Processing method, system and equipment for intelligent questionnaire survey and medium

Abstract

The invention discloses a processing method, system and device for intelligent questionnaire survey and a medium, and the method specifically comprises the steps: obtaining questionnaire data and evaluation parameter setting inputted by a user, calling a main model to carry out universal dimension inspection on a questionnaire, and obtaining a preliminary analysis result; according to a preliminary analysis result, distributing different types of problems to corresponding expert models for deep analysis; in combination with culture feature data inquired in real time, a culture suitability score is generated to supplement an analysis result of the expert model; based on the analysis results of the plurality of expert models, performing weighted fusion processing on the score of each dimension by using a weight coefficient dynamically adjusted through a reinforcement learning mechanism to obtain a comprehensive evaluation score; the comprehensive evaluation score is returned to the user interface, feedback data are collected after the user completes interaction, and the feedback data are used for iteratively updating model parameters and setting evaluation parameters. According to the invention, efficient and accurate examination of questionnaire contents is realized.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of artificial intelligence, and in particular to a processing method and system for intelligent questionnaire investigation, a device and a medium. BACKGROUND

[0002] In the current questionnaire editing field, questionnaire quality is the core element that determines the accuracy and effectiveness of the research results. A high-quality questionnaire can accurately obtain the required information and provide reliable basis for subsequent data analysis and decision-making; while a low-quality questionnaire may lead to data bias, information loss, and thus affect the value of the entire research.

[0003] Traditionally, questionnaire checking mainly relies on manual auditing. However, this approach has many drawbacks. On the one hand, manual auditing requires a large amount of human resources and time cost, and the auditing personnel need to check the questionnaire content word by word, which is extremely time-consuming and labor-intensive for large-scale questionnaire research projects. On the other hand, manual auditing is easily influenced by the subjective factors of the auditing personnel, and different auditing personnel may have different understanding and judgment standards for questionnaire quality, resulting in lack of stability and consistency of the auditing results, and difficulty in ensuring the objectivity and fairness of the questionnaire quality evaluation.

[0004] With the development of technology, some automatic questionnaire checking tools based on simple rules have also appeared. These tools, although to some extent, improve the efficiency of questionnaire checking, can quickly check the format and basic logic of the questionnaire, such as the number of questions, the integrity of options, the correctness of logical jumps, etc. However, their functional limitations are great, and their depth analysis ability for questionnaire content is severely insufficient. These tools cannot deeply understand the intent of the questionnaire questions, the quality of language expression, the logical coherence, and the rationality of option setting, etc. key dimensions, and are difficult to provide targeted and practical optimization suggestions, and cannot meet the demand for high-quality questionnaires.

[0005] In addition, in the aspect of using AI large models for questionnaire checking, there are also some problems to be solved. On the one hand, there is a lack of a convenient and effective way to call AI large models to check the questionnaire content, and to automatically give accurate scores and detailed modification suggestions, in order to improve the efficiency and effectiveness of questionnaire checking. On the other hand, different types of questionnaires have different characteristics and needs, but the existing way of using large models to check questionnaires often uses the same checking dimensions for all types of questionnaires, and checks according to the dimensions described by the fixed system prompt words, which cannot conduct targeted analysis and evaluation according to the specific type of the questionnaire, cannot fully exert the advantages of large models in questionnaire checking, and limits the improvement of questionnaire quality. SUMMARY

[0006] The present application aims to provide a processing method, system, device and medium for questionnaire intelligent investigation, which realizes efficient and accurate checking of questionnaire content, dynamically adjusts checking dimensions and weights according to questionnaire types, generates comprehensive evaluation results with cultural adaptability, and iteratively updates model parameters by collecting user feedback data, so as to solve at least one of the above technical problems.

[0007] In a first aspect, the present application provides a processing method for questionnaire intelligent investigation, which specifically comprises: obtaining questionnaire data and evaluation parameter settings input by a user, calling a main model through a Dify workflow engine to perform general dimension checking on the questionnaire, and obtaining preliminary analysis results; distributing different types of questions to corresponding expert models for in-depth analysis through an intelligent routing decision engine according to problem classification labels of the preliminary analysis results; generating cultural adaptability scores to supplement the analysis results of the expert models in combination with cultural feature data queried from a multi-source cultural feature database in real time; based on the analysis results of the multiple expert models, weighting and fusing the scores of each dimension by using weight coefficients dynamically adjusted through a reinforcement learning mechanism, and obtaining a comprehensive evaluation score; returning the comprehensive evaluation score to a user interface, and collecting feedback data after the user completes interaction, which is used to iteratively update model parameters and evaluation parameter settings.

