Tutoring robot processing system based on natural language processing

Abstract

The application discloses a tutoring robot processing system based on natural language processing and belongs to the technical field of robot processing. The system comprises a knowledge point text acquisition module, a knowledge point text preprocessing module, a knowledge point text vectorization processing module, a knowledge point text clustering module and a tutoring robot processing module. The application specifically introduces a chaotic adaptive wandering strategy to improve the dolphin swarm algorithm, obtains the part-of-speech influence, obtains the distribution concentration based on distribution entropy, generates the word importance feature vector, and splices the knowledge point text vector with the semantic feature vector. The initial clustering center and the dimension weight are constructed into a membership matrix according to the double-feature weighted distance, the clustering center is updated according to the credibility index, the dimension weight is updated by combining the clustering contribution, the part-of-speech influence, the distribution concentration and the variance, convergence is judged, and the clustering result is obtained, so that the recommendation result of the tutoring robot is stably focused on the core knowledge point and the accuracy of the tutoring robot recommendation is improved.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of robot processing, and particularly relates to a tutor robot processing system based on natural language processing. BACKGROUND

[0002] The tutor robot processing system is a kind of system which utilizes natural language processing and machine learning technology to realize intelligent response to user learning problems, quickly matches the most relevant knowledge points, generates a personalized recommendation list, and thus provides accurate and efficient learning assistance for users. However, in the existing tutor robot processing system, there is a problem that only relying on term frequency-inverse document frequency or only relying on semantic vectors cannot accurately reflect the semantics and core information of knowledge points, which leads to difficulty in finding relevant knowledge points when matching user queries, and affects the recommendation accuracy of the tutor robot. In the existing tutor robot processing system, there is a problem that the tutor robot cannot automatically identify and highlight the truly core knowledge points and is easily disturbed by noise texts, and the importance of different dimensions of information lacks dynamic adaptation, which leads to the recommendation results of the tutor robot often being mixed with marginal or irrelevant content, and users are difficult to quickly lock the required knowledge points. SUMMARY

[0003] In view of the above situation, in order to overcome the defects of the prior art, the tutor robot processing system based on natural language processing is provided. For the problem in the existing tutor robot processing system that only relying on term frequency-inverse document frequency or only relying on semantic vectors cannot accurately reflect the semantics and core information of knowledge points, which leads to difficulty in finding relevant knowledge points when matching user queries, and affects the recommendation accuracy of the tutor robot, the scheme calculates the term frequency and inverse document frequency, introduces a chaotic adaptive wandering strategy to improve the dolphin swarm algorithm, obtains the part-of-speech influence, obtains the distribution concentration based on the distribution entropy, generates a word importance feature vector, and splices the word importance feature vector and the semantic feature vector to obtain a knowledge point text vector. The generated vector takes into account the importance of vocabulary and deep semantics, accurately depicts the core features of knowledge points, provides a high-quality feature basis for subsequent user query matching, and improves the accuracy of the tutor robot recommendation. For the problem in the existing tutor robot processing system that the tutor robot cannot automatically identify and highlight the truly core knowledge points and is easily disturbed by noise texts, and the importance of different dimensions of information lacks dynamic adaptation, which leads to the recommendation results of the tutor robot often being mixed with marginal or irrelevant content, and users are difficult to quickly lock the required knowledge points, the scheme generates an initial clustering center based on the word importance feature vector, initializes the dimension weight, constructs a membership matrix according to the double-feature weighted distance, updates the clustering center according to the credibility index, combines the clustering contribution, the part-of-speech influence, the distribution concentration and the variance to update the dimension weight, judges the convergence based on the change rate of the objective function value and the maximum moving distance of the clustering center, obtains the clustering result, reduces the disturbance of noise texts on clustering, makes the recommendation results of the tutor robot stably focus on the core knowledge points, and users can quickly lock the required core information from the recommendation list.

[0004] The application provides a tutor robot processing system based on natural language processing, which comprises a knowledge point text acquisition module, a knowledge point text preprocessing module, a knowledge point text vectorization processing module, a knowledge point text clustering module and a tutor robot processing module.

[0005] The knowledge point text acquisition module is used for collecting knowledge point texts by the tutor robot.

[0006] The knowledge point text preprocessing module is used for preprocessing the collected knowledge point texts by the tutor robot to obtain a global vocabulary table and a knowledge point text data set.

[0007] The knowledge point text vectorization processing module calculates the word frequency and the inverse document frequency, introduces a chaotic adaptive wandering strategy to improve a dolphin swarm algorithm, obtains a part-of-speech influence, obtains a distribution concentration based on a distribution entropy, generates a word importance feature vector, and splices the word importance feature vector and a semantic feature vector to obtain a knowledge point text vector.

[0008] The knowledge point text clustering module generates an initial clustering center based on the word importance feature vector, initializes a dimension weight, constructs a membership matrix according to a double-feature weighted distance, updates the clustering center according to a credibility index, updates the dimension weight by combining a clustering contribution, the part-of-speech influence, the distribution concentration and a variance, judges convergence based on a target function value change rate and a maximum moving distance of the clustering center, and obtains a clustering result.

[0009] The tutor robot processing module is used for receiving real-time texts input by a user, selecting a target cluster based on the clustering result, and generating a recommendation list.

[0010] Further, the knowledge point text preprocessing module is used for preprocessing the collected knowledge point texts by the tutor robot, and the preprocessing comprises text cleaning, text segmentation, part-of-speech tagging and data set construction, so as to obtain the global vocabulary table and the knowledge point text data set.

[0011] Further, the knowledge point text vectorization processing module is used for vectorizing each knowledge point text in the knowledge point text data set based on a natural language processing technology, and the knowledge point text vectorization processing module is provided with a word frequency and inverse document frequency calculation unit, a part-of-speech influence calculation unit, a distribution concentration calculation unit, a word importance feature value calculation unit and a text vector generation unit.

