an adaptive node detection mechanism based on multi-dimensional dynamic error characteristics is realized, enabling path reconstruction to be triggered only in necessary local areas, improving the targeting of personalized intervention and the efficiency of system execution.

[0114] After selecting target nodes, the rhythm density of the target nodes is adaptively adjusted based on the rhythm amount of the window to be reconstructed and the knowledge absorption delay of each node. A rhythm mapping value is calculated, reflecting the relative positional deviation of the target node on the rhythm axis. This value serves as the input signal for subsequent time layers, thereby readjusting the temporal spacing between nodes without altering the knowledge structure logic, achieving dynamic rhythm control. For example, if the system detects that the learner's absorption delay at a conditional probability node is significantly longer than the average, the system will correspondingly reduce its rhythm density and extend the push interval; conversely, it will increase the rhythm density and shorten the interval, ensuring that the overall learning process is neither slowed down nor excessively stacked. This rhythm and delay coupling control mechanism allows the learning path to adaptively adjust according to the learner's comprehension rhythm, achieving a dynamic scheduling effect with variable rhythm density. Compared to traditional fixed paths, the learning path generated by this invention better matches the learner's cognitive rhythm in the time dimension and maintains the original knowledge dependencies in the logical dimension, thus achieving a balance between efficiency and robustness.

[0115] All thresholds in this invention can be obtained using the mean-standard deviation method;

[0116] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for dynamically generating personalized learning paths based on large models, characterized by: The method includes: The learner's interactive events are timestamped on the learning interface. After data preprocessing, behavioral data is generated, which includes reaction time series and knowledge absorption delay. Based on the behavioral data, the original rhythm quantity is obtained. Based on the original rhythm quantity, the rhythm quantity is obtained by analyzing the degree of constraint of the semantic state of the previous time window on the current learning rhythm. After differential calculation, the rhythm change stability index is obtained as a criterion for whether the rhythm fluctuation between windows is stable. This includes: based on the original rhythm quantity, by introducing the semantic mastery degree in the previous time window after standardization and segmentation mapping, the degree of constraint of the semantic state of the previous time window on the current learning rhythm is analyzed to obtain the rhythm quantity. The rhythm quantity is a dynamic parameter reflecting the learner's current learning rhythm state and is used to control the task push interval in the task scheduling process; by indexing and storing the rhythm quantity with time windows, a rhythm parameter sequence is formed. The rhythm parameter sequence is subtracted from adjacent rhythm values ​​to obtain the rhythm change rate, and the rhythm change stability index is calculated based on the rhythm change rate; after identifying whether the rhythm fluctuations between windows are stable, a stability indicator function is generated. By using a large model, cross-window semantic comparison is performed on the stable window, the semantic change rate is calculated, and a compensating semantic change rate is generated in the unstable window through interpolation compensation, so as to form semantic mastery in the semantic vector space. The knowledge node set is constructed and traversed. By analyzing the local error state of each knowledge node, the degree of local and global error clustering in the knowledge node set is analyzed to obtain the local error clustering strength and the global error clustering strength. Combined with the semantic mastery, the window to be reconstructed is locked, and the personalized path reconstruction program is started to generate a dynamic personalized learning path.

2. The method for dynamically generating personalized learning paths based on a large model according to claim 1, characterized in that: The learner's interactive events are timestamped on the learning interface. After data preprocessing, behavioral data is generated, including reaction time series and knowledge absorption delay. Based on the behavioral data, the raw rhythm data is obtained, including: The learning interaction interface records the timestamps of learners' answers to questions, which are then converted into a reaction time series. The reaction time series is then time-stamped to fill in missing values. The reaction time series is divided into windows to obtain several time windows. At the same time, the time difference between the learner's first exposure to the corresponding knowledge point and the first correct answer is statistically analyzed to obtain the knowledge absorption delay and construct behavioral data within each time window. Based on behavioral data, we conduct a preliminary analysis of learners' temporal rhythm status within each time window to obtain the raw rhythm quantity. The raw rhythm quantity is an original intensity index of learning rhythm calculated based on pure time characteristics.

3. The method for dynamically generating personalized learning paths based on a large model according to claim 2, characterized in that: Identify whether the rhythm fluctuations between windows are stable and generate a stability indicator function, including: Within the time window, learners, corresponding question stem texts, answer texts, and feedback explanation texts are extracted and semantically encoded through the semantic embedding layer of the large model to generate a set of semantic vectors. At the same time, the set of semantic vectors from the previous time window is read, and a time alignment table is established based on the time index and rhythm quantity. Based on the comparison between the rhythm change stability index and the preset stability threshold, a stability indicator function is generated. If the rhythm change stability index does not exceed the stability threshold, the stability indicator function is 1, and the corresponding time window is included in the stable set as a stable window; otherwise, the stability indicator function is 0.