[0008] In a second aspect, the present application provides a processing system for questionnaire intelligent investigation, which specifically comprises: a first processing module for obtaining questionnaire data and evaluation parameter settings input by a user, calling a main model through a Dify workflow engine to perform general dimension checking on the questionnaire, and obtaining preliminary analysis results; a second processing module for distributing different types of questions to corresponding expert models for in-depth analysis through an intelligent routing decision engine according to problem classification labels of the preliminary analysis results; a third processing module for generating cultural adaptability scores to supplement the analysis results of the expert models in combination with cultural feature data queried from a multi-source cultural feature database in real time; a fourth processing module for weighting and fusing the scores of each dimension by using weight coefficients dynamically adjusted through a reinforcement learning mechanism based on the analysis results of the multiple expert models, and obtaining a comprehensive evaluation score; a fifth processing module for returning the comprehensive evaluation score to a user interface, and collecting feedback data after the user completes interaction, which is used to iteratively update model parameters and evaluation parameter settings.

[0009] Thirdly, the present invention provides a computer device, comprising: a memory and a processor, and a computer program stored in the memory, wherein when the computer program is executed on the processor, it implements the processing method for intelligent questionnaire survey as described in any of the above methods.

[0010] Fourthly, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the processing method for intelligent questionnaire survey as described in any of the above methods.

[0011] Compared with the prior art, the present invention has at least one of the following technical effects: 1. This invention uses the Dify workflow engine to call the main model to perform general dimension checks on the questionnaire, realizing convenient one-click operation, quickly obtaining preliminary analysis results, improving the efficiency of questionnaire checking, and reducing the time and manpower costs of manual review.

[0012] 2. The different types of expert models in this invention can be professionally processed according to the specific characteristics of the questionnaire, providing a more in-depth and accurate evaluation, which solves the problem that existing automated tools are insufficient in their ability to deeply analyze questionnaire content.

[0013] 3. This invention enables questionnaire surveys to not only consider the quality of the questionnaire itself, but also take into account the cultural background of the target region, thereby generating more practical and adaptable assessment results and meeting the needs of questionnaire surveys in different cultural environments.

[0014] 4. The reinforcement learning mechanism of this invention can automatically adjust the weight coefficients according to the characteristics of the current assessment task, such as questionnaire type, application scenario and cultural background, so as to make the comprehensive evaluation more scientific and reasonable and accurately reflect the overall quality of the questionnaire.

[0015] 5. This invention makes targeted adjustments to the parameters of the master model and expert model, and dynamically optimizes the configuration parameters and content structure of the evaluation parameter settings, which can continuously improve the performance and accuracy of the model, enabling the questionnaire inspection method to continuously adapt to the needs of practical applications and continuously improve the quality of the questionnaire. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in the embodiments of this application, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 This is a flowchart illustrating a processing method for intelligent questionnaire surveys provided in an embodiment of the present invention; Figure 2 This is a schematic diagram of the structure of a processing system for intelligent questionnaire surveys provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of the structure of a computer device provided in an embodiment of the present invention. Detailed Implementation

[0018] In the following description, specific details such as particular system architectures and techniques are set forth for illustrative purposes and not for limitation, in order to provide a thorough understanding of the embodiments of this application. However, those skilled in the art will understand that this application may also be implemented in other embodiments without these specific details. In other instances, detailed descriptions of well-known systems, apparatuses, circuits, and methods have been omitted so as not to obscure the description of this application with unnecessary detail.

[0019] It should be understood that, when used in this application specification and the appended claims, the term "comprising" indicates the presence of the described features, integrals, steps, operations, elements and / or components, but does not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components and / or a collection thereof.

[0020] It should also be understood that the term “and / or” as used in this application specification and the appended claims means any combination of one or more of the associated listed items and all possible combinations, and includes such combinations.

[0021] As used in this application specification and the appended claims, the term "if" may be interpreted, depending on the context, as "when," "once," "in response to determination," or "in response to detection." Similarly, the phrase "if determined" or "if detected [the described condition or event]" may be interpreted, depending on the context, as meaning "once determined," "in response to determination," "once detected [the described condition or event]," or "in response to detection [the described condition or event]."

