Education knowledge base self-inspection and self-evolution system and method

Abstract

The invention provides a self-checking and self-evolution system and method for an education knowledge base, and relates to the technical field of data storage, and the system comprises an obtaining unit which is used for obtaining new knowledge entries; the vector splicing unit is used for carrying out vector mixing and splicing on the new knowledge entries to obtain corresponding first high-dimensional mixed vectors; the conflict calculation unit is used for calculating a dimension conflict between the new knowledge item and the old knowledge item according to the first high-dimensional mixed vector corresponding to the new knowledge item and the second high-dimensional mixed vector corresponding to the old knowledge item; the conflict integration unit is used for obtaining a semantic total conflict score between the new knowledge item and the old knowledge item according to the dimension conflict and the dimension weight after collaborative self-evolution optimization; and the knowledge judgment unit is used for judging whether the semantic total conflict score is smaller than a conflict threshold value after collaborative self-evolution optimization or not, and adding the new knowledge items into the knowledge base. According to the system and the method provided by the invention, self-checking, self-repairing and self-evolution of the knowledge base can be realized.

Description

Technical Field

[0001] This invention relates to the field of data storage technology, and in particular to a self-checking and self-evolution system and method for educational knowledge bases. Background Technology

[0002] An educational knowledge base is a database used in intelligent teaching systems to store knowledge in the teaching field. The quality and scale of its knowledge directly affect the system's performance.

[0003] In the process of updating the knowledge base in the education industry, such as for enrollment consultation, student status management, and policy interpretation, due to the frequent updates, there will be a large number of knowledge conflicts in the updated content. Relying solely on manual maintenance is time-consuming and costly, which has a certain impact on the management of the knowledge base. Summary of the Invention

[0004] This invention provides a self-checking and self-evolution system and method for educational knowledge bases to solve the technical problem that due to frequent knowledge updates, there are a large number of knowledge conflicts in the updated content, and manual maintenance alone is time-consuming and costly.

[0005] To achieve the above and other related objectives, this invention provides a self-checking and self-evolution system for an educational knowledge base, comprising: an acquisition unit for acquiring new knowledge entries; a vector concatenation unit for performing vector mixing and concatenation on the new knowledge entries to obtain a corresponding first high-dimensional mixed vector; a conflict calculation unit for calculating the dimensional conflict between the first high-dimensional mixed vector corresponding to the new knowledge entry and the second high-dimensional mixed vector corresponding to the old knowledge entry, wherein the dimensional conflict includes semantic distance, semantic logical conflict probability, and context relevance penalty term; a conflict synthesis unit for obtaining a total semantic conflict score between the new knowledge entry and the old knowledge entry based on the dimensional conflict and the dimensional weights optimized through collaborative self-evolution; and a knowledge judgment unit for judging whether the total semantic conflict score is less than the conflict threshold optimized through collaborative self-evolution; if so, adding the new knowledge entry to the knowledge base; if not, performing knowledge repair on the new knowledge entry to generate a repaired knowledge entry to be added to the knowledge base.

[0006] In one embodiment of the present invention, the vector concatenation unit includes: a vector encoding subunit, used to perform vector encoding on the new knowledge item to obtain a high-dimensional vector corresponding to the new knowledge item, the high-dimensional vector including a language model encoding vector, a domain label vector and a meta-feature encoding vector; and a vector mixing subunit, used to perform vector concatenation and normalization processing on the high-dimensional vector to obtain a corresponding first high-dimensional mixed vector.

[0007] In one embodiment of the present invention, the conflict calculation unit includes: a distance conflict calculation subunit, used to calculate the semantic distance between a new knowledge entry and an old knowledge entry based on a first high-dimensional mixture vector, a second high-dimensional mixture vector, the number of semantic feature layers, and the hierarchical weights corresponding to each semantic feature layer; a logical conflict calculation subunit, used to calculate the semantic logical conflict probability between a new knowledge entry and an old knowledge entry based on a first high-dimensional mixture vector, a second high-dimensional mixture vector, the number of semantic feature layers, and the hierarchical weights corresponding to each semantic feature layer; and a relevance penalty calculation subunit, used to calculate a context relevance penalty term between a new knowledge entry and an old knowledge entry based on a first similarity in the classification label dimension, an overlap in the time dimension, a second similarity in the grouping dimension, a first weight coefficient corresponding to the first similarity, a second weight coefficient corresponding to the overlap, and a third weight coefficient corresponding to the second similarity.

[0008] In one embodiment of the present invention, the logical conflict calculation subunit includes: a conflict intensity calculation module, used to input new knowledge entries and old knowledge entries into a constructed logical conflict discrimination model to obtain the conflict intensity of logical conflict between new knowledge entries and old knowledge entries; and a conflict probability calculation module, used to obtain the semantic logical conflict probability between new knowledge entries and old knowledge entries based on the conflict intensity.

[0009] In one embodiment of the present invention, the relevance penalty calculation subunit includes: a context similarity calculation module, used to calculate the context similarity between the new knowledge entry and the old knowledge entry based on the first similarity between the new knowledge entry and the old knowledge entry in the category label dimension, the degree of overlap in the time dimension, the second similarity in the grouping dimension, the first weight coefficient corresponding to the first similarity, the second weight coefficient corresponding to the degree of overlap, and the third weight coefficient corresponding to the second similarity; and a penalty calculation module, used to calculate the context relevance penalty term between the new knowledge entry and the old knowledge entry based on the context similarity and the maximum context similarity corresponding to the context similarity.

