Test question difficulty prediction method and device, storage medium and computer equipment

Abstract

The invention discloses a test question difficulty prediction method and device, a storage medium and computer equipment, relates to the technical field of intelligent education, and mainly aims to improve test question difficulty prediction precision and prediction efficiency and save prediction cost. The method comprises the following steps: collecting video stream data when a user answers a target test question; based on the video stream data, determining writing process data when the user answers the target test question; based on the writing process data, determining a difficulty coefficient of the target test question for the user; and based on the difficulty coefficient, determining the difficulty of the target test question for the user. The method is suitable for a difficulty prediction scene of the test question to the user.

Description

Technical Field

[0001] The present invention relates to the field of intelligent education technology, and in particular to a method, device, storage medium and computer equipment for predicting the difficulty of test questions. Background Art

[0002] Understanding the difficulty of test questions helps achieve personalized learning. By analyzing students' performance on test questions of varying difficulty levels, teachers can better understand students' learning needs and weaknesses, thereby developing more targeted teaching strategies and providing personalized learning resources.

[0003] Currently, users are usually asked to use a pressure-sensitive stylus to answer questions, and the difficulty of the questions is predicted by the writing pressure reflected by the pressure-sensitive stylus. However, this method requires each user to be equipped with a pressure-sensitive stylus, and the stylus needs to be calibrated and maintained before the user writes, which wastes time and cost. In addition, the pressure-sensitive stylus can only reflect the user's writing pressure, which is affected by many factors such as the user's personal writing habits and psychological state. Therefore, this method of predicting the difficulty of the test questions based on writing pressure results in low prediction accuracy of the test question difficulty. Summary of the Invention

[0004] The present invention provides a method, device, storage medium and computer equipment for predicting the difficulty of test questions, which are mainly capable of improving the prediction accuracy and preset efficiency of the test question difficulty and saving prediction costs.

[0005] According to a first aspect of the present invention, there is provided a method for predicting the difficulty of a test question, comprising:

[0006] Collect video stream data when users answer target test questions;

[0007] Determining, based on the video stream data, writing process data of the user when answering the target test question;

[0008] Determining a difficulty coefficient of the target test question for the user based on the writing process data;

[0009] Based on the difficulty coefficient, the difficulty of the target test question for the user is determined.

[0010] Optionally, determining the difficulty coefficient of the target test question for the user based on the writing process data includes:

[0011] determining at least one data feature score based on the writing process data;

[0012] The score weight coefficient corresponding to each of the data feature scores is determined respectively, and based on each of the data feature scores and the corresponding score weight coefficient, the difficulty coefficient of the target test question for the user is determined.

[0013] Optionally, respectively determining the score weight coefficient corresponding to each of the data feature scores includes:

[0014] Determine multiple weighted assessment terminals;

[0015] sending weight evaluation criterion information including each of the data feature scores as reference information to each of the weight evaluation terminals, so that each of the weight evaluation terminals performs a terminal zero interaction score based on the reference information to obtain a zero interaction score result;

[0016] receiving the zero-interaction scoring result, and sending the zero-interaction scoring result to each of the weight evaluation terminals, respectively, so that each of the weight evaluation terminals performs terminal interaction scoring based on the zero-interaction scoring result and the reference information to obtain an interaction scoring result;

[0017] The interactive scoring result is received, and based on the interactive scoring result, a score weight coefficient corresponding to each of the data feature scores is determined.

[0018] Optionally, determining the difficulty coefficient of the target test question for the user based on the writing process data includes:

[0019] Based on the video stream data, acquiring facial expression data of the user when answering the target test question;

[0020] The expression data and the writing process data are input into a large language model to perform difficulty prediction, and a difficulty coefficient of the target test question for the user is obtained.

[0021] Optionally, inputting the expression data and the writing process data into a large language model to perform difficulty prediction to obtain a difficulty coefficient of the target test question for the user includes:

[0022] respectively determining a writing feature vector corresponding to the writing process data and an expression feature vector corresponding to the expression data;

[0023] Performing feature alignment on the writing feature vector and the expression feature vector, and splicing the aligned writing feature vector and the aligned expression feature vector to obtain a spliced ​​feature vector;

[0024] The concatenated feature vector is input into the large language model to perform difficulty prediction, and a difficulty coefficient of the target test question for the user is obtained.

[0025] Optionally, aligning the writing feature vector and the expression feature vector includes:

[0026] Constructing a fusion feature matrix based on the writing feature vector and the expression feature vector, and determining a covariance matrix corresponding to the fusion feature matrix;

[0027] Performing eigenvalue decomposition on the covariance matrix to obtain eigenvalues ​​and eigenvectors corresponding to the eigenvalues;

[0028] Selecting a principal component eigenvector whose eigenvalue is greater than a preset threshold value from each of the eigenvectors, and constructing a transformation matrix based on each of the principal component eigenvectors;

[0029] Based on the transformation matrix, linear transformation is performed on the writing feature vector and the expression feature vector respectively to obtain the aligned writing feature vector and the aligned expression feature vector.

[0030] Optionally, the step of splicing the aligned writing feature vector and the aligned expression feature vector to obtain a spliced ​​feature vector comprises:

[0031] Determine a feature splicing order corresponding to the aligned handwriting feature vector and the aligned expression feature vector;

[0032] Based on the feature splicing order, the aligned writing feature vector and the aligned expression feature vector are horizontally spliced ​​to obtain a spliced ​​feature vector.

[0033] Optionally, determining the writing process data of the user when answering the target test question based on the video stream data includes:

[0034] intercepting a video frame containing the user's writing process from the video stream data, and performing text recognition on the video frame to obtain text content;

[0035] Performing semantic analysis on the text content, and identifying the answer content for the target test question in the text content based on the semantic analysis result;

[0036] Based on the video stream data, the text content, and the answer content, the writing process data of the user when answering the target test question is determined.

[0037] Optionally, determining the difficulty of the target test question for the user based on the difficulty coefficient includes:

[0038] The coefficient interval to which the difficulty coefficient belongs is determined, and the difficulty corresponding to the coefficient interval is determined as the difficulty of the target test question for the user, wherein different coefficient intervals correspond to different test question difficulties.