[0012] The word frequency and inverse document frequency calculation unit is used for, for each knowledge point text d in the knowledge point text data set A, counting the number of times each word t in the global vocabulary table Z appears in d and the number of times t appears in A, counting the total number of words in d and the total number of texts in A, and obtaining the word frequency of t in d and the inverse document frequency of t in A.

[0013] Part-of-speech influence calculation unit; obtains the optimal part-of-speech influence for each part of speech by improving the dolphin swarm algorithm, and then obtains the part-of-speech influence for each word in the global vocabulary; includes the following:

[0014] Initial dolphin positions: Establish a search space for all parts of speech categories in the global vocabulary, randomly initialize Q dolphin positions within the search space, and use each dolphin position to represent a combination of part-of-speech influences of a set of parts of speech categories. Vectorize each knowledge point text in the knowledge point text dataset based on each combination of part-of-speech influences, and then use the K-Means algorithm to cluster all knowledge point text vectors. Use the F1 score of the clustering results as the fitness value of the corresponding dolphin position.

[0015] Dolphin position update; the dolphin pod algorithm simulates the foraging behavior of dolphins, including four key stages: searching, calling, receiving, and hunting. A chaotic adaptive walk strategy is added after the foraging stage. This strategy is activated when the fitness value of the globally optimal position remains unchanged after five consecutive iterations. The chaotic adaptive walk strategy includes the following:

[0016] Generate a chaotic adaptive walking path; sort all dolphin positions in the dolphin pod in descending order of fitness value, select the top 20% of dolphin positions as high-quality positions, pre-set the walking step number as M, generate a direction factor based on the Logistic chaos value, and generate an adaptive step size based on the Euclidean distance between the high-quality position and the global optimal position. Combine the direction factor and the adaptive step size to obtain the chaotic adaptive walking path corresponding to each high-quality position.

[0017] Update: Set a dynamic boundary, map the chaotic adaptive walk path to the dynamic boundary, generate a new position for each high-quality position. If the fitness value of the new position is greater than the fitness value of the corresponding high-quality position, then replace the corresponding high-quality position with the new position; otherwise, retain the original high-quality position.

[0018] The optimal part-of-speech influence is determined. When the fitness value of a dolphin position is greater than the preset fitness threshold γ, the part-of-speech influence combination represented by that dolphin position is taken as the optimal part-of-speech influence combination. The part-of-speech influence of each word in the global vocabulary is obtained based on the optimal part-of-speech influence combination. Otherwise, if the maximum number of iterations is reached, the iteration count is reset to 0 and the dolphin position is re-initialized. Otherwise, the iteration count is incremented by 1 and the dolphin position is updated.

[0019] Distribution concentration calculation unit; for each word t in the global vocabulary Z, count the number a of knowledge point texts containing t in the knowledge point text dataset A. If knowledge point text d contains t, then calculate the probability of d appearing in all knowledge point texts containing t. for ,otherwise The distribution entropy of t is calculated based on probability, and then normalized to obtain the distribution concentration of t.

[0020] The word importance feature value calculation unit integrates word frequency, inverse document frequency, part-of-speech influence, and distribution concentration to obtain the word importance feature value of each word in the global vocabulary in each knowledge point text.

[0021] Text vector generation unit: For each knowledge point text d in the knowledge point text dataset A, the word importance feature values ​​of all words in d are arranged according to the word order in the global vocabulary Z to form the word importance feature vector corresponding to d. Then, the semantic feature vector of each d is generated based on the BERT model. The word importance feature vector and the semantic feature vector are concatenated to obtain the knowledge point text vector of d. The knowledge point text vector dataset is constructed based on the knowledge point text vectors corresponding to all knowledge point texts in A.

[0022] Furthermore, the knowledge point text clustering module clusters all knowledge point text vectors in the knowledge point text vector dataset. The knowledge point text clustering module includes a clustering parameter initialization unit, a membership matrix construction unit, a cluster center update unit, a dimension weight update unit, and a convergence judgment unit; specifically, it includes the following:

[0023] Clustering parameter initialization unit: Calculate the sampling probability of each knowledge point text vector based on the word importance feature vector, and perform weighted sampling G times from the knowledge point text vector dataset B according to the sampling probability to obtain G initial cluster centers, and set the weight of each dimension of the word importance feature vector and semantic feature vector respectively;

[0024] Membership matrix construction unit; simultaneously considering the weight of each dimension of word importance feature vector and semantic feature vector, calculate the dual-feature weighted distance from each knowledge point text vector to each current cluster center, and assign each knowledge point text vector to the cluster center with the smallest dual-feature weighted distance to obtain the membership matrix;

[0025] Cluster center update unit; calculate the credibility index of each knowledge point text vector based on distribution concentration, and calculate the center l of the word importance feature vector and semantic feature vector of all knowledge point text vectors in each cluster according to the credibility index. wo and l se , will l wo and l se The new cluster centers for each cluster are obtained by splicing them together.

[0026] Dimension weight update unit: Calculate the clustering contribution of each dimension in the word importance feature vector and semantic feature vector respectively, then combine the part-of-speech influence and distribution concentration to update the weight of each dimension in the word importance feature vector, and update the weight of each dimension in the semantic feature vector based on the variance of the dimension.

[0027] Convergence judgment unit; calculates the objective function value based on the dual-feature weighted distance and the confidence index; if the rate of change of the objective function value is less than or equal to 10... -5 And the maximum movement distance among all cluster centers is less than or equal to 10. -3 If the clustering has converged, the clustering process ends and the clustering results are output; otherwise, if the maximum number of updates is reached, the update count is reset to 0 and the process is moved to the clustering parameter initialization unit; otherwise, the update count is incremented by 1 and the process is moved to the membership matrix construction unit.

