Robust test question coding scoring method and system based on diversity and interpretable modeling

Abstract

The invention provides a robust test question coding scoring method and system based on diversity and interpretable modeling, and the method comprises the steps: collecting a test question answering text set, carrying out the error correction of the test question answering text set based on a preset model, and obtaining a text set after error correction; based on the text set after error correction, separating high-frequency and low-frequency subject terms in a set subject range according to keyword long tail distribution in a subject to obtain a long tail discovery result; on the basis of the obtained long tail discovery result, interpretable coding suggestions are made through three steps of coding, model training and low-frequency word enhancement; and performing automatic scoring in combination with coding suggestions and answer point digests. Compared with a traditional coding evaluation method, the method has the advantages that the self-adaptive topic modeling method is constructed, the topic discovery method for protecting the low-frequency long-tail topic words is established, automatic scoring is achieved through interpretable coding, innovativeness is protected, and meanwhile a more robust examinee answering coding method is achieved.

Description

Technical Field

[0001] The present invention relates to the field of educational examination data mining, and in particular to a robust examination question coding and scoring method and system based on diversity and explainable modeling. Background Art

[0002] Intelligent review technology for open-ended examinations aims to achieve automated analysis and evaluation of subjective open-ended examination questions through artificial intelligence. Although online examination systems and intelligent question banks are now widely used in the field of educational examinations, the scoring of open-ended examinations still faces multiple technical bottlenecks: First, existing scoring models rely too much on matching standard answer templates, lack an effective recognition mechanism for innovative viewpoints that reflect the long-tail distribution of disciplinary thinking in examinees' answers, and tend to suppress diverse cognitive expressions; second, automatic text error correction technology lacks tolerance for unconventional grammatical structures and variants of domain terminology, which may cause semantically complete innovative answers to be misjudged as incorrect statements; third, scoring methods based on shallow feature encoding have difficulty parsing complex semantic logic and are easily disturbed by keyword stacking, making it impossible to accurately assess the depth of argumentation and knowledge transfer capabilities of answers. To address the above shortcomings, this paper proposes a robust question encoding and scoring method based on diversity and interpretable modeling.

[0003] Existing open-ended test analysis methods include an intelligent test grading method and device based on natural language processing, which is used to achieve automated test scoring. Specifically, the method involves: first, obtaining a test image and identifying the text content using OCR technology. After segmenting the text using sentence segmentation rules, the segmented sentences are standardized using a text error correction model. Next, the corrected text is semantically parsed using natural language processing technology, and a score is calculated based on the semantic match. During the text analysis process, feature words are extracted using a word segmentation tool and converted into vector representations, which are then combined with a clustering algorithm to generate semantic feature centers. Finally, the test score is determined by calculating the vector space distance between the answer text and the standard answer. Word segmentation uses a maximum matching algorithm or a hidden Markov model, and word vector clustering preferably uses the K-Means algorithm. The efficiency of semantic analysis is improved by filtering invalid words and constructing mathematical representations.

[0004] The aforementioned NLP-based intelligent grading method has two significant limitations. First, its evaluation mechanism, which calculates the cosine similarity between the semantic center of the answer and the standard answer, is mathematically constrained by the semantic distribution of mainstream answers. This makes it difficult to reasonably evaluate low-frequency but innovative answers due to their deviation from the main semantic components. Second, its scoring mechanism, based on vector space calculations, lacks clear logical connections and traceability, essentially acting as a black-box decision model. These two flaws together inhibit innovation and hinder interpretability in the scoring system. Summary of the Invention

[0005] The purpose of the present invention is to provide a robust test question coding and scoring method and system based on diversity and interpretable modeling to solve the above problems.

[0006] To achieve the above object, the present invention adopts the following technical solutions: In a first aspect, the present invention provides a robust test question coding and scoring method based on diversity and interpretable modeling, comprising: Collecting a test answer text set, and correcting the test answer text set based on a preset model to obtain a corrected text set; Based on the corrected text collection within the set topic range, according to the long-tail distribution of keywords within the topic, high-frequency and low-frequency keywords are separated to obtain long-tail discovery results; Based on the long-tail discovery results, explainable coding suggestions are made through three steps: coding, model training, and low-frequency word enhancement. Automatic scoring is performed by combining coding suggestions and answer point summaries.