[0039] According to a second aspect of the present invention, there is provided a device for predicting the difficulty of a test question, comprising:

[0040] A collection unit, used to collect video stream data when the user answers the target test questions;

[0041] a data determining unit, configured to determine, based on the video stream data, writing process data of the user when answering the target test question;

[0042] a coefficient determination unit, configured to determine a difficulty coefficient of the target test question for the user based on the writing process data;

[0043] A difficulty determination unit is configured to determine the difficulty of the target test question for the user based on the difficulty coefficient.

[0044] According to a third aspect of the present invention, there is provided a computer-readable storage medium having a computer program stored thereon, which implements the above method for predicting the difficulty of test questions when executed by a processor.

[0045] According to a fourth aspect of the present invention, there is provided a computer device comprising a memory, a processor and a computer program stored in the memory and executable on the processor, wherein the processor implements the above method for predicting the difficulty of test questions when executing the program.

[0046] According to the present invention, a method, device, storage medium and computer equipment for predicting the difficulty of a test question are provided. Compared with the current method of having users use a pressure-sensitive pen to answer questions and predicting the difficulty of the test question through the writing pressure reflected by the pressure-sensitive pen, the present invention collects video stream data of the user answering the target test question; and based on the video stream data, determines the writing process data of the user answering the target test question; then, based on the writing process data, determines the difficulty coefficient of the target test question for the user; and finally, based on the difficulty coefficient, determines the difficulty of the target test question for the user. In this way, the difficulty of the test question is predicted by the video stream data of the user answering the target test question. All users can share a video stream acquisition device, saving costs. At the same time, there is no need to calibrate the video stream acquisition device in advance, thereby improving prediction efficiency. In addition, the video stream data can capture all-round data of the user during the writing process. These writing process data help to more comprehensively understand the user's problem-solving status and thinking process and other behaviors, so that the difficulty of the test question can be predicted more accurately. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] The drawings described herein are used to provide a further understanding of the present invention and constitute a part of this application. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute an improper limitation of the present invention. In the drawings:

[0048] Figure 1A flow chart of a method for predicting the difficulty of a test question provided by an embodiment of the present invention is shown;

[0049] Figure 2 A flow chart of another method for predicting the difficulty of a test question provided by an embodiment of the present invention is shown;

[0050] Figure 3 A schematic diagram showing the structure of a device for predicting the difficulty of a test question provided by an embodiment of the present invention is shown;

[0051] Figure 4 A schematic diagram showing the structure of another device for predicting the difficulty of test questions provided by an embodiment of the present invention is shown;

[0052] Figure 5 A schematic diagram of the physical structure of a computer device provided by an embodiment of the present invention is shown. DETAILED DESCRIPTION

[0053] The present invention will be described in detail below with reference to the accompanying drawings and in combination with embodiments. It should be noted that, unless there is a conflict, the embodiments and features in the embodiments of the present application can be combined with each other.

[0054] Currently, having users use a pressure-sensitive stylus to answer questions and then using the pressure they write to predict the difficulty of each question is expensive because it requires each user to have a pressure-sensitive stylus. Furthermore, the pressure they write with is affected by various factors, such as the user's writing habits and psychological state, which can affect the accuracy of the predicted difficulty.

[0055] In order to solve the above problems, the embodiment of the present invention provides a method for predicting the difficulty of test questions, such as Figure 1 As shown, the method includes:

[0056] 101. Collect video stream data of users answering target test questions.

[0057] Target questions can be from any domain, such as science (including mathematics, physics, chemistry, biology, languages, geography, history, etc.), professional fields (including computers, medicine, law, economics, etc.), professional skills (including accounting, marketing, human resources management, etc.), general skills (including reasoning and analysis, etc.), etc. Target questions can be any of a variety of question types, including short-answer questions, fill-in-the-blank questions, essay questions, and calculation questions.

[0058] In the embodiment of the present invention, in order to predict the difficulty of the target test question for the user, a video acquisition device is pre-installed in an area where multiple users are present, or in order to improve the acquisition effect, each adjacent user within a preset range shares a video acquisition device (wherein the video acquisition device can be a camera or other device). Later, when the user answers the target test question, the video acquisition device is started to collect the video stream data of the user answering the target test question, and then the difficulty of the target test question for the user is analyzed by the video stream data. In this way, the difficulty of the test question can be predicted by a small number of video acquisition devices, avoiding the hardware cost of equipping each user with a pressure-sensitive pen containing a high-precision sensor, and also avoiding the time wasted by constantly calibrating and maintaining the pressure-sensitive pen. Thus, the embodiment of the present invention saves prediction cost and prediction time when predicting the difficulty of the entity.

[0059] 102. Based on the video stream data, determine the writing process data of the user when answering the target test questions.

[0060] The writing process data includes at least one of the following: answering time data, answering behavior data, answering area data, answering results data, etc. Answering time data includes: total answering time for the target test question, benchmark answering time (a value preset according to actual needs), total pause time, etc.; answering behavior data includes: number of erasures and modifications, benchmark pause times (a value preset according to actual needs), number of answering pauses, etc.; answering area data includes: actual answering area, total available answering area, total writing distance, etc.; answering results data includes: answer score, total test score, etc.

[0061] For the embodiment of the present invention, after collecting the video stream data of the time when the user answers the target test question, it is necessary to extract the writing process data of the user for the target test question from the video stream data. Based on this, step 102 specifically includes: intercepting the video frame containing the user's writing process in the video stream data, and performing text recognition on the video frame to obtain text content; performing semantic analysis on the text content, and based on the semantic analysis result, identifying the answer content for the target test question in the text content; based on the video stream data, the text content, and the answer content, determining the writing process data of the user when answering the target test question.

