Classification method and device based on question answering, storage medium and program product

Abstract

The embodiment of the invention discloses a question answering-based classification method and device, a storage medium and a program product. The method comprises the following steps: constructing an exclusive classification model of each candidate scale in a candidate scale set, wherein each candidate scale comprises a plurality of candidate questions; screening the candidate scale and the candidate problems in the candidate scale based on the performance of the exclusive classification model to obtain a plurality of optimal problems and construct an optimal problem set; taking the exclusive classification model corresponding to each optimal problem as a target model, and constructing a fusion model based on the number of classification categories of the target model; all the target models and the fusion model form a fusion classification model; and executing an actual classification task based on the optimal problem set and the fusion classification model. According to the method, redundant and irrelevant information can be removed by screening problems based on scale exclusive classification model performance, and meanwhile, a fusion classification model is constructed to integrate reasoning information of a plurality of models, so that an overfitting problem is solved, and an accurate and high-rate classification effect is achieved.

Description

Technical Field

[0001] This disclosure relates to the field of artificial intelligence technology, and in particular to a classification method, apparatus, storage medium, and program product based on question answering. Background Technology

[0002] In various fields, especially in areas such as mental health assessment and assisting doctors in disease screening, there is an urgent need for efficient and accurate classification methods. Existing methods typically use scales to complete classification tasks. Scales are tools used to measure certain psychological traits, behavioral performance, health status, etc., consisting of a series of standardized questions scored according to certain rules to obtain quantitative assessment results. When combined with AI models, a common practice is to have users answer questions using a single scale, and then input the responses into the model for classification.

[0003] However, the development and use of scales have certain limitations. Scale design is often based on specific theories and sample groups, which may not fully and accurately reflect the true situation of different individuals, and its applicability is poor for some special groups or newly emerging problem scenarios. Secondly, the method is cumbersome, requiring users to spend a lot of time and effort to complete the scale, making it extremely inefficient for large-scale screening. Moreover, the implementation and interpretation of the scale usually rely on professionals, which not only increases labor costs but also limits the widespread application of the scale. Furthermore, this single classification method lacks flexibility and dynamism, making it difficult to adapt to complex and changing real-world situations. When faced with new problem types or data characteristics, the model's generalization ability is poor, failing to accurately classify users, resulting in a significant reduction in the reliability and effectiveness of the classification results, and failing to meet the needs of efficient and accurate classification in practical applications. Summary of the Invention

[0004] In view of this, the present disclosure provides a question-answering-based classification method, apparatus, storage medium, and program product, which can screen questions based on the performance of a scale-specific classification model, accurately remove redundant and irrelevant information to obtain the optimal question set, and simultaneously construct a fusion classification model to integrate the reasoning information of multiple models to solve the overfitting problem. This achieves the important effect of saving information collection time and improving classification efficiency while ensuring classification accuracy.

[0005] In a first aspect, embodiments of this disclosure provide a classification method based on question answering, employing the following technical solution: Obtain a candidate scale set, which contains multiple candidate scales, and each candidate scale contains multiple candidate questions; Build a dedicated classification model for each candidate scale; Based on the performance of the dedicated classification model, the candidate scale and the candidate questions in the candidate scale are screened to obtain multiple optimal questions and construct an optimal question set; For each optimal problem, the corresponding dedicated classification model is used as the target model. Based on the number of classification categories in the target model, a fusion model is constructed. Combine all target models and fusion models into a fusion classification model; The actual classification task is performed based on the optimal problem set and the fusion classification model.

[0006] Optionally, based on the performance of the dedicated classification model, the candidate scale and candidate questions in the candidate scale are screened to obtain multiple optimal questions and construct an optimal question set, including: Performance evaluation is performed on each dedicated classification model to obtain the performance metrics for each dedicated classification model; The candidate scales used by the dedicated classification model that meets the performance indicators are taken as the effective scales; Based on the dedicated classification model for each valid scale, the first correlation coefficient among all valid scales is obtained; Based on the first correlation coefficient, multiple target scales are selected from all valid scales; Obtain the independent answers of the subjects to each candidate question in each target scale, input each independent answer into the corresponding model, and obtain the classification performance metric value of each candidate question in each target scale; Based on the classification performance metric, multiple valid questions are selected from the candidate questions in all target scales; Based on the dedicated classification model corresponding to each valid question, the second correlation coefficient among all valid questions is obtained; Based on the second correlation coefficient, multiple optimal problems are selected from all valid problems, and all optimal problems are combined into an optimal problem set.

[0007] Optionally, the step of obtaining the first correlation coefficient among all valid scales based on the dedicated classification model for each valid scale includes: If the number of categories in the dedicated classification model is 2, then obtain the first risk probability sequence output by the dedicated classification model for each valid scale on the corresponding test dataset; Based on the first risk probability sequence, obtain the first correlation coefficient among all valid scales; If the number of categories in the dedicated classification model is greater than 2, then obtain the first risk probability matrix output by the dedicated classification model for each valid scale on the corresponding test dataset. Based on the first risk probability matrix, obtain the covariance matrix of each effective scale and the covariance matrix between every two effective scales; By solving the generalized eigenvalue problem, the first correlation coefficient among all valid scales is obtained based on the covariance matrix of each valid scale and the covariance matrix between every two valid scales.

[0008] Optionally, the expression for the first risk probability sequence is: In the formula, Indicates the first The first risk probability sequence output by the dedicated classification model of each effective scale on the corresponding test dataset; Indicates the first A dedicated classification model for each valid scale is used for the first valid scale in the corresponding test dataset. The risk probability output for each sample; This indicates the number of samples in the test dataset. ; The formula for calculating the first correlation coefficient is: In the formula, Indicates the first The first effective scale and the first The first correlation coefficient between the effective scales; Indicates the first The mean of the first risk probability sequence output by the dedicated classification model of each effective scale on the corresponding test dataset; Indicates the first A dedicated classification model for each valid scale is used for the first valid scale in the corresponding test dataset. The risk probability output for each sample; Indicates the first The mean of the first risk probability sequence output by the dedicated classification model of each effective scale on the corresponding test dataset.

[0009] Optionally, obtaining the second correlation coefficient among all valid questions based on the dedicated classification model corresponding to each valid question includes: If the number of categories in the exclusive classification model is 2, then the independent answers of each subject to each valid question are input into the corresponding exclusive classification model, and a second risk probability sequence is formed based on the output of the exclusive classification model. Based on the second risk probability sequence, obtain the second correlation coefficient among all valid questions; If the number of categories in the exclusive classification model is greater than 2, then the independent answers of each subject to each valid question are input into the corresponding exclusive classification model, and a second risk probability matrix is ​​formed based on the output of the exclusive classification model. Based on the second risk probability matrix, obtain the covariance matrix of each valid problem and the covariance matrix between every two problems; By solving the generalized eigenvalue problem, the second correlation coefficient among all valid problems is obtained based on the covariance matrix of each valid problem and the covariance matrix between every two valid problems.

[0010] Optionally, the output layer of the dedicated classification model includes a linear transformation function and an activation function; The linear transformation function is used to generate Logits values, and the activation function is used to convert the Logits values ​​into a classification output format.

[0011] Optionally, the fusion model includes an input layer, a hidden layer, and an output layer; When the number of classification categories of the target model is 2, the input layer of the fusion model is used to generate Logits values ​​of all target models to form an N-dimensional Logits feature vector, and transmit the Logits feature vector to the hidden layer of the fusion model. When the number of classification categories of the target model is greater than 2, the input layer of the fusion model is used to generate Logits values ​​of all target models to form an M×N dimensional Logits feature vector, and transmit the Logits feature vector to the hidden layer of the fusion model. Regardless of whether the number of classification categories of the target model is 2 or greater, the hidden layer of the fusion model is used to convert the received Logits feature vector into a comprehensive Logits value and transmit the comprehensive Logits value to the output layer of the fusion model; The output layer of the fusion model is used to convert the received comprehensive Logits values ​​into a classification output format.