[0010] In one embodiment of the present invention, the knowledge judgment unit, in the process of performing knowledge repair on new knowledge entries to generate repaired knowledge entries to be added to the knowledge base, includes: generating repaired knowledge entries to be added to the knowledge base through a large language model based on the new knowledge entries, old knowledge entries, and the knowledge base context; wherein, the repaired knowledge entries and old knowledge entries simultaneously satisfy the following conditions: the total repair semantic conflict score between the repaired knowledge entries and old knowledge entries is less than a conflict threshold; the semantic deviation between the repaired high-dimensional mixed vector corresponding to the repaired knowledge entries and the first high-dimensional mixed vector is less than a deviation threshold; and the coherence between the repaired knowledge entries and the knowledge base is greater than a coherence threshold.

[0011] In one embodiment of the present invention, the invention further includes: a knowledge supplementation unit; the knowledge supplementation unit includes: a coverage detection subunit, used to extract key tasks from new knowledge entries and perform coverage detection on the key tasks; and a supplementation generation subunit, used to generate supplementary knowledge entries and add them to the knowledge base based on the key tasks and the knowledge base through a large language model when the coverage of the key tasks is less than the coverage threshold.

[0012] In one embodiment of the present invention, the invention further includes: a parameter update unit, configured to perform co-evolutionary analysis on the dimension weights and conflict thresholds based on the total semantic conflict score to obtain update weights and update thresholds; the parameter update unit includes: obtaining a corresponding state vector representation based on the dimension weights, conflict thresholds, and the corresponding total semantic conflict score, wherein the state vector representation includes an action space representation; obtaining a reward function representation based on the consistency reward, accuracy reward, knowledge coverage reward, threshold stability reward between the conflict threshold and the previous conflict threshold corresponding to the last update, a first reward weight corresponding to the consistency reward, a second reward weight corresponding to the accuracy reward, a third reward weight corresponding to the knowledge coverage reward, and a fourth reward weight corresponding to the threshold stability reward; and updating the dimension weights and conflict thresholds based on the action space representation, the reward function representation, the learning rate, and the discount factor to obtain update weights and update thresholds.

[0013] In one embodiment of the present invention, the invention further includes: a global evaluation unit; the global evaluation unit includes: a re-scoring subunit, used to recalculate the corresponding updated semantic total conflict score based on the update weight, update threshold, new knowledge entries, and old knowledge entries; and a knowledge base update subunit, used to obtain an evolutionary objective function based on the updated semantic total conflict score, update threshold, and the knowledge base before the update; and to perform self-evolution processing on the knowledge base before the update by maximizing the evolutionary objective function to obtain the updated knowledge base.

[0014] To achieve the above and other related objectives, this invention also provides a self-checking and self-evolution method for an educational knowledge base, comprising: acquiring new knowledge entries through an acquisition unit; performing vector mixing and splicing on the new knowledge entries through a vector splicing unit to obtain a corresponding first high-dimensional mixed vector; calculating the dimensional conflict between the new knowledge entries and the old knowledge entries based on the first high-dimensional mixed vector corresponding to the new knowledge entries and the second high-dimensional mixed vector corresponding to the old knowledge entries through a conflict calculation unit, wherein the dimensional conflict includes semantic distance, semantic logical conflict probability, and context relevance penalty term; obtaining the total semantic conflict score between the new knowledge entries and the old knowledge entries through a conflict synthesis unit based on the dimensional conflict and the dimensional weights optimized through collaborative self-evolution; and determining whether the total semantic conflict score is less than the conflict threshold optimized through collaborative self-evolution through a knowledge judgment unit. If so, the new knowledge entries are added to the knowledge base; otherwise, the new knowledge entries are repaired to generate repaired knowledge entries to be added to the knowledge base.

[0015] The beneficial effects of this invention are as follows: This invention proposes a self-checking and self-evolution system and method for an educational knowledge base. It acquires knowledge data and forms new knowledge entries. Then, it vectorizes the new knowledge entries acquired by the data acquisition agent, encoding them as high-dimensional vectors. These vectors are then concatenated and normalized to generate a first high-dimensional mixed vector. By utilizing the new knowledge entry, the old knowledge entry, the first high-dimensional mixed vector corresponding to the new knowledge entry, and the second high-dimensional mixed vector corresponding to the old knowledge entry, the dimensional conflict for each conflict dimension is calculated. This dimensional conflict includes semantic distance, semantic logical conflict probability, and contextual relevance penalty. Furthermore, by utilizing the dimensional conflict and the dimensional weights optimized through collaborative self-evolution, a comprehensive semantic conflict score between the new and old knowledge entries is calculated. Finally, by comparing the total semantic conflict score with the corresponding collaboratively self-evolution optimized conflict threshold, new knowledge entries with a total semantic conflict score less than the threshold are added to the knowledge base. Conversely, when the total semantic conflict score is greater than the conflict threshold, knowledge repair can be performed on the new knowledge entry, generating a repaired knowledge entry to be added to the knowledge base. The above methods ensure the reliability of newly added knowledge and avoid false additions; they effectively detect conflicts between new and old knowledge entries, and can select updates based on timestamps when conflicts exist; in addition, a manual review entry is provided for a small number of cases that cannot be automatically determined. Thus, a knowledge base system capable of automatic knowledge vectorization and structured representation, cross-source knowledge conflict detection, dynamic threshold adjustment, knowledge repair, knowledge addition, and global evolution goal optimization can be constructed. Therefore, through the multi-agent collaboration and reinforcement learning threshold control of this invention, the knowledge base can achieve self-inspection, self-repair, and self-evolution, improving the reliability and stability of the system. Attached Figure Description

[0016] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application. It is obvious that the drawings described below are merely some embodiments of this application, and those skilled in the art can obtain other drawings based on these drawings without any inventive effort.

[0017] In the attached diagram: Figure 1 A structural block diagram of the self-checking and self-evolution system for the educational knowledge base provided in this embodiment of the invention; Figure 2 The diagram shows a flowchart of a self-checking and self-evolution method for an educational knowledge base provided in an embodiment of the present invention.

