AI-based intelligent question-setting system and method

By segmenting the document feature and screening with historical exam data, the AI ​​question setting system solves the problem of inability to automatically select topics and lack of targetedness, achieving efficient and targeted test questions.

CN118113844BActive Publication Date: 2024-11-26GUANGXI ZHONGGUANG TALENT EDUCATION & TRAINING CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202410262251.0
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-03-07
Publication Date
2024-11-26
Estimated Expiration
2044-03-07

AI Technical Summary

Technical Problem

The AI ​​question setting system cannot automatically select topics based on historical exam data, and cannot ensure that the selected question type is of certain targetedness.

Method used

Characteristic segmentation of relevant documents through the document uploading end and document segmentation end, feature text is generated, and the question setting model is used to select relevant question types from the question bank, and the difficulty ratio parameters are determined based on the past exam data to filter out highly targeted ones. Test questions.

Benefits of technology

It improves the efficiency of question setting, ensures that the generated test questions are targeted and considerate, and improves the overall question setting effect of AI.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118113844B_ABST
    Figure CN118113844B_ABST
Patent Text Reader

Abstract

The present invention discloses an AI-based intelligent question-setting system and method, which relates to the field of AI question-setting technology, and solves the problem that the existing AI technology cannot automatically select questions according to historical test data, and cannot ensure that the selected question types have a certain degree of pertinence. The present invention performs screening processing again on the subsequent completed question selection, preferentially determines the number of difficult questions based on past test data, prioritizes difficult questions from the selected questions, and then determines subsequent related question types, so that the generated test questions have a certain degree of pertinence and consideration, and in the subsequent question-setting process, question types are distinguished according to knowledge points, so that the knowledge points involved can be fully considered, thereby improving the overall question-setting effect of AI and ensuring the corresponding pertinence of the test questions.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the field of AI question setting technology, and specifically to an AI-based intelligent question setting system and method. Background Art

[0002] The AI ​​question-generating function is a tool for automatically generating test questions. The goal of this function is to provide enterprises and institutions using the China Broadcasting Training online education system with a more convenient method for generating a large number of questions, improve the efficiency of question management, and provide students with more abundant questions for online practice and examinations.

[0003] The application with publication number CN109241534B discloses a method for automatically generating test questions based on text AI learning, including: obtaining the text of the test question material; performing feature extraction on the text to generate a text feature vector; using a pre-trained vector matching model to match the text with samples in a sample library according to the text feature vector, wherein the sample includes a sample test question and a sample test question material corresponding to the sample test question; using a pre-trained question setting regularity pattern determination model to determine the question setting regularity pattern according to the text feature difference between the target sample test question material and the corresponding target sample test question; according to the question setting regularity pattern, the text of the test question material is converted into a test question. The present application also provides an automatic test question generation device based on text AI learning. The present application generates test questions through artificial intelligence, which saves manpower and time costs, reduces the cost of generating test questions, and can be applied to the automatic construction of a massive question bank for Internet testing.

[0004] When setting questions, intelligent AI selects and displays question types corresponding to relevant knowledge points based on the input knowledge points. The operator selects the question types and determines the test questions. However, this method will consume the operator's time. Intelligent AI cannot automatically select questions based on historical test data, nor can it ensure that the selected question types are targeted. Summary of the invention

[0005] In view of the deficiencies in the prior art, the present invention provides an AI-based intelligent question-setting system and method, which solves the problem that AI cannot select questions on its own based on past test data and ensure that the selected question types are targeted to a certain extent.

[0006] To achieve the above objectives, the present invention is implemented through the following technical solutions: an AI-based intelligent question-setting system, comprising:

[0007] The document upload end is operated by an operator to upload relevant documents and transmit the uploaded relevant documents to the document segmentation end;

[0008] The document segmentation end performs feature segmentation on the relevant documents, identifies several feature texts from the relevant documents, and transmits the identified feature texts to the question model, including:

[0009] According to the documents uploaded by the user, the documents are first segmented according to the delimiter; the segmented text is converted into a sentence vector list [sentence1, sentence2, ..., sentenceN]; then the cosine similarity between the sentences is calculated, and the segmentation points of each round are determined by the silhouette coefficient to segment the documents;

[0010] The specific calculation formula of cosine similarity is as follows:

[0011]

[0012] Where A and B are two n-dimensional vectors; A1 to An are the values ​​of any dimension of vector A, and B1 to Bn are the values ​​of any dimension of vector B; A is [A1, A2, ..., An]; B is [B1, B2, ..., Bn]; the cosine similarity value of two vectors with exactly the same direction is 1, and the cosine similarity value of two vectors opposite to each other is -1;

[0013] The specific method of determining the segmentation point of each round by the silhouette coefficient is as follows:

[0014] Traverse each split point k and divide the document into two clusters, containing sentence vector lists [sentence1, sentence2, ..., sentencek] and [sentencek+1, sentencek+2, ..., sentenceN] respectively; for each sample, calculate the average cosine distance a with other samples in the same cluster; for each sample, calculate the average cosine distance b with samples in another cluster; calculate the silhouette coefficient: The value range of the silhouette coefficient is between -1 and 1; its meaning is: a silhouette coefficient close to 1 indicates that the sample clustering is reasonable, the distance within the cluster is small and the distance between clusters is large; a silhouette coefficient close to 0 indicates that the sample clusters overlap; a silhouette coefficient close to -1 indicates that the sample is incorrectly assigned to an adjacent cluster;

[0015] Therefore, for k separation points, the point with the largest silhouette coefficient is selected as the document segmentation point; if the length of the separated document is lower than the threshold, no separation is performed; if it is greater than the threshold, the above operation is repeated;

[0016] Segment the relevant documents based on the segmentation points to obtain feature texts;

[0017] The question-generating model selects relevant question types corresponding to the knowledge points tested from the question bank according to the input feature text, and transmits the selected relevant question types to the question type preliminary screening end. Before transmission, manual intervention is performed to determine the total transmission amount and difficulty of the relevant question types, and mark the more difficult question types as difficult questions.

[0018] At the initial screening end, based on past exam data, the frequency of occurrence of relevant question types is confirmed, and then based on the proportion of difficult questions and the proportion of relatively good scores in past exams, the difficulty ratio parameters of this exam are determined. Based on the frequency of occurrence of corresponding difficult questions and the difficulty ratio parameters, the selected difficult questions for this question are screened out, including:

[0019] Based on the extracted relevant question types, determine the frequency P of the relevant question types appearing in past exams from past exam data. i , where i represents different related question types, N groups of test questions are determined from past test data, and the number of occurrences C of the related question types is determined from the determined N groups of test questions, using P i =C÷N Determine the frequency of occurrence of related question types P i , where N is a preset value;

[0020] Then, from the past N groups of test data, from the single group of test data, determine the test score CJ k , CJ k ≥YS×80% of the test scores are marked as excellent scores, where YS is the total score of the corresponding test, k represents the test scores of different students, and based on the total number of excellent scores and the total number of test scores, the proportion of excellent scores in this group of test data ZB is determined, and then the proportion of difficult questions in this group of test data TZ is determined. The proportion of excellent scores generated in N groups of test data is marked as ZB t , the problem ratio is marked as TZ t , where t = 1, 2, ..., N, N represents the total number of different test data;

[0021] Adopt ZB t ÷TZ t =ZH t Lock conversion parameters ZH t , set N groups of conversion parameters ZH t Perform mean processing, determine the conversion factor ZY, and use BJ÷ZY=NY to determine the best difficult problem ratio NY of this test, where BJ is the preset best excellent score ratio. Based on the total number of questions proposed this time and the best difficult problem ratio NY, determine the number of difficult problems GS. If GS is not an integer, round it up;

[0022] Determine the marked difficult questions from several related question types and sort them according to the frequency of occurrence P iSort the puzzles from small to large to generate a puzzle sequence, and based on the determined puzzle number GS, extract the same number of puzzles from the puzzle sequence from the front to the back, and mark them as selected puzzles;

[0023] The second-order screening end determines the total number of remaining questions based on the total number of selected difficult questions, and then determines the remaining questions from the remaining related question types, and transmits the selected difficult questions and the remaining questions to the display end. The specific method is as follows:

[0024] Based on the total number of selected difficult questions GS and the total number of questions to be prepared, determine the total number of remaining questions ZZ;

