Language learning evaluation system and method based on data fusion model

Abstract

The invention discloses a language learning evaluation system and method based on a data fusion model, and relates to the technical field of data processing, and the method comprises the steps: obtaining language writing text data and eye movement fixation data of a learner; language writing text data is disassembled into a plurality of semantic function units, the semantic function units comprise a core carrier unit, a logic connection unit and a support expansion unit, and a dynamic logic density value is marked for each semantic function unit; dynamically binding the eye movement fixation data with a semantic function unit to generate a cognitive load vector comprising a unit type, a dynamic logic density, a grammar depth, a fixation duration and a jump frequency; an avoidance index of the core carrier unit is calculated based on the cognitive load vector, and when the avoidance index exceeds a preset avoidance threshold value and the activation rate of the support expansion unit exceeds a preset activation threshold value, a high-order thinking escape strategy judgment instruction is triggered; and in response to the high-order thinking escape strategy judgment instruction, determining a comprehensive evaluation report.

Description

Technical Field

[0001] This application relates to the field of data processing technology, and more specifically, to a language learning assessment system and method based on a data fusion model. Background Technology

[0002] In today's increasingly globalized world, language learning is of paramount importance for personal development, international exchange, and cultural dissemination. Language learning assessment, as a key means of measuring learners' language proficiency, helps teachers accurately grasp students' learning progress in education, thereby adjusting teaching strategies and improving teaching effectiveness. In the professional field, it helps companies select talent with the necessary language skills, ensuring smooth communication in the workplace. Therefore, accurate and comprehensive language learning assessment is crucial.

[0003] However, existing language learning assessment methods have certain limitations. On the one hand, assessments often rely on information from a single dimension, making it difficult to comprehensively reflect learners' language usage status by integrating different types of data. On the other hand, they do not pay enough attention to the deep thinking characteristics exhibited by learners during language output, making it difficult to effectively determine whether learners are avoiding complex expressions or exhibiting mental inertia. As a result, the depth and relevance of assessment results need to be improved.

[0004] There is currently no effective solution to the above problems. Summary of the Invention

[0005] This application provides a language learning assessment system and method based on a data fusion model to solve the above-mentioned technical problems.

[0006] This application provides a language learning assessment system based on a data fusion model, comprising: a data acquisition module for acquiring learners' language writing text data and eye-tracking fixation data; and a decomposition module for decomposing the language writing text data into multiple semantic functional units, each semantic functional unit including a core carrier unit, a logical cohesion unit, and a supporting extension unit, and labeling each semantic functional unit with a dynamic logical density value; wherein, the core carrier unit includes sentence stem components for indicating subject-verb-object structures; the logical cohesion unit includes connecting components for expressing causal, adversative, and referential relationships; and the supporting extension unit includes components for... The system includes: a modifier module for the core carrier unit's attributive, adverbial, or complement components; a binding module for dynamically binding the eye-tracking fixation data to the semantic functional unit, generating a cognitive load vector including unit type, dynamic logical density, grammatical depth, fixation duration, and jump frequency; a judgment module for calculating the avoidance index of the core carrier unit based on the cognitive load vector, triggering a higher-order thinking avoidance strategy judgment instruction when the avoidance index exceeds a preset avoidance threshold and the activation rate of the supporting extension unit exceeds a preset activation threshold; and a report determination module for determining a comprehensive evaluation report in response to the higher-order thinking avoidance strategy judgment instruction.

[0007] This application provides a language learning assessment method based on a data fusion model, comprising: acquiring learners' language writing text data and eye-tracking fixation data; decomposing the language writing text data into multiple semantic functional units, each semantic functional unit including a core carrier unit, a logical cohesion unit, and a supporting extension unit, and labeling each semantic functional unit with a dynamic logical density value; wherein, the core carrier unit includes sentence stem components indicating subject-verb-object structures; the logical cohesion unit includes connecting components expressing causal, adversative, and referential relationships; the supporting extension unit includes attributive, adverbial, or complement components modifying the core carrier unit; dynamically binding the eye-tracking fixation data with the semantic functional units to generate a cognitive load vector including unit type, dynamic logical density, grammatical depth, fixation duration, and jump frequency; calculating the avoidance index of the core carrier unit based on the cognitive load vector; when the avoidance index exceeds a preset avoidance threshold and the activation rate of the supporting extension unit exceeds a preset activation threshold, triggering a higher-order thinking avoidance strategy judgment instruction; and determining a comprehensive assessment report in response to the higher-order thinking avoidance strategy judgment instruction.

[0008] Based on the embodiments provided in this application, by acquiring learners' language writing text data and eye-tracking fixation data and integrating and analyzing the two, the limitations of single-data-dimensional assessment are overcome. This approach can more comprehensively reflect learners' language learning status from two levels: the content of the language output itself and the learners' cognitive processes when processing language, thus providing richer evidence for the assessment results.

[0009] By decomposing language writing text data into core carrier units, logical cohesion units, and supporting extension units, and labeling them with dynamic logical density values, and then dynamically binding them with eye-tracking data to generate cognitive load vectors, the analysis of language texts goes beyond the surface and delves into the composition and logical relationships of semantic functional units. Combined with cognitively relevant data such as fixation duration and jump frequency, this allows for a more accurate capture of learners' characteristics in language use. Based on the cognitive load vector, an avoidance index for the core carrier units is calculated. When specific conditions are met, a higher-order thinking avoidance strategy judgment instruction is triggered, and a comprehensive assessment report is generated. This allows the assessment to specifically identify whether learners avoid core expressions or rely on extended components to circumvent higher-order thinking, resulting in a more targeted comprehensive assessment report that better serves the guidance and improvement of language learning. Attached Figure Description

[0010] The accompanying drawings, which are included to provide a further understanding of embodiments of the invention and form part of this application, illustrate exemplary embodiments of the invention and, together with their description, serve to explain the invention and do not constitute an undue limitation thereof. In the drawings:

[0011] Figure 1 This is a structural diagram of an optional language learning assessment system based on a data fusion model according to an embodiment of this application;

[0012] Figure 2 A flowchart illustrating an optional language learning evaluation method based on a data fusion model according to an embodiment of this application;

[0013] Figure 3 A flowchart illustrating another optional language learning evaluation method based on a data fusion model according to an embodiment of this application;

[0014] Figure 4 This is a schematic diagram of the structure of an optional electronic device according to an embodiment of this application.

[0015] The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. 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 a part of the embodiments of the present invention, and not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0017] According to one aspect of an embodiment of this application, a language learning assessment system based on a data fusion model is provided. For example... Figure 1 As shown, the system includes:

[0018] Data acquisition module 101 is used to acquire learners' language writing text data and eye movement fixation data;

[0019] The decomposition module 102 is used to decompose the language writing text data into multiple semantic functional units. These semantic functional units include core carrier units, logical connection units, and supporting extension units, and each semantic functional unit is labeled with a dynamic logical density value. Among them, the core carrier unit includes the main components of the sentence that indicate the subject-verb-object structure; the logical connection unit includes the connecting components that express causal, adversative, and referential relationships; and the supporting extension unit includes the attributive, adverbial, or complement components that modify the core carrier unit.

[0020] The binding module 103 is used to dynamically bind eye-tracking gaze data with semantic functional units to generate a cognitive load vector that includes unit type, dynamic logical density, syntactic depth, gaze duration and jump frequency.

[0021] The judgment module 104 is used to calculate the avoidance index of the core carrier unit based on the cognitive load vector. When the avoidance index exceeds the preset avoidance threshold and the activation rate of the supporting extension unit exceeds the preset activation threshold, a higher-order thinking avoidance strategy judgment instruction is triggered.

[0022] The report identifies module 105, which is used to determine a comprehensive evaluation report in response to the higher-order thinking avoidance strategy judgment instruction.