[0018] The attached figures are labeled as follows: Acquisition Unit 111; Vector Concatenation Unit 112; Conflict Calculation Unit 113; Conflict Synthesis Unit 114; Knowledge Judgment Unit 115. Detailed Implementation

[0019] The following specific examples illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments. Various details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. In the absence of conflict, the following embodiments and features in the embodiments can be combined with each other.

[0020] It should be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of the present invention. The drawings only show the components related to the present invention and are not drawn according to the actual number, shape and size of the components in the actual implementation. In the actual implementation, the form, quantity and proportion of each component can be arbitrarily changed, and the layout of the components may also be more complex.

[0021] In the following description, numerous details are explored to provide a more thorough explanation of embodiments of the invention. However, it will be apparent to those skilled in the art that embodiments of the invention may be practiced without these specific details. In other embodiments, well-known structures and devices are shown in block diagram form rather than in detail to avoid obscuring embodiments of the invention.

[0022] Please see Figure 1This invention provides a self-checking and self-evolution system for an educational knowledge base, comprising: an acquisition unit 111 for acquiring new knowledge entries; a vector splicing unit 112 for performing vector mixing and splicing on the new knowledge entries to obtain a corresponding first high-dimensional mixed vector; a conflict calculation unit 113 for calculating the dimensional conflict between the first high-dimensional mixed vector corresponding to the new knowledge entry and the second high-dimensional mixed vector corresponding to the old knowledge entry, wherein the dimensional conflict includes semantic distance, semantic logical conflict probability, and context relevance penalty; a conflict synthesis unit 114 for obtaining a total semantic conflict score between the new knowledge entry and the old knowledge entry based on the dimensional conflict and the dimensional weights optimized through collaborative self-evolution; and a knowledge judgment unit 115 for judging whether the total semantic conflict score is less than the conflict threshold optimized through collaborative self-evolution; if so, adding the new knowledge entry to the knowledge base; otherwise, performing knowledge repair on the new knowledge entry to generate a repaired knowledge entry to be added to the knowledge base.

[0023] It is clear from the above that the self-checking and self-evolution system for the educational knowledge base of this invention can be based on the collaborative execution of multiple intelligent agents. Specifically, this can be achieved by collecting data from intelligent agents. In other words, it acquires knowledge data through unit 111 and forms new knowledge entries. Then, it uses a vector agent... In other words, the vector concatenation unit 112 performs vectorization processing on the new knowledge entries acquired by the acquisition agent, encoding the new knowledge entries into high-dimensional vectors for vector concatenation and normalization, and generating a first high-dimensional hybrid vector. This is achieved through a conflict detection agent. The system can include a conflict calculation unit 113, a conflict synthesis unit 114, and a knowledge judgment unit 115. Specifically, the conflict calculation unit 113 first calculates the dimensional conflict for each conflict dimension using new knowledge entries, old knowledge entries, the first high-dimensional mixture vector corresponding to the new knowledge entry, and the second high-dimensional mixture vector corresponding to the old knowledge entry. This dimensional conflict includes semantic distance, semantic logical conflict probability, and a contextual relevance penalty term; it can also include dimensional conflicts under other conflict dimensions. Then, the conflict synthesis unit 114 uses the dimensional conflict and the dimensional weights optimized through collaborative self-evolution to comprehensively calculate the total semantic conflict score between new and old knowledge entries. Finally, the knowledge judgment unit 115 compares the total semantic conflict score with the corresponding collaborative self-evolutionary optimized conflict threshold to determine whether to add new knowledge entries with a total semantic conflict score less than the threshold to the knowledge base. Conversely, when the total semantic conflict score is greater than the conflict threshold, the agent can be optimized. This approach involves knowledge repair of new knowledge entries, generating repaired knowledge entries to be added to the knowledge base. By doing so, the reliability of newly added knowledge can be ensured, avoiding false additions; conflicts between new and old knowledge entries can be effectively detected, and updates can be selected based on timestamps when conflicts exist; additionally, a manual review entry point is provided for a small number of cases that cannot be automatically determined. This allows for the construction of a knowledge base system capable of automatic knowledge vectorization and structured representation, cross-source knowledge conflict detection, dynamic threshold adjustment, knowledge repair, knowledge supplementation, and global evolution goal optimization.

[0024] In the self-checking and self-evolution system of the educational knowledge base of the present invention, the vector splicing unit 112 includes: a vector encoding subunit, used to perform vector encoding on new knowledge items to obtain a high-dimensional vector corresponding to the new knowledge items, the high-dimensional vector including language model encoding vector, domain label vector and meta-feature encoding vector; and a vector mixing subunit, used to perform vector splicing and normalization processing on the high-dimensional vector to obtain a corresponding first high-dimensional mixed vector.

[0025] In the process of knowledge vectorization, new knowledge items can be vectorized using vector encoding sub-units to obtain high-dimensional vectors corresponding to the new knowledge items. Specifically, knowledge item k can be... i Encoding is done as a high-dimensional vector, and the formula can be expressed as follows: .in, This can be represented as a language model encoding vector. Examples include BERT / RoBERTa / SimCSE encoding; it can also be represented as a domain label vector. For example, it can be embedded through domain labels, such as "admission policy" or "enrollment plan"; it can also be represented as a meta-feature encoding vector. For example, meta-feature encoding such as source credibility, year, and school type. Then, through the vector mixing sub-unit, the high-dimensional vector is concatenated and normalized to obtain the corresponding first high-dimensional mixed vector, which can be expressed by the formula: ,in, The expression represents vector concatenation, and Norm represents normalization. By implementing fine-grained semantics, more accurate conflict judgment can be achieved. By implementing structured semantics, misjudgments of "synonyms in different scenarios" can be avoided. Through interpretability processing, the source and context of each piece of knowledge can be traced.

