An intelligent test paper-generating method for medical theory examination based on artificial intelligence technology

Abstract

The present invention relates to the technical field of question composition, and discloses an intelligent composition method for medical theory examinations based on artificial intelligence technology, comprising the steps of: question information preprocessing, question similarity calculation, and intelligent composition. In the question processing link, questions are decomposed and structured according to multiple modalities such as text, image, and case analysis, and features are extracted by combining techniques such as keyword extraction and local feature descriptor methods. In question similarity calculation, cosine similarity and weighted summation are used to accurately quantify the similarity of questions of different question types, effectively avoiding question duplication in the examination paper. When composing the examination paper, questions are screened according to preset question type structure, knowledge point coverage, difficulty level and other parameters, and similarity threshold judgment and replacement mechanism are combined to ensure a reasonable difficulty distribution and comprehensive coverage of knowledge points. The replacement object is determined according to the number of historical usage of the question, thereby improving the scientific nature of the composition and the utilization rate of the question bank resources, and optimizing the composition process of the medical theory examination.

Description

Technical Field

[0001] The present invention relates to the technical field of question composition, and in particular to an intelligent composition method for medical theory examinations based on artificial intelligence technology. Background Art

[0002] In the field of medical education and talent assessment, medical theory exams are a key step in evaluating medical students' knowledge base, clinical thinking, and practical skills. The quality of these exams directly impacts the scientific nature and effectiveness of the assessment results. Traditionally, medical theory exams rely on manual compilation. Compilers must manually select questions from a massive question bank based on the exam syllabus. This not only consumes considerable time and effort, but also makes it difficult to precisely control the distribution of question types, knowledge coverage, and difficulty level. This can easily lead to problems such as unbalanced exam structure, repeated questions, and omitted knowledge points.

[0003] With the vigorous development of artificial intelligence technology, some intelligent test paper generation methods have emerged, but they still have significant defects. Existing technologies often do not fully consider the complexity and particularity of medical examinations. They lack systematic classification processing and structured storage of multimodal questions such as text, image and case analysis. It is difficult to accurately extract the characteristics of various questions, resulting in deviations in the calculation of question similarity, and the test paper generation results cannot meet the diverse examination needs. In addition, in the process of test paper generation, highly similar questions cannot be effectively avoided, and problems such as repeated examination content and insufficient discrimination are prone to occur, making it impossible to comprehensively and objectively evaluate the medical knowledge level and comprehensive application ability of candidates. Therefore, there is an urgent need for a scientific, efficient and intelligent test paper generation method suitable for medical theory examinations to improve the quality of examinations and the effectiveness of assessments. Summary of the Invention

[0004] The purpose of the present invention is to provide an intelligent test paper-forming method for medical theory examinations based on artificial intelligence technology, which solves the technical problems raised in the background technology.

[0005] The purpose of the present invention can be achieved through the following technical solutions:

[0006] An intelligent test paper-generating method for medical theory examination based on artificial intelligence technology comprises the following steps:

[0007] Step 1: Preprocessing the question information: Break down the questions into text, image, and case analysis multimodal questions by question type, and extract features from them. Feature extraction includes:

[0008] Text feature extraction: Segment the text of text and case analysis questions, calculate word frequency, inverse document frequency, and determine the keyword set;

[0009] Image feature extraction: For image and case analysis questions, the image parts are divided into blocks, color and texture features are extracted, and a set of image feature vectors is constructed;

[0010] Step 2: Calculate topic similarity: Based on the keyword set and feature vector set, calculate text similarity, image similarity, and overall multimodal topic similarity.

[0011] Step 3: Intelligent test paper composition: Filter questions according to the preset question structure, knowledge point coverage and difficulty level requirements, and optimize the combination to obtain a set of test paper questions.

[0012] As a further solution of the present invention: for text-based questions: directly extract the text content and represent it with a text set W = {W (i1, j1)};

[0013] Where i1=1, 2, ..., n1, n1 represents the number of text-type questions, j1=1, 2, ..., m1, m1 represents the number of characters in the i1-th text-type question; and W(i1, j1) represents the j1-th character in the i1-th text-type question;

[0014] For image-related questions: directly extract the original image data and represent it with the image set X = {X(i2, j2)};

[0015] Where i2=1, 2, ..., n2, n2 represents the number of image-type questions, j2=1, 2, ..., m2, m2 represents the number of images in the i-th image-type question; and X(i2, j2) represents the j2-th image in the i2-th image-type question;

[0016] For case analysis multimodal questions: separate the text part and the image part, and pass the text part through WX={WX(i3,j a )}, the image part is represented by XW={X(i3,j b )}express;

[0017] Among them, i3=1, 2, ... n3, n3 represents the number of case analysis multimodal questions, j a =1, 2, ...m a , m a represents the number of characters in the i3th case analysis multimodal question; j b =1, 2, ...m b , m b represents the number of images in the i3th case analysis multimodal question; and X(i3,j a ) represents the jth case analysis multimodal question in the i3th case analysis a characters, X(i3,j b ) represents the jth case analysis multimodal question in the i3th case analysis b images.