[0012] Optionally, the expression for the loss function of the fusion model is: In the formula, This represents the value of the loss function for the fusion model; This represents the total number of samples in the training dataset of the fusion model; Indicates the number of classification categories in the fusion model; The fusion model represents the first... The weights of each category, ; The fusion model is pre-set to the first The sample belongs to the first The probability of each category. ; Indicates the focus parameter; Indicates the fusion model for the first The sample belongs to the first Indicator functions for each category; It is a regulating factor.

[0013] Secondly, this disclosure also provides a question-and-answer-based classification system, employing the following technical solution: The set acquisition module is used to acquire a candidate scale set, which contains multiple candidate scales, and each candidate scale contains multiple candidate questions; The first building module is used to build a dedicated classification model for each candidate scale; The question filtering module is used to filter the candidate scale and the candidate questions in the candidate scale based on the performance of the dedicated classification model, obtain multiple optimal questions and construct an optimal question set; The second building module is used to take the dedicated classification model corresponding to each optimal problem as the target model and build a fusion model based on the number of classification categories of the target model. The model composition module is used to combine all target models and fusion models into a fusion classification model; The task execution module is used to perform the actual classification task based on the optimal problem set and the fusion classification model.

[0014] Thirdly, this disclosure also provides a computer device, which adopts the following technical solution: The computer device includes: At least one processor; and, A memory communicatively connected to the at least one processor; wherein, The memory stores instructions that can be executed by the at least one processor to enable the at least one processor to perform any of the above-described problem-solving-based classification methods.

[0015] Fourthly, embodiments of this disclosure also provide a computer-readable storage medium storing computer instructions for causing a computer to perform any of the above-described question-and-answer-based classification methods.

[0016] Fifthly, embodiments of this disclosure also provide a computer program product, including a computer program / instructions that, when executed by a processor, implement the steps of any of the methods described above.

[0017] The question-and-answer-based classification method provided in this disclosure, after obtaining a set of candidate scales, filters the candidate scales and their constituent candidate questions, ultimately selecting the optimal questions from a large number of candidate scales to obtain a streamlined question set. This process effectively removes redundant and irrelevant information, making the final question set used for classification more accurately reflect the characteristics related to cognitive impairment. Furthermore, while ensuring accuracy, reducing the number of questions saves patients' total answering time. Additionally, filtering based on the performance of the dedicated classification model accurately identifies those questions that are truly valuable for classification, ensuring that the final optimal question set more accurately reflects the characteristics related to the target classification. For example, in the cognitive impairment classification scenario, it accurately focuses on questions closely related to disease characteristics, making the classification basis more targeted. Moreover, model performance is an important indicator of its classification ability; therefore, performance-based filtering ensures that the dedicated classification models corresponding to the retained questions have high accuracy. When using these optimal questions for actual classification, it provides reliable assurance for the classification results, reducing classification errors caused by invalid or low-quality questions, thereby improving the overall accuracy of the classification. After using the dedicated classification model corresponding to each optimal problem as the target model, a fusion model is constructed based on the number of classification categories in the target model. All target models and the fusion model are then combined to form a fusion classification model. This fusion classification model can quickly integrate information from different target scales, resolving the overfitting problem and making this method more advantageous in practical applications. Through the preceding steps, an optimal problem set and a fusion classification model are obtained. Performing actual classification tasks based on this model integrates the accuracy of the problem set, the effective integration of information by the model, and the advantages over overfitting, thus completing actual classification tasks more efficiently and accurately. This approach plays a significant role in classification applications in cognitive impairment diseases, drug treatment efficacy evaluation, and even other fields.

[0018] The above description is merely an overview of the technical solution disclosed herein. In order to better understand the technical means of this disclosure and to implement it in accordance with the contents of the specification, and to make the above and other objects, features and advantages of this disclosure more apparent and understandable, preferred embodiments are described below in detail with reference to the accompanying drawings. Attached Figure Description

[0019] To more clearly illustrate the technical solutions of the embodiments of this disclosure, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this disclosure. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0020] Figure 1 A flowchart illustrating the question-answering-based classification method provided in this embodiment of the disclosure; Figure 2 A flowchart illustrating the method for constructing a proprietary classification model provided in this embodiment of the disclosure; Figure 3 A flowchart illustrating the optimal problem-solving method provided in this embodiment of the disclosure; Figure 4 A flowchart illustrating the first correlation coefficient acquisition method provided in this embodiment of the disclosure; Figure 5 Another schematic flowchart of the method for obtaining the first correlation coefficient provided in the embodiments of this disclosure; Figure 6 A flowchart illustrating the second correlation coefficient acquisition method provided in this embodiment of the disclosure; Figure 7 This is another schematic flowchart illustrating the second correlation coefficient acquisition method provided in the embodiments of this disclosure; Figure 8 A schematic diagram illustrating the process of the fusion classification model provided in this embodiment performing an actual classification task; Figure 9 A block diagram illustrating the principle of a question-answering-based classification system provided in this embodiment of the disclosure; Figure 10 This is a schematic diagram of the structure of a computer device provided in an embodiment of the present disclosure. Detailed Implementation

[0021] The embodiments of this disclosure will now be described in detail with reference to the accompanying drawings.

[0022] It should be understood that the following specific examples illustrate the implementation of this disclosure, and those skilled in the art can easily understand other advantages and effects of this disclosure from the content disclosed in this specification. Obviously, the described embodiments are only a part of the embodiments of this disclosure, and not all of them. This disclosure can also be implemented or applied through other different specific implementation methods, and the details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of this disclosure. It should be noted that, in the absence of conflict, the following embodiments and features in the embodiments can be combined with each other. Based on the embodiments in this disclosure, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this disclosure.

[0023] It should be noted that various aspects of embodiments within the scope of the appended claims are described below. It will be apparent that the aspects described herein can be embodied in a wide variety of forms, and any particular structure and / or function described herein is merely illustrative. Based on this disclosure, those skilled in the art will understand that one aspect described herein can be implemented independently of any other aspect, and two or more of these aspects can be combined in various ways. For example, any number of aspects set forth herein can be used to implement the device and / or practice the method. Additionally, this device and / or method can be implemented using structures and / or functionalities other than one or more of the aspects set forth herein.

[0024] It should also be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of this disclosure. The drawings only show the components related to this disclosure and are not drawn according to the number, shape and size of the components in actual implementation. In actual implementation, the form, quantity and proportion of each component can be arbitrarily changed, and the layout of the components may also be more complex.

[0025] Furthermore, specific details are provided in the following description to facilitate a thorough understanding of the examples. However, those skilled in the art will understand that the described aspects can be practiced without these specific details.

[0026] Reference Figure 1 This disclosure provides a classification method based on question-and-answer format, comprising the following steps: S1: Obtain a set of candidate scales, which contains multiple candidate scales, and each candidate scale contains multiple candidate questions; S2: Construct a dedicated classification model for each candidate scale; S3: Based on the performance of the dedicated classification model, the candidate scales and candidate questions in the candidate scales are filtered to obtain multiple optimal questions and construct an optimal question set; S4: Take the dedicated classification model corresponding to each optimal problem as the target model, and construct a fusion model based on the number of classification categories in the target model; S5: Combine all target models and fusion models into a fusion classification model; S6: Perform the actual classification task based on the optimal problem set and the fusion classification model.

