An educational evaluation target data processing method and system

Abstract

The application discloses an education evaluation target data processing method and system, and is based on the basic semantic vector generated by the DistilBERT model to calculate the fragment density value, automatically identifies the peak point and the boundary point through the nonparametric density peak value clustering algorithm, dynamically generates the binning with the peak point as the core and by giving consideration to the theme category consistency and the density value standard, does not need to preset the clustering quantity, can adapt to the multi-peak / scattered distribution scene, avoids the binning fragmentation or improper combination of key information, generates the basic semantic vector by the DistilBERT model, is matched with the single-branch multi-scale convolution model, replaces the complex shared operator architecture, quickly screens the primary information through the BM25 algorithm, reduces invalid calculation, greatly reduces the hardware configuration requirement, enables ordinary education institutions to realize efficient deployment and real-time processing without high-end equipment, and guarantees the accuracy of extraction of core evaluation information.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to a method and system for processing educational evaluation target data. Background Technology

[0002] Educational evaluation data refers to various types of information collected, processed, and analyzed during the educational evaluation process, which are used to make value judgments on the degree of achievement of educational goals, the effectiveness of the teaching process, and the level of student development.

[0003] In traditional educational evaluation data processing methods, evaluation texts often rely on manual reading and summarization, which is time-consuming and labor-intensive, and is easily affected by subjective factors, resulting in the evaluation results needing to be improved in terms of objectivity and accuracy.

[0004] With the development of artificial intelligence technology, people have proposed using AI to process educational evaluation data. Existing processing solutions address the problem of inaccurate information extraction caused by the random distribution of information of interest within educational evaluation texts by employing a "binning and segmentation-multi-granularity recognition" approach.

[0005] However, information binning generally relies on clustering algorithms (such as K-means), requires pre-determining the number of clusters, and is sensitive to the initial cluster centers. Furthermore, if the information of interest in the educational evaluation text is scattered or has a multi-peak distribution, it is prone to problems such as excessive fragmentation of binning or improper merging of key information. At the same time, multi-granularity recognition involves complex structures such as multi-granularity implicit representation, shared operators (including multi-level stage modules), and hierarchical recognition algorithms, which require a large amount of computing resources. This may make it difficult for ordinary educational institutions to achieve efficient deployment and real-time processing due to insufficient hardware configuration.

[0006] Therefore, there is an urgent need for a more lightweight method for processing educational evaluation target data, which can overcome the limitations of binning algorithms while addressing the problems of high model complexity and high resource consumption. Summary of the Invention

[0007] The technical problem to be solved by this invention is to provide a method and system for processing educational evaluation target data, which can overcome the limitations of binning algorithms while solving the problems of high model complexity and high resource consumption.

[0008] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows:

[0009] A method for processing educational evaluation target data, comprising the following steps:

[0010] S1. Using the DistilBERT model, basic semantic vectors for sentence-level text fragments are generated for educational evaluation texts.

[0011] S2. Based on the basic semantic vector, the fragment density value is matched with the preset evaluation keyword library by cosine similarity calculation to obtain the topic category, and density peak points and boundary points are automatically identified by non-parametric density peak clustering algorithm.

[0012] S3. Using the peak point as the core, merge adjacent text segments with a segment density value greater than a preset threshold and the same topic category to form a bin. Based on the position of the text segments at the beginning and end of each bin, extract the corresponding sub-text from the educational evaluation text to generate a sub-text set.

[0013] S4. Based on the BM25 algorithm, calculate the BM25 score of each evaluation keyword in the evaluation keyword library with the sub-text set and the BM25 score with the educational evaluation text, and determine the integrated initial information set based on the BM25 score.

[0014] S5. Input the basic semantic vector corresponding to each text segment in the initial information set into a pre-trained single-branch multi-scale convolutional model. The single-branch multi-scale convolutional model generates the topic category and confidence of each evaluation keyword and calculates the bin density value of the bin to which the evaluation keyword belongs.

[0015] S6. Calculate the weights based on the bin density value, the topic category, and the confidence level, and sort and filter according to the weights to obtain the core evaluation information.

[0016] To solve the above-mentioned technical problems, another technical solution adopted by the present invention is as follows:

[0017] An educational evaluation target data processing system, comprising:

[0018] The data acquisition module is used to acquire the educational evaluation text to be processed;

[0019] The data processing module implements the steps in the above-described method for processing educational evaluation target data.