[0018] As a further solution of the present invention: the text feature extraction method is as follows:

[0019] For W = {W (i1, j1)} and WX = {WX (i3, j a )}, use keyword extraction technology to obtain text features, the specific steps are as follows:

[0020] Step 1.2.1, word segmentation processing:

[0021] Use word segmentation technology in natural language processing to split the text into independent words;

[0022] Step 1.2.2, calculate word frequency:

[0023] Count the number of times each word appears in each question containing text content C(i, j);

[0024] Among them, i∈{i1, i3}, j∈{j1, j a};

[0025] pass , calculate the frequency TF(i, j) of the jth word appearing in the i-th question containing text content;

[0026] Among them, m∈{m1,m a};

[0027] Step 1.2.3. Calculate the inverse document frequency:

[0028] Then, select each word one by one and count the number of documents containing each word in the title containing text content, and record it as D(j);

[0029] Then through , calculate the inverse document frequency IDF(j) of each word;

[0030] Step 1.2.4, determine the keywords:

[0031] Afterwards, , calculate the term frequency - inverse document frequency TF-IDF (i, j) of each word;

[0032] Then, words whose TF-IDF (i, j) is higher than the corresponding preset TF-IDF threshold are selected to form the keyword set K.

[0033] As a further solution of the present invention: the image feature extraction method is as follows:

[0034] For X = {X (i2, j2)} and XW = {X (i3, j b )}Use the local feature descriptor method to extract image features. The specific steps are as follows:

[0035] Step 1.3.1: Image segmentation:

[0036] Select an image and divide it into s local regions;

[0037] Step 1.3.2, color feature extraction:

[0038] Convert the image from RGB color space to HSV color space, and then calculate the color mean and variance of each local area:

[0039] Specifically include:

[0040] Mean HP(k) and variance HF(k) of hue H;

[0041] The mean SP(k) and variance SF(k) of saturation S;

[0042] The mean VP(k) and variance VF(k) of brightness V;

[0043] Where k = 1, 2, ... s;

[0044] Step 1.3.3, texture feature extraction:

[0045] Gray-level co-occurrence matrix is ​​used to calculate the texture parameters of each local area in the image;

[0046] Specifically include:

[0047] Energy E(k), entropy H(k), contrast C(k), correlation R(k);

[0048] Step 1.3.4, construct feature vector:

[0049] The color features and texture features extracted in the local area are used to form the feature vector of the local area:

[0050] That is, F(k)=[HP(k), HF(k), SP(k), SF(k), VP(k), VF(k), E(k), H(k), C(k), R(k)];

[0051] Then, the feature vectors corresponding to all local areas contained in the image are combined into an image feature vector set, that is, F=[F(k)]=[F1,F2,…Fs].

[0052] As a further solution of the present invention: the text similarity is calculated as follows:

[0053] The cosine similarity calculation method based on the keyword set is as follows:

[0054] Select two text-type topics and extract their corresponding keyword sets Ka={Ka(g1)} and Kb={Ka(g2)} and convert them into binary vectors;

[0055] Where g2=1, 2, ..., v1, v1 represents the number of all keywords in the keyword set Ka, g2=1, 2, ..., v2; v2 represents the number of all keywords in the keyword set Kb;

[0056] The length of the binary vector is the number r of all non-repeated keywords in the two texts, where the occurrence of a keyword is marked as 1 and the absence of a keyword is marked as 0.

[0057] Get binary vectors: Va=[Va(f)] and Vb=[Vb(f)];

[0058] Where f = 1, 2, ..., r;

[0059] pass: , calculate the text similarity SK between two text-type topics.

[0060] As a further solution of the present invention: the image similarity is calculated as follows:

[0061] The cosine similarity calculation method based on the image feature vector set is as follows:

[0062] Select two image-related topics and extract their corresponding keyword sets Fa={Fa(k)} and Fb={Fb(k)} and convert them into binary vectors;

[0063] pass: , calculate the image similarity SF between the two image class topics.

[0064] As a further solution of the present invention: the similarity calculation method of case analysis multimodal questions is as follows:

[0065] Based on the text and image similarity calculation method, a weighted summation method is used. The steps are as follows:

[0066] Select two case analysis multimodal questions, calculate the similarity SK0 of the corresponding text parts of the two using the text similarity calculation method, and at the same time, calculate the similarity SF0 of the corresponding image parts of the two using the image similarity calculation method;

[0067] pass: , calculate the overall similarity SZ between two case analysis multimodal questions;

[0068] Where, α K and α F is the corresponding preset weight coefficient, and α K+α F =1.

[0069] As a further solution of the present invention: the intelligent test paper forming method is as follows:

[0070] Step R1: Extract the test paper generation requirements:

[0071] Extract the following parameters that are pre-set according to the purpose of the exam:

[0072] Question structure: The total number of questions is N, and the question type distribution is Nx for text questions, Ny for image questions, and Nz for case analysis questions;

[0073] And it satisfies N=Nx+Ny+Nz;

[0074] Knowledge point coverage: The knowledge point set S = {S(q)}, and each knowledge point S(q) is covered by at least e(q) questions, where e(q) is a preset value;

[0075] Difficulty level: Questions are divided into three levels: low, medium, and high. The number of questions in each level is N1, N2, and N3, respectively, satisfying N=N1+N2+N3;

[0076] Step R2, topic screening:

[0077] Step R2.1, Question Type Screening:

[0078] Select text, image, and case analysis questions that meet the quantity requirements from the pre-established question bank and form a question set Tx, Ty, and Tz that meets the following requirements:

[0079] ∣Tx∣=Nx,∣Ty∣=Ny,∣Tz∣=Nz;

[0080] Among them, |Tx|, |Ty|, and |Tz| refer to the number of text, image, and case analysis questions in the question set, respectively;

[0081] Step R2.2, knowledge point screening:

[0082] For each knowledge point S(q), select questions related to this knowledge point from Tx, Ty, and Tz, and at the same time, the knowledge point S(q) must be covered by at least e(q) questions;

[0083] Step R2.3, Difficulty Screening:

[0084] According to the preset difficulty level of the questions, select the questions that meet the difficulty quantity requirements from the screened questions to form the question sets T1, T2, and T3, and they meet the following requirements:

[0085] ∣T1∣=N1, ∣T2∣=N2, ∣T3∣=N3;