[0025] According to the relevant question types corresponding to different knowledge points, extract the frequency P of each group of knowledge points i The smallest related question type, and the extracted related question types are used to generate a first group of question type sequences, and then the subsequent question types are processed in the same way to generate a second group of question type sequences, a third group of question type sequences, and so on;

[0026] Based on the total number of remaining questions ZZ, the first group of question type sequences are selected first. When the first group of question type sequences are completed, the second group of question type sequences are selected until the total number of remaining questions ZZ is completed, and the selected ZZ remaining questions are determined;

[0027] The selected questions and remaining questions are transmitted to the display terminal for display;

[0028] Preferably, it also includes:

[0029] The knowledge point input end is operated by external operators to input the knowledge points required for the test and directly transmit the input test knowledge points to the question-setting model;

[0030] The question-setting model selects relevant question types corresponding to the test knowledge points from the question bank according to the input test knowledge points, and transmits the selected relevant question types to the question type initial screening end.

[0031] Preferably, the AI-based intelligent question-setting method comprises the following steps:

[0032] Step 1: Input the test knowledge points or upload relevant documents and segment them to determine the feature text, and transfer the test knowledge points or feature text to the question model;

[0033] Step 2: The question-generating model selects relevant question types corresponding to the test knowledge points from the question bank according to the input test knowledge points or feature texts, and transmits the selected relevant question types to the question type preliminary screening end. Before the transmission, manual intervention is performed to determine the total transmission amount and difficulty of the relevant question types, and mark the more difficult question types as difficult questions.

[0034] Step 3: Based on past exam data, confirm the frequency of occurrence of relevant question types, and then determine the difficulty ratio parameters of this exam question based on the ratio of difficult questions and the ratio of relatively good scores in past exam questions. Based on the frequency of occurrence of corresponding difficult questions and the difficulty ratio parameters, select the selected difficult questions for this exam question.

[0035] Step 4: Based on the total number of marked selected difficult questions, determine the total number of remaining questions, then determine the remaining questions from the remaining related question types, and transmit the selected difficult questions and remaining questions to the display terminal.

[0036] The present invention provides an AI-based intelligent question-setting system and method. Compared with the prior art, it has the following advantages:

[0037] Beneficial effects:

[0038] The present invention can directly extract corresponding feature items from relevant documents by performing feature segmentation on uploaded relevant documents, without the need for operators to input corresponding knowledge points. In this way, the knowledge points to be tested can be quickly determined, and questions can be selected through the corresponding question-setting model, which can effectively improve the efficiency of the question-setting process without manual input.

[0039] For the subsequent completed questions, screening will be carried out again, and the number of difficult questions will be determined based on past test data. The difficult questions will be prioritized from the questions, and then the subsequent related question types will be determined, so that the generated test questions have a certain degree of pertinence and consideration. In the subsequent question-setting process, the question types are differentiated according to the knowledge points, so that the knowledge points involved can be fully considered, thereby improving the overall question-setting effect of the AI ​​and ensuring the corresponding pertinence of the test questions. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] Figure 1 It is a schematic diagram of the principle framework of the present invention;

[0041] Figure 2 It is a schematic diagram of the feature text segmentation process of the present invention. DETAILED DESCRIPTION

[0042] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0043] Embodiment 1

[0044] See also Figure 1, the present application provides an AI-based intelligent question-setting system, including a knowledge point input terminal, a question-setting model, a question type initial screening terminal, a second-order screening terminal, and a display terminal, as well as a document upload terminal and a document segmentation terminal;

[0045] The knowledge point input terminal, the question model, the question type initial screening terminal, the second-order screening terminal and the display terminal are electrically connected in sequence from the output node to the input node, and the document upload terminal, the document segmentation terminal and the question model are electrically connected in sequence from the output node to the input node;

[0046] Among them, the knowledge point input end is operated by the external operator to input the knowledge points required for the test, and directly transmit the input test knowledge points to the question model. Specifically, when inputting the knowledge points, the input knowledge points have an impact on the quality of the questions, and the scope of the required questions should be described as much as possible, or a relatively complete text should be used;