[0020] The data output module outputs the processing results from the data processing module.

[0021] The beneficial effects of this invention are as follows: The educational evaluation target data processing method and system of this invention calculates fragment density values ​​based on basic semantic vectors generated by the DistilBERT model. It automatically identifies peak points and boundary points through a parameter-free density peak clustering algorithm. Using peak points as the core, it dynamically generates bins according to rules that balance theme category consistency and density value compliance. This eliminates the need to pre-determine the number of clusters and adapts to multi-peak / scattered distribution scenarios, avoiding bin fragmentation or improper merging of key information. The basic semantic vectors generated by the DistilBERT model, combined with a single-branch multi-scale convolutional model, replace complex shared operator architectures. Furthermore, the BM25 algorithm quickly filters initial information, reducing invalid computation and significantly lowering hardware configuration requirements. This allows ordinary educational institutions to achieve efficient deployment and real-time processing without high-end equipment, while ensuring the accuracy of core evaluation information extraction. Attached Figure Description

[0022] Figure 1 This is a simplified flowchart of an educational evaluation target data processing method according to an embodiment of the present invention;

[0023] Figure 2 This is a schematic diagram of an educational evaluation target data processing system according to an embodiment of the present invention;

[0024] Label Explanation:

[0025] 1. An educational evaluation target data processing system; 2. A data acquisition module; 3. A data processing module; 4. A data output module. Detailed Implementation

[0026] To explain in detail the technical content, objectives, and effects of the present invention, the following description is provided in conjunction with the embodiments and accompanying drawings.

[0027] The educational evaluation target data processing method and system described above are applicable to the processing of educational evaluation data. The following specific embodiments illustrate this method:

[0028] Please refer to Figure 1 One embodiment of the present invention is as follows:

[0029] A method for processing educational evaluation target data, comprising the following steps:

[0030] S0. Based on natural language processing technology and a stop word list, preprocessing of educational evaluation texts is performed, including word segmentation, stop word filtering, and part-of-speech tagging.

[0031] Through preprocessing steps such as word segmentation, stop word filtering, and词性 tagging, purify the educational evaluation text data and剔除 invalid noise information. Provide high-quality input for subsequent basic semantic vector generation, segment density calculation, etc., reduce the interference of noise on the subsequent processing results, and improve the accuracy and efficiency of the overall data processing from the source.

[0032] In this embodiment, first preprocess the educational evaluation text to convert it into "computable and high-semantic" standardized data, providing accurate input for subsequent density analysis. It includes:

[0033] (1)Text cleaning:

[0034] Removal of invalid information: Remove HTML tags (such as ), special symbols (such as @, #, ★), and consecutive repeated characters (such as "哈哈哈" retained as "哈哈"), and retain Chinese, English, numbers, and common punctuation (。,;!?).

[0035] Format unification: Convert the text to UTF-8 encoding uniformly, the line break character to \n uniformly, and the case to be the same (convert English evaluation vocabulary to lowercase, such as "Creative" → "creative").

[0036] (2)Semantic standardization:

[0037] Word segmentation: Use jieba word segmentation (accurate mode) to segment the cleaned text and output a list of words (example: "小明课堂参与积极,作业完成质量高" → ["小明","课堂","参与","积极",",","作业","完成","质量","高"]).

[0038] Stop word filtering: Load a special stop word library for the education field (with a total of more than 800 words, including 3 subsets: more than 300 general stop words (的、了、是), more than 300 evaluation scenario stop words (本次、综上、特此说明), more than 200 meaningless function words (哦、呢、啊)), filter and retain the core words (nouns, verbs, adjectives, adverbs), and output a list of core words.

[0039] Word frequency screening: Remove words with a word frequency ≤ 1 in the list of core words (to avoid the interference of isolated words), and finally output a standardized word sequence (data format: List[str], length ≥ 50 (except for short texts)).

[0040] S1. Use the DistilBERT model to generate basic semantic vectors for sentence-level text segments of the educational evaluation text.

[0041] In this embodiment, the preprocessed educational evaluation text (standardized vocabulary sequence) is split into sentences (using .; !? as clause delimiters). If each sentence is less than 10 words long, it is merged with the next sentence. If it is more than 50 words long, it is split according to semantic breakpoints (based on punctuation + conjunctions "and / with / but") to generate sentence-level text fragments. Each fragment is input into the model, and a sentence-level semantic vector is output (dimension 768, data format: np.array([float]*768)).