[0023] According to another aspect of the embodiments of this application, such as Figure 2 As shown, this application also provides a language learning assessment method based on a data fusion model, including:

[0024] S201, Obtain learners' language writing text data and eye movement fixation data;

[0025] The data fusion model is used to collaboratively process language writing text data and eye movement fixation data, specifically including S202 and S203;

[0026] S202 decomposes language writing text data into multiple semantic functional units, which include core carrier units, logical connection units, and supporting extension units, and marks each semantic functional unit with a dynamic logical density value.

[0027] The core carrier unit includes the main sentence components that indicate the subject-verb-object structure; the logical connection unit includes the connecting components that express causal, adversative, and referential relationships; and the supporting extension unit includes the attributive, adverbial, or complement components that modify the core carrier unit.

[0028] Among them, the core carrier unit constitutes the main components of the sentence's basic meaning and is the core information carrier of language expression. For example, in "He completed his homework," "he" (subject), "completed" (predicate), and "homework" (object) together constitute the core carrier unit, directly expressing the basic meaning of "who did what."

[0029] Logical cohesive units connect different semantic units and embody logical relationships, ensuring the coherence of language expression. For example, in "Although he was very tired, he still insisted on completing his homework," "Although...but..." is a logical cohesive unit expressing a contrast; in "Because it was raining, he brought an umbrella," "Because...therefore..." is a logical cohesive unit expressing a cause-and-effect relationship.

[0030] Supporting extension units include components that modify and supplement the core carrier unit, enriching the details of the expression. For example, in the sentence "He quickly completed the complex task," "quickly" (adverbial) modifies "completed," and "complex" (attributive) modifies "task." Both are supporting extension units, supplementing the state of "completed" and the characteristics of "task."

[0031] It should be explained that the dynamic logical density value refers to a comprehensive quantitative value of the tightness of logical connections and syntactic structural complexity within a semantic functional unit. It is used to reflect the logical carrying capacity of the unit in language expression, and its value changes dynamically with the specific content of the unit. The dynamic logical density value is equal to the number of logical connectors within the unit multiplied by the depth of the syntax tree.

[0032] In this context, grammar tree depth refers to the number of levels from the sentence stem (core carrier unit) to the semantic functional unit, reflecting the complexity of the nested grammatical structure. For example, in the sentence "Xiaoming is carefully reading the extracurricular book assigned by the teacher yesterday," the core carrier unit is "Xiaoming is reading the book" (grammar tree depth is 1); the supporting extension unit "carefully" (adverbial) directly modifies "reading," with a grammar tree depth of 2; and "assigned by the teacher yesterday" (attributive) modifies "extracurricular book," nested under the level of "book," with a grammar tree depth of 3.

[0033] If a logical connector unit contains two logical connectives (such as "not only... but also..."), and its syntax tree depth is 2, then the dynamic logical density value is 2×2=4; if another supporting extension unit contains one logical connective (such as "because"), and its syntax tree depth is 3, then the dynamic logical density value is 1×3=3.

[0034] S203, dynamically binds eye-tracking fixation data with semantic functional units to generate a cognitive load vector that includes unit type, dynamic logical density, syntactic depth, fixation duration and jump frequency;

[0035] S204, based on the cognitive load vector, calculates the avoidance index of the core carrier unit. When the avoidance index exceeds the preset avoidance threshold and the activation rate of the supporting extended unit exceeds the preset activation threshold, a higher-order thinking avoidance strategy judgment instruction is triggered.

[0036] The avoidance index of the core carrier unit reflects the degree to which learners avoid using the core carrier unit (or simplify its complexity) in language expression. It is calculated using features related to the core carrier unit in the cognitive load vector. The index is then compared with the activation status of supporting extension units, taking into account the frequency of use of the core carrier unit, dynamic logical density, and corresponding eye-tracking data (e.g., shorter fixation duration may indicate simplification).

[0037] For example, if a learner uses only 5 complete core carrier units in 10 sentences (the rest are replaced with omissions or simple sentences), and the average dynamic logic density of these core carrier units is 2 (lower than the normal level of 3), then the avoidance index = (1-5 / 10)×(1-2 / 3) = 0.5×0.33≈0.17; if the fixation time of most core carrier units is 40% shorter than the average, then the avoidance index may be increased to 0.45.

[0038] The preset avoidance threshold is a critical value used to determine whether there is any behavior of avoiding core carrier units, and it is determined by statistical analysis of a large amount of language learning data. For example, the preset avoidance threshold can be set to 0.6 or 0.7. If a learner's avoidance index is calculated to be 0.65 (above 0.6), it is preliminarily determined that they may have a tendency to avoid core carrier units.

[0039] The activation rate of supporting extension units refers to the proportion of supporting extension units actually used by learners to the total number of reasonably usable supporting extension units in their language expression, reflecting the degree of dependence on modifiers. For example, if there are a total of 10 reasonably usable supporting extension units (such as attributives and adverbs) in a text, and the learner actually uses 8, then the activation rate = 8 / 10 = 0.8.

[0040] The preset activation threshold is a critical value used to determine whether the support extension unit is overactivated. For example, the preset activation threshold can be set to 0.8 or 0.85. If the activation rate of the support extension unit is 0.82 (exceeding 0.8) and the avoidance index exceeds the preset threshold, it is further determined that there may be higher-order thinking avoidance.

[0041] In one specific implementation, the avoidance index is calculated based on the following formula:

[0042] AI=ω1×(1-DLDstandard / DLD actual )+ω2×(1-GD standard / GD actual )+ω3×(1-FD base / FD actual )+ω4×(JT base / JT actual )

[0043] Among them, AI is the avoidance index (0 to 1) of the core carrier unit, and the larger the value, the more significant the avoidance tendency; ω1, ω2, ω3, and ω4 are weight coefficients (summing up to 1). In this embodiment, the values ​​are ω1 = 0.35, ω2 = 0.35, ω3 = 0.15, and ω4 = 0.15, reflecting the core position of text structural features; DLD actual The dynamic logical density (logical nodes / sentences) of the core carrier units actually used by learners, such as the ratio of the number of subject-verb-object relationships to the sentence length; DLD standard The standard dynamic logical density (logical nodes / sentences) of the core carrier units in this learning stage is based on the pre-set teaching syllabus; GD actual The grammatical depth (levels) of the core carrier units actually used by learners, such as the number of nested clauses (levels 1 to 5); GD standard This represents the standard grammatical depth (level) of the core carrier unit in this learning stage; FD actual The actual fixation time (seconds) of the learner on the core carrier unit; FD base The baseline fixation duration (seconds) for learners of similar skill levels on similar core carrier units is based on historical data statistics; JT actual This refers to the actual number of jumps made by learners on the core carrier unit (times / minute, times / min); JT base The baseline jump frequency (times / minute, times / min) of learners at the same level to the same core carrier unit.

[0044] S205, in response to the higher-order thinking avoidance strategy judgment instruction, determines the comprehensive assessment report.

[0045] It should be noted that when both the avoidance index and the activation rate of the supporting extension unit exceed the corresponding threshold, a signal is used to trigger the subsequent comprehensive evaluation, indicating that the learner may be trying to escape deep logical expression by relying on the supporting extension unit and avoiding the core carrier unit.

[0046] Furthermore, such as Figure 3 As shown, before determining the comprehensive evaluation report in response to the higher-order thinking avoidance strategy assessment instruction, the method also includes:

[0047] S301, construct a dynamic knowledge topology network, where nodes represent grammatical knowledge points, edges represent logical dependencies between knowledge points, and edge weights are dynamically adjusted according to the frequency of error co-occurrence.

[0048] The nodes in the dynamic knowledge topology network include basic knowledge point nodes and related higher-order knowledge point nodes.