[0007] Optionally, the collected test answer text set includes: Use OpenCV to preprocess and segment the collected test paper images and extract the standardized answer area image set ; Then use deep learning OCR technology to detect and recognize text on the standardized image to generate the initial answer text set .

[0008] Optionally, the step of correcting the test answer text set based on a preset model to obtain a corrected text set includes: Based on the answer text collection , adopts an end-to-end Encoder-Decoder model based on Transformer; the Encoder part uses MLM, randomly masks some words during training, understands the semantics and grammatical structure of the text, and identifies error patterns; the Decoder part uses Transformer Decoder, based on the semantic representation generated by the Encoder, combined with the self-attention mechanism and feedforward network, gradually predicts and generates the corrected text sequence until the complete and accurate text is output; finally, the corrected text set is obtained .

[0009] Optionally, a Transformer-based Encoder-Decoder architecture is used to implement end-to-end error correction. The Encoder uses an MLM pre-training strategy to randomly mask local vocabulary in the input text to learn language patterns and contextual features and identify potential errors. The Decoder uses a self-attention mechanism and a hierarchical decoding strategy based on the semantic representation of the Encoder output to generate correction results layer by layer, and finally outputs a set of corrected texts. .

[0010] Optionally, the error-corrected text set is based on a set topic scope, and according to the long-tail distribution of keywords within the topic, high-frequency and low-frequency subject words are separated to obtain long-tail discovery results, including: exist On this basis, based on the improved adaptive LDA topic modeling, the candidate's answer text is analyzed; first, the number of topics is set , for each candidate The values are used for topic modeling, and the quality of the model is evaluated by the dual indicators of perplexity and silhouette coefficient; the standardized joint scoring formula is used:

[0011] Determining the optimal number of topics ,to prevent low-frequency words from being overly dispersed or insufficiently clustered; Then, based on the long-tail distribution of keywords within the topic, high-frequency and low-frequency topic words are separated; the former reflects the mainstream answer ideas, while the latter reflects innovation. Cosine similarity in the semantic vector space is used to remove duplicates within the cluster and retain the associated original answer text. The final topic model fully records the high- and low-frequency words of each topic and their corresponding answer text. The long-tail discovery results are expressed as follows: .

[0012] Optionally, based on the obtained long-tail discovery results, interpretable coding suggestions are made through three steps: coding, model training, and low-frequency word enhancement, including: (1) Coding: Based on high-frequency keywords, a set of keyword-related documents-coding suggestion triplets is constructed. Through manual analysis, a standardized logic description of the problem-solving process is formed, and a set of expert coding suggestions is obtained. ; (2) Model training: Using the T5 model architecture, a keyword-driven summary generation model is trained based on triple data. , to achieve automatic coding suggestion generation; (3) Low-frequency word adaptation and encoding generation: Calculate the semantic representations of high-frequency words and low-frequency words belonging to the same topic through Bert, and calculate the mean of the difference between the high-frequency word representation and the low-frequency word representation , in the use of When generating suggestions for low-frequency keywords, feature f is spliced and fused into the input of the encoding layer of the T5 model. After the coding suggestions are generated and reviewed in combination with the original answer content, the effective suggestions are added to the coding knowledge base, forming a closed-loop collection mechanism for innovative problem-solving ideas.

[0013] Optionally, the automatic scoring based on the coding suggestion and the answer summary includes: First, a set of encoding suggestions Formulate scoring rules and determine each Then use a small amount of test-taker answer text to train Model for point summary; when scoring, the candidate's answer point summary and Perform semantic vector cosine similarity matching and set the threshold : If there is a matching item with a similarity greater than p at a certain point, it will be scored, otherwise it will not be scored.