[0042] Extract the token-level attention weights from the model output (each word corresponds to a weight value, ranging from 0 to 1). Take the arithmetic average of the word weights for each text segment to obtain the attention weight of the text segment (formula: sentence weight = Σ (word weight) / number of words, data format: float, retain 4 decimal places).

[0043] The basic semantic vector, attention weights, and character indices of the sentence output in this step are the sole data source for subsequent "density calculation and position mapping," ensuring that subsequent steps do not require secondary processing of the original text and reducing redundant calculations.

[0044] The DistilBERT model uses the HuggingFace open-source distilbert-base-uncased pre-trained model (6-layer Transformer, 66M parameters), loading its pre-trained weights and configuration file.

[0045] In this embodiment, the model is fine-tuned to better suit educational scenarios. The parameters of the first four Transformer layers are frozen (to avoid compromising general semantic understanding capabilities), while only the parameters of the last two Transformer layers and the output layer are unfrozen (for adaptation to educational scenarios). A fine-tuning task is constructed, employing a sentence pair classification task—inputting educational evaluation text and a set of topic keywords, and outputting a similarity classification (high / medium / low) between the educational evaluation text and the topic category, allowing the model to learn the semantic association between evaluation sentences and core topics.

[0046] Training parameter settings:

[0047] Optimizer: AdamW (learning rate = 2e-5, weight decay = 1e-4);

[0048] Batch size: 32;

[0049] Training epochs: 3 (to avoid overfitting);

[0050] Loss function: Cross-entropy loss;

[0051] Early stopping strategy: Stop training if the accuracy on the validation set does not improve for two consecutive rounds.

[0052] Save after fine-tuning: Save the fine-tuning weights of the thawed layer. When using it later, load "pre-trained weights + fine-tuning weights" to ensure both general semantics and adaptability to the educational scenario are considered.

[0053] S2. Based on the basic semantic vector, match it with a preset evaluation keyword library by calculating the cosine similarity, calculate the segment density value, obtain the topic category, and automatically identify the density peak point and boundary point through the density peak clustering algorithm without parameters.

[0054] In this embodiment, the evaluation keyword library includes topic categories and corresponding core keyword libraries, and topic vectors are built based on the core keyword libraries. A simple example is shown in Table 1 below:

[0055] Table 1

[0056] Theme Categories Core keyword library Topic Vectors Classroom performance Class participation, raising hands to answer, focus, collaborative discussion, attitude in class... Using the DistilBERT model based on a core keyword database, a 768-dimensional topic vector is output. After-school learning Self-study, homework completion, after-class exercises, parental supervision, extended learning... Same as above Competency and qualities Logical thinking, hands-on skills, innovation, expression, problem-solving strategies... Same as above Attitude and Character Positive attitude, conscientiousness, helpfulness, sense of responsibility, self-discipline... Same as above Discipline and regulations Being late, maintaining a regular schedule, following rules, classroom discipline... Same as above

[0057] In this embodiment, for each text segment (sentence), calculate the cosine similarity between its basic semantic vector and 5 types of topic vectors, and take the topic category with the highest similarity as the topic category of the text segment.

[0058] At the same time, for each keyword in the core keyword library, multiple synonyms are associated (e.g., "diligent" → ["industrious", "sedulous", "assiduous"]), and they are uniformly converted into 768-dimensional vectors (generated by the DistilBERT model) to obtain keyword vectors. Calculate the cosine similarity between the basic semantic vector and all keyword vectors, take the preset number of cosine similarities with the highest values, and calculate the mean as the keyword matching degree.

[0059] The segment density value is obtained by fusing the keyword matching degree and the attention weight of the text segment:

[0060] Segment density value = keyword matching degree × sentence attention weight × 100.

[0061] The purpose of multiplying by 100 is to normalize it to the 0 - 100 interval. Perform threshold calibration. For example, through statistics of 1000 educational evaluation text training sets, the mean density value is 45 and the standard deviation is 15. Therefore, set the high-density threshold to 60 and the low-density threshold to 30).

[0062] Automatically identifying the density peak point and boundary point through the density peak clustering algorithm without parameters includes the steps:

[0063] For all the text segments, calculate the Euclidean distance between the segment density values of each pair to generate a distance matrix.