[0062] Specifically, the video stream data is processed to extract image data from each frame. Using techniques such as optical character recognition, the text information within each frame is extracted to form editable text content. By analyzing changes in the text content, as well as the user's hand movements and eye expressions in the video stream, the user's pauses and the number of pauses can be detected. For example, if the recognized text content remains unchanged for a period of time, and the user's hand movements in the video stream slow down or stop, the user can be considered to be in a pause state. Total pause duration: The total pause duration is calculated by summing the durations of all detected pauses. Number of answer pauses: The number of pauses directly related to the answer is counted among the detected pauses. Number of erasures and modifications: The number of erasures and modifications can be calculated by analyzing changes in the text content, particularly detecting the deletion and re-entry of text data. The actual answer area is determined by analyzing the user's handwriting in the video stream. For example, image processing techniques can be used to identify the pixel area of ​​the user's handwriting and calculate its area. Total available answer area: This is the preset answer area, which can be determined based on the answer requirements of the test question. Total Writing Distance: Image processing and motion trajectory analysis techniques are used to analyze the user's hand movement trajectory as they write in the video stream to calculate the total writing distance. Answer Score: The answer score is determined based on the user's answer content and the test question's scoring criteria. For example, natural language processing and other technologies are used to semantically understand and analyze the user's answer content, and then compare it with the reference answer to calculate the answer score. The total score for the test question is a preset value, determined based on the difficulty level of the test question and the scoring criteria. The start and end times of the video stream data are analyzed to determine the total time from the start to the end of the user's answer. Statistics and analysis are combined with a large amount of historical data to determine a reasonable benchmark answer time. In this way, the user's answer time data, answer behavior data, answer area data, answer results data, and other data when answering the target test question can be determined according to the above data. Ultimately, the difficulty of the target test question for the user is determined based on this data. The embodiment of the present invention can determine the difficulty of the target test question for the user based on only a simple device, a video acquisition device, and can avoid the hardware cost wasted when each user is equipped with a pressure-sensitive pen to perform a writing pressure test. Therefore, the embodiment of the present invention can solve the prediction cost of the test question difficulty.

[0063] 103. Based on the writing process data, determine the difficulty coefficient of the target test questions for the user.

[0064] For the embodiment of the present invention, in order to improve the prediction efficiency and accuracy of the difficulty coefficient, a pre-built large language model can be used to determine the difficulty coefficient of the target test question for the user. Based on this, step 103 specifically includes: based on the video stream data, obtaining the facial expression data of the user when answering the target test question; inputting the facial expression data and the writing process data into the large language model for difficulty prediction, and obtaining the difficulty coefficient of the target test question for the user. Among them, the method of inputting the facial expression data and the writing process data into the large language model for difficulty prediction includes: respectively determining the writing feature vector corresponding to the writing process data and the facial expression feature vector corresponding to the facial expression data; aligning the writing feature vector and the facial expression feature vector, and splicing the aligned writing feature vector and the aligned facial expression feature vector to obtain a spliced ​​feature vector; inputting the spliced ​​feature vector into the large language model for difficulty prediction, and obtaining the difficulty coefficient of the target test question for the user.

[0065] Among them, expression data includes: emotional data of users when answering target test questions, such as happiness, sadness, anger, surprise, fear, disgust, etc., as well as the intensity, duration, and change frequency of the above emotions. At the same time, emotion data can also include facial feature point data of users, such as the position, movement trajectory and morphological change data of key facial feature points such as eyes, mouth, nose, eyebrows, etc., and the above data can be analyzed to obtain information such as the intensity, duration, and change frequency of each emotion. Specifically, facial expression recognition technology based on deep learning algorithms (such as convolutional neural networks) can be used to collect expression data. The algorithm judges the expression state by identifying changes in key facial feature points, such as confusion (frown, confused eyes, etc.), confidence (smiling, firm eyes, etc.), and anxiety (biting lips, frequent blinking, etc.).

[0066] Specifically, word embedding and other methods are used to determine the writing feature vector corresponding to the writing process data and the expression feature vector corresponding to the expression data, and then the writing feature vector and the expression feature vector are feature aligned in the following manner: the writing feature vector and the expression feature vector are respectively standardized to obtain the standardized writing feature vector and the standardized expression feature vector; a fusion feature matrix is ​​constructed based on the standardized writing feature vector and the standardized expression feature vector; the covariance matrix corresponding to the fusion feature matrix is ​​determined; the covariance matrix is ​​decomposed into eigenvalues ​​to obtain eigenvalues ​​and eigenvectors corresponding to the eigenvalues; the principal component eigenvector whose eigenvalue is greater than a preset threshold is selected from each of the eigenvectors, and a transformation matrix is ​​constructed from each of the principal component eigenvectors; based on the transformation matrix, the writing feature vector and the expression feature vector are respectively linearly transformed to obtain the aligned writing feature vector and the aligned expression feature vector. Among them, since the writing process data includes at least one of the answer time data, answer behavior data, answer area data, answer result data, etc., the writing feature vector should include at least one of the time feature vector corresponding to the answer time data, the behavior feature vector corresponding to the answer behavior data, the area feature vector corresponding to the answer area data, the result feature vector corresponding to the answer result data, etc.

[0067] Specifically, since the dimensions and value ranges of each eigenvector may differ, all eigenvectors need to be normalized before feature alignment, so that each feature has a mean of 0 and a variance of 1. Standardization helps eliminate dimensional differences between different features, ensuring that they contribute equally to the alignment process. After feature normalization, the covariance matrix between the eigenvectors needs to be calculated. For the handwriting eigenvectors and expression eigenvectors, they can be combined into a large data matrix, and the covariance matrix of this data matrix is ​​then calculated. The covariance matrix is ​​subjected to eigenvalue decomposition to obtain eigenvalues ​​and corresponding eigenvectors. The eigenvalues ​​reflect the variance of each principal component, while the eigenvectors indicate the direction of each principal component. Based on the eigenvalues, the eigenvectors corresponding to the eigenvalues ​​greater than a preset threshold (the preset threshold is set according to actual needs) or the eigenvectors corresponding to the first k (the k value is set according to actual needs) eigenvalues ​​are selected as the principal component eigenvectors. A transformation matrix is ​​then formed from the principal component eigenvectors. Each eigenvector is then multiplied by the transformation matrix to obtain the aligned handwriting eigenvectors and aligned expression eigenvectors. By converting all eigenvectors into the same coordinate system (i.e., principal component space) in this way, the correlation between eigenvectors is reduced or eliminated in the new coordinate system, making subsequent analysis and processing easier, and thus obtaining a higher accuracy difficulty coefficient.