[0049] In some embodiments, edge weights are used to characterize the tightness of the logical dependency between two syntactic knowledge points, and the specific logic for dynamically adjusting them according to the frequency of error co-occurrence is as follows:

[0050] Initial settings: Based on the inherent logical relationships of the grammatical system (such as the strong dependency between "noun singular and plural" and "subject-verb agreement"), the initial weights are set to a base value of 1 to 5 (the stronger the dependency, the higher the initial weight).

[0051] Adjustment rules: When the system detects that two knowledge points co-occur simultaneously in a learner's written text (i.e., errors co-occur), the edge weight increases by a fixed increment (e.g., 0.1) for each co-occurrence. If no co-occurrence occurs in multiple consecutive learning tasks, the edge weight decreases at a certain decay rate (e.g., 0.05 per task) until it returns to its initial base value. For example, the initial weight for "simple past tense" and "verb past tense conjugation" is 3. If both are incorrect in three consecutive tasks, the edge weight is adjusted to 3 + 0.1 × 3 = 3.3. If no co-occurrence occurs in the subsequent five tasks, the edge weight decreases to 3.3 - 0.05 × 5 = 3.05, dynamically reflecting the actual strength of the error association between knowledge points.

[0052] S302, when the higher-order thinking escape strategy judgment instruction is triggered, high-risk defect nodes are marked through the cognitive impulse dynamics model;

[0053] S303, for the marked high-risk defect nodes, retrieve the historical cognitive load vectors of all semantic functional units associated with the node; if the average gaze duration exceeds the first threshold and the jump frequency is lower than the second threshold, it is confirmed as a root knowledge defect; otherwise, it is judged as an accidental error.

[0054] The first threshold corresponds to the average fixation duration (reflecting the degree of attention paid to the knowledge point), and the second threshold corresponds to the jump frequency (reflecting the coherence of thought). The average fixation duration and jump frequency are determined based on the historical cognitive load vector of all semantic functional units associated with the node.

[0055] For example, the first threshold can be set to 2 seconds or 2.5 seconds (if the average gaze duration of the semantic functional unit associated with a high-risk defect node is 3 seconds, which exceeds 2 seconds); the second threshold can be set to 3 times or 4 times (if the jump frequency of the unit associated with the node is 2 times, which is less than 3 times).

[0056] If all of the above conditions are met, it is judged as a fundamental knowledge deficiency (such as the learner not having mastered the grammar point of "subject-verb agreement"); if the average fixation time is 1.5 seconds (less than 2 seconds) or the jump frequency is 4 times (more than 3 times), it is judged as an accidental error (such as careless spelling).

[0057] It should be explained that the determination of root knowledge deficits is based on the correlation between eye movements and thought processes in cognitive psychology:

[0058] Long fixation duration: When learners encounter knowledge points that they have not mastered or find difficult to understand, the brain needs to invest more cognitive resources in processing (such as recalling and analyzing). This is reflected in eye movements by an extended fixation time on the semantic functional units associated with the knowledge point, which is a direct manifestation of cognitive effort.

[0059] Low jump frequency: If there are fundamental defects in the knowledge points, learners' thinking is likely to get stuck in that unit and it is difficult to switch smoothly to other related units (due to a lack of sufficient knowledge reserves to support logical connections), which leads to a decrease in the jump frequency of eye movement trajectory.

[0060] The combination of these two factors (long fixation + low jump) indicates that the learner's mistake was not accidental (accidental mistakes are usually accompanied by short fixation and high jump because the thinking did not stop), but rather that there was a stable obstacle to understanding or applying the knowledge point, i.e., a fundamental knowledge deficiency.

[0061] S304, when the cognitive load value of the basic knowledge point node is higher than the cognitive load value of the associated higher-order knowledge point node in three consecutive learning tasks, and the difference ratio exceeds the preset ratio of the total load, a cognitive resource crowding warning is generated; wherein, the comprehensive evaluation report is determined, including: generating a comprehensive evaluation report based on the judgment result of the avoidance strategy, the confirmation result of the root knowledge deficiency and the cognitive resource crowding warning.

[0062] Cognitive load is a quantitative representation of the cognitive burden learners bear when processing specific grammatical knowledge points. Its calculation is based on the core features of the cognitive load vector and is achieved by integrating language task complexity and cognitive process indicators. The specific logic is as follows:

[0063] Input features: Extract the dynamic logical density (reflecting the logical complexity of the knowledge point), grammatical depth (reflecting the difficulty of structural nesting), average fixation duration (reflecting the level of cognitive input), and jump frequency (reflecting the fluency of thought) of the semantic functional units associated with the target grammatical knowledge points from the cognitive load vector.

[0064] The computational logic is as follows: First, each feature is standardized (e.g., dynamic logic density and grammatical depth are mapped to a complexity range of 0-5, and fixation duration and jump frequency are mapped to a cognitive process range of 0-5). Then, based on the weighted impact of each feature on cognitive load (e.g., dynamic logic density and grammatical depth have higher weights because they are directly related to the difficulty of the knowledge point), a weighted sum is calculated to obtain the cognitive load value. For example, if a certain grammatical knowledge point has a dynamic logic density of 3, a grammatical depth of 2, an average fixation duration of 4, and a jump frequency of 1, and the weights are 0.3, 0.3, 0.2, and 0.2 respectively, then the cognitive load value = 3 × 0.3 + 2 × 0.3 + 4 × 0.2 + 1 × 0.2 = 0.9 + 0.6 + 0.8 + 0.2 = 2.5. The higher this value, the more cognitive resources the learner needs to invest in understanding or applying the knowledge point, and the heavier the burden.

[0065] Total workload refers to the sum of the cognitive load values ​​of basic knowledge point nodes and related higher-order knowledge point nodes in a learning task.

[0066] The preset percentage of the total load is used to determine whether basic knowledge points are crowding out cognitive resources for higher-level knowledge points; for example, the preset percentage can be set to 30% or 40%. If, in three consecutive tasks, the cognitive load value of basic knowledge points (1.2) is 0.5 higher than that of higher-level knowledge points (2.8) (difference of 0.5), and the difference accounts for 12.5% ​​of the total load (4) (less than 30%), then no warning is issued; if the difference is 1.6, accounting for 1.6 / 4 = 40% (reaching the preset percentage), then a cognitive resource crowding warning is generated (indicating that the learner is consuming too much energy on basic knowledge points, affecting the learning of higher-level knowledge points).

[0067] The preset ratio is based on statistical calibration of the cognitive resource allocation patterns in language learning. The process is as follows: Cognitive load values ​​of learners at different levels on basic knowledge points and related higher-order knowledge points are collected, and the difference between the two and the corresponding learning outcomes (such as the speed of mastering higher-order knowledge points and the accuracy of application) are recorded. When the cognitive load value of basic knowledge points is consistently higher than that of higher-order knowledge points, and the difference accounts for a small proportion of the total load (e.g., <20%), learners can still allocate resources normally to higher-order knowledge points, and the learning outcome is not significantly affected. When the proportion exceeds a certain critical value (e.g., 30%), basic knowledge points consume too many cognitive resources, leading to insufficient learning input on higher-order knowledge points, and a significant decrease in their mastery speed and accuracy. The preset ratio (e.g., 30% or 40%) is used as the "critical proportion where learning outcomes show a significant decline," and is continuously verified and adjusted through a large amount of subsequent learning data to ensure that the warning accurately reflects the resource crowding out of higher-order knowledge points by basic knowledge points.