[0014] In a second aspect, the present invention provides a robust test question coding and scoring system based on diversity and interpretable modeling, comprising: A data collection module is used to collect a text set of test answer answers, and perform error correction on the text set of test answer answers based on a preset model to obtain a text set after error correction; The keyword separation module is used to separate high-frequency and low-frequency keywords based on the long-tail distribution of keywords within the set topic based on the corrected text collection, and obtain long-tail discovery results; The coding suggestion module is used to make explainable coding suggestions based on the long-tail discovery results through three steps: coding, model training, and low-frequency word enhancement; Scoring module, used to automatically score by combining coding suggestions and answer summary.

[0015] In a third aspect, the present invention provides a computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the steps of the robust test question coding and scoring method based on diversity and interpretable modeling when executing the computer program.

[0016] In a fourth aspect, the present invention provides a computer-readable storage medium, wherein the computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the steps of the robust test question coding and scoring method based on diversity and interpretable modeling are implemented.

[0017] Compared with the prior art, the present invention has the following technical effects: The purpose of the present invention is to provide an open test question analysis method based on a domain-based large model, which realizes the analysis of answers to open test questions by constructing a domain model, fine-tuning with a few sample prompts, mining answer viewpoints, expert feedback on innovative viewpoints, and automatic scoring. First, based on the open source basic large model, large-scale incremental pre-training is carried out using subject resources and other field knowledge, a multi-task template set is designed for instruction fine-tuning, and a domain-based large model is constructed; secondly, a template for viewpoint mining and answer summary prompt learning is designed to fine-tune the domain model with a few sample prompts; then, the viewpoint information of the answer is analyzed and summarized, and the new viewpoints that emerge are submitted to experts for review; finally, based on the viewpoint set composed of new and old viewpoints, automatic scoring is performed with reference to the matching degree between the answer viewpoint and the viewpoint set. The present invention fully mines the semantic information of the answers to open test questions and simplifies it into the form of a viewpoint set, which is convenient for extracting and reviewing innovative viewpoints and provides more fine-grained automatic scoring at the semantic level. It has the advantages of sufficient semantic understanding and accurate automatic scoring, which makes the present invention have obvious advantages over other methods based on simple text feature pattern recognition. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 This is a block diagram of the robust test question coding and scoring method based on diversity and explainability modeling of the present invention.

[0019] Figure 2 It is a structured flowchart of the test questions.

[0020] Figure 3 It is a flowchart of automatic text correction.

[0021] Figure 4 This is a flowchart of long-tail learning topic discovery Figure 5 It is a flowchart generated by interpretable coding.

[0022] Figure 6 This is a flowchart of automatic scoring. DETAILED DESCRIPTION

[0023] The following describes the embodiments of the present invention in detail with reference to the accompanying drawings and examples. It should be noted that the embodiments described herein are only intended to explain the present invention and are not intended to limit the present invention. In addition, the technical features of the embodiments of the present invention may be combined with each other unless there is a conflict.

[0024] The specific implementation process of the present invention includes test question structuring, text automatic error correction, long-tail learning topic discovery, interpretable coding generation, and automatic scoring process. Figure 1 This is a block diagram of the robust test question coding and scoring method based on diversity and explainable modeling of the present invention.

[0025] 1. Question Structuring Process Question paper structuring refers to the process of identifying and processing candidates' test paper information. The implementation steps are as follows: First, obtain the digital image file of the candidates' answer sheets through an image acquisition system. Based on the OpenCV computer vision library, perform preprocessing and region segmentation operations on the original image, accurately extract the answer regions of each candidate, and perform resolution standardization processing on all segmented images. Finally, obtain a set of n standardized answer region images , and then adopt the Chinese optical character recognition (OCR) technology based on deep learning to perform multi-dimensional text feature extraction processing on the standardized images, thereby generating an initial answer text set corresponding to n candidates .

[0026] The above question paper structuring process flowchart is as Figure 2 shown 2. Text automatic error correction process To solve the problem of handwritten OCR text error correction in educational examinations, an end-to-end Encoder-Decoder model based on Transformer is adopted. It can be divided into three steps: constructing a common OCR confusion library, MLM fine-tuning, and end-to-end training.