[0064] Euclidean distance d The calculation is expressed as:

[0065] d ( i,j ) =|density ( i ) -density ( j ) | ;

[0066] in, d ( i,j ) represents the Euclidean distance between text fragment i and text fragment j. density ( i ) represents the fragment density value of text fragment i. density ( j Let represent the segment density value of text segment j. Then, construct an N×N distance matrix, where N represents the total number of text segments (sentences).

[0067] Calculate the arithmetic mean of all non-zero values ​​in the distance matrix to obtain the global average distance D:

[0068] D = Σd (i,j) / (N×(N-1)).

[0069] For each of the text segments, a determination is made: if a certain text segment exists:

[0070] ;

[0071] The text segment is then the peak point;

[0072] If one of the aforementioned text fragments exists:

[0073] ;

[0074] or:

[0075] ;

[0076] The text fragment is then a boundary point;

[0077] If one of the aforementioned text fragments exists:

[0078] ;

[0079] The text fragment is then an isolated point;

[0080] in, This represents the fragment density value of text fragment s. Indicates the high-density threshold. Indicates the low density threshold. D Indicates the global average distance. This represents the preset independent determination coefficient for the peak point. This represents the preset outlier rejection coefficient. This represents the minimum Euclidean distance between the text segment s and the segment density values ​​of the identified peak points. The mean Euclidean distance between text segment s and all text segments with higher density values ​​is expressed by the formula:

[0081] ;

[0082] Here, represents the number of text segments with a segment density value higher than that of text segment s.

[0083] This algorithm accurately distinguishes peak points, boundary points, and isolated points using a quantification formula, eliminating the need to pre-determine the number of clusters and avoiding the sensitivity of traditional algorithms to initial values. It effectively adapts to scenarios where interesting information is scattered or multi-peaked in educational evaluation texts, providing accurate core and auxiliary sentence segmentation for subsequent binning and reducing interference from irrelevant information.

[0084] In this embodiment, referring to the above rules for determining peak points, boundary points, and isolated points, three types of sentence labels are output. The core rules are briefly summarized as follows:

[0085] Peak points: ① Sentence density value ≥ high density threshold; ② The average distance between this sentence and all sentences with higher density values ​​≥ 1.2 × D (ensuring it is a local density core).

[0086] Boundary points: Density value ≥ low density threshold and < high density threshold, or density value < but distance from the nearest peak point ≤ D.

[0087] Isolated point: density value < and distance from the nearest peak point > 2×D (meaningless scattered information).

[0088] S3. Using the peak point as the core, merge adjacent text segments with a segment density value greater than a preset threshold and the same topic category to form a bin. Based on the position of the text segments at the beginning and end of each bin, extract the corresponding sub-text from the educational evaluation text to generate a sub-text set.

[0089] Using peak points as the core, merging adjacent text fragments with a fragment density value greater than a preset threshold and the same topic category to form bins includes the following steps:

[0090] For the target peak point of binning, query and merge text fragments that are adjacent in both the forward and backward directions and have the same topic category to determine the binning range:

[0091] If it is a boundary point, continue to expand the binning range in the current direction;

[0092] If it is a peak point, then determine the number of text segments between the peak point and the target peak point. If it is less than a preset threshold, then continue to expand the binning range in the current direction; otherwise, stop expanding the binning range in the current direction.

[0093] If it is an isolated point, stop expanding the binning range in the current direction;

[0094] Based on the binning range after expansion stops in both directions, the text fragments are merged to form bins.

[0095] Using peak points as the core, bins are dynamically generated by combining merging rules based on topic category, density value, and segment spacing. This ensures both the semantic coherence and topic purity of text segments within each bin, while avoiding excessive fragmentation or improper merging of key information. This allows the extracted sub-texts to focus on independent clusters of information of interest, improving the efficiency and accuracy of subsequent information recognition.

[0096] In this embodiment, the core rules for dynamic binning include extension conditions and termination conditions, as shown in the following examples:

[0097] Extended conditions (must be met simultaneously):

[0098] The sentences to be merged are boundary points, non-peak points with a density ≥ 30, or peak points with a spacing of less than a preset threshold.

[0099] The mean density distance between the sentence to be merged and all sentences within the bin is less than or equal to the global average distance D.

[0100] The topic category of the sentence to be merged is the same as the topic category of the binning (the topic category of the peak point is used as the topic category of the binning; if there are multiple peak points, the topic category with the highest frequency is taken) (in this embodiment, it can be extended to semantic similarity ≥ semantic similarity threshold).

