Test question management method based on multi-modal adaptive similarity learning

Abstract

The invention discloses a test question management method based on multi-modal adaptive similarity learning, and relates to the technical field of data management, and the method comprises the steps: collecting an original test question set, generating an evolutionary test question set through a preset large language model, generating a variable test question set through a data enhancement technology, and carrying out the fusion processing of the test question set to form a test question ecological library, the diversity and coverage of test question resources are enriched; the test question ecological library is input into a preset semantic similarity model and a preset syntactic similarity model to obtain a semantic similarity score and a syntactic similarity score of each test question, so that the similarity of the test questions can be comprehensively measured, and the screening accuracy is improved; the multi-modal similarity score is obtained according to the semantic similarity score and the syntactic similarity score to be compared with the preset threshold value, then the test questions are imported into the auditing module or the question bank storage module according to the comparison result, intelligent distribution of the test questions is facilitated, and the timeliness and quality of the question bank are ensured.

Description

Technical Field

[0001] This application relates to the technical field of data management, and in particular to a question management method based on multi-modal adaptive similarity learning. Background Art

[0002] With the rapid development of artificial intelligence and big data technologies, the intelligent transformation in the education field is advancing rapidly, and intelligent question banks play a key role in this process. By automatically managing question data, intelligent question banks not only optimize the allocation of educational resources but also promote the development of personalized teaching. Intelligent question banks are important for improving teaching quality, ensuring educational fairness, and meeting different educational needs.

[0003] Although significant progress has been made in intelligent question bank management technologies, however, when faced with a large amount of question data sets, challenges still exist in the accurate assessment of question similarity and the optimization of the resource warehousing process. Therefore, to ensure the timeliness, accuracy of question bank resources, and reduce knowledge point redundancy, it is urgent to perform multi-modal similarity analysis on questions and refine the question warehousing process. Summary of the Invention

[0004] The main purpose of this application is to provide a question management method based on multi-modal adaptive similarity learning, aiming to solve the technical problem that in the existing question management process, it is difficult to accurately evaluate question similarity, which affects the timeliness of question updates and the quality of questions in the question bank.

[0005] To achieve the above objective, this application proposes a question management method based on multi-modal adaptive similarity learning, and the method includes:

[0006] Collect the original question set, generate an evolved question set through a preset large language model, generate a mutated question set using data augmentation technology, and perform fusion processing on the original question set, the evolved question set, and the mutated question set to form a question ecosystem library;

[0007] Input the question ecosystem library into a preset semantic similarity model and a preset syntactic similarity model respectively to obtain the semantic similarity scores and syntactic similarity scores of each question in the question ecosystem library;

[0008] Obtain multi-modal similarity scores based on the semantic similarity scores and syntactic similarity scores, compare the multi-modal similarity scores with a preset threshold, and import each question into an audit module or a question bank storage module according to the comparison result.

[0009] In one embodiment, the steps of collecting the original test question set, generating an evolved test question set through a preset large language model, generating a mutated test question set using data augmentation techniques, and fusing the original test question set, the evolved test question set, and the mutated test question set to form a test question ecological library include:

[0010] Set the data collection scope and collect the original test question set from the data source based on the data collection scope;

[0011] Construct a prompt template containing semantic constraint conditions and use the prompt template to guide the preset large language model to generate an evolved test question set based on the original test question set;

[0012] Expand and mutate the original test question set through data augmentation methods such as synonym replacement, sentence pattern recombination, and knowledge point expansion to generate a mutated test question set;

[0013] Fuse the original test question set, the evolved test question set, and the mutated test question set to obtain fused test question data, and perform deduplication processing on the fused test question data to form a standardized test question ecological library.

[0014] In one embodiment, the steps of inputting the test question ecological library into a preset semantic similarity model to obtain the semantic similarity scores of each test question in the test question ecological library include:

[0015] Input the test question ecological library into a preset semantic similarity model, where the preset semantic similarity model includes a SemGloVe module, a preset BERT module, a graph neural network module, and a similarity scoring module;

[0016] In the SemGloVe module, establish the semantic association between the vocabulary in the test question text through co-occurrence matrix analysis, and combine the attention mechanism to extract word-level similarity features to construct a word similarity matrix;

[0017] Integrate the word similarity matrix through the preset BERT module to perform multi-level semantic representation and generate word embedding vectors with semantic information;

[0018] Use the graph neural network module to construct an adjacency matrix based on the word embedding vectors, and aggregate the adjacent node information in the adjacency matrix to generate sentence vectors;

[0019] In the similarity scoring module, process the sentence vectors sequentially through a fully connected layer and a Softmax layer to obtain the semantic similarity scores of each test question.

[0020] In one embodiment, the steps of establishing the semantic association between the vocabulary in the test question text through co-occurrence matrix analysis, and combining the attention mechanism to extract word-level similarity features to construct a word similarity matrix in the SemGloVe module include:

[0021] Obtain the test question text in the test question ecological library, and obtain the word frequency and word co-occurrence relationship of each word in the test question text through a word co-occurrence matrix, and construct a global word co-occurrence count matrix;

[0022] Average and aggregate the attention weights corresponding to each byte pair encoding under each word to obtain the attention weight corresponding to each word;

[0023] Based on the attention weights corresponding to each word, determine the semantic association between each word through the Division distance function to obtain a word similarity matrix.

[0024] In one embodiment, the step of integrating the word similarity matrix through a preset BERT module to perform multi-level semantic representation and generate a word embedding vector with semantic information includes:

[0025] Integrate the word similarity matrix into the multi-head attention mechanism of the BERT model to calculate the corresponding attention weights through the input representation vector;

[0026] Integrate the attention weights corresponding to each attention head to obtain an attention output, and perform a linear transformation on the attention output to obtain a word embedding vector with semantic information.

[0027] In one embodiment, the step of constructing an adjacency matrix according to the word embedding vector through the graph neural network module and aggregating to generate a sentence vector based on the adjacency node information in the adjacency matrix includes:

[0028] Construct an adjacency matrix according to the word embedding vector, and construct a graph structure representation according to the adjacency matrix;

[0029] Add corresponding relative position encodings to each node in the graph structure representation to obtain a graph structure with position information;

[0030] Perform graph convolution operations based on the graph structure with position information to aggregate the adjacency node information of each node to obtain updated node embeddings;

[0031] Aggregate each of the updated node embeddings to generate a sentence vector.