[0068] Furthermore, the aligned writing feature vector and the aligned expression feature vector are spliced. Based on this, the method includes: determining a feature splicing order corresponding to the aligned writing feature vector and the aligned expression feature vector; based on the feature splicing order, horizontally splicing the aligned writing feature vector and the aligned expression feature vector to obtain a spliced ​​feature vector.

[0069] Specifically, the feature splicing order can be determined according to actual needs. For example, if the writing feature vector is in front and the expression feature vector is in the back, the aligned writing feature vector and the aligned expression feature vector are horizontally spliced ​​to obtain the spliced ​​feature vector. For example, if the aligned writing feature vector is V s , the expression feature vector after alignment is V ex , then the splicing feature vector V obtained after horizontal splicing is c =[V s ,V ex ]. If the written feature vector includes the aligned time feature vector V t , aligned behavioral feature vector V qw , aligned regional feature vector V be , the resulting feature vector V after alignment qa , and the aligned expression feature vector is V ex , then the splicing feature vector V obtained after horizontal splicing is c =[V t ,V qw ,V be ,V qa ,V ex ]. If the written feature vector includes the aligned time feature vector V t , aligned behavioral feature vector V qw , and the aligned expression feature vector is V ex , then the splicing feature vector V obtained after horizontal splicing is c =[V t ,V qw ,V ex ].

[0070] In another embodiment of the present invention, feature splicing can also be performed in the following manner: the aligned writing feature vector and the aligned expression feature vector are subjected to feature-level splicing processing to obtain a feature splicing vector; the aligned writing feature vector and the aligned expression feature vector are subjected to element-level splicing processing to obtain an element splicing vector; the aligned writing feature vector and the aligned expression feature vector are subjected to low-order splicing processing to obtain a low-order splicing vector; the feature splicing vector, the element splicing vector and the low-order splicing vector are transformed using a preset transformation function to obtain a splicing feature vector.

[0071] Specifically, in order to make full use of the relationship between data, extract more implicit features, and take into account both high-order and low-order processing, so that data utilization is more sufficient, the subsequent prediction results are more accurate, and meet the needs of actual application scenarios, it is necessary to splice the aligned writing feature vectors and the aligned expression feature vectors. The specific splicing processing method is as follows: if the aligned writing feature vector is (a1, a2) and the aligned expression feature vector is (b1, b2), the specific splicing processing method includes: performing feature-level splicing between different feature vectors, that is, performing Hadamard product on all elements between vectors, and then performing convolution transformation under certain weights to obtain a feature splicing vector of f(w*( a1*b1,a2*b2)); at the same time, all feature vector data are spliced ​​at the element level, that is, after making a Hadamard product for each element between the vectors, different weight values ​​are assigned to the results of each product, and then a linear transformation is performed, and the obtained element splicing vector is f(w1*a1*b1,w2*a2*b2); in addition, all feature vectors are low-order spliced, and the results after splicing are assigned weight coefficients, and then a linear transformation is performed, and the obtained low-order splicing vector is f(w2(a1,a2,b1,b2)); finally, the above feature splicing vectors, element splicing vectors, and low-order splicing vectors are transformed using the preset transformation function to obtain the spliced ​​feature vector.In another embodiment of the present invention, if the writing process data includes: answer time data, answer behavior data, answer area data, and answer result data, then the writing feature vector includes: a time feature vector corresponding to the answer time data, a behavior feature vector corresponding to the answer behavior data, a region feature vector corresponding to the answer area data, and a result feature vector corresponding to the answer result data; if the aligned time feature vector is (a1, a2), the aligned behavior feature vector is (b1, b2), the aligned region feature vector is (c1, c2), the aligned result feature vector is (d1, d2), and the aligned expression feature vector is (e1, e2), the specific splicing processing method includes: performing feature-level splicing between different feature vectors, that is, performing a Hadamard product on all elements between the vectors, and then performing a convolution transformation under a certain weight to obtain a feature splicing vector of f(w*(a1*b1*c1*d1*e 1,a2*b2*c2*d2*e2,a3*b3*c3*d3*e3)); at the same time, all feature vector data are spliced ​​at the element level, that is, after doing the Hadamard product for each element between the vectors, different weight values ​​are assigned to the results of each product, and then a linear transformation is performed to obtain the element splicing vector f(w1*a1*b1*c1*d1*e1,w2*a2*b2*c2*d2*e2,w3*a3*b3*c3*d3*e3); in addition, all feature vectors are low-order splicing processed, and the results after splicing are assigned weight coefficients, and then a linear transformation is performed to obtain the low-order splicing vector f(w2(a1,a2,b1,b2,c1,c2,d1,d2,e1,e2)))); finally, the above feature splicing vectors, element splicing vectors, and low-order splicing vectors are transformed using the preset transformation function to obtain the splicing feature vector. Among them, the preset transformation function is actually set according to actual needs, and the embodiment of the present invention does not limit it here. It should be noted that the above examples are only illustrative and do not limit the embodiments of the present application. Therefore, by splicing the aligned writing feature vectors and the aligned expression feature vectors, different features can be automatically or explicitly combined to generate new feature combinations. These combined features may contain complex nonlinear relationships between the original features, so that the model can capture more detailed and rich information in the data, that is, it can make full use of the relationship between various data, extract more implicit features, and take into account high-order and low-order processing at the same time, so that data utilization is more sufficient, and the difficulty coefficient prediction results obtained later are more accurate, meeting the needs of actual application scenarios. Finally, the spliced ​​feature vector is input into the large language model, and the difficulty coefficient of the target test question for the user can be output through the large language model. In another embodiment of the present invention, the determination of each feature vector can be implemented based on the feature extraction module in the large language model, and the alignment between each feature can be implemented by the encoder and multi-layer perceptron in the feature alignment module.By incorporating facial expression data and comprehensively analyzing the writing process and facial expression data to determine the difficulty coefficient, the present invention improves the accuracy of difficulty coefficient prediction, thereby increasing the accuracy of predicting the difficulty of the target test question for the user. Furthermore, using a large language model to predict the difficulty coefficient further improves the efficiency and accuracy of difficulty coefficient prediction, thereby increasing the efficiency and accuracy of predicting the difficulty of the test question itself.