[0101] The consistency ratio of the topic category of sentences within the merged bins is greater than or equal to the topic consistency threshold.

[0102] Termination condition (any one of these conditions must be met to stop expansion in that direction):

[0103] Encountering an isolated point;

[0104] Another peak point was encountered (the number of text segments in intervals exceeded the preset threshold).

[0105] The number of merged sentences has reached the maximum limit of 10.

[0106] The sentences to be merged do not meet any of the above expansion conditions.

[0107] Then, based on the "character position + topic information" of the binning, the semantically complete and topic-free sub-text is extracted to provide high-quality objects for subsequent recognition.

[0108] This embodiment also adds triple verification for the sub-text set:

[0109] Content validation: Keyword frequency ≥ number of sentences in bins × 0.3, keywords are evenly distributed.

[0110] Semantic integrity check: The semantic similarity between the first and last sentences is ≥0.4.

[0111] Topic validation: The consistency ratio between the topic category of text fragments in a sub-text set and the topic category of the bin is ≥80% (corresponding to the binning rules). If it is lower, the sub-text set is removed (to avoid low-quality data with topic confusion).

[0112] S4. Based on the BM25 algorithm, calculate the BM25 score of each evaluation keyword in the evaluation keyword library with the sub-text set and the BM25 score with the educational evaluation text, and determine the integrated initial information set based on the BM25 score.

[0113] Step S4 is as follows:

[0114] Based on the BM25 algorithm, the BM25 score of each evaluation keyword in the evaluation keyword library and the sub-text set is calculated, and the first set is obtained by acquiring a first preset number of evaluation keywords with higher scores and the text segments in which they are located.

[0115] Calculate the BM25 score of each evaluation keyword and the educational evaluation text, obtain a second preset number of evaluation keywords with higher scores and the text segments in which they are located, and obtain a second set;

[0116] Integrate the first set and the second set to generate a primary information set.

[0117] The BM25 algorithm is used to filter keywords in both the sub-text set and the original educational evaluation text, and then the two types of results are integrated to generate a preliminary information set. This captures the core local information within the bins while also taking into account the global key information of the original text, avoiding information omissions caused by local filtering, and quickly narrowing the recognition range to reduce the computational burden on subsequent models.

[0118] In this embodiment, the keyword library is used as the "query term set", and each sub-text set is used as the "document". The BM25 score of each query term and sub-text set is calculated. The first preset number of keywords with higher BM25 scores (Top 10 in this embodiment) and the sentences containing them are extracted from each sub-text set to form the initial "initial information set", which is the first set.

[0119] Perform the same initial screening on the original text, extract the top 5 scores, and mark them as "global supplementary information", which is the second set (to avoid missing global key information not covered by subtext).

[0120] By integrating the first set and the second set, the final initial information set is obtained.

[0121] S5. Input the basic semantic vector corresponding to each text segment in the initial information set into the pre-trained single-branch multi-scale convolutional model. The single-branch multi-scale convolutional model generates the topic category and confidence of each evaluation keyword, and calculates the bin density value of the bin to which the evaluation keyword belongs.

[0122] The single-branch multi-scale convolutional model includes an input layer, a multi-scale convolutional layer, a pooling layer, a fully connected layer, and an output layer.

[0123] The multi-scale convolutional layer contains three parallel convolutional kernels of different scales.

[0124] The single-branch multi-scale convolutional model uses three parallel convolutional kernels of different scales to capture multi-granularity semantic features in a lightweight architecture. This avoids the high resource consumption of complex shared operators while comprehensively capturing the semantic information of evaluation keywords. It reduces hardware requirements while maintaining recognition accuracy, enabling efficient deployment in ordinary educational institutions.

[0125] In this embodiment, the model structure is shown below:

[0126] Input layer: Input a 768-dimensional vector (basic semantic vector).

[0127] Multi-scale convolutional layer: 1 convolutional layer containing 3 parallel convolutional kernels (3×768, 5×768, 7×768), each convolutional kernel outputs 32 feature maps (activation function: ReLU), capturing semantics of different granularities (phrases, short sentences, long sentences).

[0128] Pooling layer: Perform max pooling (pooling kernel size = 2) on the output of each convolution kernel to reduce the dimensionality and output a 3×32=96-dimensional feature vector.

[0129] Fully connected layer: 2 fully connected layers. The first layer has a 96-dimensional input and a 64-dimensional output (activation function: ReLU). The second layer has a 64-dimensional input and a 4-dimensional output (3 information types + 1 confidence level).