[0026] In the self-checking and self-evolution system of the educational knowledge base of the present invention, the conflict calculation unit 113 includes: a distance conflict calculation subunit, used to calculate the semantic distance between new knowledge entries and old knowledge entries based on a first high-dimensional mixing vector, a second high-dimensional mixing vector, the number of semantic feature layers, and the hierarchical weights corresponding to each layer of semantic features; a logical conflict calculation subunit, used to calculate the semantic logical conflict probability between new knowledge entries and old knowledge entries based on a first high-dimensional mixing vector, a second high-dimensional mixing vector, the number of semantic feature layers, and the hierarchical weights corresponding to each layer of semantic features; and a relevance penalty calculation subunit, used to calculate the context relevance penalty term between new knowledge entries and old knowledge entries based on the first similarity in the classification label dimension, the degree of overlap in the time dimension, the second similarity in the grouping dimension, the first weight coefficient corresponding to the first similarity, the second weight coefficient corresponding to the degree of overlap, and the third weight coefficient corresponding to the second similarity.

[0027] When performing dimensional conflict calculation using the conflict calculation unit 113, the semantic distance between new and old knowledge entries can be calculated through the distance conflict calculation subunit. The formula can be expressed as follows: , Indicates a new knowledge entry. This indicates entries representing prior knowledge. This represents the first high-dimensional hybrid vector corresponding to the new knowledge entry. Let L represent the second high-dimensional mixture vector corresponding to the old knowledge entry, L be the multi-level semantic features (such as word-level, sentence-level, paragraph-level), and wl be the hierarchical weight, with ∑wl=1. Additionally, the semantic logical conflict calculation subunit can be used to calculate the semantic logical conflict probability between the new and old knowledge entries by combining the first high-dimensional mixture vector, the second high-dimensional mixture vector, the number of semantic feature layers, and the hierarchical weights corresponding to each semantic feature layer. The contextual relevance penalty calculation subunit can then be used to calculate the contextual relevance penalty term between the new and old knowledge entries by combining the first similarity in the classification label dimension, the degree of overlap in the time dimension, the second similarity in the grouping dimension, the first weight coefficient corresponding to the first similarity, the second weight coefficient corresponding to the degree of overlap, and the third weight coefficient corresponding to the second similarity.

[0028] The logical conflict calculation subunit includes: a conflict intensity calculation module, which is used to input new knowledge entries and old knowledge entries into the constructed logical conflict discrimination model to obtain the conflict intensity of logical conflict between new knowledge entries and old knowledge entries; and a conflict probability calculation module, which is used to obtain the semantic logical conflict probability between new knowledge entries and old knowledge entries based on the conflict intensity.

[0029] In the process of calculating the semantic logical conflict probability, the logical conflict calculation subunit can use the constructed logical conflict discrimination model through the conflict intensity calculation module to process new and old knowledge entries, thereby determining the conflict intensity between them. The formula for the logical scoring function for calculating the conflict intensity can be expressed as follows: Here, hᵢⱼ is the cross-sentence interactive semantic representation encoded by the language model (LLM / Transformer); W1, W2, b1, b2 are trainable parameters in the neural network; ReLU() represents the non-linear activation function; σ(gϕ) represents the final logical conflict probability. Additionally, g... ϕ ( ) is a logical conflict discrimination function. Its core task is to take two knowledge texts (kᵢ, kⱼ) as input and output a continuous value to measure the strength of the logical conflict between them.

[0030] The two knowledge texts (kᵢ, kⱼ) are shown in the table below: For the logical scoring function g ϕ During the construction process, knowledge pairs (k) can be constructed by using a pre-trained Natural Language Inference (NLI) model. i , k j The system outputs the conflict probability, then generates weakly labeled data by combining it with a Large Language Model (LLM), self-trains a binary conflict discriminator, and obtains L(k) by mapping the output value to a Sigmoid function. i ,k j ) ∈ [0,1].

[0031] Since semantic similarity and logical conflict are independent of each other, an additional logical scoring function is needed to identify "hard conflicts" such as numerical conflicts, conditional contradictions, and temporal inconsistencies, and to reduce scenarios that pure semantic models cannot handle. After obtaining the conflict intensity of a logical conflict through the logical scoring function, a further conflict probability calculation module can be used to calculate the semantic logical conflict probability between new and old knowledge items. The formula can be expressed as follows: g ϕ ( ) is the language model's logical scoring function, and σ( ) is the Sigmoid function, which maps the output to [0,1]. If the large model considers the probability of conflict to be high, L approaches 1; otherwise, it approaches 0.

[0032] In addition, the relevance penalty calculation subunit includes: a context similarity calculation module, used to calculate the context similarity between new and old knowledge entries based on the first similarity between the new and old knowledge entries in the category label dimension, the degree of overlap in the time dimension, the second similarity in the grouping dimension, the first weight coefficient corresponding to the first similarity, the second weight coefficient corresponding to the degree of overlap, and the third weight coefficient corresponding to the second similarity; and a penalty calculation module, used to calculate the context relevance penalty term between new and old knowledge entries based on the context similarity and the maximum context similarity.

[0033] In the process of calculating the context relevance penalty term through the relevance penalty calculation subunit, the context similarity calculation module can calculate the context similarity between new and old knowledge entries based on the first similarity in the category label dimension, the degree of overlap in the time dimension, the second similarity in the grouping dimension, the first weight coefficient corresponding to the first similarity, the second weight coefficient corresponding to the degree of overlap, and the third weight coefficient corresponding to the second similarity. The formula can be expressed as follows: , , , Let be the weight coefficients for each context dimension, and satisfy: ; This represents the first similarity between knowledge entries ki and kj on their category tags; This indicates the degree of overlap between two knowledge entries in terms of their effective time / publication time (date); This indicates the second similarity between two knowledge items in terms of applicable groups or policy groups.