[0086] Among them, |T1|, |T2|, and |T3| refer to the number of questions of each difficulty level in the question set;

[0087] Step R3, question combination optimization:

[0088] Step R3.1, similarity extraction:

[0089] For a set of questions of the same question type, extract the similarity S0 between each question;

[0090] Among them, S0∈{SK, SF, SZ};

[0091] Step R3.2, threshold determination:

[0092] Extracting a preset similarity threshold S0y;

[0093] Where S0y∈{SKy, SFy, SZy}; SKy refers to the text threshold, SFy refers to the image threshold, and SZy refers to the multimodal threshold;

[0094] Between two questions in the same question type, if their similarity S0 is greater than the similarity threshold S0y, they are judged as a highly similar question pair;

[0095] Then, a question is randomly selected from the remaining questions of the original screening criteria, that is, a question that meets the requirements of question structure, knowledge point coverage, and difficulty level but is not selected, and replaces the question in the highly similar question pair;

[0096] Then, the similarity between the replaced question and other questions is recalculated until the similarity S0 of all question pairs is less than or equal to the similarity threshold S0y, and the set of test questions is obtained.

[0097] As a further solution of the present invention: wherein, the question to be replaced in the highly similar question pair is determined by its historical usage count in the question bank, and specifically the question with the largest historical usage count is replaced.

[0098] Beneficial effects of the present invention:

[0099] Efficient question processing and feature extraction: By carefully breaking down and structurally storing medical exam questions into multiple modalities such as text, images, and case analysis, combined with advanced text feature extraction technologies (such as keyword extraction) and image feature extraction methods (such as local feature descriptor methods), we can accurately obtain the core information of the questions, providing a solid data foundation for subsequent intelligent test paper composition, and ensuring a deep understanding and effective use of various questions.

[0100] Accurate question similarity calculation: Based on the cosine similarity calculation method between keyword sets and image feature vector sets, and a weighted summation method for multimodal questions such as case analysis, this method accurately quantifies the similarity of different question types. This helps avoid the occurrence of a large number of similar questions in the test paper, ensures the richness and diversity of the test paper content, and enables the comprehensive assessment of candidates' knowledge.

[0101] Intelligent and Flexible Test-Generating Strategy: Based on pre-defined parameters such as question structure, knowledge point coverage, and difficulty level, this multi-dimensional screening, combined with similarity threshold judgment and an optimized replacement mechanism, not only rapidly generates test papers that meet exam requirements, but also ensures a reasonable distribution of question difficulty, comprehensive knowledge point coverage, and reasonable similarity between questions. Furthermore, replacements are determined based on the number of times a question has been used historically, further enhancing the rationality and scientific nature of test-generating and improving the efficiency of question bank resource utilization. BRIEF DESCRIPTION OF THE DRAWINGS

[0102] The present invention will be further described below with reference to the accompanying drawings.

[0103] Figure 1 This is a flow chart of an intelligent test paper-forming method for medical theory examinations based on artificial intelligence technology according to the present invention;

[0104] Figure 2 The present invention is a flowchart of an intelligent test paper composition method for medical theory examination based on artificial intelligence technology. DETAILED DESCRIPTION

[0105] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. 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 making any creative efforts shall fall within the scope of protection of the present invention.

[0106] As embodiment 1 of the present invention:

[0107] See also Figure 1 and Figure 2 As shown, the present invention is an intelligent test paper-forming method for medical theory examination based on artificial intelligence technology, comprising the following steps:

[0108] Step 1: Preprocessing of topic information:

[0109] For Chinese text questions in medical examinations;

[0110] Directly extract the text content and represent it with a text set W = {W (i1, j1)};

[0111] Where i1=1, 2, ..., n1, n1 represents the number of text-type questions, j1=1, 2, ..., m1, m1 represents the number of characters in the i1-th text-type question; and W(i1, j1) represents the j1-th character in the i1-th text-type question;

[0112] For W = {W (i1, j1)}, keyword extraction technology is used to obtain text features. The specific steps are as follows:

[0113] Step 1.2.1, word segmentation processing:

[0114] Use word segmentation technology in natural language processing to split the text into independent words;

[0115] Step 1.2.2, calculate word frequency:

[0116] Count the number of times each word appears in each question containing text content C(i, j);

[0117] Among them, i∈{i1, i3}, j∈{j1, j a};

[0118] pass , calculate the frequency TF(i, j) of the jth word appearing in the i-th question containing text content;

[0119] Among them, m∈{m1,m a};

[0120] Step 1.2.3. Calculate the inverse document frequency:

[0121] Then, select each word one by one and count the number of documents containing each word in the title containing text content, and record it as D(j);

[0122] Then through , calculate the inverse document frequency IDF(j) of each word;

[0123] Step 1.2.4, determine the keywords:

[0124] Afterwards, , calculate the term frequency - inverse document frequency TF-IDF (i, j) of each word;

[0125] Then, words with TF-IDF (i, j) higher than the corresponding preset TF-IDF threshold are selected to form the keyword set K, which is the feature vector of the text;

[0126] Step 2: Calculate the similarity of the title:

[0127] Text similarity calculation:

[0128] The cosine similarity calculation method based on the keyword set is as follows:

[0129] Select two text-type topics and extract their corresponding keyword sets Ka={Ka(g1)} and Kb={Ka(g2)} and convert them into binary vectors;

[0130] Where g2=1, 2, ..., v1, v1 represents the number of all keywords in the keyword set Ka, g2=1, 2, ..., v2; v2 represents the number of all keywords in the keyword set Kb;

[0131] The length of the binary vector is the number r of all non-repeated keywords in the two texts, where the occurrence of a keyword is marked as 1 and the absence of a keyword is marked as 0.