[0032] In one embodiment, the step of inputting the test question ecological library into a preset syntactic similarity model to obtain the syntactic similarity scores of each test question in the test question ecological library includes:

[0033] Input the test question ecological library into a preset syntactic similarity model, and the preset syntactic similarity model includes: a Chinese word segmentation module, a part-of-speech judgment and termization module, and a PT tree kernel module;

[0034] Through the Chinese word segmentation module, segment the test question text into a sequence of words according to the keyword vocabulary and context in the test question information;

[0035] Through the part-of-speech judgment and term representation module, perform part-of-speech tagging on each word in the sequence of words, and perform term representation on each tagged word to obtain a number of term triples;

[0036] In the construction of the PT tree kernel module, construct a syntactic parsing tree CPT based on each of the term triples, and calculate the node similarity through the PT kernel, and obtain the syntactic similarity score according to the similarity calculation result.

[0037] In one embodiment, the Chinese word segmentation module is constructed based on a CNN-BiGRU-CRF composite neural network model. The step of segmenting the test question text into a sequence of words through the Chinese word segmentation module according to the keyword vocabulary and context in the test question information includes:

[0038] Map the text sequence to a word vector matrix through the CNN layer combined with the embedding layer to obtain the word vector matrix corresponding to the test question text;

[0039] Capture the context information in the test question text through the BIGRU layer, and convert the word vector matrix into a sequence representation;

[0040] Obtain the sequence of words by obtaining the dependency relationship of each label in the sequence representation through the CRF layer to optimize the prediction result of the sequence labeling task.

[0041] In one embodiment, the step of performing part-of-speech tagging on each word in the sequence of words through the part-of-speech judgment and term representation module, and performing term representation on each tagged word to obtain a number of term triples includes:

[0042] Use the natural language processing toolkit to perform part-of-speech tagging on each word in the sequence of words to obtain the sequence of words after part-of-speech tagging;

[0043] Confirm a number of key terms based on the sequence of words after part-of-speech tagging, and perform concept mapping on each of the key terms to obtain a list of terms after concept mapping;

[0044] Based on the list of terms after concept mapping, construct the semantic vector representation of each key term through the dynamic vector space model;

[0045] Calculate the semantic similarity between each of the key terms, and integrate each of the key terms according to the semantic similarity between each of the key terms to form a combined term set, and the combined term set consists of a number of term triples.

[0046] In one embodiment, the step of obtaining the multimodal similarity score according to the semantic similarity score and the syntactic similarity score, and comparing the multimodal similarity score with a preset threshold, and importing each of the test questions into the review module or the question bank storage module according to the comparison result includes:

[0047] Non-linearly fuse the semantic similarity score and the syntactic similarity score based on a dynamic weighting mechanism to generate a multimodal similarity score;

[0048] Compare the multimodal similarity score with a preset threshold;

[0049] If the multimodal similarity score is higher than the preset threshold, import the test question into the review module;

[0050] If the multimodal similarity score is not higher than the preset threshold, import the test question into the question bank storage module.

[0051] This application discloses a test question management method based on multimodal adaptive similarity learning. By collecting an original test question set, generating an evolved test question set through a preset large language model, generating a mutated test question set using data augmentation technology, and fusing the original test question set, the evolved test question set, and the mutated test question set to form a test question ecological library; inputting the test question ecological library into a preset semantic similarity model and a preset syntactic similarity model respectively to obtain the semantic similarity score and the syntactic similarity score of each test question in the test question ecological library; obtaining a multimodal similarity score according to the semantic similarity score and the syntactic similarity score, and comparing the multimodal similarity score with a preset threshold, and importing each test question into the review module or the question bank storage module according to the comparison result. Since this application constructs a test question ecological library by fusing the original test question set, the evolved test question set, and the mutated test question set, it enriches the diversity and coverage of test question resources; uses a semantic similarity model and a syntactic similarity model to evaluate the semantic and syntactic similarities of test questions respectively, comprehensively measures the similarity of test questions, and improves the accuracy of screening; through the comparison of the multimodal similarity score with a preset threshold, it realizes the intelligent diversion of test questions and ensures the timeliness and quality of the question bank. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] The accompanying drawings here are incorporated into the specification and form a part of this specification, showing embodiments consistent with this application, and are used together with the specification to explain the principles of this application.

[0053] In order to more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the following will briefly introduce the accompanying drawings required for describing the embodiments or the prior art. Obviously, for those of ordinary skill in the art, other drawings can also be obtained based on these drawings without creative efforts.

[0054] Figure 1 This is a schematic flowchart of the first embodiment of the test question management method based on multi-modal adaptive similarity learning in this application;

[0055] Figure 2 This is a schematic diagram of the acquisition process of the original test question set;

[0056] Figure 3 This is a schematic diagram of the adaptive similarity fusion process;

[0057] Figure 4 This is a schematic flowchart of the second embodiment of the test question management method based on multi-modal adaptive similarity learning in this application;

[0058] Figure 5 This is a schematic diagram of the module structure of the semantic similarity model SemBert-GCN;

[0059] Figure 6 This is a schematic flowchart of the third embodiment of the test question management method based on multi-modal adaptive similarity learning in this application;

[0060] Figure 7 This is a schematic diagram of the module structure of the syntactic similarity model TE-PTK;

[0061] Figure 8 This is a schematic flowchart of the full process of the test question management method based on multi-modal adaptive similarity learning in this application. Detailed implementation manners

[0062] It should be understood that the specific embodiments described herein are only used to explain the technical solutions of this application and are not used to limit this application.

[0063] To better understand the technical solutions of this application, the following will be described in detail in combination with the accompanying drawings of the specification and specific implementation manners.

[0064] The embodiments of this application provide a test question management method based on multi-modal adaptive similarity learning. Refer to Figure 1 , Figure 1 This is a schematic flowchart of the first embodiment of the test question management method based on multi-modal adaptive similarity learning in this application. In this embodiment, the method includes: Steps S10 to S30:

[0065] Step S10: Collect the original test question set, generate an evolved test question set through a preset large language model, generate a mutated test question set using data augmentation technology, and perform fusion processing on the original test question set, the evolved test question set, and the mutated test question set to form a test question ecological library.

[0066] It should be noted that this embodiment can be applied to the scenario of building an intelligent question bank. The execution subject of this embodiment can be a computing service device with data processing, network communication, and program running functions, such as a tablet computer, a personal computer, a mobile phone, a question bank server, etc., or other electronic devices that can implement functions as shown in the figure or similar functions.

[0067] It should be understood that the original question set can be publicly available question data collected from the network. For example, initial question data can be obtained from structured data sources such as university official websites and MOOC platforms through automated crawler technology, including specific technical means such as API interface calls and web page parsing (BeautifulSoup / Selenium).

[0068] It can be understood that a data collection scope can be set in advance to ensure that data collection is carried out within the framework of legality, morality, and effectiveness. For example, to ensure legality, explicit authorization from the data owner or relevant regulatory agency must be obtained. At the same time, the robots.txt protocol should be strictly adhered to, focusing on web pages of public information and avoiding touching sensitive or private data. The storage and use of data should ensure security and transparency and be limited to predefined purposes.