[0068] Based on the embodiments provided in this application, a dynamic knowledge topology network is constructed, with grammatical knowledge points as nodes and logical dependencies between knowledge points as edges, the edge weights being dynamically adjusted according to the frequency of error co-occurrence. When a higher-order thinking avoidance strategy judgment instruction is triggered, a cognitive impulse dynamics model is used to mark high-risk defect nodes. Then, combined with historical cognitive load vectors, root-cause knowledge deficiencies and accidental errors are distinguished. Simultaneously, the cognitive load values ​​of basic and higher-order knowledge points are monitored to generate a cognitive resource crowding-out warning. Finally, a report is generated by integrating multiple results. This technical solution allows assessment to go beyond superficial language performance and delve into the structural level of the knowledge system. It can accurately locate weak links and cognitive resource allocation problems in the learner's knowledge system, making the assessment report more instructive and helping learners to specifically address knowledge deficiencies and optimize cognitive resource allocation.

[0069] Furthermore, after marking high-risk defect nodes, a logical fault depth tracing operation is performed, including:

[0070] Starting with the core carrier unit that is avoided, the dynamic knowledge topology network is traversed, and all logical connection units that directly depend on the starting node are extracted to form a primary dependency path.

[0071] It should be noted that the avoided core carrier units refer to the core carrier units (sentence cores) that learners should use in language expression but do not use fully (or use in a simplified way). This is usually manifested as avoiding complex core structures, replacing them with incomplete cores, or reducing their frequency of use.

[0072] If a topic (such as "describe the experimental process") should normally include a complete core carrier unit of "subject (experimenter) + predicate (operation) + object (experimental object)", but the learner only uses "did the experiment" (omitting the subject and simplifying the predicate), and this type of simplification occurs frequently (the avoidance index exceeds the preset threshold), then the core carrier unit "experimenter operates the experimental object" is considered to be avoided; or if the dynamic logical density of a core carrier unit is significantly lower than the average of learners at the same level, it is also considered to be avoided.

[0073] In a dynamic knowledge topology network, logical connecting units have a direct logical dependency on the starting node (the grammatical knowledge point node corresponding to the avoided core carrier unit) (without intermediate nodes). If the starting node is a "subject-verb-object core structure" (corresponding to the core carrier unit), the logical connecting units that directly depend on it include "referential words that follow the preceding text (such as 'this' and 'its')" and "causal words that connect the preceding and following sentences (such as 'therefore')"—these logical connecting units must be based on the "subject-verb-object core structure" to function (without a core structure, the connection is meaningless), hence the direct dependency.

[0074] For each logical connection unit in the primary dependency path, retrieve its downstream supporting extension units in the dynamic knowledge topology network to form a complete dependency link;

[0075] It should be explained that the downstream support extension unit refers to the support extension unit located after the current logical connection unit in the logical dependency chain of the dynamic knowledge topology network, and is directly or indirectly supported by it (that is, the existence or function of the support extension unit depends on the current logical connection unit).

[0076] For example, if the logical connecting unit is "because" (expressing a causal relationship), its downstream supporting extension unit may be "due to the cold weather" (an adverbial phrase used to explain the reason). "Due to the cold weather" is a supporting extension unit, which needs to be associated with the core carrier unit (such as "he wore thick clothes") through the logical connecting unit "because", so it is a downstream supporting extension unit of "because".

[0077] Obtain eye-tracking fixation data corresponding to each semantic functional unit in the complete dependency chain, and calculate the focus maintenance rate of each semantic functional unit in the complete dependency chain; wherein, the focus maintenance rate is obtained by dividing the fixation duration by the sum of the fixation duration and the total time of jumping to units outside the chain;

[0078] In this context, "out-of-chain units" refer to semantic functional units outside the complete dependency chain (excluding non-semantic areas such as whitespace outside text and images). A complete dependency chain consists of the avoided core carrier unit, the directly dependent logical connection unit, and the downstream supporting extension unit, in that order. Out-of-chain units are other core carrier units, logical connection units, or supporting extension units that do not belong to this chain. For example, if the complete dependency chain is "(Avoided core) Xiaoming eats fruit → (Logical connection) Because → (Downstream support) He likes sweet things," then out-of-chain units could be "Xiaohong drinks juice" (other core carrier unit), "But" (other logical connection unit), or "Very thirsty" (other supporting extension unit), excluding whitespace areas at the page edges.

[0079] Identify semantic functional units in the complete dependency chain whose focus maintenance rate is lower than the average value of the link unit and also lower than the average value of the successor node as logical breakpoints.

[0080] In this context, a successor node refers to the adjacent unit (the next unit immediately following the current semantic functional unit) in a complete dependency chain. For example, if the complete dependency chain is A (avoided core carrier unit) → B (logical connection unit) → C (support extension unit) → D (support extension unit), then the successor node of B is C, the successor node of C is D, and D has no successor node.

[0081] The link unit average refers to the average focus maintenance rate of all semantic functional units (including the breakpoint itself) in a complete dependency chain. Breakpoint data is included in the calculation to ensure an accurate reflection of the overall link performance. For example, if a complete dependency chain has 3 units with focus maintenance rates of 0.6 (A), 0.4 (B, the breakpoint), and 0.7 (C), then the link unit average = (0.6 + 0.4 + 0.7) / 3 = 0.57 (including data from breakpoint B).

[0082] A fault impact map, including logical breakpoint locations and complete dependency links, is generated and stored in the evaluation database. The fault impact map is visualized in the form of a topological network, with breakpoint semantic functional units highlighted in red and complete dependency links connected by bold arrows.

[0083] In one specific implementation, the focus maintenance rate of the semantic functional unit is determined based on the following formula:

[0084]

[0085] Among them, FMR u FD represents the focus maintenance rate (0 to 1) of the u-th semantic functional unit; a higher value indicates more stable attention. u FC is the learner's total fixation time (in seconds) on the u-th unit; u OT represents the number of fixations made by the learner on the u-th unit; u The total time (in seconds) for a learner to jump from the u-th unit to a unit outside the complete dependency chain; IT u The total time (in seconds) for the learner to jump from the u-th unit to other units within the complete dependency chain; k is the in-chain jump correction coefficient, which is set to k = 0.3 in this embodiment to distinguish between normal logical jumps and irrelevant distractions.

[0086] Based on the embodiments provided in this application, after marking high-risk defect nodes, a deep logical fault tracing operation is performed. Starting with the avoided core carrier unit as the initial node, the dynamic knowledge topology network is traversed to extract primary dependency paths and form complete dependency links. The focus maintenance rate of each semantic functional unit is calculated, logical breakpoints are identified, and a fault impact map is generated and stored. This achieves accurate tracing of the location and scope of influence of logical faults from two dimensions: knowledge dependency relationships and learner eye-tracking focus changes. It clearly presents the specific links of logical breaks in language expression, providing a clear direction for subsequent targeted repair of logical defects and helping learners improve the logic and coherence of their language expression.

[0087] Furthermore, based on the stored fault impact map, logical chain repair verification is performed in subsequent learning tasks, including:

[0088] When the complete dependency link recorded in the fault impact map reappears, monitor in real time the change in focus maintenance rate of the breakpoint semantic functional unit that matches the logical breakpoint.

[0089] By analyzing whether the semantic functional unit of the breakpoint contains the recurrence or referential relationship of the core words of its direct upstream semantic functional unit, the semantic association strength between the semantic functional unit of the breakpoint and its direct upstream semantic functional unit is detected.

[0090] It should be noted that a direct upstream semantic functional unit refers to the adjacent unit (the unit immediately preceding the current breakpoint semantic functional unit) in a complete dependency chain, which has a direct logical connection with the current unit. For example, in a complete dependency chain A→B→C→D, where C is the breakpoint semantic functional unit, the direct upstream semantic functional unit of C is B (B is directly connected to C, and B's logic directly supports C).

[0091] If the focus maintenance rate of the breakpoint semantic functional unit is improved to above the average value of the link units and the semantic association strength exceeds the preset repair threshold, the path logic coherence repair will be marked as qualified in the comprehensive evaluation report.