[0072] 104. Based on the difficulty coefficient, determine the difficulty of the target test questions for the user.

[0073] For the embodiment of the present invention, after determining the difficulty coefficient, it is necessary to determine the difficulty of the target test question based on the difficulty coefficient. Based on this, step 104 specifically includes: determining the coefficient interval to which the difficulty coefficient belongs, and determining the difficulty corresponding to the coefficient interval as the difficulty of the target test question for the user, wherein different coefficient intervals correspond to different test question difficulties.

[0074] Specifically, for example, if the difficulty coefficient is W di , when 0<W di When 0.3<W di When W is less than a2, the difficulty level of the target question for the user is determined to be medium, that is, the target question is of medium difficulty for the user; when W di＞ At a2, it is determined that the difficulty level of the target question for the user is high, that is, the target question is relatively difficult for the user. It should be noted that the above examples are only illustrative and do not specifically limit the embodiments of the present invention.

[0075] According to a method for predicting the difficulty of a test question provided by the present invention, compared with the current method of having users use a pressure-sensitive pen to answer questions and predicting the difficulty of the test question through the writing pressure reflected by the pressure-sensitive pen, the present invention collects video stream data when the user answers the target test question; and based on the video stream data, determines the writing process data when the user answers the target test question; then, based on the writing process data, determines the difficulty coefficient of the target test question for the user; and finally, based on the difficulty coefficient, determines the difficulty of the target test question for the user. In this way, the difficulty of the test question is predicted by the video stream data when the user answers the target test question. All users can share a video stream acquisition device, saving costs. At the same time, there is no need to calibrate the video stream acquisition device in advance, thereby improving prediction efficiency. In addition, the video stream data can capture all-round data of the user during the writing process. These writing process data help to more comprehensively understand the user's problem-solving status and thinking process and other behaviors, so that the difficulty of the test question can be predicted more accurately.

[0076] Furthermore, in order to better illustrate the above process of predicting the difficulty of the test questions, as a refinement and extension of the above embodiment, the embodiment of the present invention provides another method for predicting the difficulty of the test questions, such as Figure 2 As shown, the method includes:

[0077] 201. Collect video stream data of users answering target test questions.

[0078] 202. Based on the video stream data, determine the writing process data of the user when answering the target test question.

[0079] Specifically, a video capture device is used to capture video stream data of the user answering a target question, and character recognition methods and natural language processing technologies are used to extract writing process data of the user answering the target question from the video stream data. For example, at least one writing process data is extracted, including answering time data, answering behavior data, answering area data, answering result data, etc.

[0080] 203. Determine at least one data feature score based on the writing process data.

[0081] The writing process data includes at least one of the following data: the user's answer time data for the target test question, answer behavior data, answer area data, and answer outcome data. Answer time data includes: total answer time, benchmark answer time, and total pause time. Answer behavior data includes: number of erasures and revisions, benchmark pauses, and answer pauses. Answer area data includes: actual answer area, total available answer area, and total writing distance. Answer outcome data includes: answer score and total test score. Data feature scores include at least one of: time feature score, behavior feature score, area feature score, and outcome feature score.

[0082] Specifically, if the time feature score is to be determined, it is necessary to determine the time feature score of the user when answering the target test question based on the total answering time, the benchmark answering time, the total pause time, the number of pauses, and the total writing distance. The specific method for determining the time feature score includes: dividing the total answering time by the benchmark answering time to obtain the time trend force; dividing the total pause time by the number of pauses to obtain the average pause time; dividing the total writing distance by the total answering time to obtain the average writing speed; respectively determining the time weight coefficients corresponding to the time trend force, the average pause time, and the average writing speed, and based on the time weight coefficients, adding the time trend force, the average pause time, and the average writing speed to obtain the time feature score of the user when answering the target test question.

[0083] Specifically, the time feature score w is calculated according to the following formula:st :

[0084] W st =ω1*T v +ω2*A p +ω3*V s

[0085] Among them, the time trend force Average pause time Average writing speed T t BT is the total time of answering. t The benchmark answer time, T p is the total pause time, N p is the total number of pauses, w d is the total writing distance, ω1 is the time weight coefficient corresponding to the time trend force, ω2 is the time weight coefficient corresponding to the average pause duration, and ω3 is the time weight coefficient corresponding to the average writing speed.

[0086] Furthermore, to determine the behavioral feature score, it is necessary to determine the behavioral feature score of the user when answering the target test question based on the number of pauses in answering, the total answering time, the baseline number of pauses, and the number of erasures and modifications. The specific method for determining the behavioral feature score includes: dividing the number of pauses in answering by the total answering time to obtain the pause frequency; dividing the number of pauses in answering by the baseline number of pauses to obtain the pause tendency force; respectively determining the behavioral weight coefficients corresponding to the number of erasures and modifications, the pause frequency, and the pause tendency force, and based on the behavioral weight coefficients, adding the number of erasures and modifications, the pause frequency, and the pause tendency force to obtain the behavioral feature score of the user when answering the target test question.

[0087] Specifically, the behavior feature score w is determined according to the following formula: sb :

[0088] W sb =ω4*N e +ω5*F p +ω6*N v

[0089] Among them, N e Erasure modification times, pause frequency Pause Trend Force T t is the total time of answering, BN p is the number of baseline pauses, N p is the total number of pauses, ω4 is the behavioral weight coefficient corresponding to the number of erase and modification times, ω5 is the behavioral weight coefficient corresponding to the pause frequency, and ω6 is the behavioral weight coefficient corresponding to the pause trend force.