[0069] Specifically, reference can be made to Figure 2 , Figure 2 which is a schematic diagram of the process of collecting the original question set. As Figure 2 shown, when collecting the original question set, web scraping technology can be used to collect the question corpus of artificial intelligence. In this process, a script system is used to obtain the links of relevant program pages, and HTML parsers such as Selenium's WebDriver and Python libraries (Requests and BeautifulSoup) are used to parse the web page content and organize it into a structured data framework. After extracting the question information, it is compiled into a single data frame, and the compiled data frame is stored as a CSV format file with a file size of approximately 10MB. Considering the server load, the data collection rate is set to collect once every 5 - 10 seconds.

[0070] Next, the collected original question data can be processed to ensure the quality and uniformity of the generated corpus. Therefore, a standardized terminology system can be introduced to standardize core entities such as program names, department names, and course codes, thereby eliminating the confusion caused by naming differences and enhancing the consistency and comparability of the data. In addition, a series of data consistency verification processes can be implemented. By cross - verifying and carefully comparing with the original website information, non - standard and conflicting information in the corpus can be effectively identified and corrected to ensure the accuracy and integrity of the collected original question data.

[0071] Furthermore, the test questions can be finely classified and multi-labeled according to multiple dimensions such as the nature, type, field of application, and specific knowledge points of the test questions, so as to form an original test question set Set with "gene coding". Ori 。

[0072] It should be noted that the pre-set large language model can be constructed based on general LLMs such as GPT-4 and PaLM, or a dedicated model fine-tuned on educational domain corpora. A set of refined prompt templates can be set up to guide the LLMs to mine the knowable information in the original test questions, so as to form an evolved test question set Set. evo 。

[0073] Exemplarily, the specific requirements of the prompt module can be: 1) objectively and accurately answer the known information about the given test questions; 2) ensure that your answer is concise and complete, not exceeding four sentences; 3) generate new test questions containing important semantic information; 4) do not change the knowledge point characteristics covered by the original test questions when generating new test questions; The term "objectively accurate" aims to reduce hallucinations and ambiguities in LLMs. The purpose of requirement 3 is to keep the semantic information of the test questions generated by LLMs from being damaged and to emphasize that the answers should not be irrelevant information.

[0074] Data augmentation techniques can include operations such as synonym replacement, sentence restructuring, and conditional variation. By making controllable modifications to the original test question set through the above different technical enhancement means, the data diversity of the original test question set can be enhanced, and a variant test question set Set can be obtained. mut 。

[0075] In the above data augmentation process, not only can the difficulty gradient of the test questions under the same knowledge point be adjusted, the conditions be changed, and the scenarios be simulated to generate hierarchical variants, but also by introducing interdisciplinary elements or combining current affairs hotspots, new test question scenarios can be created to further broaden the coverage of the test questions. In addition, the expression of sentences can be modified using synonym and near-synonym replacement strategies to increase the diversity of sentences, and the word order and sentence structure of sentences can be changed through sentence swapping and random swapping to enrich the presentation of test questions.

[0076] It should be understood that using the large language model as the engine of "natural selection" can deeply mine the knowable information of the test questions. Introducing data augmentation techniques to simulate "genetic variation" can enrich the data diversity of the test questions. By integrating the collected original test question set, the evolved test question set generated by the large model, and the variant test question set after data augmentation, a structured test question ecological library can be formed.

[0077] Furthermore, the standardized question bank ecosystem can also be achieved through data cleaning and filtering techniques. This process involves identifying and removing candidate keywords that match specific patterns, such as removing HTML tags, clearing non-text characters (e.g., special symbols and emojis), trimming excessive spaces and line breaks, and standardizing text encoding formats. In addition, it also includes appropriate handling of numbers and special formats, as well as performing stop-word filtering.

[0078] In specific implementation, first set the data collection scope, and collect the original question set from the data source based on the data collection scope; then construct a prompt template containing semantic constraint conditions, and guide the preset large language model to generate an evolved question set based on the original question set through the prompt template; then expand and mutate the original question set through data augmentation methods such as synonym replacement, sentence restructuring, and knowledge point expansion to generate a mutated question set; finally, fuse the original question set, the evolved question set, and the mutated question set to obtain fused question data, and perform deduplication processing on the fused question data to form a standardized question bank ecosystem.

[0079] Step S20: Input the question bank ecosystem into a preset semantic similarity model and a preset syntactic similarity model respectively to obtain the semantic similarity scores and syntactic similarity scores of each question in the question bank ecosystem.

[0080] It should be noted that the question bank ecosystem can be placed into a preset semantic similarity model SemBert-GCN and a preset syntactic similarity model TE-PTK respectively, so as to obtain the similarity of the questions at the semantic and syntactic structure levels respectively.

[0081] It should be understood that the semantic similarity model SemBert-GCN is used to obtain the semantic similarity score S of the questions g ; the syntactic similarity model TE-PTK is used to obtain the semantic similarity score S of the questions s 。

[0082] Among them, SemBert-GCN can include a SemGloVe module, a preset BERT module, and a graph neural network (GCN) module. The SemGloVe module takes the standardized and structured "question bank ecosystem" as input, maps the vocabulary to a high-dimensional vector space to capture the global semantic relationship between words; the preset BERT module can generate a score matrix and token embeddings containing rich semantic features by integrating the word similarity matrix. The GCN module then uses the obtained scoring matrix and token embeddings carrying semantic information to obtain the vector representation of similar questions, and then performs pooling activation processing to obtain the sentence semantic similarity score S q 。

[0083] Among them, the TE-PTK model can include a Chinese word segmentation module, a part-of-speech judgment and termization module, and a PT tree kernel construction module. The Chinese word segmentation module takes the multi-attribute label test question information sequence as the input, adopts the CNN-BIGRU-CRF network model, and cuts the continuous character sequence into a word sequence with clear meaning and boundaries according to the keywords and context in the test question information. The part-of-speech judgment and termization module can use word segmentation tools such as Stanford corenlp to automatically assign corresponding grammatical function labels to each word to achieve accurate part-of-speech recognition. To solve the ambiguity problem of text terms, the CN-Probase knowledge graph developed by the Knowledge Factory Laboratory of Fudan University can also be used to conceptualize the terms, screen and exclude useless information terms, and ensure the clear and accurate semantics of the terms in the sentence. The PT tree kernel construction model can use terms as basic semantic units to construct a constituency parsing tree for short texts, enhance the expressive power of the semantics of the words themselves, and then calculate the number of common substructures between the two trees using the PT kernel to obtain the syntactic similarity score S of the test question text s 。

[0084] Step S30: Obtain the multi-modal similarity score according to the semantic similarity score and the syntactic similarity score, compare the multi-modal similarity score with a preset threshold, and import each test question into the review module or the question bank storage module according to the comparison result