[0132] Get binary vectors: Va=[Va(f)] and Vb=[Vb(f)];

[0133] Where f = 1, 2, ..., r;

[0134] For example: Suppose there are two sets of keywords in the text:

[0135] K1={"heart disease","treatment","symptoms"};

[0136] K2={"diabetes","treatment","complications"};

[0137] All non-repeated keywords in texts K1 and K2 are: {"heart disease", "treatment", "symptoms", "diabetes", "complications"}, that is, the vector length is r = 5;

[0138] For the binary vector of text K1:

[0139] “heart disease”∈K1, “treatment”∈K1, “symptoms”∈K1, “diabetes”∉K1, “complications”∉K1;

[0140] Then its binary vector is V1=[1, 1, 1, 0, 0];

[0141] "heart disease"∉K1, "treatment"∈K1, "symptoms"∉K1, "diabetes"∈K1, "complications∈K1;

[0142] Then its binary vector is V1=[0, 1, 0, 0, 0];

[0143] pass: , calculate the text similarity SK between two text-type questions;

[0144] Among them, the larger the SK value, the higher the text similarity, and the value range is [0,1];

[0145] Step 3: Intelligent test paper generation:

[0146] Step R1: Extract the test paper generation requirements:

[0147] Extract the following parameters that are pre-set according to the purpose of the exam:

[0148] Question structure: Number of text questions Nx;

[0149] Knowledge point coverage: The knowledge point set S = {S(q)}, and each knowledge point S(q) is covered by at least e(q) questions, where e(q) is a preset value;

[0150] Difficulty level: Questions are divided into three levels: low, medium, and high. The number of questions in each level is N1, N2, and N3, respectively, satisfying N=N1+N2+N3;

[0151] Step R2, topic screening:

[0152] Step R2.1, Question Type Screening:

[0153] Select text questions that meet the quantity requirements from the pre-established question bank and form a question set Tx, which satisfies: |Tx|=Nx;

[0154] Among them, |Tx| refers to the number of text-related questions in the question set;

[0155] Step R2.2, knowledge point screening:

[0156] For each knowledge point S(q), select questions related to this knowledge point from Tx, and at the same time, the knowledge point S(q) must be covered by at least e(q) questions;

[0157] Step R2.3, Difficulty Screening:

[0158] According to the preset difficulty level of the questions, select the questions that meet the difficulty quantity requirements from the screened questions to form the question sets T1, T2, and T3, and they meet the following requirements:

[0159] ∣T1∣=N1, ∣T2∣=N2, ∣T3∣=N3;

[0160] Among them, |T1|, |T2|, and |T3| refer to the number of questions of each difficulty level in the question set;

[0161] Step R3, question combination optimization:

[0162] Step R3.1, similarity extraction:

[0163] For a set of questions of the same question type, extract the similarity SK between each question;

[0164] Step R3.2, threshold determination:

[0165] Extract the pre-set similarity threshold SKy between two questions in the same question type. When their similarity SK is greater than the similarity threshold SKy, they are judged as highly similar question pairs.

[0166] Then, a question is randomly selected from the remaining questions of the original screening criteria, that is, a question that meets the requirements of question structure, knowledge point coverage, and difficulty level but is not selected, and replaces any question in the highly similar question pair;

[0167] Then recalculate the similarity between the replaced question and other questions until the similarity SK of all question pairs is less than or equal to the similarity threshold SKy, and obtain the set of test questions;

[0168] Example 1 proposes an intelligent test paper composition method for medical theory examinations based on artificial intelligence technology. In the preprocessing of question information, text features are extracted through steps such as word segmentation, word frequency calculation, and inverse document frequency, which can accurately mine the key information of text-type questions and lay a solid foundation for subsequent question similarity calculation. In the question similarity calculation, the cosine similarity calculation method based on the keyword set converts the keyword set into a binary vector for calculation, which can intuitively and efficiently measure the similarity between text-type questions. In the process of intelligent test paper composition, question type screening, knowledge point screening, and difficulty screening are carried out in turn to ensure that the test paper composition meets multiple requirements. Finally, the question combination is optimized through similarity extraction and threshold judgment, which effectively reduces the question similarity, improves the quality of the test paper, and makes the test paper more reasonable in question type structure, knowledge point coverage and difficulty distribution, meeting the professional and scientific requirements of medical examinations.

[0169] As the second embodiment of the present invention:

[0170] See also Figure 1 and Figure 2 As shown, in the specific implementation of this application, compared with Example 1, the technical solution of this embodiment is different from that of Example 1 only in that:

[0171] The question to be replaced in a pair of highly similar questions is determined by its historical usage count in the question bank, and the question with the largest historical usage count is replaced;

[0172] Compared to Example 1, Example 2 replaces highly similar question pairs by replacing them instead of randomly selecting them. Instead, it determines the replacement based on the number of times a question has been used historically in the question bank, replacing the question with the one with the highest number of uses. This improvement makes the test paper compilation process more targeted and rational, prioritizing the replacement of frequently used questions. This helps improve the utilization rate of questions in the question bank, avoids the repetition of frequently used questions, and increases the freshness and diversity of questions on the test paper. It also allows for a more equitable assessment of candidates' mastery of different questions, further optimizing the test paper compilation process and enhancing the quality and practicality of the test paper.