[0028] Furthermore, the teaching assistant robot processing module receives real-time text input by the user, performs text cleaning, word segmentation, part-of-speech tagging, and vectorization on the real-time text to obtain real-time text vectors, calculates the cosine similarity between the real-time text vectors and each cluster center in the clustering results, selects the cluster to which the cluster center with the highest cosine similarity belongs as the target cluster, then calculates the cosine similarity between the real-time text vectors and the text vectors of each knowledge point in the target cluster, and sorts all the corresponding knowledge point texts in the target cluster in descending order according to the magnitude of the cosine similarity to generate a recommendation list.

[0029] The beneficial effects achieved by adopting the above solution are as follows:

[0030] (1) To address the problem that existing teaching assistant robot processing systems rely solely on word frequency-inverse document frequency or solely on semantic vectors, which fail to accurately reflect the semantics and core information of knowledge points, making it difficult for users to find relevant knowledge points during query matching and affecting the recommendation accuracy of teaching assistant robots, this solution calculates word frequency and inverse document frequency to quantify the local occurrence frequency and global rarity of words; introduces a chaotic adaptive walk strategy to improve the dolphin swarm algorithm to obtain part-of-speech influence, enabling the feature vector to accurately reflect the differences in the contribution of part-of-speech to the semantics of knowledge points; obtains distribution concentration based on distribution entropy, enabling the feature vector to highlight the core terms with concentrated distribution; generates word importance feature vectors, enabling the vectors to comprehensively characterize the contribution of words to knowledge points; and concatenates them with semantic feature vectors to obtain knowledge point text vectors. The generated vectors take into account both word importance and deep semantics, accurately characterizing the core features of knowledge points, providing a high-quality feature foundation for subsequent user query matching, and improving the accuracy of teaching assistant robot recommendations.

[0031] (2) To address the problems in existing teaching assistant robot processing systems, such as the inability to automatically identify and highlight the true core knowledge points, susceptibility to interference from noisy text, and lack of dynamic adaptation to the importance of information in different dimensions, resulting in the teaching assistant robot's recommendation results often containing marginal or irrelevant content, making it difficult for users to quickly locate the required knowledge points, this solution generates initial cluster centers based on word importance feature vectors and initializes dimension weights to make the initial centers closer to the core knowledge points and reduce the interference of noisy text on the initial clustering; constructs a membership matrix based on the weighted distance of dual features, considering both word importance features and semantic features, which can accurately distinguish between core and marginal content and reduce irrelevant text from entering the core cluster; updates the cluster centers based on the credibility index to make the cluster centers more biased towards the features of core knowledge points; combines cluster contribution, part-of-speech influence, distribution concentration, and variance to update dimension weights, so that the clustering process can dynamically focus on the most critical dimensions for distinguishing core knowledge points; judges convergence based on the rate of change of the objective function value and the maximum moving distance of the cluster centers to obtain the clustering results, so that the teaching assistant robot's recommendation results can stably focus on core knowledge points, and users can quickly locate the required core information from the recommendation list. Attached Figure Description

[0032] Figure 1 A schematic diagram of the teaching assistant robot processing system based on natural language processing provided by the present invention;

[0033] Figure 2 This is a schematic diagram of the knowledge point text vectorization processing module;

[0034] Figure 3 This is a schematic diagram of the knowledge point text clustering module.

[0035] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention and do not constitute a limitation thereof. Detailed Implementation

[0036] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0037] In the description of this invention, it should be understood that the terms "upper", "lower", "front", "rear", "left", "right", "top", "bottom", "inner", "outer", etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.

[0038] Example 1, see Figure 1 The teaching assistant robot processing system based on natural language processing provided by the present invention includes a knowledge point text acquisition module, a knowledge point text preprocessing module, a knowledge point text vectorization processing module, a knowledge point text clustering module, and a teaching assistant robot processing module.

[0039] The knowledge point text acquisition module is used by the teaching assistant robot to collect knowledge point text and send the data to the knowledge point text preprocessing module.

[0040] The knowledge point text preprocessing module receives data sent by the knowledge point text acquisition module. The teaching assistant robot preprocesses the acquired knowledge point text to obtain a global vocabulary and a knowledge point text dataset, and then sends the data to the knowledge point text vectorization processing module.

[0041] The knowledge point text vectorization processing module receives data sent by the knowledge point text preprocessing module, calculates word frequency and inverse document frequency, introduces a chaotic adaptive walk strategy to improve the dolphin swarm algorithm, obtains part-of-speech influence, obtains distribution concentration based on distribution entropy, generates word importance feature vectors, and concatenates them with semantic feature vectors to obtain knowledge point text vectors, and sends the data to the knowledge point text clustering module.

[0042] The knowledge point text clustering module receives data sent by the knowledge point text vectorization processing module, generates initial cluster centers based on word importance feature vectors, initializes dimension weights, constructs a membership matrix based on dual-feature weighted distance, updates cluster centers based on credibility index, updates dimension weights by combining cluster contribution, part-of-speech influence, distribution concentration and variance, determines convergence based on the rate of change of objective function value and the maximum movement distance of cluster centers, obtains clustering results, and sends the data to the teaching assistant robot processing module.

[0043] The teaching assistant robot processing module receives data sent by the knowledge point text clustering module. The teaching assistant robot receives real-time text input by the user, selects target clusters based on the clustering results, and generates a recommendation list.

[0044] Example 2, see Figure 1 This embodiment is based on the above embodiment. In the knowledge point text acquisition module, the teaching assistant robot acquires knowledge point text based on textbooks, courseware and exercise sets.