[0022] Furthermore, in the description of this application and the appended claims, the terms "first," "second," "third," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0023] References to "one embodiment" or "some embodiments" as described in this specification mean that one or more embodiments of this application include a specific feature, structure, or characteristic described in connection with that embodiment. Therefore, the phrases "in one embodiment," "in some embodiments," "in other embodiments," "in still other embodiments," etc., appearing in different parts of this specification do not necessarily refer to the same embodiment, but rather mean "one or more, but not all, embodiments," unless otherwise specifically emphasized. The terms "comprising," "including," "having," and variations thereof mean "including but not limited to," unless otherwise specifically emphasized.

[0024] In this application embodiment, the entity executing the process includes a terminal device. This terminal device includes, but is not limited to, devices capable of executing the methods disclosed in this application, such as servers, computers, smartphones, and tablets. Figure 1 A flowchart illustrating a processing method for intelligent questionnaire surveys according to an embodiment of the present invention is shown below in detail: S101: Obtain the questionnaire data and evaluation parameter settings input by the user, and use the Dify workflow engine to call the main model to perform general dimension checks on the questionnaire and obtain preliminary analysis results.

[0025] In this embodiment, a user interface is constructed, comprising two main input areas: one for inputting questionnaire data and the other for setting evaluation parameters. The questionnaire data input area offers multiple input methods to accommodate different user needs. The evaluation parameter setting area provides users with a series of configurable options. These options cover various aspects, such as the rigor of the review; users can choose between strict, moderate, or lenient modes, each corresponding to different review standards. The strict mode imposes more stringent requirements on the questionnaire, while the lenient mode offers greater flexibility. The options also include settings for the degree of emphasis on different general dimensions, such as questionnaire format specifications, clarity of question wording, and logical rationality. Users can assign different weight values ​​to each dimension according to their actual needs; a higher weight value indicates that the dimension is given more importance during the review process.

[0026] After users complete the questionnaire data input and evaluation parameter settings in the corresponding areas according to their research needs, they click the "Submit" button on the interface to transmit the input data and set parameters to the backend processing system. Upon receiving the submitted data and parameters, the backend processing system analyzes the questionnaire structure, identifies key elements such as the questions, options, and question numbers, and organizes them according to a certain logical relationship to construct a structured questionnaire data model. Simultaneously, it verifies the validity of the user-set evaluation parameters, checking whether the parameter value range is reasonable and whether the sum of the weights of each dimension is 100%. If any problems are found in the parameter settings, the system promptly provides feedback to the user, requiring them to reset the parameters until they are valid.

[0027] After data preprocessing, the system invokes the Dify workflow engine. The Dify workflow engine is a highly flexible and scalable framework that coordinates the work between various modules according to predefined process rules. Within the workflow engine, predefined process nodes call the main model for general dimension checks. Based on user-defined evaluation parameters, the workflow engine transmits the structured questionnaire data to the main model. The main model is a deep learning model trained on a large amount of questionnaire data, capable of checking the general dimensions of the questionnaire. Upon receiving the questionnaire data, the main model first checks the format and compliance of the questionnaire according to predefined general dimension checking rules, such as checking the continuity of question numbers, the completeness of options, and the overall neatness of the questionnaire layout. Then, it checks the clarity of the question wording, using natural language processing technology to analyze the accuracy of vocabulary usage, the fluency of sentences, and the absence of ambiguity. Next, it checks the logical rationality of the questionnaire, including the correctness of logical relationships between questions and the reasonableness of logical transitions. After completing all general dimension checks, the main model generates preliminary analysis results, which include the questionnaire's scores on each general dimension and descriptions of any existing problems. For example, in terms of formatting compliance, the main model might score 80 points and point out several formatting errors in the question numbers; in terms of question clarity, it might score 75 points and list some questions with unclear meanings. Finally, the main model returns the preliminary analysis results to the Dify workflow engine, which then transmits the results to subsequent processing modules for further analysis and processing.

[0028] S102, based on the problem classification labels of the preliminary analysis results, the intelligent routing decision engine distributes different types of problems to the corresponding expert models for in-depth analysis.