[0173] As the third embodiment of the present invention:

[0174] See also Figure 1 and Figure 2 As shown, in the specific implementation of this application, compared with Example 1 and Example 2, the technical solution of this embodiment is to combine the solutions of Example 1 and Example 2, and the difference between the technical solution of this embodiment and Example 1 and Example 2 is only in this embodiment;

[0175] Step 1: Preprocessing of topic information:

[0176] For image-related questions in medical examinations, in this embodiment, image-related questions are such as pathological section image questions;

[0177] Directly extract the original image data and represent it with the image set X={X(i2, j2)};

[0178] Where i2=1, 2, ..., n2, n2 represents the number of image-type questions, j2=1, 2, ..., m2, m2 represents the number of images in the i-th image-type question; and X(i2, j2) represents the j2-th image in the i2-th image-type question;

[0179] For X = {X (i2, j2)} and XW = {X (i3, j b )}Use the local feature descriptor method to extract image features. The specific steps are as follows:

[0180] Step 1.3.1: Image segmentation:

[0181] Select an image and divide it into s local regions;

[0182] Step 1.3.2, color feature extraction:

[0183] Convert the image from RGB color space to HSV color space, and then calculate the color mean and variance of each local area:

[0184] Specifically include:

[0185] Mean HP(k) and variance HF(k) of hue H;

[0186] The mean SP(k) and variance SF(k) of saturation S;

[0187] The mean VP(k) and variance VF(k) of brightness V;

[0188] Where k = 1, 2, ... s;

[0189] In this embodiment, the calculation of the mean and variance of hue H, saturation S, and brightness V is a basic technology in the field of image processing. The mathematical principles behind these calculations are based on existing technologies. Each pixel contains H / S / V values, which are inherent properties of an image in the HSV color space. These values ​​can be obtained without requiring additional technical innovation and are common knowledge to those skilled in the art. Therefore, a detailed derivation is not provided.

[0190] Step 1.3.3, texture feature extraction:

[0191] Gray-level co-occurrence matrix is ​​used to calculate the texture parameters of each local area in the image;

[0192] Specifically include:

[0193] Energy E(k), entropy H(k), contrast C(k), correlation R(k);

[0194] Step 1.3.4, construct feature vector:

[0195] The color features and texture features extracted in the local area are used to form the feature vector of the local area:

[0196] That is, F(k)=[HP(k), HF(k), SP(k), SF(k), VP(k), VF(k), E(k), H(k), C(k), R(k)];

[0197] Then, the feature vectors corresponding to all local areas contained in the image are combined into an image feature vector set, that is, F=[F(k)]=[F1,F2,…Fs];

[0198] In this embodiment, texture feature extraction is a mature technology in the field of image processing and computer vision. Its core method: gray-level co-occurrence matrix and its derived texture parameter calculation are existing technologies. Its mathematical principles, parameter definitions, and application scenarios have been fully disclosed, so detailed derivation is not given here.

[0199] Step 2: Calculate the similarity of the title:

[0200] Image similarity calculation:

[0201] The cosine similarity calculation method based on the image feature vector set is as follows:

[0202] Select two image-related topics and extract their corresponding keyword sets Fa={Fa(k)} and Fb={Fb(k)} and convert them into binary vectors;

[0203] pass: , calculate the image similarity SF between the two image-type questions;

[0204] Among them, the larger the SF value, the higher the image similarity, and the value range is [0,1];

[0205] Step 3: Intelligent test paper generation:

[0206] Step R1: Extract the test paper generation requirements:

[0207] Extract the following parameters that are pre-set according to the purpose of the exam:

[0208] Question structure: the number of image-related questions Ny;

[0209] Knowledge point coverage: The knowledge point set S = {S(q)}, and each knowledge point S(q) is covered by at least e(q) questions, where e(q) is a preset value;

[0210] Difficulty level: Questions are divided into three levels: low, medium, and high. The number of questions in each level is N1, N2, and N3, respectively, satisfying N=N1+N2+N3;

[0211] Step R2, topic screening:

[0212] Step R2.1, Question Type Screening:

[0213] Select image classes that meet the quantity requirements from the pre-established question bank and form a question set Ty, which satisfies: |Ty|=Ny;

[0214] Among them, |Ty| refers to the number of image-related questions in the question set;

[0215] Step R2.2, knowledge point screening:

[0216] For each knowledge point S(q), select questions related to this knowledge point from Ty, and at the same time, the knowledge point S(q) must be covered by at least e(q) questions;

[0217] Step R2.3, Difficulty Screening:

[0218] According to the preset difficulty level of the questions, select the questions that meet the difficulty quantity requirements from the screened questions to form the question sets T1, T2, and T3, and they meet the following requirements:

[0219] ∣T1∣=N1, ∣T2∣=N2, ∣T3∣=N3;

[0220] Among them, |T1|, |T2|, and |T3| refer to the number of questions of each difficulty level in the question set;

[0221] Step R3, question combination optimization:

[0222] Step R3.1, similarity extraction:

[0223] For a set of questions of the same question type, extract the similarity SF between each question;

[0224] Step R3.2, threshold determination:

[0225] Extracting a pre-set similarity threshold SFy;

[0226] Between two questions in the same question type, if their similarity SF is greater than the similarity threshold SFy, they are judged as a highly similar question pair;

[0227] Then, a question is randomly selected from the remaining questions of the original screening criteria, that is, a question that meets the requirements of question structure, knowledge point coverage, and difficulty level but is not selected, and replaces any question in the highly similar question pair;

[0228] Then recalculate the similarity between the replaced question and other questions until the similarity SF of all question pairs is less than or equal to the similarity threshold SFy, and obtain the set of test questions;

[0229] Example 3 combines the solutions of Example 1 and Example 2. It not only effectively preprocesses information, calculates similarity, and intelligently generates test papers for text-based questions, but also targets image-based questions (such as those involving pathology slide images). This approach extracts image features through image segmentation, color feature extraction, and texture feature extraction. The similarity of image-based questions is then measured using the cosine similarity calculation method of a set of image feature vectors, making the test paper generation method applicable to image-based questions. During intelligent test paper generation, question type, knowledge point, and difficulty level are similarly screened, and question combinations are optimized. Furthermore, when replacing highly similar question pairs, the method of Example 2 is employed to determine replacement questions based on historical usage counts. This embodiment expands the scope of application of the test paper generation method, enabling it to handle both text-based and image-based questions. It also provides a more scientific approach to question replacement, further improving the comprehensiveness and rationality of the test paper generation process and meeting the requirements for generating test papers for different question types in medical examinations.