[0090] Furthermore, to determine the regional characteristic score, it is necessary to determine the regional characteristic score of the user when answering the target question based on the actual answer area, the total available answer area, and the total writing distance. The specific method for determining the regional characteristic score includes: dividing the actual answer area by the total available answer area to obtain the answer area ratio; dividing the total writing distance by the actual answer area to obtain the handwriting density; determining the regional weight coefficients corresponding to the answer area ratio and the handwriting density respectively, and based on the regional weight coefficients, adding the answer area ratio and the handwriting density to obtain the regional characteristic score of the user when answering the target question.

[0091] Specifically, the regional feature score w is determined according to the following formula: si :

[0092] W si =ω7*A r +ω8*P d

[0093] Among them, the area ratio of the answer area Handwriting density U a For the actual answer area, TU a is the total available area, w d is the total writing distance, ω7 is the regional weight coefficient corresponding to the area ratio of the answering area, and ω8 is the regional weight coefficient corresponding to the handwriting density.

[0094] Furthermore, to determine the achievement feature score, it is necessary to determine the achievement feature score of the user when answering the target question based on the answer score and the total test question score. The specific method for determining the achievement feature score includes: dividing the answer score by the total test question score to obtain the score rate; determining the achievement weight coefficient corresponding to the score rate, and multiplying the achievement weight coefficient by the score rate to obtain the achievement feature score of the user when answering the target question.

[0095] Specifically, the achievement feature score w is calculated according to the following formula: tst :

[0096] W tst =ω9*TST

[0097] Among them, the scoring rate S c To score the answer, TS c is the total score of the test question, and ω9 is the achievement weight coefficient corresponding to the score rate.

[0098] 204. Determine the score weight coefficient corresponding to each data feature score respectively, and determine the difficulty coefficient of the target test question for the user based on each data feature score and its corresponding score weight coefficient.

[0099] Specifically, if the data feature scores include time feature scores, behavior feature scores, regional feature scores, and outcome feature scores, the difficulty coefficient w is calculated according to the following formula: di :

[0100] w di =αw st +βw sb +γw si +ηw tst

[0101] Among them, α, β, γ, and η are the score weight coefficients corresponding to the time feature score, behavior feature score, regional feature score, and outcome feature score, respectively. The sum of the score weight coefficients is 1, and each time feature score, behavior feature score, regional feature score, and outcome feature score is a score that has been normalized to the interval [0,1]. If the data feature score includes a time feature score and a behavior feature score, it is only necessary to add the time feature score and the behavior feature score according to the weight coefficients corresponding to the time feature score and the behavior feature score, and determine the added result as the difficulty coefficient. It should be noted that the data feature score is not limited to the above-mentioned scores, but may also include other feature scores. The number of data feature scores may be one or more, and the embodiment of the present invention does not make specific limitations here.

[0102] Furthermore, in order to improve the accuracy of determining the difficulty coefficient, it is also necessary to reasonably set the score weight coefficient corresponding to each data feature score. Based on this, the method includes: determining multiple weight evaluation terminals; sending the weight evaluation criterion information containing each data feature score as reference information to each weight evaluation terminal, so that each weight evaluation terminal performs terminal zero interaction scoring based on the reference information to obtain a zero interaction scoring result; receiving the zero interaction scoring result, and sending the zero interaction scoring result to each weight evaluation terminal respectively, so that each weight evaluation terminal performs terminal interaction scoring based on the zero interaction scoring result and the reference information to obtain an interaction scoring result; receiving the interaction scoring result, and based on the interaction scoring result, determining the score weight coefficient corresponding to each data feature score respectively.

[0103] The various weight assessment terminals can be representative expert terminals in different fields; the weight assessment criteria information can include questionnaire information such as the weight assessment items, weight value ranges, weight jumps, and assessment considerations. For example, to facilitate weight assignment (or scoring) by the assessment terminals, a dedicated weight coefficient selection table needs to be compiled, along with detailed selection instructions to clarify the standards and requirements for the assignment. In the first round, reference information such as the reassessment criteria information is sent to each weight assessment terminal. Each weight assessment terminal then assigns weights to each data feature score based on the weight assessment criteria without any interaction. The zero-interaction assignment (scoring) results from each assessment terminal are then collected and collated (e.g., calculating averages, coefficients of variation, etc.). The collated results are then sent to each assessment terminal, who then performs interactive weight assignment based on the collated results, thereby obtaining interactive assignment results. Finally, the score weight coefficients corresponding to each data feature score are determined based on the interactive assignment results. It should be noted that the process of determining the score weight coefficients is not limited to the two-round assignment process described above; multiple rounds of weight assignment, such as three or five rounds, can also be performed.

[0104] 205. Based on the difficulty coefficient, determine the difficulty of the target test questions for the user.

[0105] Specifically, the coefficient interval to which the difficulty coefficient belongs is determined, and the difficulty corresponding to the coefficient interval is determined as the difficulty of the target test question for the user.

[0106] According to another method for predicting the difficulty of test questions provided by the present invention, compared with the method of predicting the difficulty of test questions by the writing pressure reflected by the pressure-sensitive pen, the present invention collects video stream data when the user answers the target test question; and based on the video stream data, determines the writing process data when the user answers the target test question; then determines the difficulty coefficient of the target test question for the user based on the writing process data; and finally determines the difficulty of the target test question for the user based on the difficulty coefficient. In this way, the difficulty of the test question is predicted by the video stream data when the user answers the target test question. All users can share a video stream acquisition device, saving costs. At the same time, there is no need to calibrate the video stream acquisition device in advance, thereby improving prediction efficiency. In addition, the video stream data can capture all-round data of the user during the writing process. These writing process data help to more comprehensively understand the user's problem-solving status and thinking process and other behaviors, so that the difficulty of the test question can be predicted more accurately.

[0107] Further, as Figure 1 The specific implementation of the present invention provides a device for predicting the difficulty of a test question, such as Figure 3As shown, the device includes: a collection unit 31, a data determination unit 32, a coefficient determination unit 33, and a difficulty determination unit 34.