[0029] In this embodiment, an intelligent routing decision engine is constructed. This engine is a core module with intelligent decision-making capabilities, and it internally stores a detailed routing rule base. The routing rule base is formulated based on in-depth analysis of the characteristics of various questionnaire questions and the functional characteristics of expert models. In the rule base, for each question category label, a corresponding target expert model is clearly defined. For example, for the formatting issue tag, the rule base specifies expert model A, which specializes in handling questionnaire formatting problems. This model has been trained on a large number of formatting issue samples and can accurately identify and resolve formatting-related issues such as incorrect question numbers, non-standard font formatting, and messy layout. For the semantic ambiguity tag, the corresponding expert model is B, which excels in natural language understanding and semantic analysis. It can deeply analyze the meaning of the question statements, determine whether there are ambiguities or inappropriate word choices, and provide clear modification suggestions. For the logical contradiction tag, it is associated with expert model C, which has strong logical reasoning capabilities. This model can conduct a detailed analysis of the logical relationships between questions in the questionnaire, find logical inconsistencies and contradictions, and provide reasonable correction solutions. For the unreasonable option tag, it is matched with expert model D, which focuses on option setting evaluation. It can determine whether the options are comprehensive and reasonable, and whether there are any omissions of important options or unclear options, based on the intent of the question and the purpose of the survey.

[0030] Upon receiving the preliminary analysis results and their associated problem classification labels, the intelligent routing decision engine immediately initiates the decision-making process. The engine first parses the problem classification labels to clarify the problem type represented by each label. Then, it performs a rapid matching query in the routing rule base, finding the corresponding target expert model information based on the parsed problem type. For example, if a problem is labeled as semantically ambiguous, the engine quickly determines, by querying the rule base, that the problem should be dispatched to expert model B for processing.

[0031] After determining the target expert model, the intelligent routing decision engine organizes and encapsulates detailed information related to the question, including the original text content of the question, its location information in the questionnaire, and preliminary judgments on the question from the initial analysis results. The encapsulated information is transmitted in a standardized data format to ensure that the target expert model can accurately receive and parse it. Next, the intelligent routing decision engine establishes a communication connection with the target expert model. Through pre-defined communication protocols and data interfaces, the encapsulated question information is accurately sent to the corresponding expert model. For example, when sending information about a semantically ambiguous question to expert model B, the engine will transmit it according to the communication method and data format agreed upon with model B, ensuring the integrity and accuracy of the information during transmission.

[0032] Upon receiving the question information, the target expert model immediately initiates a deep analysis process. Taking expert model B's handling of semantically ambiguous questions as an example, this model utilizes advanced natural language processing technology to perform lexical, syntactic, and semantic analysis on the question text. By analyzing the collocation relationships between words, the grammatical structure of sentences, and the deep semantics expressed by the statements, the model can accurately identify the semantic ambiguities in the question. For example, for a question like "Do you frequently use this product for work?", the model can determine that the expression "frequently" is rather vague, and different people may have different interpretations. Therefore, it suggests replacing "frequently" with a specific frequency expression (such as "at least three times a week").

[0033] After completing the in-depth analysis, the target expert model generates a detailed analysis report. This report not only includes accurate diagnostic results of the problem but also provides targeted and actionable modification suggestions. For example, for logical contradiction problems, the report clearly identifies the specific location and cause of the contradiction and provides a description of the corrected logical relationship; for problems with unreasonable options, it lists the specific problems with the unreasonable options and provides optimized option settings.

[0034] Finally, the target expert model returns the generated analysis report to the intelligent routing decision engine. The engine performs preliminary organization and format standardization of the returned analysis report to facilitate subsequent integration and processing with the analysis results of other expert models. Simultaneously, the engine records the entire process information of problem distribution and processing, including problem distribution time, target expert model, processing time, etc., providing data support for subsequent system optimization and problem tracing.

[0035] S103, combining cultural feature data queried in real time from a multi-source cultural feature database, generates a cultural suitability score to supplement the analysis results of the expert model.