[0085] It should be understood that after obtaining the semantic similarity score S g and the syntactic similarity score S s , based on the adaptive equilibrium strategy technology, the weight coefficients α, β, δ can be used to dynamically adjust the contributions of different similarity scores, and the interaction term φ(S g , S s ) can be introduced to capture the interaction between different similarity scores, realize the effective fusion of semantic and syntactic modal similarities, and thus calculate the multi-modal similarity score Score of the test question. Here, reference can be made to Figure 3 , Figure 3 which is a schematic diagram of the adaptive similarity fusion process. Based on Figure 3 , the calculation formula of the multi-modal similarity score can be:

[0086] Score = Softmax(α·σ(S g ) + β·σ(S s ) + δ·φ(S g , S s , S k )) (0)

[0087] Among them, α, β, and δ are weight coefficients used to balance the importance of different similarity scores and interaction terms, and satisfy α + β + δ = 1. σ(x) is the Sigmoid function used to limit the scores between 0 and 1; φ(S g , S s ) is used to capture the interaction between different similarity scores and can be defined as Tanh(μ1·S g + μ2·S s ), where μ1 and μ2 are parameters used to adjust the effect of the interaction term; in this formula, not only the individual contributions of each modality similarity are considered, but also the complex relationships between different modality similarities are captured through the Tanh function.

[0088] It should also be noted that the preset threshold can be the preset similarity threshold μ. Based on the obtained multi-modal similarity score Score, it is possible to effectively "screen" the storage qualification of the test questions based on the dynamic threshold shunting module. If the multi-modal similarity score Score of the test question exceeds the preset similarity threshold μ, the test question will be accurately imported into the test question review module for further evaluation or "elimination". Otherwise, it will be included in the test question bank for storage.

[0089] In a specific implementation, the semantic similarity score and the syntactic similarity score are non-linearly fused based on a dynamic weighting mechanism to generate a comprehensive similarity score; the comprehensive similarity score is compared with the preset threshold; if the comprehensive similarity score is higher than the preset threshold, the test question is imported into the review module; if the comprehensive similarity score is not higher than the preset threshold, the test question is imported into the question bank storage module.

[0090] In this embodiment, the test question ecological library is constructed by integrating the original test question set, the evolved test question set, and the mutated test question set, enriching the diversity and coverage of the test question resources; the semantic similarity model and the syntactic similarity model are used to evaluate the semantic and syntactic similarities of the test questions respectively, comprehensively measuring the similarity of the test questions, and improving the accuracy of screening; through the comparison of the multi-modal similarity score with the preset threshold, the intelligent shunting of the test questions is realized, ensuring the timeliness and quality of the question bank.

[0091] Based on the first embodiment of the present application, in the second embodiment of the present application, for the same or similar content as the above-mentioned embodiment one, reference can be made to the above introduction and will not be elaborated hereinafter. On this basis, please refer to Figure 4 , Figure 4 which is the flowchart of the second embodiment of the test question management method based on multi-modal adaptive similarity learning of the present application.

[0092] In this embodiment, to illustrate the specific process of obtaining the semantic similarity score through the preset semantic similarity model, the steps of inputting the test question ecological library into the preset semantic similarity model to obtain the semantic similarity scores of each test question in the test question ecological library include: Steps S201 to S205:

[0093] Step S201: Input the test question ecological library into the preset semantic similarity model, where the preset semantic similarity model includes a SemGloVe module, a preset BERT module, a graph neural network module, and a similarity scoring module.

[0094] It should be understood that the structure of the preset semantic similarity model SemBert-GCN in this embodiment can be described here in combination with Figure 5 Figure Figure 5 is a schematic diagram of the module structure of the semantic similarity model SemBert-GCN.

[0095] Step S202: In the SemGloVe module, establish the semantic association between the words in the test question text through the analysis of the word co-occurrence matrix, and combine the attention mechanism to extract the word-level similarity features to construct a word similarity matrix.

[0096] It should be understood that the SemGloVe model takes the standardized "test question ecological library" as input, captures the lexical frequencies and their co-occurrence relationships by analyzing the global Word-to-Word co-occurrence count matrix X. To more intuitively understand the relationship between words, the original BPE-to-BPE attention weights are converted into Word-to-Word attention weights. In the conversion process, the attention weights of all BPE tokens belonging to the same word are averaged and aggregated to obtain the attention weight AW∈R of each word to other words K×K of the target word w i and the context word w j , and the Division distance function is used to determine the distance between the target word w i and the context word w j as Div(w

[0097] Here, the data processing process of the SemGloVe module can also be described in combination with Figure 5 Figure, and the data processing process of the SemGloVe module is as follows:

[0098] First, the SemGloVe module takes the standardized "test question ecological library" as input and captures the lexical frequencies and their co-occurrence relationships by analyzing the global Word-to-Word co-occurrence count matrix X. Its entry X i，j represents a word w j ∈V in a word w iThe total number of occurrences in the context of ∈V, where V is the vocabulary of the training corpus, for the global Word-to-Word co-occurrence count w i and w j separation (i.e., the distance between them in the text), SemGloVe specifies X i，j as:

[0099]

[0100] where, P i and P j are positions in the context, and the words closer to w i get larger weights.

[0101] Next, for a given word sequence W = {w1,..., w k}, BERT converts each word (or word segment) into BPE (Byte Pair Encoding) tokens. This is because when BERT processes text, to capture more detailed semantic information, it decomposes words into multiple BPE units or byte pair encodings. To more intuitively understand the relationship between words, the SemGloVe module converts the original BPE-to-BPE attention weights into Word-to-Word attention weights. The conversion process averages and aggregates the attention weights of all BPE tokens belonging to the same word, thus obtaining the attention weights of each word to other words. The generated Word-to-Word attention weights AW ∈ R K×K , for the local window context w i of word w j , j ∈ [i - S, i + S], the attention weight AW i from word w j to word w i，j can be expressed as:

[0102]

[0103] where, m and n are the number of subwords of w i and w j respectively, and AT(k, l) represents the attention weight from BPE token t k to t1. Sort AW i，j in descending order, and select the top s words as the context words C(w i ) of w i to exclude semantically meaningless words.

[0104] Finally, the distance between the target word w i and the context word w j is determined by the Division distance function Div(w i, w j )。

[0105]

[0106] In a specific implementation, the SemGloVe module can obtain the test question text in the test question ecological library, and through the word co-occurrence matrix, obtain the word frequency and word co-occurrence relationship of each word in the test question text, and construct a global word co-occurrence count matrix; average and aggregate the attention weights corresponding to each byte pair encoding under each word to obtain the attention weights corresponding to each word; thereby, based on the attention weights corresponding to each word, determine the semantic association between each word through the Division distance function to obtain a word similarity matrix.

[0107] Step S203: Integrate the word similarity matrix through a preset BERT module to perform multi-level semantic representation and generate word embedding vectors with semantic information.