[0027] (1) Constructing a common OCR confusion library Build a dataset by manually collecting typical confusion samples, covering common handwritten error types such as glyphs, strokes, and radicals. For example: glyph confusion (such as "嬴" and "赢" resulting in "嬴政" being miswritten as "赢政"), stroke differences (such as "戌" and "戍" causing "戊戌变法" to be misspelled), radical misuse (such as the structural confusion between "汴" and "卞"), etc. for high-frequency error patterns.

[0028] (2) MLM fine-tuning Adopt two-level dynamic masking in MLM training: Error correction mask: Forcefully cover the incorrect characters in the confused data (such as "戊戍→戊戌"), and set the label to the correct character to strengthen the error correction mapping; Regularization mask: Randomly cover 15% of the correct characters (such as "汴→汴"), and maintain the original label to prevent overfitting.

[0029] Adjust the parameters through cross-entropy loss and gradient optimization iteration to gradually improve the model's error correction confidence and generation accuracy. The formal expression of the loss function is as follows:

[0030] where N is the number of samples, is the true label corresponding unnormalized score.

[0031] (3) End-to-end training The Seq2Seq loss function is used to simultaneously correct local word errors while maintaining overall semantic coherence. By minimizing the difference between the output and target sequences, the model not only locates the error point during error correction but also optimizes the global semantic structure, ensuring that the correction results are logically sound.

[0032]

[0033] in is the sample size, is the sequence length It is The first sample target vocabulary.

[0034] Combining the local error correction capability of MLM and the global generation capability of Seq2Seq, the overall objective function of the text error correction model is designed to be a linear combination of the two:

[0035] It is a balancing factor that adjusts the weight ratio of local error correction and global semantics through the verification set.

[0036] Finally, the corrected text set is obtained based on the initial text in step 1 .

[0037] The above text automatic error correction process flow chart is as follows Figure 3 shown.

[0038] 3. Long-tail learning topic discovery process First, the improved adaptive LDA topic clustering algorithm is used to cluster the initial answer texts of all candidates. Perform topic modeling and set the number of topics The value range is , for each candidate topic number Perform the following operations: (1) Perform LDA topic modeling.

[0039] (2) Synchronously calculate two evaluation indicators of the current topic model: Perplexity and Silhouette Coefficient.

[0040] The perplexity is expressed as follows:

[0041] Where D is the total number of valid documents (the number of candidates’ answer texts after removing blank papers), Representative The word sequence of a document, Refers to The total number of words in the document, Refers to the model word The perplexity measures the model’s predictive ability, which tends to select a larger number of topics but may lead to overfitting.

[0042] The form of the silhouette coefficient value is expressed as follows:

[0043] in Represents a sample The average distance to other samples in the same cluster, Represents a sample The average distance to the nearest cluster sample. This coefficient constrains the upper limit of the number of topics by clustering tightness.

[0044] After obtaining the perplexity and silhouette coefficient values, the current topic number is measured by normalizing the joint score The rationality of the scoring strategy is formally expressed as follows:

[0045] Determining the optimal number of topics through standardized joint optimization , which can effectively prevent the low-frequency keywords at the tail of the long-tail distribution from being excessively dispersed in multiple topic clusters, and can avoid the problem of aggravating the long-tail effect due to insufficient topic clustering.

[0046] Next, based on the long-tail distribution curve of keywords within each topic, we counted the high-frequency (HF) and low-frequency (LF) keywords within each topic. High-frequency keywords typically encompass the majority of candidates' answer ideas, while low-frequency keywords encompass candidates' innovative ideas, reflecting the diversity of their answers. Using the cosine similarity calculation method based on semantic vector space, we performed intra-cluster de-duplication and merging of high- and low-frequency keywords, and preserved the original answer text associated with each keyword. The final complete information representation of the topic model is as follows:

[0047] in is the total number of topics, Represents high-frequency keywords, Representing the low-frequency keywords at the tail of the long-tail distribution, the keywords of each topic provide explainable information about the candidates' answer ideas related to the topic. This method explicitly protects the low-frequency long-tail pattern while retaining high-frequency features, ensuring that answer diversity is not overwhelmed by high-frequency signals.