[0036] In this embodiment, after completing the in-depth analysis of the questionnaire using an expert model, the system initiates the cultural suitability assessment process. First, key information is extracted from the expert model's analysis results. This information includes aspects of the questionnaire that are highly relevant to the questionnaire's culture, such as the content of the questions, option settings, and question phrasing. For example, if the questionnaire includes a question about family values, the expert model analysis results will clearly specify the question's wording and options. Next, based on the extracted key information, the system determines the cultural feature dimensions to be queried. Taking the aforementioned family value question as an example, the system determines that the question involves cultural features related to family values ​​within the value category, and may also be related to different regional customs regarding family structure. Therefore, it determines to query relevant cultural feature items under the two broad categories of values ​​and customs. Then, the system sends a real-time query request to a multi-source cultural feature database. The query request includes the determined cultural feature dimensions and contextual information related to the questionnaire, such as the survey region and target audience, so that the database can more accurately return matching cultural feature data. Upon receiving the query request, the database performs rapid retrieval and matching within its vast dataset. By utilizing advanced database indexing technology and search algorithms, the database can efficiently find the cultural feature data most relevant to the query request and return this data to the system after organizing it in a certain format.

[0037] After receiving the returned cultural characteristic data, the system preprocesses this data. The preprocessing process includes data cleaning, which removes potentially erroneous data, duplicate data, and data irrelevant to the current questionnaire's cultural suitability assessment; and data standardization, which converts cultural characteristic data from different sources and in different formats into a standardized format that the system can process, such as converting textual descriptions of cultural characteristics into quantifiable indicators to facilitate subsequent analysis and scoring calculations.

[0038] A cultural fit analysis algorithm was employed to conduct in-depth analysis of the preprocessed cultural characteristic data and questionnaire content. This algorithm comprehensively considers multiple factors, including the degree of alignment between cultural characteristics and the intent of the questionnaire questions, and the impact of cultural characteristics on the rationality of the options. For example, when analyzing questions about family values, the algorithm assesses the degree of alignment between the family values ​​reflected in the questions and the mainstream family values ​​in the target survey area, as well as whether the options cover common types of family values ​​in that area. Through complex logical judgments and data analysis, the algorithm scores the matching degree of each cultural characteristic with the questionnaire content, with the scoring range set within a reasonable interval, such as 0-10 points, where a higher score indicates better cultural fit.

[0039] After scoring each cultural feature and questionnaire content, a corresponding weight is assigned to the score of each cultural feature based on the importance of different cultural features in the overall cultural fit assessment.

[0040] Based on the assigned weights, the scores for each cultural characteristic are weighted and summed to arrive at the final cultural suitability score. This score is a comprehensive indicator reflecting the degree of fit between the questionnaire content and the target cultural environment. The system integrates the generated cultural suitability score with the analysis results of the expert model to supplement and improve the expert model's analysis. For example, in the expert model's analysis report, in addition to the original evaluations of questionnaire question quality and logical rationality, a new cultural suitability score and related detailed explanations will be added, indicating which cultural characteristics the questionnaire performs well in, which aspects are lacking, and providing corresponding improvement suggestions, such as adjusting the question wording to better conform to local language habits and optimizing the option settings to cover more local cultural concepts, thereby providing users with a more comprehensive and accurate questionnaire quality assessment and analysis result.

[0041] S104, based on the analysis results of multiple expert models, uses weight coefficients dynamically adjusted through a reinforcement learning mechanism to perform weighted fusion processing on the scores of each dimension to obtain a comprehensive evaluation score.

[0042] In this embodiment, a reinforcement learning environment is constructed to simulate a real-world scenario of questionnaire quality assessment, which includes multiple virtual questionnaire assessment tasks. Each task corresponds to a questionnaire of different types and characteristics, covering multiple fields such as market research, academic research, and customer satisfaction surveys, to ensure that the environment can comprehensively reflect various situations in real-world applications. Within the environment, three key elements are defined: state, action, and reward. State represents relevant information about the current questionnaire assessment task, including questionnaire type, number of questions, and cultural background involved; action refers to the adjustment of weight coefficients assigned to the dimensional scores of each expert model analysis result; reward is given based on the closeness of the weighted fusion score to a pre-set ideal score. The closer the overall evaluation score is to the ideal score, the higher the reward, and vice versa.

[0043] Choose a suitable reinforcement learning algorithm, such as the Deep Q-Network (DQN) algorithm, and construct the corresponding neural network model. This neural network model receives state information as input and outputs suggestions for adjusting the weight coefficients based on the scores of each dimension of the expert model's analysis results. During the initialization phase, assign random values ​​to the parameters of the neural network model to enable it to perform preliminary learning and decision-making. Simultaneously, set up an experience replay buffer to store the state, action, and reward information encountered during the reinforcement learning process, for subsequent experience replay and learning optimization.