[0108] The acquisition unit 31 may be used to acquire video stream data when the user answers the target test questions.

[0109] The data determination unit 32 may be configured to determine the writing process data of the user when answering the target test question based on the video stream data.

[0110] The coefficient determination unit 33 may be used to determine the difficulty coefficient of the target question for the user using the writing process data.

[0111] The difficulty determination unit 34 may be configured to determine the difficulty of the target question for the user based on the difficulty coefficient.

[0112] In a specific application scenario, in order to determine the difficulty coefficient of the target test question for the user, the coefficient determination unit 33 can be specifically used to determine at least one data feature score based on the writing process data; determine the score weight coefficient corresponding to each of the data feature scores respectively, and determine the difficulty coefficient of the target test question for the user based on each of the data feature scores and its corresponding score weight coefficient.

[0113] In specific application scenarios, in order to determine the score weight coefficient, such as Figure 4 As shown, the coefficient determination unit 33 includes a first determination module 331 , a zero interaction scoring module 332 , and an interaction scoring module 333 .

[0114] The first determining module 331 may be configured to determine multiple weight evaluation terminals.

[0115] The zero-interaction scoring module 332 can be used to send the weight evaluation criterion information containing each data feature score as reference information to each weight evaluation terminal, so that each weight evaluation terminal can perform terminal zero-interaction scoring based on the reference information to obtain a zero-interaction scoring result.

[0116] The interaction scoring module 333 can be used to receive the zero-interaction scoring result and send the zero-interaction scoring result to each of the weight evaluation terminals, so that each of the weight evaluation terminals can perform terminal interaction scoring based on the zero-interaction scoring result and the reference information to obtain an interaction scoring result.

[0117] The first determining module 331 may be specifically configured to receive the interaction scoring result and determine the score weight coefficient corresponding to each of the data feature scores based on the interaction scoring result.

[0118] In a specific application scenario, for the difficulty coefficient, the coefficient determination unit 33 further includes an acquisition module 334 and a prediction module 335 .

[0119] The acquisition module 334 can be used to acquire the facial expression data of the user when answering the target test question based on the video stream data.

[0120] The prediction module 335 can be used to input the expression data and the writing process data into a large language model to perform difficulty prediction, and obtain the difficulty coefficient of the target test question for the user.

[0121] In a specific application scenario, in order to determine the difficulty coefficient of the target test question for the user, the prediction module 335 can be specifically used to respectively determine the writing feature vector corresponding to the writing process data and the expression feature vector corresponding to the expression data; align the writing feature vector and the expression feature vector, and splice the aligned writing feature vector and the aligned expression feature vector to obtain a spliced ​​feature vector; input the spliced ​​feature vector into the large language model for difficulty prediction to obtain the difficulty coefficient of the target test question for the user.

[0122] In a specific application scenario, in order to perform feature alignment, the prediction module 335 can also be used to construct a fusion feature matrix based on the writing feature vector and the expression feature vector, and determine the covariance matrix corresponding to the fusion feature matrix; perform eigenvalue decomposition on the covariance matrix to obtain eigenvalues ​​and eigenvectors corresponding to the eigenvalues; select the principal component eigenvectors whose eigenvalues ​​are greater than a preset threshold in each of the eigenvectors, and construct a transformation matrix from each of the principal component eigenvectors; based on the transformation matrix, perform linear transformation on the writing feature vector and the expression feature vector respectively to obtain the aligned writing feature vector and the aligned expression feature vector.

[0123] In a specific application scenario, in order to splice the feature vectors, the prediction module 335 can also be used to determine the feature splicing order corresponding to the aligned writing feature vector and the aligned expression feature vector; based on the feature splicing order, the aligned writing feature vector and the aligned expression feature vector are horizontally spliced ​​to obtain a spliced ​​feature vector.

[0124] In a specific application scenario, in order to determine the writing process data, the data determination unit 32 includes a capture module 321 , a semantic analysis module 322 , and a second determination module 323 .

[0125] The interception module 321 can be used to intercept the video frames containing the user's writing process in the video stream data, and perform text recognition on the video frames to obtain text content.

[0126] The semantic parsing module 322 may be configured to perform semantic parsing on the text content, and identify the answer content for the target test question in the text content based on the semantic parsing result.

[0127] The second determination module 323 can be used to determine the writing process data of the user when answering the target test question based on the video stream data, the text content, and the answer content.

[0128] In a specific application scenario, in order to determine the difficulty of the target test question for the user, the difficulty determination unit 34 can be specifically used to determine the coefficient interval to which the difficulty coefficient belongs, and determine the difficulty corresponding to the coefficient interval as the difficulty of the target test question for the user, wherein different coefficient intervals correspond to different test question difficulties.

[0129] It should be noted that for other corresponding descriptions of the functional modules involved in the device for predicting the difficulty of a test question provided by the embodiment of the present invention, please refer to Figure 1 The corresponding description of the method shown will not be repeated here.

[0130] Based on the above Figure 1 The method shown, accordingly, an embodiment of the present invention further provides a computer-readable storage medium, on which a computer program is stored, which implements the following steps when executed by a processor: collecting video stream data of a user answering a target test question; based on the video stream data, determining the writing process data of the user answering the target test question; based on the writing process data, determining the difficulty coefficient of the target test question for the user; based on the difficulty coefficient, determining the difficulty of the target test question for the user.

[0131] Based on the above Figure 1 The method shown and Figure 3 The embodiment of the device shown in the figure, the embodiment of the present invention also provides a physical structure diagram of a computer device, such as Figure 5As shown, the computer device includes: a processor 41, a memory 42, and a computer program stored in the memory 42 and executable on the processor, wherein the memory 42 and the processor 41 are both arranged on a bus 43. When the processor 41 executes the program, the following steps are implemented: collecting video stream data of a user answering a target test question; based on the video stream data, determining the writing process data of the user answering the target test question; based on the writing process data, determining the difficulty coefficient of the target test question for the user; based on the difficulty coefficient, determining the difficulty of the target test question for the user.