[0230] As the fourth embodiment of the present invention:

[0231] See also Figure 1 and Figure 2As shown, in the specific implementation of this application, compared with Example 1, Example 2 and Example 3, the difference between this embodiment and Example 1, Example 2 and Example 3 is that this embodiment also includes:

[0232] Step 1: Preprocessing of topic information:

[0233] For multimodal questions on case analysis in medical examinations;

[0234] In this embodiment, case analysis multimodal questions are such as questions on text and image fusion;

[0235] For case analysis multimodal questions: separate the text part and the image part, and pass the text part through WX={WX(i3,j a )}, the image part is represented by XW={X(i3,j b )}express;

[0236] Among them, i3=1, 2, ... n3, n3 represents the number of case analysis multimodal questions, j a =1, 2, ...m a , m a represents the number of characters in the i3th case analysis multimodal question; j b =1, 2, ...m b , m b represents the number of images in the i3th case analysis multimodal question; and X(i3,j a ) represents the jth case analysis multimodal question in the i3th case analysis a characters, X(i3,j b ) represents the jth case analysis multimodal question in the i3th case analysis b images;

[0237] For the text and image parts of the case analysis multimodal questions, the corresponding keyword set is extracted according to the text feature extraction method in Example 1, and the corresponding image feature vector set is extracted according to the image feature extraction method in Example 2;

[0238] Step 2: Calculate the similarity of the title:

[0239] Similarity calculation for multimodal questions related to case analysis:

[0240] Based on the text and image similarity calculation method, a weighted summation method is used. The steps are as follows:

[0241] Select two case analysis multimodal questions, calculate the similarity SK0 of the corresponding text parts of the two using the text similarity calculation method, and at the same time, calculate the similarity SF0 of the corresponding image parts of the two using the image similarity calculation method;

[0242] pass: , calculate the overall similarity SZ between two case analysis multimodal questions;

[0243] Where, α K and α F is the corresponding preset weight coefficient, and α K +α F =1;

[0244] In this embodiment, the weight coefficients of text and image can be adjusted according to actual needs;

[0245] Step 3: Intelligent test paper generation:

[0246] Step R1: Extract the test paper generation requirements:

[0247] Extract the following parameters that are pre-set according to the purpose of the exam:

[0248] Question structure: Number of case analysis questions Nz;

[0249] Knowledge point coverage: The knowledge point set S = {S(q)}, and each knowledge point S(q) is covered by at least e(q) questions, where e(q) is a preset value;

[0250] Difficulty level: Questions are divided into three levels: low, medium, and high. The number of questions in each level is N1, N2, and N3, respectively, satisfying N=N1+N2+N3;

[0251] Step R2, topic screening:

[0252] Step R2.1, Question Type Screening:

[0253] Select case analysis questions that meet the quantity requirements from the pre-established question bank and form a question set Tz, which satisfies: |Tz|=Nz;

[0254] Among them, |Tz| refers to the number of text, image, and case analysis questions in the question set;

[0255] Step R2.2, knowledge point screening:

[0256] For each knowledge point S(q), select questions related to this knowledge point from Tz, and at the same time, the knowledge point S(q) must be covered by at least e(q) questions;

[0257] Step R2.3, Difficulty Screening:

[0258] According to the preset difficulty level of the questions, select the questions that meet the difficulty quantity requirements from the screened questions to form the question sets T1, T2, and T3, and they meet the following requirements:

[0259] ∣T1∣=N1, ∣T2∣=N2, ∣T3∣=N3;

[0260] Among them, |T1|, |T2|, and |T3| refer to the number of questions of each difficulty level in the question set;

[0261] Step R3, question combination optimization:

[0262] Step R3.1, similarity extraction:

[0263] For a set of questions of the same question type, extract the similarity SZ between each question;

[0264] Step R3.2, threshold determination:

[0265] Extracting a preset similarity threshold SZy;

[0266] Between two questions in the same question type, if their similarity SZ is greater than the similarity threshold SZy, they are judged as a highly similar question pair;

[0267] Then, a question is randomly selected from the remaining questions of the original screening criteria, that is, a question that meets the requirements of question structure, knowledge point coverage, and difficulty level but is not selected, and replaces any question in the highly similar question pair;

[0268] Then recalculate the similarity between the replaced question and other questions until the similarity SZ of all question pairs is less than or equal to the similarity threshold SZy, and obtain the set of test questions;

[0269] In Example 4, for multimodal questions related to case analysis (such as questions that combine text and image), during the preprocessing of the question information, the text part and the image part are separated, and the text features and image features are extracted according to the methods of Example 1 and Example 2, respectively. In the question similarity calculation, a weighted summation method is used to combine the similarities of text and image to obtain the overall similarity, which can more comprehensively and accurately measure the similarity of multimodal questions. In the intelligent test paper compilation process, the question type, knowledge points and difficulty level are screened for case analysis questions, and the question combination is optimized through similarity extraction and threshold judgment. This embodiment solves the information processing and similarity calculation problems of multimodal questions in the test paper compilation process, enables the test paper compilation method to be applicable to more complex question types, enriches the question types of the test paper compilation, improves the test paper's ability to test actual medical application scenarios, and enables the test paper to better reflect the examinee's ability to comprehensively apply knowledge.