4. The method for dynamically generating personalized learning paths based on a large model according to claim 3, characterized in that: Perform cross-window semantic alignment on stable windows, calculate semantic change rates, and generate compensated semantic change rates through interpolation in unstable windows to form semantic mastery in the semantic vector space, including: When the stability indicator function is 1, the average semantic difference of semantic vector pairs with the same number in the semantic vector sets of adjacent stable windows is calculated, and the average semantic difference of adjacent stable windows is differentially divided to obtain the change in semantic difference. Using rhythm quantity as the time scale benchmark, the rate of semantic change is calculated; When the stability indicator function is 0, the compensation mechanism is executed after the stable window appears. If the compensation mechanism is executed, the corresponding unstable window is recorded as the compensation window. Specifically, if there are adjacent stable windows before and after the unstable window and the time interval between the stable windows does not exceed the preset time interval threshold, interpolation compensation is used to obtain the compensation semantic change rate. The compensation window and the stability window are combined to generate an analysis window set; In the semantic vector space, the contextual dependency of any two semantic items within the semantic vector set of the analysis window set is calculated to obtain the contextual dependency matrix of the corresponding time window. By calculating the difference in the contextual dependency matrices of adjacent time windows, the stability of the learner's semantic structure at different stages is evaluated to obtain the magnitude of change in semantic dependency structure between adjacent time windows. By integrating the magnitude of change in semantic dependency structure between all adjacent windows in the analysis window set over time, the dependency difference integral is obtained. At the same time, the semantic change rate of all stable windows and the compensating semantic change rate of the compensating windows are included in the time integration to obtain the semantic change integral. Based on the semantic change integral and the dependency difference integral, the semantic mastery is obtained.

5. The method for dynamically generating personalized learning paths based on a large model according to claim 4, characterized in that: Construct and traverse the knowledge node set. Analyze the degree of local and global error clustering in the knowledge node set based on the local error state of each knowledge node, and obtain the local error clustering strength and global error clustering strength, including: The learning content of learners is analyzed at the knowledge unit level to form a set of knowledge nodes. During the process of learners performing tasks, the number of errors, the index of the first correct answer time, the pace of the current time window, and the degree of semantic mastery of each knowledge node are recorded. Traverse all knowledge nodes in the knowledge node set, calculate the local error change and centrality parameter for each knowledge node, and obtain the local error clustering strength by multiplying the local error change of each knowledge node in each time window by the corresponding centrality parameter. Summing the local error clustering strengths and taking their mean values ​​yields the global error clustering strength, which reflects the clustering strength of errors within the knowledge structure.

6. The method for dynamically generating personalized learning paths based on a large model according to claim 5, characterized in that: Based on the global error clustering strength, the window to be reconstructed is locked, including: The global error clustering intensity is normalized and compared based on the numerical trend of semantic mastery. The semantic correction factor is calculated, and the corrected global error clustering intensity is obtained by multiplying the semantic correction factor by the global error clustering intensity. This is used to describe the inhibitory effect of semantic mastery level on error clustering intensity. If the corrected global erroneous clustering strength exceeds the preset clustering threshold, the personalized path reconstruction program will be started; otherwise, the personalized path reconstruction program will not be started.

7. The method for dynamically generating personalized learning paths based on a large model according to claim 1, characterized in that: Initiate the personalized route refactoring process, including: The time window corresponding to the start of the personalized path reconstruction program is determined as the reconstruction window, and the knowledge nodes involved in the reconstruction window are taken as target nodes. The target nodes are then selected according to the constraints, including: the local error clustering strength of the target node in the current reconstruction window exceeds the preset strength threshold, and the difference between the local error clustering strength of the target node in the current reconstruction window and the previous detection exceeds the preset difference threshold. Take all target nodes that satisfy at least one constraint condition and find their union to obtain the target set. Based on the rhythm amount of the window to be reconstructed, determine the basic rhythm allocation amount of each target node. Combined with the knowledge absorption delay of each target node, adaptively adjust the rhythm density of each target node in the target set to obtain the rhythm mapping value. The rhythm mapping value is used to reflect the relative position deviation of the target node on the rhythm axis. Using the rhythm mapping value as the time layer input signal, the temporal node spacing of each target node in the knowledge path can be readjusted to realize a dynamic personalized learning path.

Images