[0092] The preset repair threshold is a critical value (ranging from 0 to 1, with higher values ​​indicating stronger association) used to determine whether the semantic association strength between the breakpoint semantic functional unit and its direct upstream unit meets the standard. For example, the preset repair threshold can be set to 0.7 or 0.8. If the semantic association strength between the two is calculated to be 0.75 (e.g., the breakpoint unit contains a repetition of the core vocabulary of the upstream unit), exceeding 0.7, then the association is considered to meet the standard.

[0093] If the target is not met, an enhanced intervention mechanism will be activated, inserting a mandatory writing task between the breakpoint semantic functional unit and its direct upstream semantic functional unit in subsequent tasks and shortening the task interval.

[0094] Among them, the mandatory associative writing task refers to a specific writing task designed for the breakpoint unit that has not met the repair standard and its direct upstream unit. Learners are required to use both units simultaneously in the task to strengthen the logical connection between them. For example, if the breakpoint unit is "fast" (supporting extension unit) and the direct upstream unit is "run" (predicate of the core carrier unit), the mandatory associative writing task could be "Write a sentence using 'run' and 'fast' to explain 'run'."

[0095] Based on the embodiments provided in this application, and using the stored fault impact map, logical chain repair verification is performed in subsequent learning tasks. The success of logical coherence repair is determined by monitoring changes in focus maintenance rate and detecting semantic association strength; if it fails to meet the standards, a reinforcement intervention mechanism is activated. This achieves dynamic tracking and verification of the logical chain repair process, enabling timely confirmation of repair effectiveness and targeted reinforcement intervention for cases that do not meet the standards, ensuring effective repair of logical faults and helping learners gradually improve the logical chains in their language expression.

[0096] Furthermore, before determining higher-order thinking avoidance strategies, cognitive intention verification is performed, including:

[0097] The detection function checks whether the gaze duration of the support extension unit is lower than a specific percentage of the historical baseline.

[0098] The historical baseline refers to the average fixation duration of learners when processing similar supporting extension units in the past (based on historical eye movement data statistics). For example, if a learner's average fixation duration for processing 5 similar adverbial units (such as "carefully" and "slowly") was 2 seconds, then the historical baseline is 2 seconds.

[0099] The specific percentage refers to the threshold used to determine whether the current fixation duration is abnormally low. For example, the specific percentage can be set to 30% or 40%, meaning that subsequent verification is triggered when the current fixation duration is lower than 70% (2 × 70% = 1.4 seconds) or 60% (2 × 60% = 1.2 seconds) of the historical baseline value.

[0100] When the fixation duration is detected to be lower than a certain percentage of the historical baseline, a visual stimulus marker is injected at the end of the text in that unit and a timer is started; the eye movement focus is monitored within a preset time window to see if it returns to the marked area.

[0101] Among them, visual stimulus markers include visually distinctive markers added at the end of the supporting extension unit text (distinct from the text content, used to attract attention). The form can be based on the interactive interface design, such as red asterisks (), yellow underlines, short flashing borders, etc.

[0102] For example, if the supporting extension unit is "in a bright classroom", then add the label "in a bright classroom" at the end of it, where "*" is a visual stimulus label, which is distinguished from the text by color or dynamic effects.

[0103] Monitoring whether the eye movement focus returns to the marked area refers to detecting whether the learner's eye movement trajectory moves from other areas (such as other parts of the text, the edge of the interface) to the location of the visual stimulus mark (i.e., the mark at the end of the unit text) within a preset time window (e.g., 3 seconds). For example, if the mark is at the end of "in a bright classroom*", if the learner's eye movement focus first falls on "bright", then moves away to "outside the window", and returns to "" within 3 seconds, it is considered "returning to the marked area"; if the learner does not look at "*" at all, it is considered not to have returned.

[0104] If the focus returns and the dwell time exceeds the time threshold, it is marked as attention drift behavior and the unit is excluded from the decision process for escape strategy.

[0105] The duration threshold is a critical value used to determine whether the eye movement's focus returning to the marked area constitutes a "valid return" (reflecting whether the learner actually noticed the mark). For example, the duration threshold can be set to 0.5 seconds or 0.8 seconds. If the focus returns to the marked area and remains for 0.6 seconds (more than 0.5 seconds), it is considered a valid return; if it only remains for 0.3 seconds (less than 0.5 seconds), it is considered an invalid return.

[0106] It should be noted that if the fixation duration of the extended unit is lower than the historical baseline (possibly due to distraction and insufficient attention), but the eye focus returns to the visual stimulus marker and stays for more than the threshold, it indicates that the learner is not intentionally ignoring the unit, but rather that their attention has temporarily deviated (such as a brief lapse in concentration), which is attention drift (rather than active avoidance).

[0107] The determination of higher-order thinking avoidance strategies should be based on learners’ “active avoidance” behavior. Attention drift is a passive distraction, not an active strategy. Including it in the determination will lead to misjudgment. Therefore, this unit should be excluded to ensure the accuracy of the assessment.

[0108] If no valid return behavior is detected, the unit is retained to participate in the decision-making process for the escape strategy.

[0109] Based on the embodiments provided in this application, cognitive intent verification is performed before determining higher-order thinking avoidance strategies. This is achieved by detecting the fixation duration of supporting extended units and monitoring eye movement focus return by injecting visual stimulus markers, thus distinguishing between attention drift behavior and units involved in the avoidance strategy determination. This eliminates misjudgments caused by inattention, more accurately identifies units truly involved in higher-order thinking avoidance strategies, makes the determination results of higher-order thinking avoidance strategies more reliable, and improves the accuracy of the assessment.

[0110] Furthermore, when an attention drift behavior is identified, the presentation duration of the unit in subsequent tasks is extended to a specific multiple of the baseline value; a grammatical structure hint layer is dynamically generated in the interactive interface to cover the text area of ​​the unit.

[0111] The specific multiple by which the presentation duration is extended to the baseline value refers to a fixed proportional multiple by which the display time of the semantic functional unit in subsequent learning tasks is extended to its historical baseline duration. This multiple is set according to the severity of attention drift, and is usually 1.5-2 times. For example, if the historical baseline presentation duration of a support extension unit is 2 seconds (the average time for learners to normally process this unit), and the specific multiple is set to 1.5 times, then the extended presentation duration is 3 seconds; if the drift is more severe (such as multiple occurrences), the multiple can be set to 2 times, that is, 4 seconds of presentation.

[0112] It should be noted that attention drift indicates that learners have not paid enough attention to the unit (short fixation time). Extending the presentation time can give them more time to process information, reduce comprehension bias caused by insufficient time, strengthen cognitive memory of the unit, and make up for the learning gap caused by attention drift.

[0113] When retaining the decision-making process for participating in the avoidance strategy, the preset avoidance threshold is lowered based on the error propagation weight of the unit in the dynamic knowledge topology network; the occurrence interval of similar semantic functional units in subsequent learning tasks is shortened; and the influence weight value of related high-order knowledge point nodes in the dynamic knowledge topology network is increased.

[0114] Among them, error propagation weight refers to the probability that an error in a certain semantic functional unit will cause errors in other related units. The higher the weight, the greater the impact of the error in that unit on the overall language expression logic (e.g., an error in a key logical connecting unit can easily lead to a break in the logic of the entire sentence).

[0115] In one embodiment, the initial preset avoidance threshold is assumed to be 0.6. If the error propagation weight of a certain supporting extension unit (such as "although...but...") is 0.8 (high weight, its error is likely to cause logical confusion in subsequent sentences), the avoidance threshold is lowered to 0.5 (i.e., a judgment is triggered when the avoidance index exceeds 0.5); if the error propagation weight is 0.3 (low weight), it is lowered to 0.55. By differentially lowering the threshold, the system becomes more sensitive to the avoidance behavior of high-impact units, avoiding missed judgments due to the avoidance of key units.