[0048] The construction process flow chart of the above long tail learning topic discovery process is as follows Figure 4 shown.

[0049] 4. Interpretable encoding generation process The interpretable coding process can be divided into three steps: expert coding of high-frequency keywords, training of summary generation model, and enhancement processing of low-frequency keywords.

[0050] (1) Based on the complete information of topic modeling obtained in step 3, subject experts ,in accordance with The correct high-frequency keywords are extracted from the questionnaire, and keywords with the ability to represent the disintegration path are extracted. For each keyword, a text is written to explain the correct answer logic under the current answer ideas and keywords, which is used as the candidate data coding suggestion. A (keyword, related document, coding suggestion) triple is constructed, where the related documents come from in , summarize the triple information of each topic and get the expert coding suggestion set , For coding suggestions.

[0051] (2) Based on of Model based on the triple set pair The model is trained to obtain a keyword-guided text summary generation model The input of the model is keywords and documents related to keywords, and the output is the documents that meet the Standard encoding suggestion text based on this keyword.

[0052] (3) First, the semantic representations of high-frequency words and low-frequency words belonging to the same topic are calculated by Bert, and the mean of the difference between the high-frequency word representation and the low-frequency word representation is calculated. , for each topic, low-frequency keywords and candidate answer texts related to keywords, using When generating suggestions for low-frequency keywords, the feature Splice and merge to Among the inputs of the model encoding layer, remember the corresponding encoding suggestion data output These coding suggestions generated by low-frequency keywords represent the candidates' innovative answer ideas and are generated by subject experts through The relevant keywords and students’ original answers are reviewed. If the coding suggestion is correct, a (keyword, related document, coding suggestion) triple is constructed and added to the expert coding suggestion set. middle.

[0053] The above flowchart can explain the encoding generation process as follows Figure 5 shown.

[0054] 5. Automatic scoring process The automatic scoring process includes three steps: scoring rule formulation, summary model construction, and automatic scoring.

[0055] First, subject experts set expert coding suggestions Customize scoring rules, that is, specify each coding suggestion rule score.

[0056] Then, a small number of original texts of candidates' answers are selected, the text summary information of their points is marked, and training for the point-by-point text summary task is carried out based on T5.

[0057] Finally, for each candidate's answer, first conduct a summary of the text points, and for each point, compare it with the coding suggestions in the expert coding suggestion set. Calculate the cosine similarity based on the semantic vector space one by one, and define the similarity threshold as , if there is a similarity greater than If the point data is correct, the point is judged to be correct and scored, otherwise the point is not scored.

[0058] The above automatic scoring process is as follows Figure 6 shown.

[0059] The robust test question coding and scoring method based on diversity and interpretable modeling disclosed in the present invention, compared with traditional coding and scoring methods, constructs an adaptive topic modeling method, establishes a topic discovery method that protects low-frequency long-tail keywords, and realizes automatic scoring through interpretable coding, thereby achieving a more robust candidate answer coding method while protecting innovation.

[0060] In yet another embodiment of the present invention, a robust test question coding and scoring system based on diversity and interpretable modeling is provided, which can be used to implement the above-mentioned robust test question coding and scoring method based on diversity and interpretable modeling. Specifically, the system includes: A data collection module is used to collect a text set of test answer answers, and perform error correction on the text set of test answer answers based on a preset model to obtain a text set after error correction; The keyword separation module is used to separate high-frequency and low-frequency keywords based on the long-tail distribution of keywords within the set topic based on the corrected text collection, and obtain long-tail discovery results; The coding suggestion module is used to make explainable coding suggestions based on the long-tail discovery results through three steps: coding, model training, and low-frequency word enhancement; Scoring module, used to automatically score by combining coding suggestions and answer summary.

[0061] The module division in the embodiments of the present invention is illustrative and represents only one logical functional division. In actual implementation, other division methods may be used. Furthermore, the functional modules in various embodiments of the present invention may be integrated into a single processor, exist physically as separate modules, or two or more modules may be integrated into a single module. The integrated modules may be implemented in either hardware or software functional modules.