[0027] The question-and-answer-based classification method disclosed herein, after obtaining a set of candidate scales, filters the candidate scales and their constituent candidate questions, ultimately selecting the optimal questions from a large number of candidate scales to obtain a streamlined question set. This process effectively removes redundant and irrelevant information, making the final question set used for classification more accurately reflect the characteristics related to cognitive impairment. Furthermore, while ensuring accuracy, reducing the number of questions saves patients' total answering time. Additionally, filtering based on the performance of the dedicated classification model accurately identifies those questions that are truly valuable for classification, ensuring that the final optimal question set more accurately reflects the characteristics related to the target classification. For example, in the cognitive impairment classification scenario, it accurately focuses on questions closely related to disease characteristics, making the classification basis more targeted. Moreover, model performance is an important indicator of its classification ability; therefore, performance-based filtering ensures that the dedicated classification models corresponding to the retained questions have high accuracy. When using these optimal questions for actual classification, it provides reliable assurance for the classification results, reducing classification errors caused by invalid or low-quality questions, thereby improving the overall accuracy of classification.

[0028] After using the dedicated classification model corresponding to each optimal problem as the target model, a fusion model is constructed based on the number of classification categories in the target model. All target models and the fusion model are then combined to form a fusion classification model. This fusion classification model can quickly integrate information from different target scales, resolving the overfitting problem and making this method more advantageous in practical applications. Through the preceding steps, an optimal problem set and a fusion classification model are obtained. Performing actual classification tasks based on this model integrates the accuracy of the problem set, the effective integration of information by the model, and the advantages over overfitting, thus completing actual classification tasks more efficiently and accurately. This approach plays a significant role in classification applications in cognitive impairment diseases, drug treatment efficacy evaluation, and even other fields.

[0029] In S1, a large number of candidate scales are collected according to the application scenarios of the classification tasks required. For example, for the application scenario of assisting doctors in screening for cognitive impairment, the goal is to collect existing, clinically validated cognitive assessment scales, such as the Brief Mental State Examination (MMSE), Montreal Cognitive Assessment (MoCA), Clinical Dementia Rating Scale (CDRSU), Digit Symbol Transition Test (DSPT), Verbal Fluency Test (VF), Auditory Word Learning Test (RAVLT), etc., through cooperation with multiple medical institutions. These candidate scales are then integrated into a set. In addition, each candidate scale contains multiple candidate questions for users to engage in question-and-answer interaction.

[0030] In S2, refer to Figure 2The flowchart illustrating the method for building a dedicated classification model is shown. "Building a dedicated classification model for each candidate scale" includes the following steps: S21: For each candidate scale, a unified initial model with the same architecture and parameter configuration is preset; S22: Train the corresponding initial model for each candidate scale separately to obtain a dedicated classification model for each candidate scale.

[0031] In the above steps, to ensure fairness and consistency in the subsequent training and evaluation of classification models specific to different candidate scales, it is necessary to pre-define an initial model with the same architecture and parameter configuration for each candidate scale. First, a high-quality and representative training dataset is constructed for each scale. This dataset can be built based on existing large-scale, high-quality cognitive assessment speech databases in the medical industry; this database is the cornerstone of all subsequent model training and optimization. A large number (e.g., thousands) of participants' full-length recordings of their answers to candidate questions within different candidate scales are collected. These participants need to cover the entire sample, including healthy controls (HC), subjective cognitive decline (SCD), mild cognitive impairment (MCI), and patients with different stages of dementia, forming a complete spectrum of aging and cognitive decline. Such a large and diverse sample is necessary because sample size is crucial to the model's generalization ability and robustness. Research shows that in cognitive diagnostic models, 500 samples are considered the minimum requirement for achieving high accuracy. However, depending on actual needs, training datasets for each scale can also be constructed using participants' written answers to candidate questions. Next, the model architecture needs to be determined based on the number of classification types required and the specific application scenario of the classification task. For example, for a binary classification task with two classification types, a regression model can be used. Then, considering the specific application scenario of assisting doctors in screening for cognitive impairment, a suitable regression model type can be further selected. For instance, a neural network architecture suitable for processing sequence data, such as a Convolutional Neural Network (CNN), a Long Short-Term Memory Network (LSTM), or a Transformer, can be used, followed by fully connected layers and a sigmoid activation function to output regression probabilities. This same architecture should be used to pre-set the initial model for each candidate scale. Simultaneously, the parameter configurations of the initial models should be standardized, including hyperparameters such as learning rate, batch size, and number of training epochs, to ensure that all initial models begin training under the same conditions.

[0032] After pre-setting the initial model, the training dataset for each scale is labeled, and the corresponding initial model is trained using the labeled training dataset. For example, training can be performed using collected complete speech data and corresponding clinical diagnostic labels. Each recording is accompanied by precise metadata, where the clinical diagnostic labels are "gold standard" risk labels given by neurology experts based on standard clinical guidelines (e.g., 0 represents no risk / health, 1 represents high risk / diagnosed MCI or dementia). This provides accurate supervision information for the initial model training. Segment labeling is also required, that is, accurately labeling the start and end times of the speech segments corresponding to the subjects' answers to each scale question in each recording. The input of the initial model is features extracted from speech. Advanced speech processing technology is used to extract deep acoustic features from each complete scale recording. This can include traditional prosodic features such as Mel-frequency cepstral coefficients (MFCCs), pitch, speech rate, and pauses, or it can use pre-trained large speech models (such as Wav2Vec, Whisper, etc.) to extract more representative embedding vectors. During training, the initial model continuously adjusts its parameters to minimize the error between the predicted risk probability and the actual clinical diagnostic label. For a binary regression model, the output is a continuous value between 0 and 1, representing the risk probability that the subject has cognitive impairment; a value greater than 0.5 indicates the condition, and a value less than or equal to 0.5 indicates no condition. In this way, the initial model trained for each candidate scale gradually adapts to the characteristics and data distribution of that scale, ultimately resulting in a dedicated classification model (also known as a weak-risk classification model) for each candidate scale. While these dedicated classification models are not the final screening tool, they serve as "probes" and "metrics" for subsequent screening of scales and questions.

[0033] In S3, refer to Figure 3 The flowchart illustrating the optimal question acquisition method, which involves "filtering candidate scales and candidate questions within them based on the performance of the dedicated classification model to obtain multiple optimal questions and construct an optimal question set," includes: S31: Evaluate the performance of each dedicated classification model and obtain the performance metrics for each dedicated classification model; S32: The candidate scales used by the dedicated classification model that meets the performance index criteria shall be used as the effective scales; S33: Based on the dedicated classification model for each valid scale, obtain the first correlation coefficient among all valid scales; S34: Based on the first correlation coefficient, select multiple target scales from all valid scales; S35: Obtain the independent answers of the subjects to each candidate question in each target scale, input each independent answer into the corresponding model, and obtain the classification performance metric value of each candidate question in each target scale; S36: Based on the classification performance metric, select multiple valid questions from the candidate questions in all target scales; S37: Based on the dedicated classification model corresponding to each valid question, obtain the second correlation coefficient among all valid questions; S38: Based on the second correlation coefficient, select multiple optimal problems from all valid problems and combine all optimal problems into an optimal problem set.

[0034] In S31 and S32, the same group of subjects answered the candidate questions in each specific classification model. Based on the subjects' answer data, a test dataset for each specific classification model was constructed. The performance of the corresponding specific classification model was evaluated using the test dataset, and the performance index of each specific classification model was obtained. The performance index can be AUC (Area Under the Curve).