[0130] Output layer: The softmax activation function is used to output the probability distribution of information type (learning performance / ability / attitude / other) and confidence score (formula: confidence = max(type probability distribution), range 0-1).

[0131] The training of the single-branch multi-scale convolutional model includes:

[0132] Label the training data with topic types;

[0133] The weights and biases of the convolutional layer and the fully connected layer are initialized using the He initialization method;

[0134] The single-branch multi-scale convolutional model is trained using the Adam optimizer and joint loss function, and overfitting is prevented by a regularization strategy.

[0135] A targeted training process was designed for single-branch multi-scale convolutional models. Through topic labeling, He initialization, Adam optimizer, and regularization strategies, the model is made accurately adaptable to educational evaluation scenarios. This improves the model's accuracy in recognizing evaluation keywords, topic categories, and confidence levels, enhances its generalization ability, avoids overfitting, and ensures stable and reliable output results in different educational evaluation text scenarios.

[0136] In this embodiment, the sentences in the training set are used to generate 768-dimensional vectors through the fine-tuned DistilBERT, which are then bound to the labeled topic types to form the model input data (vectors) and labels (topic types).

[0137] The weights and biases of the convolutional and fully connected layers are initialized using the He initialization method to ensure training stability.

[0138] Training parameter settings:

[0139] Optimizer: Adam (learning rate = 1e-4, weight decay = 1e-5);

[0140] Batch size: 64;

[0141] Training epochs: 5;

[0142] Loss function: Joint loss function (cross-entropy loss (type classification) + MSE loss (confidence regression));

[0143] Regularization strategy: Dropout (dropoutrate=0.2, to prevent overfitting) + L2 regularization.

[0144] After each training round, the validation set is used for evaluation (accuracy, recall, F1 score). If the F1 score decreases for two consecutive rounds, the learning rate is reduced (multiplied by 0.5).

[0145] Save the complete trained model weights (including parameters of convolutional and fully connected layers) for later use in recognition tasks.

[0146] The calculation of the density value of the sub-box is as follows:

[0147] The binning density value of each evaluation keyword corresponding to the first set is the average of the fragment density values ​​of each text fragment in the sub-text set;

[0148] The binning density value of each evaluation keyword corresponding to the second set is the average of the binning density values ​​of all the sub-text sets.

[0149] For evaluation keywords in different sets, differentiated bin density value calculation methods are developed to ensure that the density value accurately reflects the information density of the bin in which the keyword belongs. This provides an accurate and reliable quantitative basis for subsequent weight calculations, making the selection of core information more objective and avoiding misjudgment or omission of important information due to deviations in density value calculation.

[0150] Combining the model's output, the final data format is List[Dict], where each data entry contains "keyword", "sentence", "info_type", "confidence", "sub_text_id", and "density_mean".

[0151] Wherein, the binning density value density_mean corresponding to each of the evaluation keywords in the first set is the mean of the fragment density values ​​of each of the text fragments in the sub-text set, and the binning density value density_mean corresponding to each of the evaluation keywords in the second set is the mean of the binning density values ​​of all the sub-text sets.

[0152] S6. Calculate the weights based on the bin density value, the topic category, and the confidence level, and sort and filter according to the weights to obtain the core evaluation information;

[0153] The weight calculation based on the density value and the confidence level includes the following steps:

[0154] Calculate the base weight value based on the density value and the confidence level;

[0155] Based on the topic category corresponding to the evaluation keywords, obtain the preset adjustment coefficient corresponding to the topic category;

[0156] The final weight value is calculated based on the base weight value and the adjustment coefficient.

[0157] The weighting calculation integrates binning density values, confidence levels, and topic adjustment coefficients to measure the importance of evaluation keywords from multiple dimensions. This not only reflects the value and reliability of binning information but also aligns with the thematic dimension requirements of educational evaluation, making the priority of the selected core information clear and more suitable for practical application scenarios in educational evaluation.

[0158] In this embodiment, the weight calculation includes:

[0159] Basic weight = density_mean / 100 × confidence;

[0160] Here, confidence is the level of confidence.

[0161] Meanwhile, this embodiment takes into account the evaluation needs of different topic types of data in different scenarios, and sets adjustment coefficients for different topic types in order to obtain core evaluation information that is more inclined to the target preferred topic type.

[0162] Therefore, the final weight = base weight × adjustment coefficient.