[0047] Among them, the question model selects the relevant question types corresponding to the test knowledge points from the question bank according to the input test knowledge points, and transmits the selected relevant question types to the question type preliminary screening end. Before transmission, manual intervention is performed to determine the total transmission amount and difficulty of the relevant question types, and mark the more difficult question types as difficult questions. The reason for determining and marking the difficulty of the question types here is to select difficult questions later, so as to ensure that the generated specific test questions have certain test capabilities;

[0048] Among them, the initial screening of question types is based on past test data to confirm the frequency of occurrence of relevant question types, and then based on the proportion of difficult questions and the proportion of better scores in past test questions, the difficulty ratio parameters of this test questions are determined. Based on the frequency of occurrence of corresponding difficult questions and the difficulty ratio parameters, the difficult questions of this selection are screened out and marked as selected difficult questions. Specifically, the relevant question types given by the question bank are definitely more than the question types required this time, so it is necessary to select the corresponding question types of this test from several related question types, and integrate the selected question types to generate all the test question types of this test. Assuming that this test only needs 100 questions, and its related question types include more than 200 questions, it is necessary to screen them step by step to determine 100 questions;

[0049] Among them, the specific methods for screening out the difficult problems in this topic selection include:

[0050] Based on the extracted relevant question types, determine the frequency P of the relevant question types appearing in past exams from past exam data. i , where i represents different related question types, N groups of test questions are determined from past test data, and the number of occurrences C of the related question types is determined from the determined N groups of test questions, using P i =C÷N Determine the frequency of occurrence of related question types P i, where N is a preset value, and its specific value is determined by the operator based on experience;

[0051] Then, from the past N groups of test data, from the single group of test data, determine the test score CJ k , CJ k ≥YS×80% of the test scores are marked as excellent scores, where YS is the total score of the corresponding test, k represents the test scores of different students, and based on the total number of excellent scores and the total number of test scores, the excellent score ratio ZB in this group of test data is determined, where ZB=total number of excellent scores÷total number of test scores, and then the difficult problem ratio TZ in this group of test data is determined, and the difficult problem ratio TZ=total number of difficult problems÷total number of test questions in a single group, and the excellent score ratio generated in N groups of test data is marked as ZB t , the problem ratio is marked as TZ t , where t = 1, 2, ..., N, N represents the total number of different test data;

[0052] Adopt ZB t ÷TZ t =ZH t Lock conversion parameters ZH t , set N groups of conversion parameters ZH t Perform mean processing, determine the conversion factor ZY, and use BJ÷ZY=NY to determine the best difficult problem ratio NY of this test, where BJ is the preset best excellent score ratio (its value is between 20%-30%, and the specific value is determined by the operator based on experience). Based on the total number of questions to be prepared and the best difficult problem ratio NY, determine the number of difficult problems GS. If GS is not an integer, perform rounding. For example, if GS is 5.1, directly take 6;

[0053] Determine the marked difficult questions from several related question types and sort them according to the frequency of occurrence P i The puzzles are sorted from small to large to generate a puzzle sequence. Based on the determined puzzle number GS, the same number of puzzles are extracted from the puzzle sequence from the front to the back and marked as selected puzzles. If the number of extracted puzzles is insufficient, an error signal is generated and displayed on the external display terminal, and the operator intervenes to make up the number of puzzles.

[0054] Specifically, since the difficult questions have been marked, in order to ensure that the difficulty of the test questions can better evaluate the candidates and make the evaluation results more examineable, it is necessary to determine the corresponding number of difficult questions. If the number of difficult questions is too large, the overall score will be low. If the number of difficult questions is too small, the overall score will be high, and both are not examineable. Therefore, a more standard number of corresponding difficult questions must be determined.

[0055] The second-order screening end determines the total number of remaining questions based on the total number of selected difficult questions that have been marked, and then determines the remaining questions from the remaining related question types, and transmits the selected difficult questions and the remaining questions to the display end. The specific method of determining the remaining questions is:

[0056] Based on the total number of selected difficult questions GS and the total number of questions to be prepared, determine the total number of remaining questions ZZ;