[0035] If the number of categories in the specific classification model is 2, it indicates that the specific classification model is a binary regression model, and the calculation of AUC is relatively straightforward. First, the specific classification model outputs a predicted probability for each sample in the test dataset. This probability represents the likelihood that the sample belongs to the positive class. Then, these predicted probabilities are compared with the true labels of the samples (usually 0 or 1), and the samples are sorted from high to low according to the predicted probabilities. Next, a participant operating characteristic (ROC) curve is plotted, with the false positive rate (FPR) on the horizontal axis and the true positive rate (TPR) on the vertical axis. By calculating the area under the ROC curve, the AUC value of this binary specific classification model can be obtained. The closer the AUC value is to 1, the better the model's performance.

[0036] If the number of categories in the dedicated classification model is greater than 2, it indicates that the dedicated classification model is a multi-class model. In this case, a "category comparison strategy" is needed to transform the multi-class problem into a binary classification problem, and then an "aggregation strategy" is used to synthesize the results. The specific operation is as follows: The "category comparison strategy" refers to, for a multi-class problem with n (n > 2) categories, treating each category as the positive class sequentially, and merging the remaining n-1 categories as the negative class, thus obtaining n binary classification problems. For each binary classification problem, the model outputs the predicted probability of a sample belonging to the current positive class, and then calculates the AUC value for these n binary classification problems according to the binary classification AUC calculation method. The "aggregation strategy" refers to, after obtaining the n binary classification AUC values, combining them to obtain the AUC value of the multi-class model. Common aggregation methods include macro-averaging and micro-averaging. Macro-averaging directly averages the n binary classification AUC values; micro-averaging first merges the predicted probabilities and true labels of all samples, and then recalculates a total AUC value. The choice of which aggregation method to use depends on the specific application scenario and requirements.

[0037] After obtaining the AUC performance metric for each dedicated classification model, the AUC of each dedicated classification model can be compared with a pre-set AUC threshold. Dedicated classification models with an AUC greater than the AUC threshold are judged as meeting the standard, and the candidate scales used by these compliant models are the valid scales. Dedicated classification models with an AUC less than the AUC threshold are judged as not meeting the standard, and the candidate scales used by the non-compliant models are deleted.

[0038] To illustrate the above method, AUC is an indicator that comprehensively measures the model's ability to balance the "true positive rate" and "false positive rate" across all possible classification thresholds. It is particularly suitable for evaluating medical screening tools. According to the expert consensus on digital screening tools, an AUC of 0.8 or higher is considered a marker of good discriminative power. Therefore, AUC is used as the primary screening criterion to exclude candidate scales with low corresponding model AUC values ​​(e.g., below 0.75), because this indicates that the scale has a weak overall ability to distinguish speech signals of cognitive impairment. In contrast, effective scales can classify samples more accurately in application, improving the reliability and accuracy of classification.

[0039] In S33, the calculation method for the first correlation coefficient differs depending on whether the number of categories in the dedicated classification model is 2 or greater. (Refer to...) Figure 4 The flowchart illustrating the method for obtaining the first correlation coefficient shows that, when the number of categories in the dedicated classification model is 2, "obtaining the first correlation coefficient between all valid scales based on the dedicated classification model for each valid scale" includes the following steps: S331: Obtain the first risk probability sequence output by the dedicated classification model for each valid scale on the corresponding test dataset; S332: Based on the first risk probability sequence, obtain the first correlation coefficient among all valid scales.

[0040] The expression for the first risk probability sequence is: In Formula 1, Indicates the first The first risk probability sequence output by the dedicated classification model of each effective scale on the corresponding test dataset; Indicates the first A dedicated classification model for each valid scale is used for the first valid scale in the corresponding test dataset. The risk probability output for each sample; This indicates the number of samples in the test dataset. .

[0041] The formula for calculating the first correlation coefficient is: In Formula 2, Indicates the first The first effective scale and the first The first correlation coefficient between the effective scales; Indicates the first The mean of the first risk probability sequence output by the dedicated classification model of each effective scale on the corresponding test dataset; Indicates the first A dedicated classification model for each valid scale is used for the first valid scale in the corresponding test dataset. The risk probability output for each sample; Indicates the first The mean of the first risk probability sequence output by the dedicated classification model of each effective scale on the corresponding test dataset.

[0042] Reference Figure 5 Another flowchart illustrating the method for obtaining the first correlation coefficient shows that when the number of categories in the dedicated classification model is greater than 2, "obtaining the first correlation coefficient between all valid scales based on the dedicated classification model for each valid scale" includes the following steps: S333: Obtain the first risk probability matrix output by the dedicated classification model for each valid scale on the corresponding test dataset; S334: Based on the first risk probability matrix, obtain the covariance matrix of each valid scale and the covariance matrix between every two valid scales; S335: By solving the generalized eigenvalue problem, the first correlation coefficient among all valid scales is obtained based on the covariance matrix of each valid scale and the covariance matrix between every two valid scales.

[0043] In S333, for a multi-class classification model, what is obtained is the probability of each sample belonging to each class. These probabilities form a probability matrix, namely the first risk probability matrix, where each row represents a sample and each column represents the probability of a class.

[0044] In S334, the first risk probability matrix is ​​transformed into a new set of variables, each representing the probability of a class. For example, if there are three classes, each sample will be transformed into a 3-dimensional vector, with each dimension corresponding to the probability of a class. The transformed data is then standardized to ensure that each variable has a mean of 0 and a variance of 1. This is a prerequisite for CCA analysis, aiming to eliminate dimensional differences between different valid scales (or variables) and make them comparable in the analysis. The covariance matrix describes the linear relationship between different variables; therefore, in CCA, it is necessary to construct the covariance matrix of each valid scale itself and the cross-covariance matrix between every two valid scales.

[0045] In S335, a first generalized eigenvalue problem function is constructed based on the covariance matrix of each pair of valid scales and the covariance matrix between each pair of valid scales. By solving the first generalized eigenvalue problem function, the largest eigenvalue between each pair of valid scales is obtained. The square root of the largest eigenvalue between each pair of valid scales is then calculated to obtain the first correlation coefficient between each pair of valid scales. The expression for the first generalized eigenvalue problem function is as follows: In formula 3, Indicates the first The inverse matrix of the covariance matrix of each effective scale; Indicates the first The first effective scale and the first Covariance matrix among the valid scales; Indicates the first The inverse matrix of the covariance matrix of each effective scale; Indicates the first The first effective scale and the first Covariance matrix among the valid scales; Indicates the first The first effective scale and the first The eigenvectors of the generalized eigenvalue problem corresponding to each effective scale, in CCA, represent the direction corresponding to the maximum correlation in the original space. It is a solution to the generalized eigenvalue problem, indicating which combination of variables best explains the correlation between two sets of variables in the multidimensional space. Indicates the first The first effective scale and the first The eigenvalues ​​of a generalized eigenvalue problem corresponding to an effective scale are also known as canonical correlation coefficients. In CCA, the eigenvalues ​​represent the strength of the correlation between two sets of variables. The larger the eigenvalue, the stronger the correlation between the two sets of variables.

[0046] In step S34, the clinical importance score and the performance index of the specific classification model for each valid scale are obtained. The clinical importance score is the average score of multiple experts on the valid scale, while the performance index is the AUC obtained in step S31. The clinical importance score and the performance index of the specific classification model for each valid scale are weighted and summed to obtain the comprehensive score for each valid scale. When the first correlation coefficient between two valid scales is greater than a preset first correlation coefficient threshold (e.g., 0.9), it indicates that the two valid scales may be assessing similar cognitive dimensions, resulting in information redundancy. Therefore, the scale with the higher comprehensive score is selected as the target scale. Through this pairwise comparison and screening method, multiple complementary target scales can be obtained.