[0108] It should be understood that the preset BERT module can be a fine-tuned BERT model. By fine-tuning the baseline BERT model, the word similarity matrix S constructed from the evolutionary population S1 and the original population S2 generated by LLMs is integrated into the multi-head attention mechanism of BERT. Using the similarity matrix S as additional input information, the BERT model can more accurately evaluate the semantic relationship between words in its self-attention layer, thereby enhancing the model's ability to understand complex semantic structures. The multi-head attention mechanism of BERT linearly transforms the query (Θ), key (K), and value (Ψ) vectors, applies scaled dot-product attention, and concatenates the results of multiple "heads" and linearly transforms again to generate the output vector MultiHead(Θ, K, Ψ). Since the model injects the word similarity matrix S to calculate the Hadamard product, the BERT attention using scaled dot-product calculation can be expressed as Attention(Θ, K, Ψ).

[0109] Here, it can also be combined with Figure 5 to illustrate the data processing process of the preset BERT module. The data processing process of the preset BERT module is as follows:

[0110] First, by fine-tuning the baseline BERT model, the original population S1 = {p1,... p i ... p l (S1)} and the evolutionary population S2 = {p1,... p i ... p l (S2)} constructed word similarity matrix S = {p 1，1 ,... p i，j ... p l，l(S1, S2)} is integrated into the multi - head attention mechanism of BERT. By taking the similarity matrix S as additional input information, the BERT model can more accurately evaluate the semantic relationships between words in its self - attention layer, thereby enhancing the model's ability to understand complex semantic structures. The multi - head attention mechanism of BERT linearly transforms the query (Θ), key (K), and value (Ψ) vectors, then applies scaled dot - product attention, and finally concatenates the results of multiple "heads" and linearly transforms again to generate the output vector, which can be expressed as:

[0111]

[0112] MultiHead(Θ, K, Ψ) = Concat(head1,..., head h )W O (5)

[0113] where and are the parameter matrices corresponding to the query, key, and value of the i - th attention head respectively, and W O is the weight matrix for concatenation of the h - th attention head.

[0114] Finally, since the model injects the word similarity matrix S to calculate the Hadamard product, making the model more focused on word pairs with higher similarity in the sentence pair, the BERT attention is calculated using scaled dot - product as:

[0115] Scores = ΘK T *S + MASK (6)

[0116]

[0117] In a specific implementation, the preset BERT module can integrate the word similarity matrix into the multi - head attention mechanism of the BERT model to calculate the corresponding attention weights through the input representation vectors; integrate the attention weights corresponding to each attention head, obtain the attention output, and perform a linear transformation on the attention output to obtain the word embedding vectors with semantic information.

[0118] Step S204: Construct an adjacency matrix according to the word embedding vectors through the graph neural network module, and aggregate to generate a sentence vector based on the adjacent node information in the adjacency matrix.

[0119] It should be understood that taking the word embedding h i of each token carrying semantic information obtained by BERT as the input and passing it to the subsequent GCN model. Different from the standard GCN model, taking the scores in the score matrix as the adjacency matrix A i，j, each token serves as each node in the GCN, and relative position encoding is added to the GCN to enable it to learn the relative position information of the tokens. Based on the adjacency matrix A i，j , for a given node i, the GCN collects the relevant semantic information carried by its context words in A i，j and obtains the output representation of node i through calculation After being processed by the GCN module, the vector of each token is obtained After average pooling, the sentence vector h is obtained s .

[0120] Similarly, here it can be combined with Figure 5 to illustrate the data processing process of the graph neural network (GCN) module. The data processing process of this GCN module is as follows:

[0121] First, taking the word embedding h i of each token carrying semantic information obtained by BERT as the input and passing it to the subsequent GCN model. Different from the standard GCN model, the score matrix scores are used as the adjacency matrix A i，j , each token serves as each node in the GCN, and relative position encoding is added to the GCN to enable it to learn the relative position information of the tokens. Based on the adjacency matrix A i，j , for a given node i, the GCN collects the relevant semantic information carried by its context words in A i，j and obtains the output representation of node i through calculation

[0122]

[0123] Second, after being processed by the GCN module, the vector of each token is obtained After average pooling, the sentence vector is obtained:

[0124]

[0125] where m is the total number of tokens, that is, the length of the sentence sequence.

[0126] In a specific implementation, the GCN module can construct an adjacency matrix according to the word embedding vectors, and construct a graph structure representation according to the adjacency matrix; add corresponding relative position encodings to each node in the graph structure representation to obtain a graph structure with position information; perform graph convolution operations based on the graph structure with position information to aggregate the adjacent node information of each node and obtain updated node embeddings; aggregate each updated node embedding to generate a sentence vector.

[0127] Step S205: In the similarity scoring module, process the sentence vectors through a fully connected layer and a Softmax layer in sequence to obtain the semantic similarity scores of each test question.

[0128] In a specific implementation, the sentence vector h obtained after pooling s is processed through a fully connected layer and then through a Softmax layer to finally obtain the sentence semantic similarity score S g .

[0129] In this embodiment, the semantic similarity model captures the global semantic relationships of words, the semantic associations between words, and the semantic information of sentences layer by layer through the SemGloVe module, the fine-tuned BERT module, and the GCN module, and finally outputs the semantic similarity score. This is beneficial for accurately evaluating the semantic similarity of test questions, capturing deep semantic relationships, screening out high-quality and non-repetitive test questions, improving the quality of the question bank, and optimizing intelligent test paper generation.

[0130] Based on the first and second embodiments of the present application, in the third embodiment of the present application, the same or similar content as in the above-mentioned first and second embodiments can be referred to the above introduction and will not be repeated hereinafter. On this basis, please refer to Figure 6 , Figure 6 which is a schematic flowchart of the third embodiment of the test question management method based on multi-modal adaptive similarity learning of the present application.

[0131] In this embodiment, in order to illustrate the specific process of obtaining the syntactic similarity score through a preset syntactic similarity model, the steps of inputting the test question ecological library into the preset syntactic similarity model to obtain the syntactic similarity scores of each test question in the test question ecological library include: Steps S206 to S209:

[0132] Step S206: Input the test question ecological library into a preset syntactic similarity model, and the preset syntactic similarity model includes: a Chinese word segmentation module, a part-of-speech judgment and termization module, and a PT tree kernel construction module.

[0133] It should be understood that this can be combined with Figure 7 the structure of the preset syntactic similarity model TE-PTK of this embodiment for explanation. Figure 7 which is a schematic diagram of the module structure of the syntactic similarity model TE-PTK.

[0134] Step S207: Through the Chinese word segmentation module, segment the test question text into a sequence of words according to the keyword vocabulary and context in the test question information.

[0135] It should be noted that the Chinese word segmentation module uses a CNN-BIGRU-CRF network model to perform the word segmentation task on the test question information sequence. The Chinese word segmentation module mainly consists of three core layers: the CNN layer, the BIGRU layer, and the CRF layer.