[0062] In another embodiment of the present invention, a computer device is provided, which includes a processor and a memory, wherein the memory is used to store a computer program, the computer program includes program instructions, and the processor is used to execute the program instructions stored in the computer storage medium. The processor may be a central processing unit (CPU), or may be other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA) or other programmable logic devices, discrete gates or transistor logic devices, discrete hardware components, etc. It is the computing core and control core of the terminal, which is suitable for implementing one or more instructions, specifically suitable for loading and executing one or more instructions in the computer storage medium to implement the corresponding method flow or corresponding function; the processor described in the embodiment of the present invention can be used for the operation of the robust test question coding and scoring method based on diversity and interpretable modeling.

[0063] In another embodiment of the present invention, a storage medium is provided, specifically a computer-readable storage medium (Memory). The computer-readable storage medium is a memory device in a computer device, used to store programs and data. It is understood that the computer-readable storage medium herein may include both built-in storage media in the computer device and, of course, extended storage media supported by the computer device. The computer-readable storage medium provides storage space, which stores the terminal's operating system. Furthermore, the storage space also stores one or more instructions suitable for being loaded and executed by a processor. These instructions may be one or more computer programs (including program code). It should be noted that the computer-readable storage medium herein may be a high-speed RAM memory or a non-volatile memory, such as at least one disk storage device. The processor may load and execute the one or more instructions stored in the computer-readable storage medium to implement the corresponding steps of the robust test question coding and scoring method based on diversity and interpretable modeling described in the above-mentioned embodiment.

[0064] Those skilled in the art will appreciate that embodiments of the present invention may be provided as methods, systems, or computer program products. Thus, the present invention may take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to magnetic disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0065] The present invention is described with reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowcharts and / or block diagrams, as well as combinations of processes and / or blocks in the flowcharts and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowcharts and / or block diagrams. Figure 1 a process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0066] These computer program instructions may also be stored in a computer readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1a process or multiple processes and / or boxes Figure 1 The function specified in one or more boxes.

[0067] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operational steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing the instructions executed on the computer or other programmable device for implementing the process. Figure 1 a process or multiple processes and / or boxes Figure 1 A step that specifies a function in one or more boxes.

[0068] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, ordinary technicians in the field should understand that the specific implementation methods of the present invention can still be modified or replaced by equivalents. Any modification or equivalent replacement that does not depart from the spirit and scope of the present invention should be covered by the scope of protection of the claims of the present invention.

Claims

1. A robust test question coding and scoring method based on diversity and interpretable modeling, characterized by: include: Collecting a test answer text set, and correcting the test answer text set based on a preset model to obtain a corrected text set; Based on the corrected text collection within the set topic range, according to the long-tail distribution of keywords within the topic, high-frequency and low-frequency keywords are separated to obtain long-tail discovery results; Based on the long-tail discovery results, explainable coding suggestions are made through three steps: coding, model training, and low-frequency word enhancement. Automatic scoring is performed by combining coding suggestions and answer point summaries.

2. The robust test question coding and scoring method based on diversity and interpretable modeling according to claim 1 is characterized in that: The collected test answer text set includes: Use OpenCV to preprocess and segment the collected test paper images and extract the standardized answer area image set ; Then use deep learning OCR technology to detect and recognize text on the standardized image to generate the initial answer text set .

3. The robust test question coding and scoring method based on diversity and interpretable modeling according to claim 1 is characterized in that: Correcting the test answer text set based on the preset model to obtain the corrected text set includes: Based on the answer text collection , adopts an end-to-end Encoder-Decoder model based on Transformer; the Encoder part uses MLM, randomly masks some words during training, understands the semantics and grammatical structure of the text, and identifies error patterns; the Decoder part uses Transformer Decoder, based on the semantic representation generated by the Encoder, combined with the self-attention mechanism and feedforward network, gradually predicts and generates the corrected text sequence until the complete and accurate text is output; finally, the corrected text set is obtained .