[0047] In S35 and S36, the classification performance metric can still be AUC. If the dedicated classification model has two categories, it will output a predicted probability for each participant's independent answer to each candidate question in the target scale, and the specific AUC value will be obtained based on these predicted probabilities. If the dedicated classification model has more than two categories, the "category comparison strategy" will be used to transform the multi-class problem into a binary classification problem, and then the results will be combined using the "aggregation strategy" to obtain the total AUC value.

[0048] It can sort each candidate question in each target scale from largest to smallest classification performance metric value and select the top K candidate questions as valid questions; or it can compare the classification performance metric value with a preset classification performance metric threshold and select candidate questions with a classification performance metric value greater than the preset classification performance metric threshold as valid questions.

[0049] In S37, the calculation method for the second correlation coefficient differs depending on whether the number of categories in the specific classification model is 2 or greater than 2. (Refer to...) Figure 6The flowchart illustrating the method for obtaining the second correlation coefficient shows that, when the number of categories in the dedicated classification model is 2, "obtaining the second correlation coefficient between all valid questions based on the dedicated classification model corresponding to each valid question" includes the following steps: S371: Input each subject's independent answer to each valid question into the corresponding dedicated classification model, and form a second risk probability sequence based on the output of the dedicated classification model; S372: Based on the second risk probability sequence, obtain the second correlation coefficient among all valid questions.

[0050] The expression for the second risk probability sequence is: In formula 4, Indicates the first The second risk probability sequence is formed by the output of the dedicated classification model for each valid problem; This indicates that the first The first valid question The risk probability is obtained by inputting each individual answer into a dedicated classification model; This indicates the number of independent responses to a valid question. .

[0051] The formula for calculating the second correlation coefficient is: In Formula 5, Indicates the first The first valid question and the first The second correlation coefficient between the effective questions; Indicates the first The mean of the second risk probability sequence formed by the output of the dedicated classification model for each valid problem; This indicates that the first The first valid question The risk probability is obtained by inputting each individual answer into a dedicated classification model; Indicates the first The mean of the second risk probability sequence formed by the output of the dedicated classification model for each valid problem.

[0052] Reference Figure 7 Another flowchart illustrating the method for obtaining the second correlation coefficient shows that when the number of categories in the dedicated classification model is greater than 2, "obtaining the second correlation coefficient between all valid questions based on the dedicated classification model corresponding to each valid question" includes the following steps: S373: Input each subject's independent answer to each valid question into the corresponding dedicated classification model, and form a second risk probability matrix based on the output of the dedicated classification model; S374: Based on the second risk probability matrix, obtain the covariance matrix of each valid problem and the covariance matrix between every two problems; S375: By solving the generalized eigenvalue problem, the second correlation coefficient between all valid problems is obtained based on the covariance matrix of each valid problem and the covariance matrix between every two valid problems.

[0053] Steps S373-S375 are based on the same principle as steps S333-S335 above. A second generalized eigenvalue problem function is constructed based on the covariance matrix of every two valid problems and the covariance matrix between every two valid problems. By solving the second generalized eigenvalue problem function, the maximum eigenvalue between every two valid problems is obtained. The square root of the maximum eigenvalue between every two valid problems is then calculated to obtain the second correlation coefficient between every two valid problems. The expression for the second generalized eigenvalue problem function is: In Formula 6, Indicates the first The inverse matrix of the covariance matrix of an effective problem; Indicates the first The first valid question and the first The covariance matrix among the effective problems; Indicates the first The inverse matrix of the covariance matrix of an effective problem; Indicates the first The first valid question and the first The covariance matrix among the effective problems; Indicates the first The first valid question and the first The eigenvectors of the generalized eigenvalue problem corresponding to each valid problem; Indicates the first The first valid question and the first The eigenvalues ​​of the generalized eigenvalue problem corresponding to a valid problem.

[0054] In S38, when the second correlation coefficient between two valid questions is greater than a preset second correlation coefficient threshold (e.g., 0.8), it indicates that there is redundancy between the two valid questions. Therefore, the question with the higher classification performance metric is selected as the optimal question. For example, if the prediction results of question X ("What year is it?") and question Y ("What season is it now?") are highly correlated, it means that they may be detecting the same time orientation ability. Only one of them needs to be retained. In this way, it can be ensured that the final question set covers multiple key cognitive domains (such as memory, attention, language, executive function, orientation, etc.) while minimizing internal information redundancy. It is evident that by using this pairwise comparison and selection method, multiple optimal questions with strong complementarity can be obtained. In the field of cognitive impairment recognition, the final number of optimal questions is usually 10-12. Furthermore, these optimal questions can form a standardized and rapid screening set of optimal questions that is concise, efficient, has a short total response time (e.g., within 2 minutes), and possesses the strongest comprehensive discrimination ability. This process ensures that the input content analyzed by the AI ​​model is homogeneous, standardized, and has high information entropy, fundamentally solving the performance uncertainty problem caused by unstructured speech input.

[0055] In S4, after obtaining the optimal problem, we trace back to which candidate scale the optimal problem originated from and use the specific classification model of that candidate scale as the target model.

[0056] To facilitate the construction of fusion models, the output layer of the dedicated classification model needs to include a linear transformation function and an activation function. The linear transformation function is used to generate Logits values, and the activation function is used to convert the Logits values ​​into a classification output form (probability or other suitable classification output form).

[0057] A fusion model is built based on a feedforward neural network. The fusion model includes an input layer, hidden layers, and an output layer. The number of classes in the target model determines the dimension of the Logits feature vector constructed by the input layer of the fusion model, and also determines the number of neurons in the output layer. When the target model has two classes, the input layer of the fusion model generates Logits values ​​from all target models to form an N-dimensional Logits feature vector, which is then transmitted to the hidden layer of the fusion model. When the target model has more than two classes, the input layer generates Logits values ​​from all target models to form an M×N-dimensional Logits feature vector, which is then transmitted to the hidden layer of the fusion model. Regardless of whether the target model has two or more classes, the hidden layer of the fusion model transforms the received Logits feature vector into a comprehensive Logits value, which is then transmitted to the output layer of the fusion model. The hidden layer contains one or more fully connected (dense) layers. The number of nodes and layers in the fully connected layers are hyperparameters that can be optimized experimentally. Each fully connected layer is followed by a non-linear activation function such as ReLU. Through the combination of fully connected layers and non-linear activation functions, multi-dimensional Logits values ​​can be transformed and combined to extract higher-level feature representations, ultimately integrating them into one or more comprehensive Logits values. For binary classification problems, the hidden layers typically integrate a single comprehensive Logits value; for multi-class classification problems, they typically integrate multiple comprehensive Logits values. In short, the number of comprehensive Logits values ​​equals the number of classification categories. Comprehensive Logits values ​​can more effectively represent the key features of the input data, improving the model's ability to recognize and classify complex patterns. In short, this process enhances the model's non-linear fitting ability to the input data. The output layer of the fusion model is used to convert the received comprehensive Logits values ​​into a classification output form. The output layer contains one or more neurons, specifically the number of neurons equal to the number of classification categories in the target model, and a pre-defined activation function is used after each neuron to output the classification result.For binary classification problems, the output layer contains one neuron (the target model has two categories). The default activation function is the sigmoid activation function. For a given logits value, the neuron is converted into a probability value between 0 and 1 by the sigmoid activation function, representing the probability of belonging to the positive class. For multi-class classification problems (the target model has more than two categories), the output layer contains multiple neurons. The default activation function is the softmax activation function. For multiple combined logits values, the softmax activation function is converted into a probability distribution, with each category corresponding to a probability value, and the sum of the probabilities of all categories is 1.