[0116] It's important to explain that shortening the intervals between similar semantic functional units (such as logical connectors like "because...therefore...") essentially strengthens learners' cognitive and application abilities of these units through "high-frequency repetition." If a unit is retained in the avoidance strategy assessment, it indicates that learners may have a tendency to actively avoid it. Shortening the interval (e.g., from once in five tasks to once in three tasks) increases practice opportunities, forcing learners to engage with and use it more, reducing avoidance space, and gradually improving their ability to apply these units.

[0117] When the influence weight of related higher-order knowledge points (such as "complex sentence logic" versus "simple sentence structure") is increased, their priority in the dynamic knowledge topology network rises, and the system will place greater emphasis on these knowledge points in assessment and subsequent task design. This is because learners may avoid higher-order thinking by avoiding core carrier units. Increasing the weight of related higher-order knowledge points can guide assessment and learning tasks to focus more on training higher-order thinking skills, forcing learners to confront higher-order knowledge points directly and reducing the motivation to avoid them.

[0118] Based on the embodiments provided in this application, when an attention drift behavior is identified, the subsequent presentation duration of the unit is extended and a grammatical structure prompt layer is generated; when the avoidance strategy judgment process is retained, the preset avoidance threshold is lowered, the interval between similar units is shortened, and the influence weight of related higher-order knowledge points is increased. This technology achieves differentiated processing based on different judgment results. For attention drift behavior, the learner's attention and understanding of the unit are strengthened by adjusting the presentation duration and providing prompts; for cases that may involve avoidance strategies, the training of such units and related knowledge points is enhanced by adjusting parameters, thereby improving learning effectiveness.

[0119] Furthermore, high-risk defect nodes are labeled using a cognitive impulse dynamics model, including:

[0120] The nodes of the dynamic knowledge topology network are transformed into spiking neurons. The basic firing frequency of a spiking neuron is jointly determined by the number of historical errors and the grammatical depth of the node. The connecting edges between nodes are transformed into synapses. The synaptic strength is dynamically adjusted by the error co-occurrence frequency of the associated nodes.

[0121] It needs to be explained that a spiking neuron refers to each grammatical knowledge point node (such as "subject-verb agreement" or "tense change") in a dynamic knowledge topology network, which is transformed into a computational unit that simulates a biological neuron. Its "firing" (pulse firing) represents that the knowledge point is activated or an error occurs that needs to be processed.

[0122] The logical dependencies (edges) between nodes are transformed into synapses. The higher the synapse strength, the more frequently errors co-occur between the two knowledge points (e.g., errors in "past tense" and "verb inflection" often occur simultaneously, indicating high synapse strength).

[0123] The basic firing frequency is determined by the number of historical errors and the grammatical depth of a node: the more historical errors (indicating that the learner is more prone to making mistakes) and the deeper the grammatical depth (the more complex the structure), the higher the basic firing frequency. For example, the "present perfect tense" (grammatical depth 3) has 10 historical errors, so the basic firing frequency is set to 5 times / second; the "simple present tense" (grammatical depth 1) has 3 historical errors, so the basic firing frequency is set to 2 times / second.

[0124] The gaze duration corresponding to the semantic functional unit is positively encoded as the neural stimulation intensity, and the jump frequency is negatively encoded as the inhibition signal.

[0125] It should be explained that a long fixation duration (e.g., 3 seconds) indicates that the learner has invested a lot of cognitive resources in this unit, encoded as a high-intensity stimulus (e.g., 5); a short fixation duration (e.g., 0.5 seconds) is encoded as a low-intensity stimulus (e.g., 1). A high jump frequency (e.g., 5 times) indicates that the learner's thinking is disjointed and lacks coherence, encoded as a strong inhibitory signal (e.g., -3); a low jump frequency (e.g., 1 time) is encoded as a weak inhibitory signal (e.g., -1).

[0126] Pulse dynamics were run within a 500-millisecond simulation time. Each neuron received upstream pulses and eye-tracking stimulation signals, and triggered its own pulse when the accumulated potential exceeded the potential threshold. Neurons that fired high-frequency pulses were marked as hyperactive nodes.

[0127] If the pulse emission frequency of a node exceeds a preset multiple of the average frequency of the entire network, it is determined to be a high-risk defective node.

[0128] The trigger pulse, which is the critical value used to determine whether a neuron is activated, is triggered when the accumulated neuronal potential exceeds a potential threshold. For example, the potential threshold can be set to -55mV or -60mV (simulating the resting potential of a biological neuron). If the accumulated neuronal potential rises from -70mV (resting) to -50mV (exceeding -55mV), a pulse is triggered; if it only rises to -58mV (below -55mV), no pulse is triggered.

[0129] In high-frequency firing pulses, "high frequency" refers to the actual firing frequency, which is more than 1.5 times the baseline firing frequency of the neuron. For example, a neuron with a baseline firing frequency of 5 Hz that fires 8 Hz (8 > 5 × 1.5 = 7.5) is considered to be firing at a high frequency; a neuron with a baseline firing frequency of 2 Hz that fires 4 Hz (4 > 2 × 1.5 = 3) is also considered to be firing at a high frequency.

[0130] The average frequency across the entire network refers to the average firing frequency of all spiking neurons in the dynamic knowledge topology network. It is obtained by summing the real-time firing frequencies of all neurons and dividing by the number of neurons. For example, if the total firing frequency of 10 neurons is 50 times / second, the average frequency across the entire network is 5 times / second. The preset multiple is a critical multiple used to determine whether a node is a high-risk defect node; for example, it can be set to 1.5 times or 2 times. If a neuron fires at a frequency of 9 times / second, and the average frequency across the entire network is 5 times / second, then 9 > 5 × 1.5 = 7.5, and it is determined to be a high-risk defect node.

[0131] In one specific implementation, the fundamental firing frequency of the spiking neuron is calculated based on the following formula:

[0132] BFRn =α×HEC n ×GD n ×NIC n

[0133] Among them, BFR n The basic firing frequency (times per second) of the spiking neuron corresponding to the nth node in the dynamic knowledge topology network. -1 ); α is the proportionality coefficient (s) -1 • Level-1), in this embodiment, the value α = 0.02 is used to convert the parameter product into a physiologically reasonable frequency unit; HEC n This represents the historical error count (times) for the nth node, such as the cumulative error record for "subjunctive mood"; GD n The grammatical depth (level) corresponding to the nth node is 1 to 5 (e.g., "simple present tense" is level 1, "non-restrictive relative clause" is level 3); NIC n The network importance coefficient of the nth node (ranging from 0.5 to 1.5) is calculated based on the node's connectivity in the dynamic knowledge topology network (the denser the connections, the higher the value).

[0134] Based on the above formula, a network importance coefficient is introduced to solve the problem of distinguishing nodes with the same number of errors but different importance (e.g., "subject-verb agreement" is more important than "punctuation"), making the labeling results more aligned with teaching priorities; the formula output is a physiologically reasonable firing frequency, consistent with the actual firing characteristics of biological neurons. BFR n The value directly determines the high-risk node marker, providing a quantitative basis for identifying root knowledge deficiencies and early warning of cognitive resource crowding, thus upgrading the assessment report from generalized recommendations to targeted intervention plans.