[0163] In this embodiment, the data is sorted from highest to lowest according to the final weight, and the Top 20 are selected as the core evaluation information. The data is then categorized by topic type and output. An example of the output data in JSON format is shown below:

[0164] {

[0165] "core_evaluation_info": {

[0166] Classroom performance: [

[0167] {

[0168] Keyword: "Actively raising hands, high participation",

[0169] "sentence": "Actively raises hand to answer questions in class, demonstrating high participation."

[0170] "info_type": "Learning performance",

[0171] "final_weight": 1.122,

[0172] "topic_match_score": 0.85,

[0173] "source": "Subtext ID_1, Bin ID_1"

[0174] }

[0175] ],

[0176] "After-class learning": [

[0177] {

[0178] Keyword: "Self-study, no supervision required"

[0179] Sentence: "Can review key points independently after class without parental supervision"

[0180] "info_type": "Learning performance",

[0181] "final_weight": 1.056,

[0182] "topic_match_score": 0.92,

[0183] "source": "Subtext ID_2, Bin ID_2"

[0184] } ]

[0186] },

[0187] "topic_statistics": [

[0188] {"topic": "Classroom performance", "count": 5, "avg_weight": 0.98, "topic_consistency": 95%},

[0189] {"topic": "After-school learning", "count": 4, "avg_weight": 0.92, "topic_consistency": 90%}

[0190] ],

[0191] "category_statistics": [

[0192] {"info_type": "Learning performance", "count": 6, "avg_weight": 0.95},

[0193] {"info_type": "ability", "count": 3, "avg_weight": 0.88}

[0194] ],

[0195] "data_traceability": {"total_processed_sentences": 15, "valid_bins":2, "effective_sub_texts": 2}

[0196] }

[0197] According to another aspect of the invention, Figure 2 This is a schematic diagram illustrating an educational evaluation target data processing system according to an embodiment of the present invention. Embodiment two of the present invention is as follows:

[0198] An educational evaluation target data processing system 1, comprising:

[0199] Data acquisition module 2 is used to acquire the educational evaluation text to be processed;

[0200] Data processing module 3 implements the steps in the educational evaluation target data processing method described in Embodiment 1 above;

[0201] Data output module 4 outputs the processing results of the data processing module.

[0202] This invention discloses a method and system for processing educational evaluation target data. By employing parameter-free density peak binning and topic association merging rules, it completely eliminates the dependence of traditional clustering on preset quantities and initial values, adapting to multi-peak / scattered distribution scenarios and avoiding binning fragmentation. The lightweight combination of DistilBERT and a single-branch multi-scale convolutional model replaces the complex shared operator architecture, significantly reducing hardware requirements and enabling efficient deployment by ordinary educational institutions. Simultaneously, preprocessing and purifying the data, improving efficiency through BM25 initial screening, and enhancing quality through multi-dimensional weight ranking form a closed loop of "precise binning - efficient identification - scientific screening," ensuring both the accuracy and comprehensiveness of core evaluation information extraction while achieving full-process automation and low resource consumption, perfectly adapting to the practical application needs of educational evaluation scenarios.

[0203] The above description is merely an embodiment of the present invention and does not limit the patent scope of the present invention. Any equivalent modifications made based on the content of the present invention specification and drawings, or direct or indirect applications in related technical fields, are similarly included within the patent protection scope of the present invention.

Claims

1. A method for processing educational evaluation target data, characterized in that, Including the following steps: S1. Using the DistilBERT model, basic semantic vectors for sentence-level text fragments are generated for educational evaluation texts. S2. Based on the basic semantic vector, the fragment density value is matched with the preset evaluation keyword library by cosine similarity calculation to obtain the topic category, and density peak points and boundary points are automatically identified by non-parametric density peak clustering algorithm. S3. Using the peak point as the core, merge adjacent text segments with a segment density value greater than a preset threshold and the same topic category to form a bin. Based on the position of the text segments at the beginning and end of each bin, extract the corresponding sub-text from the educational evaluation text to generate a sub-text set. S4. Based on the BM25 algorithm, calculate the BM25 score of each evaluation keyword in the evaluation keyword library with the sub-text set and the BM25 score with the educational evaluation text, and determine the integrated initial information set based on the BM25 score. Step S4 is as follows: Based on the BM25 algorithm, calculate the BM25 score of each evaluation keyword in the evaluation keyword library and the sub-text set, and obtain the first set by the first preset number of evaluation keywords with the highest scores and the text segments in which they are located; Calculate the BM25 score of each evaluation keyword and the educational evaluation text, and obtain the second set of evaluation keywords with the highest scores (a second preset number) and the text segments in which they are located. Integrate the first set and the second set to generate a primary information set; S5. Input the basic semantic vector corresponding to each text segment in the initial information set into a pre-trained single-branch multi-scale convolutional model. The single-branch multi-scale convolutional model generates the topic category and confidence of each evaluation keyword and calculates the bin density value of the bin to which the evaluation keyword belongs. The calculation of the density value of the sub-box is as follows: The binning density value of each evaluation keyword corresponding to the first set is the average of the fragment density values ​​of each text fragment in the sub-text set; The binning density value of each evaluation keyword corresponding to the second set is the average of the binning density values ​​of all the sub-text sets; S6. Calculate the weights based on the bin density value, the topic category, and the confidence level, and sort and filter according to the weights to obtain the core evaluation information.