[0270] As the fifth embodiment of the present invention:

[0271] See also Figure 1 and Figure 2As shown, when the present application is implemented, compared with Example 1, Example 2, Example 3 and Example 4, the technical solution of this embodiment is to combine and implement the solutions of the above-mentioned Example 1, Example 2, Example 3 and Example 4.

[0272] Example 5 combines the solutions of Examples 1, 2, 3, and 4 for implementation, and has the ability to process multiple types of questions, such as text-based, image-based, and case analysis-based multimodal questions. In terms of question information preprocessing, corresponding feature extraction methods can be used for different types of questions; when calculating question similarity, appropriate calculation methods can be selected according to the question type, such as text similarity calculation, image similarity calculation, and multimodal question overall similarity calculation; in the intelligent test paper composition process, on the basis of meeting the requirements of question structure, knowledge point coverage, and difficulty level, the replacement of highly similar question pairs can be determined based on the number of historical uses, making the test paper composition process more scientific and reasonable. This embodiment combines the advantages of each embodiment, greatly expands the application scope of the test paper composition method, improves the flexibility and comprehensiveness of the test paper composition, and can meet the diverse test paper composition needs in medical examinations, generating high-quality test papers with more reasonable structure, richer content, and more comprehensive examinations.

[0273] It should be stated that all user data collected in this application is collected with the user's consent and authorization, and the use of user data is legal and compliant, and the use and processing of user data complies with the relevant laws, regulations and standards of the relevant regions.

[0274] The above formulas are all dimensionless and numerical calculations. The formulas are obtained by collecting a large amount of data and performing software simulation to obtain the most recent real situation. The preset parameters and thresholds in the formulas are set by technicians in this field according to actual conditions.

[0275] The above description is merely a specific embodiment of the present application, but the scope of protection of the present application is not limited thereto. Any changes or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in this application should be included in the scope of protection of this application. Therefore, the scope of protection of this application should be based on the scope of protection of the claims.

Claims

1. An intelligent test paper-generating method for medical theory examination based on artificial intelligence technology, characterized in that: The following steps are involved: Step 1: Preprocessing the question information: Break down the questions into text, image, and case analysis multimodal questions by question type, and extract features from them. Feature extraction includes: Text feature extraction: Segment the text of text and case analysis questions, calculate word frequency, inverse document frequency, and determine the keyword set; Image feature extraction: For image and case analysis questions, the image parts are divided into blocks, color and texture features are extracted, and a set of image feature vectors is constructed; Step 2: Calculate topic similarity: Based on the keyword set and feature vector set, calculate text similarity, image similarity, and overall multimodal topic similarity. Step 3: Intelligent test paper composition: Questions are selected based on the preset question structure, knowledge point coverage, and difficulty level requirements, and optimized and combined based on the calculated text similarity, image similarity, and overall similarity of multimodal questions. The optimization combination method is: for the same question type, the similarity between each question is extracted, namely, text similarity, image similarity, and overall similarity of multimodal questions. Then, highly similar question pairs are judged and replaced through the preset similarity threshold to obtain the test paper question set.

2. The method for intelligently composing medical theory examination papers based on artificial intelligence technology according to claim 1, characterized in that: For text questions: directly extract the text content and represent it with a text set W = {W (i1, j1)}; Where i1=1, 2, ..., n1, n1 represents the number of text-type questions, j1=1, 2, ..., m1, m1 represents the number of characters in the i1-th text-type question; and W(i1, j1) represents the j1-th character in the i1-th text-type question; For image-related questions: directly extract the original image data and represent it with the image set X = {X(i2, j2)}; Where i2=1, 2, ..., n2, n2 represents the number of image-type questions, j2=1, 2, ..., m2, m2 represents the number of images in the i-th image-type question; and X(i2, j2) represents the j2-th image in the i2-th image-type question; For case analysis multimodal questions: separate the text part and the image part, and pass the text part through WX={WX(i3,j a )}, the image part is represented by XW={X(i3,j b )}express; Among them, i3=1, 2, ... n3, n3 represents the number of case analysis multimodal questions, j a =1, 2, ...m a , m a represents the number of characters in the i3th case analysis multimodal question; j b =1, 2, ...m b , m b represents the number of images in the i3th case analysis multimodal question; and X(i3,j a ) represents the jth case analysis multimodal question in the i3th case analysis a characters, X(i3,j b ) represents the jth case analysis multimodal question in the i3th case analysis b images.

3. The method for intelligently composing medical theory examination papers based on artificial intelligence technology according to claim 2, characterized in that: The text feature extraction method is as follows: For W = {W (i1, j1)} and WX = {WX (i3, j a )}, use keyword extraction technology to obtain text features, the specific steps are as follows: Use word segmentation technology in natural language processing to split the text into independent words; Count the number of times each word appears in each question containing text content C(i, j); where i∈{i1,i3}, j∈{j1,j a }; pass , calculate the frequency TF(i, j) of the jth word in the i-th question containing text content; where m∈{m1,m a Then, select each word one by one and count the number of documents containing each word in the title containing text content, and record it as D(j); Then through , calculate the inverse document frequency IDF (j) of each word; then pass , calculate the term frequency - inverse document frequency TF-IDF (i, j) of each word; then select words whose TF-IDF (i, j) is higher than the corresponding preset TF-IDF threshold to form the keyword set K.