[0034] After obtaining the context similarity, the penalty calculation module can be used to combine the maximum context similarity in the context similarity to calculate the context relevance penalty between the new and old knowledge entries. The formula can be expressed as follows: ,in, The maximum possible contextual similarity score is used to evaluate the current similarity. Normalize. It is actually measuring the degree of difference in context.

[0035] After obtaining the semantic distance, semantic logic conflict probability, and context relevance penalty term, we can further combine them with their respective dimensional weights. , , The semantic total conflict score is calculated using the following formula: , Here, semantic distance D can be understood as "whether they are talking about the same thing," logical conflict L can be understood as "whether the content is contradictory," and contextual difference R can be understood as "whether there are conflicts with other coordinates such as labels / regions / time." Specifically, R... context As a context-related penalty term, multi-dimensional context similarity can be defined by weighting the above dimensions. .

[0036] Through the above and The design mechanism of the computation process can be used when When approaching the maximum value (extremely high context similarity), Approaching 0 (the difference is extremely small), at this point Item for overall conflict score The impact is minimal. When Extremely small (extremely low context similarity). Approaching 1 (the difference is extremely large), at this point Item pair The punishment (because) exist The positive contribution item is the largest. The key point is... exist The role of `<meta>` is to regulate, that is, when the context differences are large ( When the value is large, it indicates that the two should not be strongly correlated. The score should be lowered to avoid false alarms (such as Beijing vs. Shanghai). Therefore, in In the formula, It should be a negative contribution, or It has been redefined as context similarity.

[0037] The comprehensive calculation method of semantic conflict score described above can solve the problem of regional / scenario differences: two pieces of knowledge may be semantically similar, but if the context is different, they are not considered to be in conflict. For example: kᵢ: "Beijing high school enrollment age ≤ 18", kⱼ: "Shanghai high school enrollment age ≤ 19". The D function may be similar, and the L function considers them to be in conflict, but because the context is different, they should not be considered to be in conflict.

[0038] The comprehensive calculation method of the semantic total conflict score described above can solve the problem of penalizing highly relevant contexts: the larger the Rcontext, the smaller the contextual differences, combined with αD + βL + γ Rcontext This can reduce false alarms.

[0039] Therefore, the comprehensive calculation method of the above semantic total conflict score can reflect multi-dimensional conflict measurement, that is, the method can be used to detect whether there are significant differences in actual semantics.

[0040] In the self-checking and self-evolution system of the educational knowledge base of the present invention, the knowledge judgment unit 115, in the process of repairing new knowledge entries through a repair agent to generate repaired knowledge entries and adding them to the knowledge base, includes: Based on new knowledge entries, old knowledge entries, and the knowledge base context, repair knowledge entries are generated and added to the knowledge base using a large language model; Among them, the repaired knowledge entry and the old knowledge entry must simultaneously meet the following conditions: The total semantic conflict score between the repaired knowledge entry and the old knowledge entry is less than the conflict threshold; The semantic deviation between the repaired high-dimensional mixture vector corresponding to the repaired knowledge entry and the first high-dimensional mixture vector is less than the deviation threshold; Repair knowledge entries and knowledge bases where the coherence exceeds the coherence threshold.

[0041] During the knowledge repair process, if there is a conflict between the old and new knowledge entries (k) i ,k j )satisfy , This represents the conflict threshold. At this point, new knowledge entries, old knowledge entries, and the knowledge base context can be used to generate repair knowledge entries through a large language model and add them to the knowledge base. The generation of new repair knowledge entries can be represented as... , This represents the knowledge base context. Furthermore, in constructing knowledge repair constraints, it is also necessary to satisfy the conflict constraint that the total semantic conflict score between the repaired knowledge entry and the old knowledge entry is less than the conflict threshold. , This indicates that the knowledge bar has been repaired. This represents an old knowledge entry; the semantic deviation between the repaired high-dimensional mixture vector and the first high-dimensional mixture vector corresponding to the repaired knowledge entry is less than the deviation threshold, i.e., a semantic deviation constraint. , This indicates the repair of a high-dimensional mixed vector. Represents the first high-dimensional mixture vector. This indicates a deviation from the threshold; it corrects consistency constraints where the coherence between knowledge entries and the knowledge base exceeds the coherence threshold. , This represents the continuity threshold, and δ and τ can be adaptively adjusted. Furthermore, if the above constraints cannot be met, the repair is considered a failure and the case is transferred to manual processing.

[0042] The self-checking and self-evolution system for the educational knowledge base of the present invention further includes: a knowledge supplementation unit; the knowledge supplementation unit includes: a coverage detection subunit, used to extract key tasks from new knowledge entries and perform coverage detection on the key tasks; and a supplementation generation subunit, used to generate supplementary knowledge entries and add them to the knowledge base based on the key tasks and the knowledge base through a large language model when the coverage of the key tasks is less than the coverage threshold.

[0043] New supplementary knowledge entries can be generated and added to the knowledge base through a knowledge supplementation mechanism. Specifically, the key tasks of the new knowledge entries are first extracted through a coverage detection subunit, and the key tasks t with insufficient coverage are identified. i That is, it can be done through the formula To identify critical tasks with insufficient coverage i ,in, This represents the coverage threshold. Further sub-units are generated by supplementing the coverage of key tasks t where the coverage of new knowledge items is insufficient. i， The large language model is used to process the data and generate supplementary knowledge entries, expressed by the following formula: Finally, it is added to the knowledge base, that is... , This represents the knowledge base. Furthermore, when adding knowledge entries to the knowledge base, a repair agent can further perform constraint checks on the new knowledge entries. If the constraints cannot be met, the addition is considered a failure and the matter is transferred to manual processing.