[0045] Example 3, see Figure 1 This embodiment is based on the above embodiment. In the knowledge point text preprocessing module, the teaching assistant robot preprocesses the collected knowledge point text. The preprocessing includes text cleaning, text segmentation, part-of-speech tagging, and dataset construction to obtain a global vocabulary and a knowledge point text dataset.

[0046] The text cleaning mentioned above involves removing special symbols, stop words, and low-frequency words that appear less than 3 times from the text containing knowledge points.

[0047] The text segmentation involved using the Jieba tool and the NLTK tool to segment Chinese and English text respectively, breaking down each knowledge point text into a word sequence.

[0048] The part-of-speech tagging is used to identify the grammatical attributes of words and to tag the part of speech of each word in the knowledge point text.

[0049] The constructed dataset is a knowledge point text dataset built based on all knowledge point texts after text cleaning, text segmentation, and part-of-speech tagging, and the vocabulary of all knowledge point texts is counted to generate a global vocabulary list.

[0050] Example 4, see Figure 1 and Figure 2 This embodiment is based on the above embodiment. In the knowledge point text vectorization processing module, each knowledge point text in the knowledge point text dataset is vectorized based on natural language processing technology. The knowledge point text vectorization processing module includes a word frequency and inverse document frequency calculation unit, a part-of-speech influence calculation unit, a distribution concentration calculation unit, a word importance feature value calculation unit, and a text vector generation unit; specifically, it includes the following:

[0051] The term frequency (TF) and inverse document frequency (IVF) calculation units are used because simply counting the number of times a word appears in a text is insufficient to distinguish its importance. TF quantifies the local frequency of a word within a single text, while IVF quantifies its global rarity across all texts, thus comprehensively reflecting word importance. For each knowledge point text d in the knowledge point text dataset A, each word t in the global vocabulary Z is traversed, and the frequency of t in d and in A is counted. The total number of words in d and the total number of texts in A are also counted to obtain the TF frequency of t in d and its IVF in A. The formulas used are as follows:

[0052] ;

[0053] ;

[0054] In the formula, d i t is the text of the i-th knowledge point in A.j It is the j-th word in Z. It is t j In d i The number of times it appears in N i It is d i Total number of words in It is t j In d i word frequency in Is A containing t? j The number of texts, N A This represents the total number of text elements in A. It is t j In A, the inverse document frequency, i and j are the knowledge point text index and the vocabulary index, respectively;

[0055] Part-of-speech (POS) influence calculation unit: Different POS have varying importance in knowledge points, but their weights are difficult to determine subjectively, and traditional optimization algorithms are prone to getting trapped in local optima. This unit utilizes an improved dolphin swarm algorithm, using the F1 score of knowledge point clustering as fitness, to optimize and obtain the optimal influence for each POS. This objectively quantifies POS importance, adapts to the semantic characteristics of knowledge points, avoids local optima, improves optimization accuracy, and ensures that POS influence meets clustering requirements. The improved dolphin swarm algorithm obtains the optimal POS influence for each POS, thereby obtaining the POS influence for each word in the global vocabulary. This includes the following:

[0056] Initial dolphin positions; establish a search space for all parts of speech categories in the global vocabulary, and randomly initialize Q dolphin positions within the search space, Q=50; use each dolphin position to represent a combination of part-of-speech influences of a set of parts of speech categories, vectorize each knowledge point text in the knowledge point text dataset based on each combination of part-of-speech influences, and then use the K-Means algorithm to cluster all knowledge point text vectors, and use the F1 score of the clustering results as the fitness value of the corresponding dolphin position;

[0057] Dolphin position update; the dolphin pod algorithm simulates the foraging behavior of dolphins, including four key stages: searching, calling, receiving, and hunting. It iteratively updates dolphin positions to find the optimal position. After the foraging stage, a chaotic adaptive walk strategy is added. This strategy is activated when the fitness value of the globally optimal position remains unchanged after five consecutive iterations. The chaotic adaptive walk strategy includes the following:

[0058] A chaotic adaptive walking path is generated. All dolphin positions in the pod are sorted in descending order of fitness value, and the top 20% of dolphin positions are selected as high-quality positions. The walking step size is pre-set to M. A direction factor is generated based on the Logistic chaos value, and an adaptive step size is generated based on the Euclidean distance between the high-quality position and the global optimum. The direction factor and adaptive step size are combined to obtain the chaotic adaptive walking path corresponding to each high-quality position. The formula used is as follows:

[0059] ;

[0060] ;

[0061] ;

[0062] In the formula, and These are the direction factor and adaptive step size for the q-th quality position to traverse at step m in the p-th iteration, respectively. m It is the Logistic chaos value corresponding to the m-th step of the walk. M=10 It is the Euclidean distance between the q-th high-quality position and the globally optimal position at the p-th iteration. The globally optimal position is the dolphin position with the highest fitness value. is the maximum Euclidean distance between the dolphin positions in the pod and the globally optimal positions at the p-th iteration, where q, m, and p are the indexes of the optimal positions, the number of steps taken, and the number of iterations, respectively. It is the chaotic adaptive walk path at the q-th quality position in the p-th iteration;

[0063] Update: Set a dynamic boundary, map the chaotic adaptive walk path to the dynamic boundary, generate a new position for each prime position. If the fitness value of the new position is greater than the fitness value of the corresponding prime position, then replace the corresponding prime position with the new position; otherwise, retain the original prime position. The formula used is as follows:

[0064] ;

[0065] ;

[0066] ;

[0067] ;

[0068] In the formula, and These are the upper and lower boundaries in the w-th dimension at the p-th iteration, respectively. and These are the upper and lower boundaries of the search space in the w-th dimension, respectively, p max It is the maximum number of iterations, p max =100, and These are the values ​​of the q-th high-quality position and the global optimal position in the w-th dimension, respectively, during the p-th iteration. yes The corresponding new position, and These are the adaptation adjustment coefficient and fitness value of the q-th high-quality position in the p-th iteration, respectively. L is the average fitness value of the positions of all dolphins in the pod at the p-th iteration. min and L max These are the minimum and maximum values ​​among all chaotic adaptive walk paths, respectively, where w is the dimension index;

[0069] The optimal part-of-speech influence is determined as follows: When the fitness value of a dolphin position is greater than the preset fitness threshold γ (γ=0.85), the part-of-speech influence combination represented by that dolphin position is taken as the optimal part-of-speech influence combination. The part-of-speech influence of each word in the global vocabulary is obtained based on the optimal part-of-speech influence combination. Otherwise, if the maximum number of iterations is reached, the iteration count is reset to 0 and the dolphin position is re-initialized. Otherwise, the iteration count is incremented by 1 and the dolphin position is updated again.