[0136] It should be noted that the Chinese word segmentation module takes the multi-attribute label test question information sequence as input, uses the CNN-BIGRU-CRF network model for Chinese word segmentation, and cuts the non-spaced character sequence into a word sequence y(M) with clear meanings and boundaries according to the keywords and context in the test question information.

[0137] Similarly, here we can combine Figure 7 to illustrate the data processing process of the Chinese word segmentation module. The data processing process of the Chinese word segmentation module is as follows:

[0138] First, the CNN layer combines with the embedding layer to map the text sequence into a word vector matrix C∈R k×d , where each word vector x i ∈R k×d , where k is the sentence length and d is the word vector dimension. Using the adaptive learning ability of the convolutional network feature extractor, it can effectively identify and capture the deep features E = [h1, h2,... h n in the text data.

[0139] Secondly, the Bi-GRU layer uses a bidirectional gated recurrent unit network to accurately capture the time series dependencies of the text, including long-term and short-term dependencies and context information. After being processed by the Bi-GRU layer, the deep feature vectors E = [h1, h2,... h n in the text data are transformed into a sequence representation M = {m1,..., m i ,..., m n}, where m i is the i-th input character vector of the CRF layer.

[0140] Finally, the CRF layer optimizes the prediction result of the sequence labeling task by considering the dependencies between the labels in the sequence. Let the label sequence be Y = {y1, y2... y n}, where y(M) is the set of labeled sequences of M. Then the conditional probability P(Y|M) of the CRF model is expressed as:

[0141]

[0142] where the function φ(y t -1, y t , M) is used to calculate the score of the input label sequence Y, A represents the label transition score matrix, is the score from label y i to yi+1 The larger the score, the greater the transfer probability; is the score matrix output by the Bi-GRU network ( Specifically expressed as the yth character of the i-th character i output matrix), matrix The size is n×k. The corresponding loss function obtained by calculating the maximum likelihood estimate through CRF can be expressed as:

[0143]

[0144] Among them, N is the number of labeled sentences for training, Y i is the true label sequence of the i-th sentence.

[0145] In the specific implementation, the Chinese word segmentation module maps the text sequence into a word vector matrix through the CNN layer combined with the embedding layer to obtain the word vector matrix corresponding to the test text; the context information in the test text is captured through the BIGRU layer, and the word vector matrix is converted into a sequence representation; the dependency relationship between each label in the sequence representation is obtained through the CRF layer to optimize the prediction results of the sequence labeling task and obtain the word sequence.

[0146] Step S208: tagging the word in the word sequence with a part of speech through the part-of-speech judgment and terminology module, and terminologically representing the tagged words to obtain a plurality of term triples.

[0147] It should be understood that after obtaining accurate word segmentation results, the part-of-speech judgment and terminology modules can use the Stanfordcorenlp1 toolkit to perform language annotations such as part-of-speech tagging for the text, and define concept classes and instance classes that distinguish noun terms. In view of the fact that short texts are mainly composed of terms, semantic vectors are constructed through a dynamic vector space model, and term sets and are extracted from short texts and texts, and terms are represented in the form of triples (terms, part-of-speech tags, concepts), and merged to form a joint term set. The semantic vectors of each term are constructed based on term similarity, and smooth inverse frequency is used as the attention weight to emphasize the different contributions of different terms to the meaning of short texts.

[0148] Here you can also combine Figure 7 The data processing process of the part-of-speech judgment and terminology module is described as follows:

[0149] First, after obtaining relatively accurate word segmentation results using the Chinese word segmentation module, the part-of-speech judgment and termization module can use the natural language processing toolkit Stanfordcorenlp1 to determine the part of speech of each word. And it can generate language annotations for the text by calling the corenlp1 library, which includes part-of-speech tagging, sentence and token boundary recognition, named entity recognition, citation attributes and their relationship recognition, etc. To clearly distinguish the part of speech of noun terms, define type(t) to divide terms into concept classes and instance classes:

[0150]

[0151] where E t / C t is a set of instances / concepts at time t, freq(e) / freq(c) is the frequency ratio of instances / concepts, and |·| is the number of items in the set.

[0152] Next, considering that short texts are mainly composed of terms, it is reasonable and effective to use terms to express the meaning of test questions. To address the sparsity problem of short texts, a dynamic vector space model is introduced to construct semantic vectors of short texts. Term sets T1 and T2 are respectively extracted from two short texts S1 and S2, and each term in the term set is represented in the form of a triple (term, part-of-speech tagging, concept). Then, semantic vectors of S1 and S2 are constructed. First, merge term sets T1 and T2 to form a joint term set T, and then construct semantic vectors for each term in T based on term similarity. Taking S1 as an example, the calculation method for the i-th dimension of its semantic vector is as follows:

[0153] semantic_vector i =sim(vector1, vector2)*W1*W2 (14)

[0154] If vector does not belong to T1, calculate the semantic similarity between vector and each item in T1, and select the score of the item with the highest similarity score between T1 and vector; if vector belongs to T1, the score is 1. Among them, sim(vector1, vector2) is the cosine similarity function of two vectors.

[0155] Finally, since different terms contribute differently to short texts, for example, stop words contribute less to the overall meaning of short texts, the attention weight W uses smoothed inverse frequency (SIF) to emphasize the role of terms, and SIF can be expressed by the following formula:

[0156]

[0157] where is the smoothing parameter, and L(node) is the word frequency of node.

[0158] In a specific implementation, the part-of-speech judgment and term extraction module uses a natural language processing toolkit to perform part-of-speech tagging on each word in the word sequence, obtaining a word sequence with part-of-speech tags; based on the word sequence with part-of-speech tags, a number of key terms are identified, and concept mapping is performed on each key term to obtain a list of terms after concept mapping; based on the list of terms after concept mapping, semantic vector representations of each key term are constructed through a dynamic vector space model; the semantic similarity between each key term is calculated, and each key term is integrated according to the semantic similarity between each key term to form a combined term set, and the combined term set consists of a number of term triples.

[0159] Step S209: In the PT tree kernel construction module, a syntactic parse tree CPT is constructed based on each of the term triples, and the node similarity is calculated through the PT kernel, and the syntactic similarity score is obtained according to the similarity calculation result.