[0044] Once the reinforcement learning environment is running, for each virtual questionnaire assessment task, the system first acquires the current task's state information, including the questionnaire's specific characteristics and relevant assessment requirements. This state information is then input into the initialized neural network model. Based on the current state, the neural network model outputs a set of weight adjustment suggestions. These suggestions determine how to allocate weights to the dimensional scores of each expert model's analysis results. For example, if the current questionnaire is a market research questionnaire on high-end consumer goods, the neural network model might, based on the questionnaire's characteristics, suggest increasing the weights of dimensions related to consumer purchasing decisions.

[0045] Based on the weight coefficient adjustment suggestions output by the neural network model, the dimensional scores of the analysis results from each expert model are weighted and fused. Specifically, each dimension score is multiplied by its corresponding weight coefficient, and then the weighted scores of all dimensions are summed to obtain a preliminary comprehensive evaluation score. This preliminary comprehensive evaluation score reflects the overall quality assessment result of the questionnaire under the current weight allocation.

[0046] Based on the difference between the preliminary comprehensive evaluation score and the pre-set ideal score, corresponding rewards are given. If the preliminary comprehensive evaluation score is very close to the ideal score, it indicates that the current weight allocation is reasonable and can accurately assess the quality of the questionnaire, and the system gives a higher reward; conversely, if the difference is large, a lower reward or even a penalty is given. The current task status, the weight coefficient adjustment actions taken, and the reward information obtained are stored in the experience replay buffer.

[0047] As the reinforcement learning environment continues to operate, the system processes a large number of virtual questionnaire assessment tasks, accumulating rich experience data. At appropriate times, the system randomly selects a batch of experience data from the experience replay buffer for learning and optimization. Specifically, the extracted state information is input back into the neural network model to obtain predicted weight adjustment actions and corresponding expected rewards. Then, the predicted expected reward is compared with the actually stored reward, and the loss function value is calculated based on the comparison result. Through backpropagation, the parameters of the neural network model are adjusted according to the loss function value, enabling the neural network model to output more reasonable weight adjustment suggestions in subsequent tasks, thereby increasing the probability of obtaining higher rewards.

[0048] Through repeated reinforcement learning training and parameter optimization, the neural network model gradually learns to dynamically adjust the weight coefficients of the dimensional scores of each expert model analysis result based on different questionnaire status information. This allows the comprehensive evaluation score obtained after weighted fusion processing to more accurately reflect the actual quality of the questionnaire. Once the reinforcement learning model reaches a stable state, in the actual intelligent questionnaire survey processing method, for the real questionnaire data input by the user, after the main model's general dimension check and expert model's in-depth analysis, the system uses the weight coefficients output by the trained reinforcement learning model to perform weighted fusion processing on the dimensional scores of each expert model analysis result. Finally, it obtains an accurate and reliable comprehensive evaluation score, which is then returned to the user interface to provide the user with a comprehensive questionnaire quality assessment result.

[0049] S105, return the comprehensive evaluation score to the user interface, and collect feedback data after the user completes the interaction. The feedback data is used to iteratively update the model parameters and evaluation parameter settings.

[0050] In this embodiment, the comprehensive evaluation score and its corresponding evaluation level description are encapsulated into a specific data packet. This data packet contains not only the score and level information but also basic questionnaire information, such as the questionnaire name, questionnaire number, and creation time, to accurately display the evaluation results related to the questionnaire on the user interface. The system sends the encapsulated data packet to the user interface through a pre-defined communication protocol and data transmission interface. After receiving the data packet, the user interface clearly displays the comprehensive evaluation score and evaluation level in the appropriate location according to the designed layout and style. For example, an evaluation result display area is set up in a prominent position on the questionnaire details page, allowing users to see the quality assessment of the questionnaire immediately.

[0051] While displaying the overall evaluation score on the user interface, the system provides users with a series of interactive options. These options include, but are not limited to, accepting or rejecting the evaluation results, such as providing "Accept Evaluation" and "Disappear Evaluation" buttons. Users can also provide specific opinions and suggestions on the questionnaire quality evaluation. A text input box is provided on the user interface, where users can enter what they believe are problems with the questionnaire quality and directions for improvement. In addition, interactive functions for evaluating the quality of specific questions in the questionnaire are also provided. Users can click on a specific question in the questionnaire and then evaluate its quality, such as selecting evaluation options like "Question is clearly stated," "Question is vaguely stated," "Options are reasonably set," and "Options are unreasonable."