[0070] Distribution concentration calculation unit: The distribution range of words in knowledge point texts affects their representativeness to specific knowledge points. Through distribution entropy and normalization calculations, the degree of concentration of word distribution in knowledge point texts is quantified, reflecting the close relationship between words and specific knowledge points. Concentratedly distributed words are more likely to be core terms of the knowledge point, providing a supplementary dimension for word importance assessment. For each word t in the global vocabulary Z, the number a of knowledge point texts containing t in the knowledge point text dataset A is counted. If knowledge point text d contains t, the probability of d appearing in all knowledge point texts containing t is calculated. for ,otherwise The distribution entropy of t is calculated based on probability, and then normalized to obtain the distribution concentration of t; the formula used is as follows:

[0071] ;

[0072] ;

[0073] In the formula, and They are t j The distribution entropy and distribution concentration, It is in the case of t j In all the knowledge points text, di The probability of occurrence, a j Is A containing t? j The number of text entries related to the knowledge points;

[0074] The word importance feature value calculation unit addresses the fact that a single dimension cannot fully reflect the importance of words in knowledge points. By integrating TF, IDF, part-of-speech influence, and distribution concentration, a comprehensive importance feature value for each word in a specific text is obtained. This multi-dimensional feature fusion more comprehensively characterizes the contribution of words to knowledge points, avoiding the one-sidedness of a single indicator and improving the accuracy of feature representation. The unit combines word frequency, inverse document frequency, part-of-speech influence, and distribution concentration to obtain the word importance feature value for each word in the global vocabulary in each knowledge point text. The formula used is as follows:

[0075] ;

[0076] In the formula, It is t j In d i Word importance feature values ​​in It is t j The influence of the part of speech;

[0077] The text vector generation unit addresses the challenge that word importance feature vectors alone cannot capture the deep semantics of text, and BERT semantic vectors alone cannot reflect the specific importance of words. By concatenating word importance feature vectors with BERT semantic feature vectors, a knowledge point text vector is formed that balances word importance and deep semantics. This preserves word-level importance information while capturing the overall semantics of the text, resulting in a more comprehensive vector representation and laying a high-quality feature foundation for subsequent clustering and matching. For each knowledge point text d in the knowledge point text dataset A, the word importance feature values ​​of all words in d are arranged according to the word order in the global vocabulary Z, forming a J-dimensional word importance feature vector for d. Then, a semantic feature vector for each d is generated based on the BERT model. The word importance feature vector and the semantic feature vector are concatenated to obtain the knowledge point text vector for d. A knowledge point text vector dataset is constructed based on the knowledge point text vectors corresponding to all knowledge point texts in A. The formulas used are as follows:

[0078] ;

[0079] In the formula, It is d i The word importance feature vector, where J is the number of words in Z. , and They are The values ​​in the 1st, jth, and Jth dimensions; if The vocabulary t corresponding to the j-th dimension j In di If it appears in the middle, then ,otherwise .

[0080] By performing the above operations, this solution addresses the problem in existing teaching assistant robot processing systems that rely solely on word frequency-inverse document frequency or semantic vectors, failing to accurately reflect the semantics and core information of knowledge points. This leads to difficulties in finding relevant knowledge points during user queries, impacting the accuracy of teaching assistant robot recommendations. The solution calculates word frequency and inverse document frequency to quantify the local frequency and global rarity of words. It introduces a chaotic adaptive walk strategy to improve the dolphin swarm algorithm, obtaining part-of-speech influence so that feature vectors accurately reflect the differences in the contribution of parts of speech to the semantics of knowledge points. Based on distribution entropy, it obtains distribution concentration, allowing feature vectors to highlight concentrated core terms. It generates word importance feature vectors, comprehensively characterizing the contribution of words to knowledge points. These vectors are then concatenated with semantic feature vectors to obtain knowledge point text vectors. The generated vectors balance word importance and deep semantics, accurately characterizing the core features of knowledge points and providing a high-quality feature foundation for subsequent user queries, thus improving the accuracy of teaching assistant robot recommendations.

[0081] Example 5, see Figure 1 and Figure 3 This embodiment, based on the above embodiment, clusters all knowledge point text vectors in the knowledge point text vector dataset within the knowledge point text clustering module. The knowledge point text clustering module includes a clustering parameter initialization unit, a membership matrix construction unit, a cluster center update unit, a dimension weight update unit, and a convergence judgment unit; specifically, it includes the following:

[0082] Clustering parameter initialization unit: Randomly initializing cluster centers may cause the clustering results to deviate from the true distribution of knowledge points, and unreasonable setting of the dimensional weights of word importance and semantic features will affect the clustering direction. Initial cluster centers are selected based on the sampling probability weighted by the word importance feature vector, and the dimensional weights are reasonably initialized to make the initial centers closer to the true knowledge point cluster centers. The initial weight values ​​are adapted to the feature importance, reducing the number of subsequent iterations and improving the quality of the starting point of clustering. The sampling probability of each knowledge point text vector is calculated based on the word importance feature vector, and the number of cluster categories G is determined based on the silhouette coefficient. Based on the sampling probability, G initial cluster centers are obtained by weighted sampling G times from the knowledge point text vector dataset B, and the weights of each dimension of the word importance feature vector and semantic feature vector are set respectively. The formulas used are as follows:

[0083] ;

[0084] ;

[0085] ;

[0086] In the formula, C v It is o v The sampling probability, and They are o v and o k The value of the corresponding word importance feature vector in the j-th dimension, o v and o k These are the v-th and k-th knowledge point text vectors in B, respectively, where v and k are the indices of the knowledge point text vectors. It is the weight of the j-th dimension of the word importance feature vector. It is the weight of the r-th dimension of the semantic feature vector. R is the maximum value of the text vector of all knowledge points in B in the j-th dimension, and R is the dimension of the semantic feature vector.

[0087] Membership matrix construction unit: Calculating distance based solely on word importance or semantic features may lead to clustering bias. Therefore, a weighted distance combining word importance and semantic features is calculated to balance lexical importance and semantic similarity, making the clustering results more consistent with the actual connections between knowledge points. Simultaneously, the weights of each dimension of both word importance and semantic feature vectors are considered. The weighted distance from each knowledge point text vector to each current cluster center is calculated, and each knowledge point text vector is assigned to the cluster with the smallest weighted distance, thus obtaining the membership matrix. The formula used is as follows:

[0088] ;

[0089] in, It is the g-th cluster center during the nth update. It is o v and The weighted distance between the two features It represents the weight of the j-th dimension of the word importance feature vector during the nth update. It represents the weight of the r-th dimension of the semantic feature vector during the n-th update. yes The value of the corresponding word importance feature vector in the j-th dimension. and They are o v and The value of the corresponding semantic feature vector in the r-th dimension, It is the membership matrix at the nth update. yes The value in row v and column g; when o v Assigned to When it belongs to a cluster, ,otherwise, ;

[0090] Cluster center update unit: Different knowledge point texts have varying credibility levels. Direct averaging can cause cluster centers to deviate from the true core. This unit calculates a text credibility index based on distribution concentration and updates the cluster centers using weighted word importance and semantic features. This makes the cluster centers more biased towards high-credibility knowledge point texts, improving the representativeness of the centers and making the texts within clusters more consistent. The credibility index of each knowledge point text vector is calculated based on distribution concentration. Based on the credibility index, the center l of the word importance feature vector and semantic feature vector of all knowledge point text vectors in each cluster is calculated. wo and l se , will l wo and l se The new cluster centers for each cluster are obtained by splicing the data; the formula used is as follows:

[0091] ;

[0092] ;

[0093] ;

[0094] in, It is o v Credibility index, d v It is o v The corresponding knowledge point text, It is an indicator function; if ,but ,otherwise, ; It is the center value of the word importance feature vector in the j-th dimension of all knowledge point text vectors in the g-th cluster at the (n+1)-th update. It is the center value of the semantic feature vector of all knowledge point text vectors in the g-th cluster at the (n+1)th update in the r-th dimension;

[0095] Dimensional weight update unit: The role of each dimension of word importance and semantic features changes iteratively during the clustering process. Fixed weights cannot adapt to dynamic needs. Based on the clustering contribution, the weights are dynamically updated by combining part-of-speech influence, distribution concentration, and variance. This highlights dimensions that are more important for cluster differentiation and suppresses irrelevant dimensions, making the weights adapt to the clustering process and improving clustering accuracy. The clustering contribution of each dimension in the word importance feature vector and semantic feature vector is calculated separately. Then, the weights of each dimension in the word importance feature vector are updated by combining part-of-speech influence and distribution concentration. Finally, the weights of each dimension in the semantic feature vector are updated based on the variance of the dimensions. The formulas used are as follows:

[0096] ;

[0097] ;

[0098] ;

[0099] ;

[0100] in, It is the clustering contribution of the j-th dimension of the word importance feature vector at the nth update. It is the clustering contribution of the r-th dimension of the semantic feature vector at the n-th update. It is the weight of the j-th dimension of the word importance feature vector at the (n+1)-th update. δ represents the weight of the r-th dimension of the semantic feature vector during the (n+1)-th update, and δ and φ are the word sparsity penalty coefficient and semantic sparsity penalty coefficient, respectively, where δ=0.1 and φ=0.05. It is the variance of the semantic feature vector of all knowledge point text vectors in B in the r-th dimension;

[0101] Convergence Judgment Unit: Cluster updates require defined termination conditions to avoid insufficient or excessive updates. Convergence is determined by the rate of change of the objective function and the maximum movement distance of cluster centers, controlling update termination to ensure stable and optimal clustering results, balancing processing efficiency and result quality, and avoiding invalid updates. The objective function value is calculated based on a dual-feature weighted distance and a confidence index. If the rate of change of the objective function value is less than or equal to 10... -5 And the maximum movement distance among all cluster centers is less than or equal to 10. -3 If the clustering has converged, the clustering process ends and the results are output; otherwise, if the preset maximum number of updates n is reached... max If the update count is zero, the process moves to the clustering parameter initialization unit; otherwise, the update count is incremented by 1 and the process moves to the membership matrix construction unit. The formula used is as follows:

[0102] ;

[0103] ;

[0104] ;

[0105] In the formula, and These are the objective function values ​​at the nth and (n-1th)th updates, respectively. It is the rate of change of the objective function value at the nth update. It is the g-th cluster center during the (n-1)-th update, where n max =50, It is the maximum movement distance among all cluster centers at the nth update. It is the L2 paradigm. It is a smoothing term. .