[0044] Furthermore, in the formula for calculating the total semantic conflict score, the dimensional weights... , , All of these are obtained based on the self-evolution mechanism of reinforcement learning. The conflict threshold is also obtained based on the self-evolution mechanism of reinforcement learning. That is, it can be continuously updated and improved through the parameter update unit, namely the threshold agent. The reward function is generated based on the consistency, coverage and stability of the repaired knowledge base, so that the weights gradually converge to the optimal configuration.

[0045] Therefore, the self-checking and self-evolution system of the educational knowledge base of the present invention further includes: a parameter update unit, used to perform collaborative self-evolution analysis on dimension weights and conflict thresholds based on the total semantic conflict score, to obtain update weights and update thresholds.

[0046] The parameter update unit includes: obtaining the corresponding state vector representation based on the dimension weights, conflict thresholds, and corresponding semantic total conflict scores, wherein the state vector representation includes the action space representation; obtaining the reward function representation based on the consistency reward, accuracy reward, knowledge coverage reward, threshold stability reward between the conflict threshold and the previous conflict threshold corresponding to the last update, the first reward weight corresponding to the consistency reward, the second reward weight corresponding to the accuracy reward, the third reward weight corresponding to the knowledge coverage reward, and the fourth reward weight corresponding to the threshold stability reward; and updating the dimension weights and conflict thresholds based on the action space representation, the reward function representation, the learning rate, and the discount factor to obtain the update weights and update thresholds.

[0047] During reinforcement learning threshold adjustment, the state vector representation of the threshold agent at each time t is obtained based on the dimensional weights, the conflict threshold, and the corresponding total semantic conflict score. The state vector representation can be: ,in, Indicates the current conflict judgment threshold. Indicates the weighting coefficient; All are based on conflict ratings Statistics derived from set calculations. The action space representation can be: ,in, , , It can be used to control whether the threshold is adjusted up, down, or remains unchanged, with a step size. It is adaptively adjustable. Furthermore, the reward can be designed as multi-objective through reinforcement learning, as expressed by the formula: To calculate the reward function representation, Indicates the first reward weight. Indicates the weight of the second reward. Indicates the third reward weight. This indicates the fourth reward weight. This represents a consistency reward, used to reduce false negatives and penalize conflicts arising from missed judgments. This represents a precision bonus, used to reduce false alarms. Indicates threshold stability reward. This represents a reward for knowledge coverage.

[0048] In the process of designing rewards for reinforcement learning, the reward weights are... , , and During the design process, it is possible to first ensure ∑λ i =1, which ensures that the magnitude of the reward signal rt remains stable across different evolutionary rounds. Secondly, some evolutionary strategies can be defined, such as setting... > > > Prioritizing the coverage and consistency of educational knowledge across sources. Furthermore, as the knowledge base evolves and stabilizes, the weights can be adjusted by introducing a decay factor. > ≈ > This improves the operational stability of the system threshold while ensuring accuracy.

[0049] The consistency reward can be expressed as: The accuracy bonus can be expressed as The threshold stability reward can be expressed as The reward for knowledge coverage can be expressed as , where T is the set of all critical tasks / problems.

[0050] In obtaining , and After that, it can be done through the formula Implement Q-learning updates, including policy selection: α represents the learning rate; γ represents the discount factor; at: action space representation (threshold adjustment).

[0051] Furthermore, the intelligent agent can be further optimized. Responsible for execution by The determined parameters are updated, ensuring parameter constraints and performing a global evaluation to update weight parameters and thresholds. That is, based on... Calculating the parameters for the next time step and updating the threshold can be expressed as: The updated weights can be represented as: .

[0052] In addition, it can also be used by The parameters are updated globally, that is, the new parameters are used for evaluation. Recalculate And evaluate the global objective function. and reward function and will Returned as feedback In this way, to achieve Focus on decision making Focusing on execution and constraints, it achieves collaborative self-evolution of thresholds and weights, which can be implemented through a global evaluation unit.

[0053] Therefore, the self-checking and self-evolution system for the educational knowledge base of the present invention further includes: a global evaluation unit; the global evaluation unit includes: a re-scoring subunit, used to recalculate the corresponding total semantic conflict score based on the update weight, update threshold, new knowledge entries, and old knowledge entries; and a knowledge base updating subunit, used to obtain an evolutionary objective function based on the total semantic conflict score, update threshold, and the knowledge base before the update; and to perform self-evolution processing on the knowledge base before the update by maximizing the evolutionary objective function to obtain the updated knowledge base. The evolutionary objective function can be expressed as: By maximizing J(K), the knowledge base can achieve self-evolution, resulting in fewer conflicts, higher problem coverage, and better system consistency. Therefore, through the mutual cooperation between various agents in the self-checking and self-evolution system of the educational knowledge base of this invention, it is possible to optimize the conflict threshold, repair knowledge, and supplement knowledge.

[0054] During the operation of the educational knowledge base self-checking and self-evolution system of this invention, new data obtained through methods such as web crawling, user provision, and APIs can be input into the system, vector-encoded, and used to calculate conflict scores with existing knowledge. The reinforcement learning agent dynamically determines the conflict threshold, and items exceeding the threshold enter the repair module. If repair fails, the system is switched to manual intervention. Tasks with insufficient knowledge coverage are automatically supplemented, a global evolution evaluation is performed, a new knowledge base is output, and the system cycles to the next round of data, achieving orderly updates of knowledge data in the knowledge base. By integrating three-dimensional conflict detection of "semantics + logic + context," the knowledge base achieves closed-loop self-evolution, reduces manual intervention, and has the ability to adjust participation and adapt.