[0132] Through the technical solution of the present invention, the present invention collects video stream data when the user answers the target test questions; and based on the video stream data, determines the writing process data of the user when answering the target test questions; then determines the difficulty coefficient of the target test questions for the user based on the writing process data; and finally determines the difficulty of the target test questions for the user based on the difficulty coefficient. In this way, the difficulty of the test questions is predicted by the video stream data when the user answers the target test questions. All users can share a video stream acquisition device, saving costs. At the same time, there is no need to calibrate the video stream acquisition device in advance, thereby improving prediction efficiency. In addition, the video stream data can capture all-round data of the user during the writing process. These writing process data help to more comprehensively understand the user's problem-solving status and thinking process and other behaviors, so as to more accurately predict the difficulty of the test questions.

[0133] Obviously, those skilled in the art will appreciate that the various modules or steps of the present invention described above can be implemented using a general-purpose computing device, centralized on a single computing device, or distributed across a network of multiple computing devices. Alternatively, they can be implemented using program code executable by a computing device, which can then be stored in a storage device and executed by the computing device. In some cases, the steps shown or described can be performed in a different order than that shown, or can be fabricated as separate integrated circuit modules, or multiple modules or steps can be fabricated as a single integrated circuit module. Thus, the present invention is not limited to any particular combination of hardware and software.

[0134] The foregoing description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Those skilled in the art will readily appreciate that various modifications and variations of the present invention are possible. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present invention shall be included within the scope of protection of the present invention.

Claims

1. A method for predicting the difficulty of a test question, characterized in that: include: Collect video stream data when users answer target test questions; Determining, based on the video stream data, writing process data of the user when answering the target test question; Determining a difficulty coefficient of the target test question for the user based on the writing process data; Based on the difficulty coefficient, the difficulty of the target test question for the user is determined.

2. The method according to claim 1, characterized in that Determining the difficulty coefficient of the target test question for the user based on the writing process data includes: determining at least one data feature score based on the writing process data; The score weight coefficient corresponding to each of the data feature scores is determined respectively, and based on each of the data feature scores and the corresponding score weight coefficient, the difficulty coefficient of the target test question for the user is determined.

3. The method according to claim 2, characterized in that The step of respectively determining the score weight coefficient corresponding to each of the data feature scores includes: Determine multiple weighted assessment terminals; sending weight evaluation criterion information including each of the data feature scores as reference information to each of the weight evaluation terminals, so that each of the weight evaluation terminals performs a terminal zero interaction score based on the reference information to obtain a zero interaction score result; receiving the zero-interaction scoring result, and sending the zero-interaction scoring result to each of the weight evaluation terminals, respectively, so that each of the weight evaluation terminals performs terminal interaction scoring based on the zero-interaction scoring result and the reference information to obtain an interaction scoring result; The interactive scoring result is received, and based on the interactive scoring result, a score weight coefficient corresponding to each of the data feature scores is determined.

4. The method according to claim 1, wherein Determining the difficulty coefficient of the target test question for the user based on the writing process data includes: Based on the video stream data, acquiring facial expression data of the user when answering the target test question; The expression data and the writing process data are input into a large language model to perform difficulty prediction, and a difficulty coefficient of the target test question for the user is obtained.

5. The method according to claim 4, characterized in that Inputting the expression data and the writing process data into a large language model to perform difficulty prediction and obtain the difficulty coefficient of the target test question for the user includes: respectively determining a writing feature vector corresponding to the writing process data and an expression feature vector corresponding to the expression data; Performing feature alignment on the writing feature vector and the expression feature vector, and splicing the aligned writing feature vector and the aligned expression feature vector to obtain a spliced ​​feature vector; The concatenated feature vector is input into the large language model to perform difficulty prediction, and a difficulty coefficient of the target test question for the user is obtained.

6. The method according to claim 5, characterized in that The step of aligning the writing feature vector and the expression feature vector comprises: Constructing a fusion feature matrix based on the writing feature vector and the expression feature vector, and determining a covariance matrix corresponding to the fusion feature matrix; Performing eigenvalue decomposition on the covariance matrix to obtain eigenvalues ​​and eigenvectors corresponding to the eigenvalues; Selecting a principal component eigenvector whose eigenvalue is greater than a preset threshold value from each of the eigenvectors, and constructing a transformation matrix based on each of the principal component eigenvectors; Based on the transformation matrix, linear transformation is performed on the writing feature vector and the expression feature vector respectively to obtain the aligned writing feature vector and the aligned expression feature vector.

7. The method according to claim 5, characterized in that The step of splicing the aligned writing feature vector and the aligned expression feature vector to obtain a spliced ​​feature vector includes: Determine a feature splicing order corresponding to the aligned handwriting feature vector and the aligned expression feature vector; Based on the feature splicing order, the aligned writing feature vector and the aligned expression feature vector are horizontally spliced ​​to obtain a spliced ​​feature vector.

8. The method according to claim 1, characterized in that The determining, based on the video stream data, the writing process data of the user when answering the target test question includes: intercepting a video frame containing the user's writing process from the video stream data, and performing text recognition on the video frame to obtain text content; Performing semantic analysis on the text content, and identifying the answer content for the target test question in the text content based on the semantic analysis result; Based on the video stream data, the text content, and the answer content, the writing process data of the user when answering the target test question is determined.

9. The method according to claim 1, characterized in that Determining the difficulty of the target test question for the user based on the difficulty coefficient includes: The coefficient interval to which the difficulty coefficient belongs is determined, and the difficulty corresponding to the coefficient interval is determined as the difficulty of the target test question for the user, wherein different coefficient intervals correspond to different test question difficulties.

10. A device for predicting the difficulty of a test question, characterized in that: include: A collection unit, used to collect video stream data when the user answers the target test questions; a data determining unit, configured to determine, based on the video stream data, writing process data of the user when answering the target test question; a coefficient determination unit, configured to determine a difficulty coefficient of the target test question for the user based on the writing process data; A difficulty determination unit is configured to determine the difficulty of the target test question for the user based on the difficulty coefficient.

11. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 9 are implemented.

12. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 9 are implemented.

Images