[0058] In S5 and S6, all target models and fusion models together form a fusion classification model. Users answer all the optimal questions in the optimal question set, and the answers are input into the corresponding target models. The target models generate Logits values, which are then input into the fusion model. By concatenating the Logits values ​​in the fusion model, a more accurate classification result can be obtained. (See reference...) Figure 8 The flowchart of the fusion classification model performing an actual classification task is shown. Taking binary classification as an example, feature fusion is performed through Logits to obtain fused features (i.e., comprehensive Logits values), which are then used for prediction to obtain classification probabilities. The classification probabilities are compared with preset thresholds to achieve discrimination and classification results.

[0059] This method extracts the Logits result sequences from the original models corresponding to each question as new feature vectors, which are then input into a specially designed fusion model for training. The core of this fusion model is to comprehensively analyze the Logits features generated by different questions (corresponding to different cognitive domains) to provide a more accurate overall cognitive impairment risk score. This architecture employs a model stacking approach, its ingenuity lying in the feature selection (Logits) and the strategy of preserving the original weak-risk classification model in the first stage. This hierarchical fusion architecture achieves a "1+1>2" effect, utilizing the powerful feature extraction capabilities of the original model while optimizing the combination and weighting of condensed features through the second-stage fusion network. This ensures that the final screening results reach or surpass existing state-of-the-art levels in key indicators such as accuracy, sensitivity, and specificity, effectively reducing the risk of missed diagnoses and misdiagnoses.

[0060] The training process of the fusion model employs an efficient and stable optimizer, such as Adam or its variant AdamW. The initial learning rate can be set to an empirical value, such as 0.001, and a learning rate decay strategy (such as cosine annealing) is used to achieve finer convergence in the later stages of training.

[0061] Considering the potential class imbalance in classification tasks, such as in cognitive impairment screening where healthy individuals (negative samples) far outnumber patients (positive samples), this method employs Focal Loss (Focal Loss for Dense Object Detection) to address this issue. This allows for customization of performance metrics, particularly in screening tasks where the weights of sensitivity (recall, the ability to correctly identify patients) and specificity (the ability to correctly identify healthy individuals) need adjustment based on the application scenario. For instance, in initial screening for the general public, high sensitivity is preferred, even accepting a certain false positive rate, rather than missing any potential patients. In clinical auxiliary diagnosis, higher specificity may be needed to avoid misdiagnosis. This method introduces a moderating factor into Focal Loss to dynamically reduce the weight of correctly classified samples (especially easily classified negative samples), allowing the model to focus more on difficult-to-distinguish samples (usually minority class positive samples) during training. By adjusting alpha (class weights) and gamma (focusing parameters) in Focal Loss, a flexible trade-off can be struck between sensitivity (recall, the ability to correctly identify patients) and specificity (the ability to correctly identify healthy individuals). This is crucial for developing a clinically practical screening tool. For example, in early screening scenarios where the goal is "better to err on the side of caution than to miss an opportunity," parameters can be adjusted to prioritize improving the model's sensitivity. Since the fusion model serves as the final decision-making model, it needs to make a final judgment on whether a subject has cognitive impairment based on the outputs of multiple target models. Focal Loss is a core tool for optimizing the performance of the fusion model and improving patient identification accuracy, and therefore, it must be employed. The loss function of the fusion model, which includes adjustment factors, is as follows: In Formula 7, This represents the value of the loss function for the fusion model; This represents the total number of samples in the training dataset of the fusion model; Indicates the number of classification categories in the fusion model; The fusion model represents the first... The weights of each category, , This is used to adjust the loss ratio for different categories. If the sample size for a certain category is extremely small, a larger ratio can be set for that category. This allows the model to prioritize the minority class during training, avoiding a bias towards the majority class due to its larger sample size, which could lead to missed diagnoses of the minority class. If the class distribution is uniform, all classes can be considered. Setting it to 1 degenerates into "loss without class weights"; The fusion model predicts the first... The sample belongs to the first The probability of each category. ; Indicates the focus parameter; Indicates the fusion model for the first The sample belongs to the first Indicator functions for each category; As a regulating factor, The fusion model's ability to "focus on difficult-to-classify samples" has two core functions: first, it targets easily classifiable samples (samples with high prediction accuracy). ,but The loss on these samples will be significantly reduced, and the fusion model will not waste training effort on samples that can be easily identified, but rather on difficult-to-distinguish samples (samples whose prediction is uncertain). (If the fusion model has only a 50% confidence level in classifying it as "mild impairment"), then it will maintain a higher value (e.g. At that time, (1−0.5) 2 =0.25), the loss for these samples will be amplified, forcing the fusion model to prioritize learning these "easily misclassified / missed" samples. It is evident that by setting an adjustment factor in the loss function used in the fusion model, the problem of class imbalance can be solved.

[0062] The loss function for the dedicated classification model can also be FocalLoss. The dedicated classification model's role in the overall process is logits feature extraction. Feature extraction should be as general as possible, minimizing manual intervention. Therefore, the simple and stable standard cross-entropy loss function is sufficient, ensuring both the generality of the logits features and providing suitable input for subsequent iterations. The loss function (standard cross-entropy loss function) for the dedicated classification model is as follows: In Formula 8, This represents the value of the loss function for the specific classification model; This represents the total number of samples in the training dataset of the dedicated classification model; Indicates the number of categories in the dedicated classification model; This indicates that the specific classification model is for the first... The sample belongs to the first Indicator functions for each category, , ; This indicates that the specific classification model predicts the first... The sample belongs to the first The probability of each category.

[0063] In summary, this disclosure proposes an innovative rapid classification method and system based on standardized interactive content and multi-model fusion analysis, which overcomes the inefficiencies of traditional scale screening and the unreliability of emerging AI voice analysis in existing technologies. The overall technical framework of this method includes a front-end data acquisition application and a back-end intelligent analysis engine. The front-end application is a user interface deployed on smartphones, tablets, or web pages, presenting questions to users in a clear and standardized manner. The user's voice response is recorded and cached separately for each question, and the recorded audio data packets are securely uploaded to the back-end server. Depending on actual needs, the recording can be modified to typed text, and the user's text response can be uploaded to the back-end server. For long-term cognitive screening, this method's use of recording allows users to conduct self-testing anytime, anywhere, completely eliminating the dependence of traditional screening on specific locations (such as hospitals and clinics) and professionals. Users do not need to make appointments or travel, greatly reducing the time and economic costs of screening and significantly improving service accessibility. This has immeasurable social value, especially for elderly people with mobility difficulties or those living in remote areas. Furthermore, the "standardized interaction" design ensures that the AI ​​model receives voice data responses to the same set of cognitive stimuli (questions) regardless of when, where, or by whom it is tested. This guarantees the homogeneity and high quality of the input data from the source, effectively eliminating interference from individual differences and environmental factors. The system records and uploads the answer to each question as an independent audio segment, which facilitates front-end process control and provides a guarantee for accurate analysis by the back-end model. It allows the model to independently analyze the voice responses for each specific cognitive domain (such as memory, computing power, and language fluency), avoiding speech aliasing and semantic cross-interference in continuous dialogue.