[0057] According to the relevant question types corresponding to different knowledge points, extract the frequency P of each group of knowledge points i The smallest related question type is extracted, and the extracted related question types are used to generate the first group of question type sequences, and then the subsequent question types are processed in the same way to generate the second group of question type sequences, the third group of question type sequences, and so on. For example: Assume that there are three knowledge points, namely A, B and C, and each knowledge point carries three groups of related question types. First, select the related question types with a frequency of occurrence P from A, B and C. i The smallest question type, assuming A{A1, A2, A3}, where the frequencies of A1, A2, and A3 are 0.1, 0.2, and 0.3 respectively, so in the first selection, A1 in the knowledge point A will be selected, and then the question type with the smallest frequency of occurrence is selected from the knowledge points B and C, assuming they are B1 and C1 respectively, then the first group of question type sequences generated is A1, B1, and C1. Since the question type with the smallest frequency of occurrence within the corresponding knowledge point has been selected, the question type with the corresponding frequency of occurrence can also be selected. If the same P appears in different question types within the same knowledge point, i , then randomly select;

[0058] Based on the total number of remaining questions ZZ, the first group of question type sequences are selected first. When the first group of question type sequences are completed, the second group of question type sequences are selected until the total number of remaining questions ZZ is completed, and the selected ZZ remaining questions are determined;

[0059] The selected questions and remaining questions are transmitted to the display terminal for display;

[0060] This way of selecting question types can fully and evenly cover the knowledge points to be tested, ensuring the comprehensiveness of the test questions.

[0061] Embodiment 2

[0062] The knowledge points in the above-mentioned embodiment are input by the operator. In this embodiment, the characteristic text (that is, the characteristic knowledge point) is determined from the transmitted document, and the determined characteristic text is transmitted to the question-setting model.

[0063] The document upload end is operated by the operator to upload the relevant documents and transfer the uploaded relevant documents to the document splitting end. The relevant document formats only support pdf, doc, docx, and txt, and the file size does not exceed 500MB.

[0064] Among them, combined Figure 2 , the document segmentation end performs feature segmentation on the relevant documents, identifies several feature texts from the relevant documents, and transmits the identified feature texts to the question model;

[0065] Among them, the specific methods of feature segmentation include:

[0066] According to the documents uploaded by users, the documents are first segmented according to the delimiters, which include the period, question mark, exclamation mark and semicolon in Chinese and English. The segmented text is converted into a sentence vector list [sentence1, sentence2, ..., sentenceN]. Then the cosine similarity between sentences is calculated, and the segmentation point of each round is determined by the silhouette coefficient to segment the documents.

[0067] The specific calculation formula of cosine similarity is as follows:

[0068]

[0069] Where A and B are two n-dimensional vectors; A1 to An are the values ​​of any dimension of vector A, and B1 to Bn are the values ​​of any dimension of vector B; A is [A1, A2, ..., An]; B is [B1, B2, ..., Bn]; the cosine similarity value of two vectors with exactly the same direction is 1, and the cosine similarity value of two vectors opposite to each other is -1;

[0070] Silhouette Coefficient is a commonly used indicator for clustering algorithm evaluation. It can help us judge the compactness and separation of clustering results. This article uses it to select the best segmentation point for text. The usage is as follows:

[0071] Traverse each split point k and divide the document into two clusters, containing sentence vector lists [sentence1, sentence2, ..., sentencek] and [sentencek+1, sentencek+2, ..., sentenceN] respectively. For each sample, calculate the average cosine distance a with other samples in the same cluster; for each sample, calculate the average cosine distance b with samples in another cluster; calculate the silhouette coefficient: The value range of the silhouette coefficient is between -1 and 1. Its meaning is: a silhouette coefficient close to 1 indicates that the sample clustering is reasonable, the distance within the cluster is small and the distance between clusters is large; a silhouette coefficient close to 0 indicates that the sample clusters overlap; a silhouette coefficient close to -1 indicates that the sample is incorrectly assigned to an adjacent cluster.

[0072] Therefore, for k separation points, the point with the largest silhouette coefficient is selected as the document segmentation point. If the length of the separated document is lower than the threshold, no further separation is performed. If it is greater than the threshold, the above operation is repeated;

[0073] The relevant documents are segmented based on the segmentation points to obtain feature texts.

[0074] The question-setting model selects relevant question types corresponding to the knowledge points tested from the question bank based on the input feature text, and transmits the selected relevant question types to the question type initial screening end. Before the transmission, manual intervention is performed to determine the total amount of relevant question types to be transmitted and their difficulty, and more difficult question types are marked as difficult questions.