4. The robust test question coding and scoring method based on diversity and interpretable modeling according to claim 3 is characterized in that: End-to-end error correction is achieved using a Transformer-based Encoder-Decoder architecture. The Encoder uses an MLM pre-training strategy to randomly mask local words in the input text to learn language patterns and contextual features and identify potential errors. Based on the semantic representation of the encoder output, the decoder uses the self-attention mechanism and hierarchical decoding strategy to generate correction results layer by layer, and finally outputs a set of error-corrected texts. .

5. The robust test question coding and scoring method based on diversity and interpretable modeling according to claim 4 is characterized in that: The text set after error correction is based on the set topic scope, and according to the long-tail distribution of keywords in the topic, high-frequency and low-frequency subject words are separated to obtain long-tail discovery results, including: exist On this basis, based on the improved adaptive LDA topic modeling, the candidate's answer text is analyzed; first, the number of topics is set , for each candidate The values are used for topic modeling, and the quality of the model is evaluated by the dual indicators of perplexity and silhouette coefficient; the standardized joint scoring formula is used: Determining the optimal number of topics ,to prevent low-frequency words from being overly dispersed or insufficiently clustered; Then, based on the long-tail distribution of keywords within the topic, high-frequency and low-frequency topic words are separated; the former reflects the mainstream answer ideas, while the latter reflects innovation. Cosine similarity in the semantic vector space is used to remove duplicates within the cluster and retain the associated original answer text. The final topic model fully records the high- and low-frequency words of each topic and their corresponding answer text. The long-tail discovery results are expressed as follows: .

6. The robust test question coding and scoring method based on diversity and interpretable modeling according to claim 1 is characterized in that: Based on the long-tail discovery results, explainable coding suggestions are made through three steps: coding, model training, and low-frequency word enhancement, including: (1) Coding: Based on high-frequency keywords, a set of keyword-related documents-coding suggestion triplets is constructed. Through manual analysis, a standardized logic description of the problem-solving process is formed, and a set of expert coding suggestions is obtained. ; (2) Model training: Using the T5 model architecture, a keyword-driven summary generation model is trained based on triple data. , to achieve automatic coding suggestion generation; (3) Low-frequency word adaptation and encoding generation: Calculate the semantic representations of high-frequency words and low-frequency words belonging to the same topic through Bert, and calculate the mean of the difference between the high-frequency word representation and the low-frequency word representation , in the use of When generating suggestions for low-frequency keywords, feature f is spliced and fused into the input of the encoding layer of the T5 model. After the coding suggestions are generated and reviewed in combination with the original answer content, the effective suggestions are added to the coding knowledge base, forming a closed-loop collection mechanism for innovative problem-solving ideas.

7. The robust test question coding and scoring method based on diversity and interpretable modeling according to claim 6 is characterized in that: The automatic scoring process combines coding suggestions and answer summary points, including: First, a set of encoding suggestions Formulate scoring rules and determine each Then use a small amount of test-taker answer text to train Model for point summary; when scoring, the candidate's answer point summary and Perform semantic vector cosine similarity matching and set the threshold : If there is a matching item with a similarity greater than p at a certain point, it will be scored, otherwise it will not be scored.

8. A robust test question coding and scoring system based on diversity and interpretable modeling, characterized by: include: A data collection module is used to collect a text set of test answer answers, and perform error correction on the text set of test answer answers based on a preset model to obtain a text set after error correction; The keyword separation module is used to separate high-frequency and low-frequency keywords based on the long-tail distribution of keywords within the set topic based on the corrected text collection, and obtain long-tail discovery results; The coding suggestion module is used to make explainable coding suggestions based on the long-tail discovery results through three steps: coding, model training, and low-frequency word enhancement; Scoring module, used to automatically score by combining coding suggestions and answer summary.

9. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the computer program, the steps of the robust test question coding and scoring method based on diversity and interpretable modeling as described in any one of claims 1 to 7 are implemented.

10. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the steps of the robust test question coding and scoring method based on diversity and interpretable modeling as described in any one of claims 1 to 7 are implemented.

Images