[0160] It should be understood that the PT tree kernel construction module can accurately construct the CPT of the short text with terms as the basic semantic units, then calculate the similarity of the corresponding nodes in the CPT using the PTK tree kernel, and then accumulate and normalize the similarity scores of all nodes to obtain a comprehensive score S reflecting the syntactic structure similarity of the two sentences. s 。

[0161] Here, it is also possible to combine Figure 7 to illustrate the data processing process of the PT tree kernel construction module, and the data processing process of the PT tree kernel construction module is as follows:

[0162] First, the PT tree kernel construction module can construct a syntactic parse tree (CPT). The CPT can reveal the phrase structure and its hierarchical syntactic relationship in the short text, and its tree structure effectively shows the structural information of the text. Given that a single word may not be sufficient to convey the complete semantics, and the short text may not fully follow the standard written grammar, traditional word segmentation methods may not be sufficient to ensure the accuracy of the CPT. Therefore, using terms as the basic semantic units can more accurately construct the CPT of the short text.

[0163] Next, the PT kernel is used to calculate the syntactic similarity. Similar to other tree kernel calculation methods, the PT tree kernel function PTK between trees T1 and T2 is defined as:

[0164] (16)

[0165] where T1 and T2 are the CPTs of S1 and S2 respectively, and Let \(T_1\) and \(T_2\) be the sets of nodes, and \(\Delta(node1, node2)\) be the number of common segments rooted at nodes \(node1\) and \(node2\), which are the cores of the tree kernel. Evaluating the common PT rooted at nodes \(node1\) and \(node2\) requires selecting a shared subset of the two nodes. Given the importance of the order of children in the syntactic structure, the subsequence kernel method is adopted to generate children nodes.

[0166] In the PT kernel, \(\Delta(node1, node2)\) is expressed as follows:

[0167]

[0168] where \(\alpha\) and \(\beta\) are decay factors: \(\alpha\) is the height of the tree, and \(\beta\) is the length of the subsequence. and are the ordered subsequences of \(node1\) and \(node2\), and \(p\) min returns and the minimum sequence length between them, and \(\Delta p\) calculates the number of common subtrees rooted at the root in the \(p\) subsequences. To avoid introducing too much noise, semantic information is integrated into the tree kernel. When both \(node1\) and \(node2\) are leaf nodes, the similarity is calculated using equation (14) to consider semantic information.

[0169] To better understand the above formula, a recursive function \(\Delta\) p is constructed to solve it:

[0170]

[0171] where \(n1\) and \(n2\) are the children of and respectively, and \(|n1|\) and \(|n2|\) represent the lengths of \(n1\) and \(n2\); \(n1[1:i]\) and \(n2[1:r]\) represent the subsequences from 1 to \(i\) in \(n1\) and the subsequences from 1 to \(r\) in \(n1\) respectively; and \(\Delta\) p-1 is calculated recursively and stops when reaching the leaf nodes.

[0172] It can be understood that to calculate the syntactic similarity of two test questions, first construct a CPT for each sentence, and then calculate the similarity of the corresponding nodes in these trees using formula (17). Then, the similarity scores of all nodes are accumulated and normalized to obtain a syntactic structure similarity \(S\) that comprehensively reflects the two sentences. s .

[0173] In a specific implementation, the PT tree kernel module can be constructed to build a corresponding syntactic parse tree CPT for each test question sentence based on each term triple; calculate the similarity of corresponding nodes in each syntactic parse tree CPT through a preset node common fragment number calculation formula; normalize the similarity of each node to obtain the syntactic similarity score between each test question.

[0174] In this embodiment, the syntactic similarity model processes the test question information layer by layer through the Chinese word segmentation module, the part-of-speech judgment and termization module, and the PT tree kernel construction module, from the word sequence to the term conceptualization, and then to the calculation of the syntactic structure similarity, and finally outputs the syntactic similarity score. It is beneficial to focus on the syntactic structure of the test questions, accurately evaluate the structural similarity, identify potential structural problems, enhance the expression diversity of the question bank, and assist in personalized recommendation.

[0175] In addition, reference can be made here Figure 8 to illustrate the full process of this application. Figure 8 It is a schematic diagram of the full process of the test question management method based on multi-modal adaptive similarity learning of this application.

[0176] It can be seen from Figure 8 that the full process can be divided into an initial construction stage, an evolution stage, a screening stage, and an extinction stage.

[0177] In the initial construction stage, the artificial intelligence test question corpus information is collected from network open-source resources, and through fine classification and multi-label annotation, these test question information are given clear gene codes, thus obtaining the original test question set.

[0178] In the evolution stage, an evolved test question set is generated through a preset large language model, a mutated test question set is generated by using data augmentation technology, and then the original test question set, the evolved test question set, and the mutated test question set are fused to form a test question ecological library.

[0179] Then, the test question data in the test question ecological library is preprocessed and input into a preset semantic similarity model SemBert-GCN and a preset syntactic similarity model TE-PTK respectively. In SemBert-GCN, through the SemGloVe module, a preset BERT module, and a graph neural network (GCN) module in sequence, the semantic similarity score S g is output; in TE-PTK, through the Chinese word segmentation module, the part-of-speech judgment and termization module, and the PT tree kernel construction module in sequence, the syntactic similarity score S s is output.

[0180] The semantic similarity score S g and the syntactic similarity score S sIt is then input into the Adaptive Similarity Fusion Module (ASIM), and an adaptive balancing strategy is adopted to achieve the fusion evaluation of the multi-modal features of the questions, obtaining the multi-modal similarity score Score of the questions.

[0181] In the screening stage, based on the multi-modal similarity score Score obtained in the previous stage, the dynamic threshold shunting module can effectively "screen" the qualification of the questions for storage. If the multi-modal similarity score of the question exceeds the preset similarity threshold μ, the question will be accurately imported into the question review module for further evaluation or deletion. Otherwise, it will be included in the question bank.

[0182] Based on the entire process of the method of this application, it can deeply integrate the semantic and syntactic information of the questions, accurately measure the similarity of the questions, automatically identify duplicate and low-quality questions, ensure the update and quality of the question bank content, and then achieve the accurate screening and quality improvement of the question resources. Thus, the structure of the question bank is optimized to better meet the high-standard requirements of the development of educational intelligence.

[0183] It should be noted that the above examples are only for understanding this application and do not constitute a limitation to the question management method based on multi-modal adaptive similarity learning of this application. Any simple transformation in more forms based on this technical concept is within the protection scope of this application.

[0184] It should be noted that in this article, the term "including", "comprising" or any other variant thereof is intended to cover non-exclusive inclusion, so that a process, method, article or system including a series of elements not only includes those elements, but also includes other elements not expressly listed, or also includes elements inherent to such process, method, article or system. Without further limitation, an element defined by the statement "including a..." does not exclude the existence of additional elements in the process, method, article or system including that element.

[0185] The serial numbers of the above embodiments of the present invention are only for description and do not represent the advantages or disadvantages of the embodiments. Moreover, they are only partial embodiments of this application and do not limit the patent scope of this application accordingly. Any equivalent structural transformation made under the technical concept of this application by using the content of the specification and drawings of this application, or directly / indirectly applied in other related technical fields, is included in the patent protection scope of this application.