[0075] Embodiment 3

[0076] The AI-based intelligent question-setting method includes the following steps:

[0077] Step 1: Input the test knowledge points or upload relevant documents and segment them to determine the feature text, and transfer the test knowledge points or feature text to the question model;

[0078] Step 2: The question-generating model selects relevant question types corresponding to the test knowledge points from the question bank according to the input test knowledge points or feature texts, and transmits the selected relevant question types to the question type preliminary screening end. Before the transmission, manual intervention is performed to determine the total transmission amount and difficulty of the relevant question types, and mark the more difficult question types as difficult questions.

[0079] Step 3: Based on past exam data, confirm the frequency of occurrence of relevant question types, and then determine the difficulty ratio parameters of this exam question based on the ratio of difficult questions and the ratio of relatively good scores in past exam questions. Based on the frequency of occurrence of corresponding difficult questions and the difficulty ratio parameters, select the selected difficult questions for this exam question.

[0080] Step 4: Based on the total number of marked selected difficult questions, determine the total number of remaining questions, then determine the remaining questions from the remaining related question types, and transmit the selected difficult questions and remaining questions to the display terminal.

[0081] Embodiment 4

[0082] The specific implementation process of this embodiment includes all the implementation processes of the above embodiments.

[0083] Some of the data in the above formulas are dimensionless and numerically calculated. Meanwhile, the contents not described in detail in this specification belong to the prior art known to those skilled in the art.

[0084] The above embodiments are only used to illustrate the technical method of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical method of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical method of the present invention.

Claims

1. The AI-based intelligent question-setting system is characterized by: include: The document upload end is operated by an operator to upload relevant documents and transmit the uploaded relevant documents to the document segmentation end; The document segmentation end performs feature segmentation on the relevant documents, identifies several feature texts from the relevant documents, and transmits the identified feature texts to the question model, including: According to the documents uploaded by the user, the documents are first segmented according to the delimiter; the segmented text is converted into a sentence vector list [sentence1, sentence2, ..., sentenceN]; then the cosine similarity between the sentences is calculated, and the segmentation points of each round are determined by the silhouette coefficient to segment the documents; The specific calculation formula of cosine similarity is as follows: Among them, A i and B i are two n-dimensional vectors; A1 to An are vectors A i Any dimension value of , B1 to Bn are vectors B i Any dimension value of A i is [A1, A2,...,An]; B i is [B1, B2, ..., Bn]; the cosine similarity value of two vectors with exactly the same direction is 1, and the cosine similarity value of two vectors opposite to each other is -1; The specific method of determining the segmentation point of each round by the silhouette coefficient is: Traverse each split point k and divide the document into two clusters, containing sentence vector lists [sentence1, sentence2, ..., sentencek] and [sentencek+1, sentencek+2, ..., sentenceN] respectively; for each sample, calculate the average cosine distance a with other samples in the same cluster; for each sample, calculate the average cosine distance b with samples in another cluster; calculate the silhouette coefficient: The value range of the silhouette coefficient is between -1 and 1; its meaning is: a silhouette coefficient close to 1 indicates that the sample clustering is reasonable, the distance within the cluster is small and the distance between clusters is large; a silhouette coefficient close to 0 indicates that the sample clusters overlap; a silhouette coefficient close to -1 indicates that the sample is incorrectly assigned to an adjacent cluster; Therefore, for k separation points, the point with the largest silhouette coefficient is selected as the document segmentation point; if the length of the separated document is lower than the threshold, no separation is performed; if it is greater than the threshold, the above operation is repeated; Segment the relevant documents based on the segmentation points to obtain feature texts; The question-generating model selects relevant question types corresponding to the knowledge points tested from the question bank according to the input feature text, and transmits the selected relevant question types to the question type preliminary screening end. Before transmission, manual intervention is performed to determine the total transmission amount and difficulty of the relevant question types, and mark the more difficult question types as difficult questions. At the initial screening end, based on past exam data, the frequency of occurrence of relevant question types is confirmed, and then based on the proportion of difficult questions and the proportion of relatively good scores in past exams, the difficulty ratio parameters of this exam are determined. Based on the frequency of occurrence of corresponding difficult questions and the difficulty ratio parameters, the selected difficult questions for this question are screened out, including: Based on the extracted relevant question types, determine the frequency P of the relevant question types appearing in past exams from past exam data. i , where i represents different related question types, N groups of test questions are determined from past test data, and the number of occurrences C of the related question types is determined from the determined N groups of test questions, using P i =C÷N Determine the frequency of occurrence of related question types P i , where N is a preset value; Then, from the past N groups of test data, from the single group of test data, determine the test score CJ k , CJ k ≥YS×80% of the test scores are marked as excellent scores, where YS is the total score of the corresponding test, k represents the test scores of different students, and based on the total number of excellent scores and the total number of test scores, determine the proportion of excellent scores in this group of test data ZB, and then confirm the proportion of difficult questions in this group of test data TZ, and mark the proportion of excellent scores in N groups of test data as ZB t , the problem ratio is marked as TZ t , where t = 1, 2, ..., N, N represents the total number of different test data; Adopt ZB t ÷TZ t =ZH t Lock conversion parameters ZH t , set N groups of conversion parameters ZH t Perform mean processing, determine the conversion factor ZY, and use BJ÷ZY=NY to determine the best difficult problem ratio NY of this test, where BJ is the preset best excellent score ratio. Based on the total number of questions proposed this time and the best difficult problem ratio NY, determine the number of difficult problems GS. If GS is not an integer, round it up; Determine the marked difficult questions from several related question types and sort them according to the frequency of occurrence P i Sort the puzzles from small to large to generate a puzzle sequence, and based on the determined puzzle number GS, extract the same number of puzzles from the puzzle sequence from the front to the back, and mark them as selected puzzles; The second-order screening end determines the total number of remaining questions based on the total number of selected difficult questions, and then determines the remaining questions from the remaining related question types, and transmits the selected difficult questions and the remaining questions to the display end. The specific method is as follows: Based on the total number of selected difficult questions GS and the total number of questions to be prepared, determine the total number of remaining questions ZZ; According to the relevant question types corresponding to different knowledge points, extract the frequency P of each group of knowledge points i The smallest related question type, and the extracted several related question types are used to generate a first group of question type sequences, and then the subsequent question types are processed in the same way to generate a second group of question type sequences and a third group of question type sequences; Based on the total number of remaining questions ZZ, the first group of question type sequences are selected first. When the first group of question type sequences are completed, the second group of question type sequences are selected until the total number of remaining questions ZZ is completed, and the selected ZZ remaining questions are determined; The selected difficult questions and the remaining questions are transmitted to the display terminal for display.