[0135] Based on the embodiments provided in this application, nodes in the dynamic knowledge topology network are transformed into spiking neurons, whose basic firing frequency is determined by the number of historical errors and grammatical depth. Connection edges are transformed into synapses, and their strength adjusts with the error co-occurrence frequency. Fixation duration is positively encoded as neural stimulation intensity, and jump frequency is negatively encoded as inhibition signals. Spiking dynamics is used to label hyperactive nodes and identify high-risk defective nodes. By drawing on the principles of neurodynamics and simulating the spiking process of neurons, combined with eye-tracking data, the activity level of knowledge nodes is quantitatively analyzed, enabling more accurate identification of high-risk defective nodes and providing a novel and effective technical means for assessing knowledge deficiencies.

[0136] Furthermore, the cognitive impulse dynamics model includes synaptic adaptive mechanisms, specifically:

[0137] When a node is marked as a high-risk defect node, the synaptic connection strength between that node and its direct parent dependent node is strengthened; where the direct parent dependent node is the node that provides input dependency to that node in the dynamic knowledge topology network.

[0138] Weaken the synaptic connection strength between the node and its direct dependent nodes; where a direct dependent node is a node in the dynamic knowledge topology network that receives the output dependency of the node.

[0139] The magnitude of enhancement and reduction is determined by the gaze duration of the semantic functional unit associated with the corresponding node.

[0140] It should be noted that a direct parent dependency node refers to an upstream node in a dynamic knowledge topology network that provides logical dependencies to the current node directly through directed edges; for example, the direct parent dependency node of "present perfect tense" is "simple past tense" ("present perfect tense" must be based on "past tense").

[0141] In a dynamic knowledge topology network, a direct child-dependent node is a downstream node that receives the output dependency of the current node through a directed edge; for example, the direct child-dependent node of "simple past tense" is "past progressive tense" ("past progressive tense" depends on the time concept of "simple past tense").

[0142] Node-associated semantic units refer to the binding of at least one semantic functional unit to each node in a dynamic knowledge topology network. The binding relationship is established by matching grammatical knowledge points with unit functional tags.

[0143] It's important to explain that errors in high-risk defective nodes (such as "present perfect tense") may stem from a weak grasp of their parent node ("simple past tense"). Strengthening synaptic strength can reinforce the connection between the two, simultaneously reinforcing the parent node during learning and addressing the defect at its root. Weakening synaptic strength with directly dependent child nodes prevents errors in high-risk nodes from excessively impacting the learning of child nodes (such as "past progressive tense"), reducing error propagation and preserving independent learning space for child nodes.

[0144] The magnitude is determined by the fixation duration of the semantic functional unit associated with the corresponding node: a long fixation duration (e.g., 3 seconds) indicates high learner attention to the unit (possibly due to effort but not yet mastery), resulting in a large enhancement / weakening magnitude (e.g., enhancement of 0.3, weakening of 0.2); a short fixation duration (e.g., 1 second) indicates low attention, resulting in a small magnitude (e.g., enhancement of 0.1, weakening of 0.05). For example, the high-risk node "present perfect tense" has a fixation duration of 3 seconds, its synaptic strength with its parent node "simple past tense" is enhanced by 0.3, and its strength with its child node "past perfect tense" is weakened by 0.2.

[0145] By dynamically adjusting synaptic strength, the dynamic knowledge topology network is made more closely aligned with the learner's actual cognitive state. The relationships between high-risk nodes are optimized in a targeted manner, which both strengthens the basic support (parent nodes) and isolates error propagation (child nodes), providing a more accurate network structure basis for subsequent knowledge defect repair and learning task design.

[0146] Based on the embodiments provided in this application, the synaptic adaptive mechanism of the cognitive impulse dynamics model strengthens the synaptic connection strength with its direct parent dependent node and weakens the connection strength with its direct child dependent node when a node is marked as a high-risk defect node, with the magnitude determined by the gaze duration. This enables the synaptic connections in the knowledge topology network to adaptively adjust according to the defect status of the node and the learner's gaze attention level, strengthening the focus on key dependencies and contributing to the optimization of knowledge system construction.

[0147] Furthermore, before dynamically binding eye-tracking gaze data to semantic functional units, an eye-tracking trajectory validity check is performed, including:

[0148] Detect whether there are trajectory interruptions in eye-tracking fixation data with consecutive jump intervals exceeding 200 milliseconds;

[0149] If there are trajectory interruptions, remove the gaze point data corresponding to the trajectory interruption segments.

[0150] When the proportion of removed gaze points exceeds 15% of the total gaze points, a data re-acquisition command is triggered; among them, eye-tracking gaze data that has passed the validity verification of eye-tracking trajectory participates in the generation of cognitive load vector.

[0151] Based on the embodiments provided in this application, before dynamically binding eye-tracking fixation data with semantic functional units, an eye-tracking trajectory validity check is performed. Fixation point data corresponding to trajectory interruptions are detected and removed. Re-acquisition is triggered when the proportion of removed data is too high. This ensures the validity and reliability of the eye-tracking data used to generate the cognitive load vector, avoids interference from invalid data in the calculation of the cognitive load vector, and thus improves the accuracy of various assessments and judgments based on this vector.

[0152] It should be noted that the embodiments implemented on the language learning evaluation system side based on the data fusion model in this application can be referenced with the embodiments implemented on the language learning evaluation method side based on the data fusion model, and will not be described in detail here.

[0153] According to another aspect of the embodiments of this application, an electronic device for implementing the above-described language learning evaluation method based on a data fusion model is also provided. This electronic device may be... Figure 4 The terminal device or server shown. This embodiment uses this electronic device as an example of a server. Figure 4 As shown, the electronic device includes a memory 402, a processor 404, and a transmission device 406. The memory 402 stores a computer program, and the processor 404 is configured to execute the steps in any of the above method embodiments through the computer program.

[0154] Optionally, in this embodiment, the aforementioned electronic device may be located in at least one of a plurality of network devices in a computer network.

[0155] Optionally, the transmission device 406 is used to receive or send data via a network. Specific examples of the network described above may include wired and wireless networks. In one example, the transmission device 406 includes a Network Interface Controller (NIC), which can be connected to other network devices and a router via a network cable to communicate with the Internet or a local area network. In another example, the transmission device 406 is a radio frequency (RF) module used to communicate with the Internet wirelessly.

[0156] In addition, the aforementioned electronic device also includes: a display 408 for displaying target identification characters contained in the identity identifier of the identified target object; and a connection bus 410 for connecting various module components in the aforementioned electronic device.

[0157] The above are merely preferred embodiments of the present invention and do not limit the scope of the patent. Any equivalent structural or procedural transformations made based on the description and drawings of the present invention, or direct or indirect applications in other related technical fields, are similarly included within the scope of patent protection of the present invention.

Claims

1. A language learning assessment system based on a data fusion model, characterized in that, include: The data acquisition module is used to acquire learners' language writing text data and eye movement fixation data; The decomposition module is used to decompose the language writing text data into multiple semantic functional units. These semantic functional units include core carrier units, logical connection units, and supporting extension units, and each semantic functional unit is labeled with a dynamic logical density value. Specifically, the core carrier unit includes sentence components indicating the subject-verb-object structure; the logical connection unit includes connecting components expressing causal, adversative, and referential relationships; and the supporting extension unit includes modifiers, adverbials, or complements that modify the core carrier unit. The binding module is used to dynamically bind the eye-tracking gaze data to the semantic functional unit, generating a cognitive load vector that includes unit type, dynamic logical density, syntactic depth, gaze duration, and jump frequency. The judgment module is used to calculate the avoidance index of the core carrier unit based on the cognitive load vector. When the avoidance index exceeds the preset avoidance threshold and the activation rate of the support extension unit exceeds the preset activation threshold, a higher-order thinking avoidance strategy judgment instruction is triggered. The report determination module is used to determine a comprehensive evaluation report in response to the higher-order thinking avoidance strategy determination instruction.