[0106] By performing the above operations, this solution addresses the problems in existing teaching assistant robot processing systems, such as the inability to automatically identify and highlight truly core knowledge points, susceptibility to interference from noisy text, and lack of dynamic adaptation to the importance of information across different dimensions. These issues often result in teaching assistant robot recommendations containing marginal or irrelevant content, making it difficult for users to quickly locate the required knowledge points. This solution generates initial cluster centers based on word importance feature vectors and initializes dimensional weights to ensure the initial centers are closer to core knowledge points, reducing interference from noisy text. A membership matrix is ​​constructed based on dual-feature weighted distance, considering both word importance and semantic features, accurately distinguishing core and peripheral content and reducing irrelevant text from entering core clusters. Cluster centers are updated based on a credibility index, making them more aligned with the features of core knowledge points. Dimensional weights are updated by combining cluster contribution, part-of-speech influence, distribution concentration, and variance, allowing the clustering process to dynamically focus on the most critical dimensions for distinguishing core knowledge points. Convergence is determined based on the rate of change of the objective function value and the maximum movement distance of the cluster centers, resulting in a stable focus on core knowledge points in the teaching assistant robot's recommendations, enabling users to quickly locate the required core information from the recommendation list.

[0107] Example 6, see Figure 1 This embodiment is based on the above embodiment. In the teaching assistant robot processing module, the real-time text input by the user needs to be quickly matched with the most relevant knowledge points; otherwise, effective assistance cannot be provided in a timely manner. After the real-time text is vectorized, the most relevant clusters and knowledge points within each cluster are matched using cosine similarity to generate a recommendation list. Based on the clustering results and similarity calculation, it is ensured that the recommended knowledge points are highly relevant to the user input. The recommendation list is sorted by similarity to improve the efficiency of the user's information acquisition and quickly meet the learning needs. The teaching assistant robot receives the real-time text input by the user, performs text cleaning, text segmentation, part-of-speech tagging, and vectorization on the real-time text to obtain real-time text vectors. The cosine similarity between the real-time text vectors and each cluster center in the clustering results is calculated. The cluster to which the cluster center with the highest cosine similarity belongs is selected as the target cluster. The cosine similarity between the real-time text vectors and the text vectors of each knowledge point in the target cluster is then calculated. All the corresponding knowledge point texts in the target cluster are sorted in descending order according to the magnitude of the cosine similarity to generate a recommendation list.

[0108] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0109] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention.

[0110] The present invention and its embodiments have been described above. This description is not restrictive, and the accompanying drawings are only one embodiment of the present invention; the actual structure is not limited thereto. In conclusion, if those skilled in the art are inspired by this description and design similar structures and embodiments without departing from the spirit of the invention, such designs should fall within the protection scope of the present invention.

Claims

1. A tutor robot processing system based on natural language processing, characterized in that: The knowledge point text acquisition module, the knowledge point text preprocessing module, the knowledge point text vectorization processing module, the knowledge point text clustering module and the tutor robot processing module are included. The knowledge point text acquisition module is a tutor robot collecting knowledge point texts. The knowledge point text preprocessing module is a tutor robot preprocessing the collected knowledge point texts to obtain a global vocabulary table and a knowledge point text data set. The knowledge point text vectorization processing module calculates the word frequency and the inverse document frequency, introduces a chaotic adaptive wandering strategy to improve the dolphin swarm algorithm, obtains the part-of-speech influence, obtains the distribution concentration based on the distribution entropy, generates the word importance feature vector, and splices the word importance feature vector and the semantic feature vector to obtain the knowledge point text vector. The knowledge point text clustering module generates an initial clustering center based on the word importance feature vector, initializes the dimension weight, constructs a membership matrix according to the double feature weighted distance, updates the clustering center according to the credibility index, updates the dimension weight by combining the clustering contribution, the part-of-speech influence, the distribution concentration and the variance, judges the convergence based on the target function value change rate and the maximum moving distance of the clustering center, and obtains the clustering result. The tutor robot processing module is a tutor robot receiving real-time texts input by a user, selecting a target cluster based on the clustering result, and generating a recommendation list. The knowledge point text vectorization processing module is provided with a part-of-speech influence calculation unit, which specifically includes the following contents: Initial dolphin position; Dolphin position update; the dolphin swarm algorithm simulates the hunting behavior of a dolphin swarm, including four key stages of searching, calling, receiving and hunting, and a chaotic adaptive wandering strategy is added after the hunting stage; when the fitness value of the global optimal position does not change for five consecutive iterations, the chaotic adaptive wandering strategy is started; the chaotic adaptive wandering strategy includes the following contents: Generating a chaotic adaptive wandering path; Updating; setting a dynamic boundary, mapping the chaotic adaptive wandering path into the dynamic boundary, generating a new position for each excellent position, if the fitness value of the new position is greater than the fitness value of the corresponding excellent position, replacing the corresponding excellent position with the new position; otherwise, keeping the original excellent position; Optimal part-of-speech influence determination; when the fitness value of the dolphin position is greater than a pre-set fitness threshold, the part-of-speech influence combination represented by the dolphin position is combined as the optimal part-of-speech influence combination, and the part-of-speech influence of each word in the global vocabulary table is obtained based on the optimal part-of-speech influence combination; otherwise, if the maximum number of iterations is reached, the number of iterations is reset to 0 and the dolphin position is reinitialized; otherwise, the number of iterations is increased by 1 and the dolphin position update is continued.