2. The AI-based intelligent question-setting system according to claim 1, characterized in that: Also includes: The knowledge point input end is operated by external operators to input the knowledge points required for the test and directly transmit the input test knowledge points to the question-setting model; The question-setting model selects relevant question types corresponding to the test knowledge points from the question bank according to the input test knowledge points, and transmits the selected relevant question types to the question type initial screening end.

3. The AI-based intelligent question-setting system according to claim 1, characterized in that: The proportion of excellent scores ZB = the total number of excellent scores ÷ the total number of test scores, and the proportion of difficult questions TZ = the total number of difficult questions ÷ the total number of test questions in a single group.

4. An AI-based intelligent question-setting method, which is applied to the AI-based intelligent question-setting system according to any one of claims 1 to 3, characterized in that: The following steps are involved: Step 1: Input the test knowledge points or upload relevant documents and segment them to determine the feature text, and transfer the test knowledge points or feature text to the question model; Step 2: The question-generating model selects relevant question types corresponding to the test knowledge points from the question bank according to the input test knowledge points or feature texts, and transmits the selected relevant question types to the question type preliminary screening end. Before the transmission, manual intervention is performed to determine the total transmission amount and difficulty of the relevant question types, and mark the more difficult question types as difficult questions. Step 3: Based on past exam data, confirm the frequency of occurrence of relevant question types, and then determine the difficulty ratio parameters of this exam question based on the ratio of difficult questions and the ratio of relatively good scores in past exam questions. Based on the frequency of occurrence of corresponding difficult questions and the difficulty ratio parameters, select the selected difficult questions for this exam question. Step 4: Based on the total number of selected difficult questions, determine the total number of remaining questions, then determine the remaining questions from the remaining related question types, and transmit the selected difficult questions and the remaining questions to the display terminal.