[0064] The backend analysis engine, deployed on a cloud server, is the "brain" of the system. Its speech preprocessing unit performs noise reduction, voice detection (VAD), and standardization on the uploaded raw audio. The weak-risk model library stores N target models that correspond one-to-one with the final question set. The Logits feature generator schedules and executes model inference, taking the speech input of each question as the corresponding model, capturing the Logits output, and aggregating it into a feature vector. The fusion classification model receives the Logits feature vector, performs the final risk probability calculation, and outputs classification results such as cognitive impairment risk assessment results. This solution adopts a carefully designed two-stage modeling strategy. First, a short question set is extracted from massive clinical scale data to standardize the front-end input. Then, an innovative Logits feature fusion neural network model is used to deeply analyze the standardized input. The Logits result is the raw output without activation function processing. It is an intermediate feature in the inference process, rather than the final probability value. Therefore, Logits retains more subtle information about the model's judgment of "confidence" and "propensity" than the final 0-1 probability value. While ensuring screening speed and convenience, it achieves accuracy and stability comparable to or even surpassing traditional clinical screening.

[0065] Furthermore, after obtaining the simplified question set, the model was not retrained based on these few questions. Instead, the original proprietary classification model trained on the complete scale data was continued to be used. The advantage of doing so is that the original model learns a richer and more comprehensive complex mapping relationship between speech features and cognitive states. Retraining directly on a small sample question set can easily lead to model overfitting and poor generalization ability.

[0066] This method redesigns the classification system, transforming it from a rigid, "black box" into a professional tool that can be fine-tuned as needed. It provides customized model versions for different medical policies, screening strategies, or clinical pathways, significantly enhancing its real-world deployment capabilities and practical value. This flexibility is unmatched by traditional fixed-threshold scales and ordinary AI models. These interconnected innovations collectively construct a rapid, convenient, accurate, and reliable new solution for cognitive impairment screening, perfectly integrating the convenience of digital platforms with standardized clinical assessments, possessing extremely high technological value and broad application prospects.

[0067] Reference Figure 9 This disclosure provides a question-and-answer-based classification system, including: The set acquisition module 101 is used to acquire a candidate scale set, which contains multiple candidate scales, and each candidate scale contains multiple candidate questions. The first building module 102 is used to build a dedicated classification model for each candidate scale; The question filtering module 103 is used to filter candidate scales and candidate questions in candidate scales based on the performance of the dedicated classification model, obtain multiple optimal questions and construct an optimal question set. The second construction module 104 is used to take the dedicated classification model corresponding to each optimal problem as the target model and construct a fusion model based on the number of classification categories of the target model. Model composition module 105 is used to combine all target models and fusion models into a fusion classification model; Task execution module 106 is used to execute actual classification tasks based on the optimal problem set and fusion classification model.

[0068] The various variations and specific examples of the question-and-answer-based classification method provided above are also applicable to the question-and-answer-based classification system provided in this disclosure. Through the foregoing detailed description of the question-and-answer-based classification method, those skilled in the art can clearly understand the implementation method of the question-and-answer-based classification system. For the sake of brevity, they will not be described in detail here.

[0069] A computer device according to embodiments of the present disclosure includes a memory and a processor. The memory is used to store non-transitory computer-readable instructions. Specifically, the memory may include one or more computer program products, which may include various forms of computer-readable storage media, such as volatile memory and / or non-volatile memory. The volatile memory may, for example, include random access memory (RAM) and / or cache memory. The non-volatile memory may, for example, include read-only memory (ROM), hard disk, flash memory, etc.

[0070] The processor may be a central processing unit (CPU) or other form of processing unit with data processing capabilities and / or instruction execution capabilities, and may control other components in the computer device to perform desired functions. In one embodiment of this disclosure, the processor is used to execute computer-readable instructions stored in the memory, causing the computer device to perform all or part of the steps of the problem-solving-based classification method of the foregoing embodiments of this disclosure.

[0071] Those skilled in the art will understand that, in order to solve the technical problem of how to achieve a good user experience, this embodiment may also include well-known structures such as communication buses and interfaces, and these well-known structures should also be included within the protection scope of this disclosure.

[0072] like Figure 10 This is a schematic diagram of a computer device provided for an embodiment of the present disclosure. It illustrates a structural schematic diagram suitable for implementing the computer device in the embodiments of the present disclosure. Figure 10The computer device shown is merely an example and should not be construed as limiting the functionality and scope of the embodiments disclosed herein.

[0073] like Figure 10 As shown, a computer device may include a processor (such as a central processing unit, graphics processing unit, etc.), which can perform various appropriate actions and processes based on programs stored in read-only memory (ROM) or programs loaded from storage devices into random access memory (RAM). The RAM also stores various programs and data required for the operation of the computer device. The processor, ROM, and RAM are interconnected via a bus. Input / output (I / O) interfaces are also connected to the bus.

[0074] Typically, the following devices can be connected to the I / O interface: input devices, such as sensors or visual information acquisition devices; output devices, such as displays; storage devices, such as magnetic tapes or hard drives; and communication devices. Communication devices allow the computer device to communicate wirelessly or wiredly with other devices (such as edge computing devices) to exchange data. Although Figure 10 A computer apparatus with various devices is shown, but it should be understood that it is not required to implement or have all of the devices shown. More or fewer devices may be implemented or included alternatively.

[0075] In particular, according to embodiments of this disclosure, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments of this disclosure include a computer program product comprising a computer program carried on a non-transitory computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via a communication device, or installed from a storage device, or installed from ROM. When the computer program is executed by a processor, all or part of the steps of the question-and-answer-based classification method of embodiments of this disclosure are performed.

[0076] For a detailed description of this embodiment, please refer to the corresponding descriptions in the foregoing embodiments, which will not be repeated here.

[0077] A computer-readable storage medium according to embodiments of the present disclosure stores non-transitory computer-readable instructions. When these non-transitory computer-readable instructions are executed by a processor, all or part of the steps of the problem-solving-based classification methods described in the foregoing embodiments of the present disclosure are performed.

[0078] The aforementioned computer-readable storage media include, but are not limited to: optical storage media (e.g., CD-ROM and DVD), magneto-optical storage media (e.g., MO), magnetic storage media (e.g., magnetic tape or portable hard drive), media with built-in rewritable non-volatile memory (e.g., memory card), and media with built-in ROM (e.g., ROM cartridge).

[0079] For a detailed description of this embodiment, please refer to the corresponding descriptions in the foregoing embodiments, which will not be repeated here.

[0080] The basic principles of this disclosure have been described above with reference to specific embodiments. However, it should be noted that the advantages, benefits, and effects mentioned in this disclosure are merely examples and not limitations, and should not be considered as essential features of each embodiment of this disclosure. Furthermore, the specific details disclosed above are for illustrative and facilitative purposes only, and are not limitations. These details do not limit the scope of this disclosure to the necessity of employing the aforementioned specific details for implementation.

[0081] In this disclosure, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. The block diagrams of devices, apparatuses, devices, and systems involved in this disclosure are merely illustrative examples and are not intended to require or imply that they must be connected, arranged, or configured in the manner shown in the block diagrams. As those skilled in the art will recognize, these devices, apparatuses, devices, and systems can be connected, arranged, and configured in any manner. Words such as "comprising," "including," "having," etc., are open-ended terms meaning "including but not limited to," and are used interchangeably with them. The terms "or" and "and" as used herein refer to the terms "and / or," and are used interchangeably with them unless the context clearly indicates otherwise. The term "such as" as used herein refers to the phrase "such as but not limited to," and is used interchangeably with it.

[0082] Additionally, as used herein, the "or" used in a list of items beginning with "at least one" indicates a separate list, such that a list of, for example, "at least one of A, B, or C" means A or B or C, or AB or AC or BC, or ABC (i.e., A and B and C). Furthermore, the word "exemplary" does not imply that the described example is preferred or better than other examples.