[0055] Furthermore, in this invention, all modules can be implemented using existing NLP models, RAG systems, and RL algorithms. Conflict scores can be calculated in real time in a vector database. Reinforcement learning only requires a small model to run, while repair and supplementation are completed by a large model. Existing LLMs can meet the requirements, and the algorithm has reliable input / output and verification methods.

[0056] Please see Figure 2 The present invention also provides a method for self-checking and self-evolution of an educational knowledge base, comprising: Step S10: Obtain new knowledge entries by acquiring unit 111; Step S20: The new knowledge items are vector-mixed and spliced ​​through the vector splicing unit 112 to obtain the corresponding first high-dimensional mixed vector; Step S30: The conflict calculation unit 113 calculates the dimensional conflict between the new knowledge entry and the old knowledge entry based on the first high-dimensional mixture vector corresponding to the new knowledge entry and the second high-dimensional mixture vector corresponding to the old knowledge entry. The dimensional conflict includes semantic distance, semantic logic conflict probability and context relevance penalty term. Step S40: Based on the dimensional conflict and the dimensional weights optimized through collaborative self-evolution, the conflict synthesis unit 114 obtains the total semantic conflict score between the new knowledge item and the old knowledge item. Step S50: The knowledge judgment unit 115 determines whether the total semantic conflict score is less than the conflict threshold optimized by collaborative self-evolution. If so, the new knowledge entry is added to the knowledge base; otherwise, the new knowledge entry is repaired to generate a repaired knowledge entry to be added to the knowledge base.

[0057] In summary, the self-checking and self-evolution system and method for an educational knowledge base disclosed in this invention acquires knowledge data and forms new knowledge entries. Then, the new knowledge entries acquired by the data acquisition agent are vectorized, encoding them into high-dimensional vectors. These vectors are then concatenated and normalized to generate a first high-dimensional mixed vector. By utilizing the new knowledge entry, the old knowledge entry, the first high-dimensional mixed vector corresponding to the new knowledge entry, and the second high-dimensional mixed vector corresponding to the old knowledge entry, the dimensional conflict corresponding to each conflict dimension is calculated. This dimensional conflict includes semantic distance, semantic logical conflict probability, and context relevance penalty. Furthermore, by utilizing the dimensional conflict and the dimensional weights optimized through collaborative self-evolution, a comprehensive semantic conflict score between the new and old knowledge entries is calculated. Finally, by comparing the total semantic conflict score with the corresponding collaborative self-evolution optimized conflict threshold, new knowledge entries with a total semantic conflict score less than the threshold are added to the knowledge base. Conversely, when the total semantic conflict score is greater than the conflict threshold, knowledge repair can be performed on the new knowledge entry, generating a repaired knowledge entry to be added to the knowledge base. By employing the above methods, the reliability of newly added knowledge can be guaranteed, avoiding false additions; conflicts between new and old knowledge entries can be effectively detected, and updates can be selected based on timestamps when conflicts exist; furthermore, a manual review entry point is provided for a small number of cases that cannot be automatically determined. This allows for the construction of a knowledge base system capable of automatic knowledge vectorization and structured representation, cross-source knowledge conflict detection, dynamic threshold adjustment, knowledge repair, knowledge addition, and global evolution goal optimization. Therefore, this invention effectively overcomes the various shortcomings of existing technologies and possesses high industrial application value.

[0058] The above embodiments are merely illustrative of the principles and effects of the present invention and are not intended to limit the invention. Any person skilled in the art can modify or alter the above embodiments without departing from the spirit and scope of the present invention. Therefore, all equivalent modifications or alterations made by those skilled in the art without departing from the spirit and technical concept disclosed in the present invention should still be covered by the claims of the present invention.

Claims

1. An educational knowledge base self-checking and self-evolution system, characterized in that, The method comprises the following steps: an acquisition unit is configured to acquire a new knowledge item; a vector splicing unit is configured to perform vector mixing and splicing on the new knowledge item to obtain a corresponding first high-dimensional mixed vector; a conflict calculation unit is configured to calculate a dimension conflict between the first high-dimensional mixed vector corresponding to the new knowledge item and a second high-dimensional mixed vector corresponding to an old knowledge item according to the dimension conflict and a dimension weight optimized through collaborative self-evolution, wherein the dimension conflict comprises a semantic distance, a semantic logic conflict probability and a context correlation penalty term; a conflict synthesis unit is configured to obtain a semantic total conflict score between the new knowledge item and the old knowledge item according to the dimension conflict and the dimension weight optimized through collaborative self-evolution; and a knowledge judgment unit is configured to judge whether the semantic total conflict score is less than a conflict threshold value optimized through collaborative self-evolution, and if yes, the new knowledge item is added to a knowledge base; if not, the new knowledge item is repaired to generate a repaired knowledge item which is added to the knowledge base. The vector splicing unit comprises:

2. The educational knowledge base self-checking and self-evolution system of claim 1, wherein, a vector encoding subunit configured to perform vector encoding on the new knowledge item to obtain a high-dimensional vector corresponding to the new knowledge item, wherein the high-dimensional vector comprises a language model encoding vector, a domain label vector and a meta-feature encoding vector; and a vector mixing subunit configured to perform vector splicing and normalization processing on the high-dimensional vector to obtain the corresponding first high-dimensional mixed vector. The conflict calculation unit comprises:

3. The educational knowledge base self-checking and self-evolution system of claim 1, wherein, a distance conflict calculation subunit configured to calculate a semantic distance between the new knowledge item and the old knowledge item according to the first high-dimensional mixed vector, the second high-dimensional mixed vector, a number of semantic feature layers and a layer weight corresponding to each layer of semantic features; a logic conflict calculation subunit configured to calculate a semantic logic conflict probability between the new knowledge item and the old knowledge item according to the first high-dimensional mixed vector, the second high-dimensional mixed vector, the number of semantic feature layers and the layer weight corresponding to each layer of semantic features; and a correlation penalty calculation subunit configured to calculate a context correlation penalty term between the new knowledge item and the old knowledge item according to a first similarity between the new knowledge item and the old knowledge item in a classification label dimension, an overlapping degree between the new knowledge item and the old knowledge item in a time dimension, a second similarity between the new knowledge item and the old knowledge item in a grouping dimension, a first weight coefficient corresponding to the first similarity, a second weight coefficient corresponding to the overlapping degree and a third weight coefficient corresponding to the second similarity. The logic conflict calculation subunit comprises:

4. The educational knowledge base self-checking and self-evolution system of claim 3, wherein, a conflict intensity calculation module configured to input the new knowledge item and the old knowledge item into a constructed logic conflict discrimination model to obtain a conflict intensity between the new knowledge item and the old knowledge item in a logic conflict; and a conflict probability calculation module configured to obtain a semantic logic conflict probability between the new knowledge item and the old knowledge item according to the conflict intensity. The correlation penalty calculation subunit comprises:

5. The educational knowledge base self-checking and self-evolution system of claim 3, wherein, ​ a context similarity calculation module, configured to calculate a context similarity between the new knowledge item and the old knowledge item according to a first similarity between the new knowledge item and the old knowledge item in a classification label dimension, an overlapping degree in a time dimension, a second similarity in a grouping dimension, a first weight coefficient corresponding to the first similarity, a second weight coefficient corresponding to the overlapping degree, and a third weight coefficient corresponding to the second similarity; and a penalty calculation module, configured to calculate a context relevance penalty term between the new knowledge item and the old knowledge item according to the context similarity and a maximum context similarity corresponding to the context similarity.

6. The educational knowledge base self-checking and self-evolution system of claim 1, wherein, The knowledge judgment unit includes the following steps in the process of repairing the new knowledge item to generate a repaired knowledge item and adding the repaired knowledge item to the knowledge base: generating a repaired knowledge item by a large language model according to the new knowledge item, the old knowledge item, and a knowledge base context, and adding the repaired knowledge item to the knowledge base; wherein the repaired knowledge item and the old knowledge item satisfy the following conditions simultaneously: a repaired semantic total conflict score between the repaired knowledge item and the old knowledge item is less than a conflict threshold value; a semantic deviation between a repaired high-dimensional mixed vector corresponding to the repaired knowledge item and the first high-dimensional mixed vector is less than a deviation threshold value; a coherence degree of the repaired knowledge item and the knowledge base is greater than a coherence threshold value.

7. The educational knowledge base self-test and self-evolution system of claim 1, wherein, Further comprising: a knowledge supplement unit; the knowledge supplement unit includes the following steps: a coverage detection subunit, configured to perform key task extraction on the new knowledge item and coverage detection on the key task; and a supplement generation subunit, configured to generate a supplement knowledge item by a large language model according to the key task and the knowledge base when the coverage of the key task is less than a coverage threshold value, and add the supplement knowledge item to the knowledge base. Further comprising:

8. The educational knowledge base self-test and self-evolution system of claim 1, wherein, a parameter updating unit, configured to perform collaborative self-evolution analysis on the dimension weight and the conflict threshold value according to the semantic total conflict score to obtain an updated weight and an updated threshold value; the parameter updating unit includes the following steps: obtaining a corresponding state vector representation according to the dimension weight, the conflict threshold value, and the corresponding semantic total conflict score, wherein the state vector representation includes an action space representation; obtaining a reward function representation according to a consistency reward, a precision reward, a knowledge coverage rate reward, a threshold stability reward between the conflict threshold value and a previous conflict threshold value corresponding to a previous update, a first reward weight corresponding to the consistency reward, a second reward weight corresponding to the precision reward, a third reward weight corresponding to the knowledge coverage rate reward, and a fourth reward weight corresponding to the threshold stability reward; performing parameter updating on the dimension weight and the conflict threshold value according to the action space representation and the reward function representation, a learning rate, and a discount factor to obtain the updated weight and the updated threshold value. Further comprising a global evaluation unit; 9. The educational knowledge base self-checking and self-evolution system of claim 8, wherein, the global evaluation unit includes the following steps: ​ a re-scoring unit configured to re-compute an updated semantic total conflict score according to the updated weights, the updated threshold, the new knowledge item and the old knowledge item; and a knowledge base updating unit configured to obtain an evolutionary objective function according to the updated semantic total conflict score, the updated threshold, the knowledge base before updating, and to perform self-evolution on the knowledge base before updating by maximizing the evolutionary objective function to obtain the knowledge base after updating.

10. An educational knowledge base self-checking and self-evolution method, characterized in that, comprising: obtaining a new knowledge item by an obtaining unit; performing vector mixing and splicing on the new knowledge item by a vector splicing unit to obtain a corresponding first high-dimensional mixed vector; computing a dimension conflict between the first high-dimensional mixed vector corresponding to the new knowledge item and a second high-dimensional mixed vector corresponding to an old knowledge item according to the dimension conflict, the dimension conflict including a semantic distance, a semantic logic conflict probability and a context correlation penalty term, by a conflict computing unit; obtaining a semantic total conflict score between the new knowledge item and the old knowledge item according to the dimension conflict and a dimension weight optimized by cooperative self-evolution, by a conflict synthesizing unit; judging whether the semantic total conflict score is less than a conflict threshold optimized by cooperative self-evolution by a knowledge judging unit, if yes, adding the new knowledge item to a knowledge base, if not, performing knowledge repair on the new knowledge item to generate a repaired knowledge item to be added to the knowledge base.

Images