2. The method for processing educational evaluation target data according to claim 1, characterized in that, Automatically identifying density peak points and boundary points using a non-parametric density peak clustering algorithm includes the following steps: For all the text segments, calculate the Euclidean distance between each pair of segment density values ​​to generate a distance matrix; Calculate the arithmetic mean of all non-zero values ​​in the distance matrix to obtain the global average distance; For each of the text segments, a determination is made: if a certain text segment exists: ； The text segment is then the peak point; If one of the aforementioned text fragments exists: ； or: ； The text fragment is then a boundary point; If one of the aforementioned text fragments exists: ； The text fragment is then an isolated point; in, This represents the fragment density value of text fragment s. Indicates the high-density threshold. Indicates the low density threshold. D Indicates the global average distance. This represents the preset independent determination coefficient for the peak point. This represents the preset outlier rejection coefficient. This represents the minimum Euclidean distance between the text segment s and the segment density values ​​of the identified peak points. The mean Euclidean distance between text segment s and all text segments with higher density values ​​is expressed by the formula: ； in, M This indicates the number of text segments with a higher segment density value than the text segment s.

3. The method for processing educational evaluation target data according to claim 2, characterized in that, Using peak points as the core, merging adjacent text fragments with a fragment density value greater than a preset threshold and the same topic category to form bins includes the following steps: For the target peak point of binning, query and merge text fragments that are adjacent in both the forward and backward directions and have the same topic category to determine the binning range: If it is a boundary point, continue to expand the binning range in the current direction; If it is a peak point, then determine the number of text segments between the peak point and the target peak point. If it is less than a preset threshold, then continue to expand the binning range in the current direction; otherwise, stop expanding the binning range in the current direction. If it is an isolated point, stop expanding the binning range in the current direction; Based on the binning range after expansion stops in both directions, the text fragments are merged to form bins.

4. The method for processing educational evaluation target data according to claim 1, characterized in that, The single-branch multi-scale convolutional model includes an input layer, a multi-scale convolutional layer, a pooling layer, a fully connected layer, and an output layer. The multi-scale convolutional layer contains three parallel convolutional kernels of different scales.

5. The method for processing educational evaluation target data according to claim 4, characterized in that, The training of the single-branch multi-scale convolutional model includes: Label the training data with topic types; The weights and biases of the multi-scale convolutional layer and the fully connected layer are initialized using the He initialization method. The single-branch multi-scale convolutional model is trained using the Adam optimizer and joint loss function, and overfitting is prevented by a regularization strategy.

6. The method for processing educational evaluation target data according to claim 1, characterized in that, The weight calculation based on the density value and the confidence level includes the following steps: Calculate the base weight value based on the density value and the confidence level; Based on the topic category corresponding to the evaluation keywords, obtain the preset adjustment coefficient corresponding to the topic category; The final weight value is calculated based on the base weight value and the adjustment coefficient.

7. The method for processing educational evaluation target data according to claim 1, characterized in that, Step S1 is preceded by the following steps: S0. Based on natural language processing technology and a stop word list, preprocessing of educational evaluation texts is performed, including word segmentation, stop word filtering, and part-of-speech tagging.

8. A data processing system for educational evaluation objectives, characterized in that, include: The data acquisition module is used to acquire the educational evaluation text to be processed; The data processing module implements the steps in the educational evaluation target data processing method described in any one of claims 1-7. The data output module outputs the processing results from the data processing module.

Images