[0083] It should also be noted that in the systems and methods of this disclosure, the components or steps can be decomposed and / or recombined. These decompositions and / or recombinations should be considered as equivalent solutions to this disclosure.

[0084] Various changes, substitutions, and modifications can be made to the technology described herein without departing from the teachings defined by the appended claims. Furthermore, the scope of the claims of this disclosure is not limited to the specific aspects of the processes, machines, manufactures, events, means, methods, and actions described above. Currently existing or later-developed processes, machines, manufactures, events, means, methods, or actions that perform substantially the same function or achieve substantially the same result as the corresponding aspects described herein can be utilized. Therefore, the appended claims include such processes, machines, manufactures, events, means, methods, or actions within their scope.

[0085] The above description of the disclosed aspects is provided to enable any person skilled in the art to make or use this disclosure. Various modifications to these aspects will be readily apparent to those skilled in the art, and the general principles defined herein may be applied to other aspects without departing from the scope of this disclosure. Therefore, this disclosure is not intended to be limited to the aspects shown herein, but rather to be carried out within the widest scope consistent with the principles and novel features disclosed herein.

[0086] The above description has been given for purposes of illustration and description. Furthermore, this description is not intended to limit the embodiments of this disclosure to the forms disclosed herein. Although numerous exemplary aspects and embodiments have been discussed above, those skilled in the art will recognize certain variations, modifications, alterations, additions, and sub-combinations therein.

Claims

1. A method of classifying based on problem solving, characterized by, The method comprises the following steps: obtaining a candidate scale set comprising a plurality of candidate scales, each candidate scale comprising a plurality of candidate questions; constructing a dedicated classification model for each candidate scale; based on the performance of the dedicated classification model, screening the candidate scales and the candidate questions in the candidate scales to obtain a plurality of optimal questions and construct an optimal question set; using the dedicated classification model corresponding to each optimal question as a target model, constructing a fusion model based on the number of classification categories of the target model; combining all target models and fusion models into a fusion classification model; based on the optimal question set and the fusion classification model, performing an actual classification task.

2. The problem-solving based classification method of claim 1, wherein, The step of screening the candidate scales and the candidate questions in the candidate scales based on the performance of the dedicated classification model to obtain a plurality of optimal questions and construct an optimal question set comprises the following steps: performing performance evaluation on each dedicated classification model to obtain a performance indicator of each dedicated classification model; using the candidate scale adopted by the dedicated classification model whose performance indicator meets the standard as an effective scale; based on the dedicated classification model of each effective scale, obtaining a first correlation coefficient between all effective scales; based on the first correlation coefficient, screening a plurality of target scales from all effective scales; obtaining the independent answer content of each candidate question in each target scale by a subject, inputting each independent answer content into the corresponding model, and obtaining a classification performance measurement value of each candidate question in each target scale; based on the classification performance measurement value, screening a plurality of effective questions from the candidate questions in all target scales; based on the dedicated classification model corresponding to each effective question, obtaining a second correlation coefficient between all effective questions; based on the second correlation coefficient, screening a plurality of optimal questions from all effective questions, and combining all optimal questions into an optimal question set.

3. The problem-solving based classification method of claim 2, wherein, The step of obtaining a first correlation coefficient between all effective scales based on the dedicated classification model of each effective scale comprises the following steps: if the number of classification categories of the dedicated classification model is 2, obtaining a first risk probability sequence output by the dedicated classification model of each effective scale on the corresponding test data set; based on the first risk probability sequence, obtaining a first correlation coefficient between all effective scales; if the number of classification categories of the dedicated classification model is greater than 2, obtaining a first risk probability matrix output by the dedicated classification model of each effective scale on the corresponding test data set; based on the first risk probability matrix, obtaining a covariance matrix of each effective scale and a covariance matrix between each two effective scales; using the generalized eigenvalue problem, obtaining a first correlation coefficient between all effective scales based on the covariance matrix of each effective scale and the covariance matrix between each two effective scales.

4. The problem-solving based classification method of claim 3, wherein, The expression of the first risk probability sequence is: In the formula, Indicates the first The first risk probability sequence output by the dedicated classification model of each effective scale on the corresponding test dataset; Indicates the first A dedicated classification model for each valid scale is used for the first valid scale in the corresponding test dataset. The risk probability output for each sample; This indicates the number of samples in the test dataset. ; The calculation formula of the first correlation coefficient is: wherein, denotes a first correlation coefficient between the first effective scale and the first effective scale; denotes a mean of a first sequence of risk probabilities output by the exclusive classification model of the first effective scale on the respective test data set; denotes a risk probability output by the exclusive classification model of the first effective scale for the first sample in the respective test data set; denotes a mean of a first sequence of risk probabilities output by the exclusive classification model of the first 5. The problem-solving based classification method of claim 2, wherein, The step of obtaining a second correlation coefficient between all effective questions based on the dedicated classification model corresponding to each effective question comprises the following steps: If the number of classification categories of the specialized classification model is 2, the independent answer content of each subject to each valid question is input into the corresponding specialized classification model, and a second risk probability sequence is formed based on the output of the specialized classification model; Based on the second risk probability sequence, a second correlation coefficient between all valid questions is obtained; If the number of classification categories of the specialized classification model is greater than 2, the independent answer content of each subject to each valid question is input into the corresponding specialized classification model, and a second risk probability matrix is formed based on the output of the specialized classification model; Based on the second risk probability matrix, a covariance matrix of each valid question and a covariance matrix between each two questions are obtained; Using the generalized eigenvalue problem, a second correlation coefficient between all valid questions is obtained based on the covariance matrix of each valid question and the covariance matrix between each two valid questions.

6. The question answer-based classification method of claim 1, wherein The output layer of the specialized classification model comprises a linear transformation function and an activation function; The linear transformation function is used to generate Logits values, and the activation function is used to convert the Logits values into a classification output form.

7. The problem-solving based classification method of claim 1, wherein, The fusion model comprises an input layer, a hidden layer, and an output layer; When the number of classification categories of the target model is 2, the input layer of the fusion model is used to form an N-dimensional Logits feature vector by using the Logits values generated by all target models, and the Logits feature vector is transmitted to the hidden layer of the fusion model; When the number of classification categories of the target model is greater than 2, the input layer of the fusion model is used to form an MxN-dimensional Logits feature vector by using the Logits values generated by all target models, and the Logits feature vector is transmitted to the hidden layer of the fusion model; Regardless of whether the number of classification categories of the target model is 2 or greater than 2, the hidden layer of the fusion model is used to convert the received Logits feature vector into a comprehensive Logits value, and the comprehensive Logits value is transmitted to the output layer of the fusion model; The output layer of the fusion model is used to convert the received comprehensive Logits value into a classification output form.

8. The problem-solving based classification method of claim 1, wherein, The expression of the loss function of the fusion model is: In the formula, represents the value of the loss function of the fusion model; represents the total number of samples of the training data set of the fusion model; represents the number of classification categories of the fusion model; represents the weight of the first classification category of the fusion model, ; represents the probability that the first sample belongs to the first classification category of the fusion model, ; represents the focusing parameter; represents the indicator function of the fusion model that the first sample belongs to the first classification category; is an adjustment factor.

9. A computer apparatus, comprising: The computer device comprises: at least one processor; and a memory connected in communication with the at least one processor; wherein The memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to perform the question answer-based classification method of any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, The computer readable storage medium stores computer instructions for causing a computer to perform the question answer-based classification method of any one of claims 1-7.

11. A computer program product comprising computer instructions, characterized in that, The computer instructions, when executed by a processor, implement the steps of the question answer-based classification method of any one of claims 1-7.

Images