Claims

1. A test question management method based on multi-modal adaptive similarity learning, characterized in that, The method includes: Collecting an original question set, generating an evolved question set through a preset large language model, generating a mutated question set using data augmentation techniques, and performing a fusion process on the original question set, the evolved question set, and the mutated question set to form a question ecological library; Inputting the question ecological library into a preset semantic similarity model and a preset syntactic similarity model respectively to obtain the semantic similarity scores and syntactic similarity scores of each question in the question ecological library; Obtaining a multi-modal similarity score based on the semantic similarity score and the syntactic similarity score, comparing the multi-modal similarity score with a preset threshold, and importing each question into an audit module or a question bank storage module according to the comparison result.

2. The method according to claim 1, characterized in that, The steps of collecting an original question set, generating an evolved question set through a preset large language model, generating a mutated question set using data augmentation techniques, and performing a fusion process on the original question set, the evolved question set, and the mutated question set to form a question ecological library include: Setting a data collection scope and collecting an original question set from a data source based on the data collection scope; Constructing a prompt template containing semantic constraint conditions and guiding a preset large language model to generate an evolved question set based on the original question set through the prompt template; Expanding and mutating the original question set through data augmentation methods such as synonym replacement, sentence pattern restructuring, and knowledge point extension to generate a mutated question set; Fusing the original question set, the evolved question set, and the mutated question set to obtain fused question data, and performing a deduplication process on the fused question data to form a standardized question ecological library.

3. The method according to claim 1, characterized in that The steps of inputting the question ecological library into a preset semantic similarity model to obtain the semantic similarity scores of each question in the question ecological library include: Inputting the question ecological library into a preset semantic similarity model, where the preset semantic similarity model includes a SemGloVe module, a preset BERT module, a graph neural network module, and a similarity scoring module; In the SemGloVe module, establishing semantic associations between words in the question text through co-occurrence matrix analysis, and extracting word-level similarity features in combination with an attention mechanism to construct a word similarity matrix; Integrating the word similarity matrix through a preset BERT module to perform multi-level semantic representation and generate word embedding vectors with semantic information; Constructing an adjacency matrix based on the word embedding vectors through the graph neural network module, and aggregating the adjacent node information in the adjacency matrix to generate sentence vectors; In the similarity scoring module, processing the sentence vectors sequentially through a fully connected layer and a Softmax layer to obtain the semantic similarity scores of each question.

4. The method according to claim 3, wherein The steps of establishing semantic associations between words in the question text through co-occurrence matrix analysis, and extracting word-level similarity features in combination with an attention mechanism to construct a word similarity matrix in the SemGloVe module include: Obtaining the question text in the question ecological library, and obtaining the word frequencies and word co-occurrence relationships of each word in the question text through a co-occurrence matrix to construct a global word co-occurrence count matrix; Average and aggregate the attention weights corresponding to each byte pair encoding under each of the said words to obtain the attention weights corresponding to each of the said words; Based on the attention weights corresponding to each of the said words, determine the semantic associations between each of the said words through the Division distance function to obtain a word similarity matrix.

5. The method according to claim 3, wherein The step of integrating the word similarity matrix through a preset BERT module to perform multi-level semantic representation and generate word embedding vectors with semantic information includes: Integrate the word similarity matrix into the multi-head attention mechanism of the BERT model to calculate the corresponding attention weights through the input representation vectors; Integrate the attention weights corresponding to each attention head to obtain an attention output, and perform a linear transformation on the attention output to obtain word embedding vectors with semantic information.

6. The method according to claim 3, wherein The step of constructing an adjacency matrix according to the word embedding vectors through the graph neural network module and aggregating based on the adjacency node information in the adjacency matrix to generate a sentence vector includes: Construct an adjacency matrix according to the word embedding vectors and construct a graph structure representation according to the adjacency matrix; Add corresponding relative position encodings to each node in the graph structure representation to obtain a graph structure with position information; Perform graph convolution operations based on the graph structure with position information to aggregate the adjacency node information of each of the said nodes to obtain updated node embeddings; Aggregate each of the updated node embeddings to generate a sentence vector.

7. The method according to claim 1, wherein The step of inputting the test question ecosystem library into a preset syntactic similarity model to obtain the syntactic similarity scores of each test question in the test question ecosystem library includes: Input the test question ecosystem library into a preset syntactic similarity model, and the preset syntactic similarity model includes: a Chinese word segmentation module, a part-of-speech judgment and termization module, and a PT tree kernel module; Through the Chinese word segmentation module, segment the test question text into a sequence of words according to the keywords and context in the test question information; Perform part-of-speech tagging on each word in the sequence of words through the part-of-speech judgment and termization module, and perform termization representation on each tagged word to obtain a number of term triples; In the PT tree kernel module, construct a syntactic parsing tree CPT based on each of the said term triples, calculate the node similarity through the PT kernel, and obtain the syntactic similarity score according to the similarity calculation result.

8. The method according to claim 7, wherein The Chinese word segmentation module is constructed based on a CNN-BiGRU-CRF composite neural network model. The step of segmenting the test question text into a sequence of words through the Chinese word segmentation module according to the keywords and context in the test question information includes: Map the text sequence to a word vector matrix through the CNN layer combined with the embedding layer to obtain the word vector matrix corresponding to the test question text; Capture the context information in the test question text through the BIGRU layer and convert the word vector matrix into a sequence representation; Obtain the dependency relationships of each label in the sequence representation through the CRF layer to optimize the prediction result of the sequence labeling task and obtain a sequence of words.

9. The method according to claim 7, wherein The step of performing part-of-speech tagging on each word in the word sequence through the part-of-speech judgment and termization module, and performing termization representation on each tagged word to obtain a number of term triples includes: Using a natural language processing toolkit to perform part-of-speech tagging on each word in the word sequence to obtain a word sequence after part-of-speech tagging; Confirming a number of key terms based on the word sequence after part-of-speech tagging, and performing concept mapping on each of the key terms to obtain a list of terms after concept mapping; Based on the list of terms after concept mapping, constructing semantic vector representations of each key term through a dynamic vector space model; Calculating the semantic similarity between each of the key terms, and integrating each of the key terms according to the semantic similarity between the key terms to form a combined term set, where the combined term set consists of a number of term triples.

10. The method according to claim 1, characterized in that, The step of obtaining a multimodal similarity score according to the semantic similarity score and the syntactic similarity score, comparing the multimodal similarity score with a preset threshold, and importing each of the test questions into an audit module or a question bank storage module according to the comparison result includes: Performing non-linear fusion on the semantic similarity score and the syntactic similarity score based on a dynamic weighting mechanism to generate a multimodal similarity score; Comparing the multimodal similarity score with a preset threshold; If the multimodal similarity score is higher than the preset threshold, importing the test question into the audit module; If the multimodal similarity score is not higher than the preset threshold, importing the test question into the question bank storage module.

Images