2. The natural language processing based teaching assistant robot processing system according to claim 1, characterized in that: The knowledge point text vectorization processing module is based on natural language processing technology to vectorize each knowledge point text in the knowledge point text data set, and the knowledge point text vectorization processing module is provided with a word frequency and inverse document frequency calculation unit, a part-of-speech influence calculation unit, a distribution concentration calculation unit, a word importance feature value calculation unit and a text vector generation unit; specifically The following contents are included: a word frequency and inverse document frequency calculation unit; for each knowledge point text d in the knowledge point text data set A, each vocabulary t in the global vocabulary table Z is traversed, the number of occurrences of t in d and the number of occurrences in A are counted, and the total number of vocabularies in d and the total number of texts in A are counted, to obtain the word frequency of t in d and the inverse document frequency in A; a part-of-speech influence calculation unit; the optimal part-of-speech influence of each part-of-speech is obtained by improving the dolphin swarm algorithm, and then the part-of-speech influence of each vocabulary in the global vocabulary table is obtained; a distribution concentration calculation unit; a word importance feature value calculation unit; the word frequency, inverse document frequency, part-of-speech influence and distribution concentration are fused to obtain the word importance feature value of each vocabulary in the global vocabulary table in each knowledge point text; a text vector generation unit; for each knowledge point text d in the knowledge point text data set A, the word importance feature values of all vocabularies in d are arranged in the order of the vocabularies in the global vocabulary table Z to form the word importance feature vector corresponding to d, and then the semantic feature vector of each d is generated based on the BERT model; the knowledge point text vector of d is obtained by splicing the word importance feature vector and the semantic feature vector, and the knowledge point text vector data set is constructed based on the knowledge point text vectors corresponding to all knowledge point texts in A.

3. The natural language processing based teaching assistant robot processing system of claim 1, wherein: The initial dolphin position is to establish a search space for all part-of-speech categories in the global vocabulary table, randomly initialize Q dolphin positions in the search space, use each dolphin position to represent the part-of-speech influence combination of a group of part-of-speech categories, perform vectorization processing on each knowledge point text in the knowledge point text data set based on each part-of-speech influence combination, and then use the K-Means algorithm to cluster all knowledge point text vectors. The F1 score of the clustering result is used as the fitness value of the corresponding dolphin position.

4. The natural language processing based teaching assistant robot processing system of claim 1, wherein: The generated chaotic adaptive wandering path is to arrange all dolphin positions in the dolphin swarm in descending order of fitness value, select the top 20% dolphin positions as high-quality positions, pre-set the wandering step number as M, generate a direction factor based on a Logistic chaotic value, and generate an adaptive step length based on the Euclidean distance between the high-quality positions and the global optimal position. The direction factor and the adaptive step length are combined to obtain the chaotic adaptive wandering path corresponding to each high-quality position.

5. The natural language processing based teaching assistant robot processing system according to claim 2, characterized in that: The distribution concentration calculation unit is to count the number of knowledge point texts containing t in the knowledge point text data set A for each word t in the global vocabulary table Z, and if the knowledge point text d contains t, the probability of the occurrence of d in all knowledge point texts containing t For , otherwise ; The distribution entropy of t is calculated based on probability, and the distribution entropy is normalized to obtain the distribution concentration of t.

6. The natural language processing based teaching assistant robot processing system according to claim 1, characterized in that: The knowledge point text clustering module is used to cluster all knowledge point text vectors in the knowledge point text vector data set. The knowledge point text clustering module is provided with a clustering parameter initialization unit, a membership matrix construction unit, a cluster center updating unit, a dimension weight updating unit and a convergence judgment unit. The following contents are included: a clustering parameter initialization unit; a membership matrix construction unit; a cluster center updating unit; a dimension weight updating unit; Calculate the credibility index of each knowledge point text vector based on the distribution concentration degree, and calculate the center l of the word importance feature vector and the semantic feature vector of all knowledge point text vectors in each cluster according to the credibility index wo And l se , l wo And l se Splice to get the new clustering center of each cluster; ​ Convergence judgment unit; based on double feature weighted distance quantity and credibility index, if the change rate of target function value is less than or equal to 10 -5 And the maximum moving distance in all cluster centers is less than or equal to 10 -3 Then it has converged, end clustering and output clustering results; Otherwise, if the preset maximum number of updates is reached, the number of updates is reset to 0 and the process is transferred to the cluster parameter initialization unit; Otherwise, the number of updates is incremented by 1 and the process is transferred to the membership matrix construction unit.

7. The natural language processing based teaching assistant robot processing system according to claim 6, characterized in that: The membership matrix construction unit calculates the double-feature weighted distance of each knowledge point text vector to each cluster center based on the weight of each dimension of the word importance feature vector and the semantic feature vector, assigns each knowledge point text vector to the cluster to which the cluster center with the smallest double-feature weighted distance belongs, and obtains the membership matrix.

8. The natural language processing based teaching assistant robot processing system according to claim 6, characterized in that: The dimension weight updating unit calculates the clustering contribution degree of each dimension in the word importance feature vector and the semantic feature vector, respectively, and then updates the weight of each dimension in the word importance feature vector by combining the part-of-speech influence and the distribution concentration, and updates the weight of each dimension in the semantic feature vector based on the variance of the dimension.

9. The natural language processing based teaching assistant robot processing system of claim 1, wherein: The knowledge point text preprocessing module preprocesses the collected knowledge point text by the teaching assistant robot, which includes text cleaning, text segmentation, part-of-speech tagging, and dataset construction, to obtain a global vocabulary and a knowledge point text dataset.

10. The natural language processing based teaching assistant robot processing system of claim 1, wherein: The teaching assistant robot processing module receives real-time text input by the user, performs text cleaning, text segmentation, part-of-speech tagging, and vectorization processing on the real-time text, obtains a real-time text vector, calculates the cosine similarity between the real-time text vector and each cluster center in the clustering result, selects the cluster to which the cluster center with the highest cosine similarity belongs as the target cluster, calculates the cosine similarity between the real-time text vector and each knowledge point text vector in the target cluster, arranges all knowledge point texts in the target cluster in descending order of cosine similarity, and generates a recommendation list.

Images