4. The method for intelligently composing medical theory examination papers based on artificial intelligence technology according to claim 3, characterized in that: Image feature extraction is as follows: For X = {X (i2, j2)} and XW = {X (i3, j b )}Use the local feature descriptor method to extract image features. The specific steps are as follows: Select an image and divide it into s local regions; Convert the image from RGB color space to HSV color space, and then calculate the color mean and variance of each local area: Gray-level co-occurrence matrix is ​​used to calculate the texture parameters of each local area in the image; The color features and texture features extracted in the local area are used to form the feature vector of the local area: Then, the feature vectors corresponding to all local areas contained in the image are combined into an image feature vector set.

5. The method for intelligently composing medical theory examination papers based on artificial intelligence technology according to claim 4 is characterized in that: The color mean and variance of each local area include: the mean HP(k) and variance HF(k) of hue H; the mean SP(k) and variance SF(k) of saturation S; the mean VP(k) and variance VF(k) of brightness V; where k = 1, 2, ... s; The texture parameters of each local area include: energy E(k), entropy H(k), contrast C(k), and correlation R(k); The eigenvector of the local area is F(k) = [HP(k), HF(k), SP(k), SF(k), VP(k), VF(k), E(k), H(k), C(k), R(k)]; The set of image feature vectors is F=[F(k)]=[F1,F2,…Fs].

6. The method for intelligently composing medical theory examination papers based on artificial intelligence technology according to claim 4, characterized in that: The text similarity is calculated as follows: The cosine similarity calculation method based on the keyword set is as follows: Select two text-type topics and extract their corresponding keyword sets Ka={Ka(g1)} and Kb={Ka(g2)} and convert them into binary vectors; Where g2=1, 2, ..., v1, v1 represents the number of all keywords in the keyword set Ka, g2=1, 2, ..., v2; v2 represents the number of all keywords in the keyword set Kb; The length of the binary vector is the number r of all non-repeated keywords in the two texts, where the occurrence of a keyword is marked as 1 and the absence of a keyword is marked as 0. Get binary vectors: Va=[Va(f)] and Vb=[Vb(f)]; Where f = 1, 2, ..., r; pass: , calculate the text similarity SK between two text-type topics.

7. The method for intelligently composing medical theory examination papers based on artificial intelligence technology according to claim 6, characterized in that: Image similarity is calculated as follows: The cosine similarity calculation method based on the image feature vector set is as follows: Select two image-related topics and extract their corresponding keyword sets Fa={Fa(k)} and Fb={Fb(k)} and convert them into binary vectors; pass: , calculate the image similarity SF between the two image class topics.

8. The method for intelligently composing medical theory examination papers based on artificial intelligence technology according to claim 7, characterized in that: The similarity calculation method for multimodal questions in case analysis is as follows: Based on the text and image similarity calculation method, a weighted summation method is used. The steps are as follows: Select two case analysis multimodal questions, calculate the similarity SK0 of the corresponding text parts of the two using the text similarity calculation method, and at the same time, calculate the similarity SF0 of the corresponding image parts of the two using the image similarity calculation method; pass: , calculate the overall similarity SZ between two case analysis multimodal questions; Where, α K and α F is the corresponding preset weight coefficient, and α K +α F =1.

9. The method for intelligently composing medical theory examination papers based on artificial intelligence technology according to claim 8, characterized in that: The intelligent test paper grouping method is as follows: Step R1: Extract the test paper generation requirements: Extract the following parameters that are pre-set according to the purpose of the exam: Question structure: The total number of questions is N, and the question type distribution is Nx for text questions, Ny for image questions, and Nz for case analysis questions, and it satisfies N=Nx+Ny+Nz; Knowledge point coverage: The knowledge point set S = {S(q)}, and each knowledge point S(q) is covered by at least e(q) questions, where e(q) is a preset value; Difficulty level: Questions are divided into three levels: low, medium, and high. The number of questions in each level is N1, N2, and N3, respectively, satisfying N=N1+N2+N3; Step R2, topic screening: Question type screening: Select text-based, image-based, and case analysis-based questions that meet the quantity requirements from the pre-established question bank and form a question set Tx, Ty, and Tz that satisfies the following conditions: |Tx|=Nx, |Ty|=Ny, |Tz|=Nz, where |Tx|, |Ty|, and |Tz| represent the number of text-based, image-based, and case analysis-based questions in the question set, respectively. Knowledge point screening: For each knowledge point S(q), select questions related to this knowledge point from Tx, Ty, and Tz, and ensure that the knowledge point S(q) is covered by at least e(q) questions; Difficulty screening: Based on the preset difficulty levels of the questions, select questions that meet the difficulty quantity requirements from the screened questions to form question sets T1, T2, and T3, and they satisfy: |T1|=N1, |T2|=N2, |T3|=N3, where |T1|, |T2|, and |T3| represent the number of questions of each difficulty level in the question set respectively; Step R3, question combination optimization: Similarity extraction: For a set of questions of the same question type, extract the similarity S0 between each question, where S0∈{SK, SF, SZ}; Threshold judgment: extract the pre-set similarity threshold S0y, where S0y∈{SKy, SFy, SZy}; SKy refers to the text threshold, SFy refers to the image threshold, and SZy refers to the multimodal threshold; Between two questions in the same question type, when their similarity S0 is greater than the similarity threshold S0y, they are determined to be a highly similar question pair; Then, a question is randomly selected from the remaining questions of the original screening criteria, that is, a question that meets the requirements of question structure, knowledge point coverage, and difficulty level but is not selected, and replaces the question in the highly similar question pair; Then, the similarity between the replaced question and other questions is recalculated until the similarity S0 of all question pairs is less than or equal to the similarity threshold S0y, and the set of test questions is obtained.

10. The method for intelligently composing medical theory examination papers based on artificial intelligence technology according to claim 9, characterized in that: in, The question to be replaced in a pair of highly similar questions is determined by its historical usage count in the question bank, and specifically the question with the largest historical usage count is replaced.

Images