2. A language learning assessment method based on a data fusion model, characterized in that, include: Acquire learners' language writing text data and eye movement fixation data; The language writing text data is decomposed into multiple semantic functional units, which include a core carrier unit, a logical connection unit, and a supporting extension unit. A dynamic logical density value is marked for each semantic functional unit. The core carrier unit includes sentence components that indicate the subject-verb-object structure; the logical connection unit includes connecting components that express causal, adversative, and referential relationships; and the supporting extension unit includes attributive, adverbial, or complement components that modify the core carrier unit. The eye-tracking fixation data is dynamically bound to the semantic functional unit to generate a cognitive load vector that includes unit type, dynamic logic density, syntactic depth, fixation duration, and jump frequency. The avoidance index of the core carrier unit is calculated based on the cognitive load vector. When the avoidance index exceeds the preset avoidance threshold and the activation rate of the support extension unit exceeds the preset activation threshold, a higher-order thinking avoidance strategy judgment instruction is triggered. In response to the higher-order thinking avoidance strategy determination instruction, a comprehensive evaluation report is generated.

3. The language learning evaluation method based on a data fusion model according to claim 2, characterized in that, Before determining the comprehensive evaluation report in response to the higher-order thinking avoidance strategy determination instruction, the method further includes: Construct a dynamic knowledge topology network, where nodes represent grammatical knowledge points, edges represent logical dependencies between knowledge points, and edge weights are dynamically adjusted according to the frequency of error co-occurrence. When the higher-order thinking escape strategy judgment instruction is triggered, high-risk defect nodes are marked through the cognitive impulse dynamics model; For high-risk defect nodes, retrieve the historical cognitive load vectors of all semantic functional units associated with the node; if the average gaze duration exceeds the first threshold and the jump frequency is lower than the second threshold, it is confirmed as a root knowledge defect; otherwise, it is judged as an accidental error. When the cognitive load value of a basic knowledge point node is higher than that of the associated higher-order knowledge point node in three consecutive learning tasks, and the difference exceeds the preset proportion of the total load, a cognitive resource crowding warning is generated. The process of determining the comprehensive evaluation report includes generating the comprehensive evaluation report based on the results of the determination of avoidance strategies, the confirmation of root knowledge deficiencies, and the early warning of cognitive resource crowding.

4. The language learning evaluation method based on a data fusion model according to claim 3, characterized in that, After marking high-risk defect nodes, perform a logical fault depth tracing operation, including: Starting with the avoided core carrier unit as the starting node, the dynamic knowledge topology network is traversed to extract all logical connection units that directly depend on the starting node to form a primary dependency path. For each logical connection unit in the primary dependency path, its downstream supporting extension units in the dynamic knowledge topology network are retrieved to form a complete dependency link; Obtain eye-tracking fixation data corresponding to each semantic functional unit in the complete dependency chain, and calculate the focus maintenance rate of each semantic functional unit in the complete dependency chain; wherein, the focus maintenance rate is obtained by dividing the fixation duration by the sum of the fixation duration and the total time of jumping to units outside the chain; Identify semantic functional units in the complete dependency chain whose focus maintenance rate is lower than the average value of the link units and also lower than the average value of the successor nodes as logical breakpoints. Generate a fault impact map including the location of logical breakpoints and the complete dependency links, and store it in the evaluation database.

5. The language learning evaluation method based on a data fusion model according to claim 4, characterized in that, Based on the stored fault impact map, logical chain repair verification is performed in subsequent learning tasks, including: When the complete dependency link recorded in the fault influence map reappears, the focus maintenance rate of the breakpoint semantic functional unit that matches the logical breakpoint is monitored in real time. By analyzing whether the breakpoint semantic functional unit contains the recurrence or referential relationship of the core words of its direct upstream semantic functional unit, the semantic association strength between the breakpoint semantic functional unit and its direct upstream semantic functional unit is detected. If the focus maintenance rate of the breakpoint semantic function unit is increased to above the average value of the link units and the semantic association strength exceeds the preset repair threshold, the path logic coherence repair is marked as qualified in the comprehensive evaluation report. If the target is not met, an enhanced intervention mechanism is activated, which inserts a forced association writing task between the breakpoint semantic functional unit and its direct upstream semantic functional unit in subsequent tasks and shortens the task interval.

6. The language learning evaluation method based on a data fusion model according to claim 3, characterized in that, Before determining higher-order thinking avoidance strategies, perform cognitive intention verification, including: The detection function checks whether the gaze duration of the support extension unit is lower than a specific percentage of the historical baseline. When the fixation duration is detected to be lower than a certain percentage of the historical baseline, a visual stimulus marker is injected at the end of the text in that unit and a timer is started; the eye movement focus is monitored within a preset time window to see if it returns to the marked area. If the focus returns and the dwell time exceeds the time threshold, it is marked as attention drift behavior and the unit is excluded from the decision process for escape strategy. If no valid return behavior is detected, the unit is retained to participate in the decision-making process for the escape strategy.

7. The language learning evaluation method based on a data fusion model according to claim 6, characterized in that, When an attention drift behavior is identified, the presentation duration of the unit in subsequent tasks is extended to a specific multiple of the baseline value; a grammar structure hint layer is dynamically generated in the interactive interface to cover the text area of ​​the unit. When retaining the decision-making process for participating in the evasion strategy, the preset avoidance threshold is lowered based on the error propagation weight of the unit in the dynamic knowledge topology network. Shorten the interval between the occurrence of similar semantic functional units in subsequent learning tasks; Increase the influence weight of related high-order knowledge point nodes in the dynamic knowledge topology network.

8. The language learning evaluation method based on a data fusion model according to claim 3, characterized in that, The method of marking high-risk defect nodes using a cognitive impulse dynamics model includes: The nodes of the dynamic knowledge topology network are transformed into spiking neurons, and the basic firing frequency of the spiking neuron is jointly determined by the number of historical errors and the grammatical depth of the node; the connecting edges between nodes are transformed into synapses, and the synaptic strength is dynamically adjusted by the error co-occurrence frequency of the associated nodes. The gaze duration corresponding to the semantic functional unit is positively encoded as the neural stimulation intensity, and the jump frequency is negatively encoded as the inhibition signal. Pulse dynamics were run within a 500-millisecond simulation time. Each neuron received upstream pulses and eye-tracking stimulation signals, and triggered its own pulse when the accumulated potential exceeded the potential threshold. Neurons that fired high-frequency pulses were marked as hyperactive nodes. If the pulse emission frequency of a node exceeds a preset multiple of the average frequency of the entire network, it is determined to be a high-risk defective node.

9. The language learning evaluation method based on a data fusion model according to claim 8, characterized in that, The cognitive impulse dynamics model includes a synaptic adaptive mechanism, specifically comprising: When a node is marked as a high-risk defect node, the synaptic connection strength between that node and its direct parent dependent node is enhanced; wherein, the direct parent dependent node is the node that provides input dependency for that node in the dynamic knowledge topology network; Weaken the synaptic connection strength between the node and its direct dependent nodes; wherein, the direct dependent nodes are nodes in the dynamic knowledge topology network that receive the output dependencies of the node. The magnitude of enhancement and reduction is determined by the gaze duration of the semantic functional unit associated with the corresponding node.

10. The language learning evaluation method based on a data fusion model according to claim 2, characterized in that, Before dynamically binding the eye-tracking gaze data to the semantic functional unit, an eye-tracking trajectory validity check is performed, including: Detect whether there are trajectory interruptions in the eye-tracking fixation data with a continuous jump interval exceeding 200 milliseconds; If the trajectory interruption segment exists, the gaze point data corresponding to the trajectory interruption segment is removed; When the proportion of removed gaze points exceeds 15% of the total gaze points, a data re-acquisition command is triggered; wherein, eye-tracking gaze data that has passed the eye-tracking trajectory validity verification participates in the generation of the cognitive load vector.

Images