[0052] Once a user completes their interaction with the user interface, the system begins collecting feedback data. The system monitors the user's various actions on the interface in real time. Once a user triggers an interaction, such as clicking the "Approve / Evaluate" button, entering content in a text input box, or evaluating a specific question, the system integrates this action information along with relevant questionnaire information and user information. For example, when a user clicks the "Disapprove / Evaluate" button and enters the opinion "The overall logic of the questionnaire is unclear," the system records the user's disapproval of the current questionnaire and stores the questionnaire number, user account, and entered opinion. For user evaluations of specific questions, the system collects information such as the question number and the user's evaluation options for that question.

[0053] After collecting feedback data, the system preprocesses the data and uses this preprocessed data to iteratively update the model and evaluation parameters. For model parameter updates, the system employs incremental learning. Taking a reinforcement learning model as an example, the system analyzes the accuracy and reasonableness of user feedback regarding questionnaire quality assessments to identify issues with the current model's weight adjustment. If users generally believe that certain dimensions should have higher weights in specific types of questionnaires, but the current model assigns lower weights, the system adjusts the neural network parameters of the reinforcement learning model based on this feedback, enabling the model to more reasonably adjust weight coefficients when processing similar questionnaires in the future. For evaluation parameter settings, the system optimizes the dimensions and standards for questionnaire quality assessment based on user feedback. For example, if user feedback indicates that the current evaluation system ignores the significant impact of cultural suitability on overall quality, the system adds cultural suitability-related evaluation dimensions and weights to the evaluation parameter settings to more comprehensively and accurately assess questionnaire quality.

[0054] During the iterative update process, the system sets certain evaluation indicators to monitor the update effect. For example, the effectiveness of the model parameter update is judged by comparing the overall evaluation scores of a batch of test questionnaires before and after the update with the scores evaluated by human experts; for updates to evaluation parameter settings, the update effect is evaluated by observing whether users' acceptance of the updated evaluation results increases. If the update effect is not ideal, the system will continue to collect more feedback data and further adjust the update strategy until a satisfactory update effect is achieved, thereby continuously improving the accuracy and reliability of the intelligent questionnaire survey processing method.

[0055] In some embodiments, step S101 above, which involves calling the main model through the Dify workflow engine to perform a general dimension check on the questionnaire and obtain preliminary analysis results, specifically includes: Input the questionnaire data and evaluation parameter settings into the Dify workflow engine; In the workflow environment of the Dify workflow engine, the master analysis model is used to evaluate the questionnaire data in multiple dimensions to obtain preliminary evaluation results. The preliminary evaluation results include question intent recognition, language expression quality detection, logical coherence analysis, and option setting rationality evaluation. Based on the pre-set evaluation criteria system and combined with the preliminary evaluation results, the quality of each dimension of the questionnaire data is analyzed in parallel through the master analysis model, generating preliminary analysis results containing detailed scores and question classification labels.

[0056] In this embodiment, after the Dify workflow engine receives the questionnaire data and evaluation parameter settings, it enters its workflow environment. Within the workflow environment, the main analytics model is activated and begins multi-dimensional feature evaluation of the questionnaire data.

[0057] For the question intent identification dimension, the main analytics model uses natural language processing technology to perform semantic analysis on each question in the questionnaire. The model identifies keywords, key phrases, and sentence structures in the questions, and determines the core intent of the questions by comparing and reasoning with a pre-built semantic knowledge base. For example, for a question like "How satisfied are you with this product?", the model can identify that its intent is to understand users' satisfaction ratings of the product.

[0058] In terms of language expression quality assessment, the model evaluates the fluency, grammatical correctness, and vocabulary accuracy of the questions. It checks for grammatical errors and whether appropriate vocabulary is used to express the intended meaning. For example, in the question "Do you think this product is good or bad?", the model can detect repetitive word usage, which affects the quality of language expression.

[0059] In the logical coherence analysis dimension, the model analyzes the logical relationships between questions in the questionnaire. It checks whether the questions are arranged in a logical order and whether there are any logical jumps or contradictions. For example, in a questionnaire about shopping experience, if the user is first asked whether they purchased a product, and then directly asked about their satisfaction with after-sales service, while omitting the question about whether the user accepted after-sales service, this would be a problem of logical inconsistency.

[0060] Under the option setting rationality assessment dimension, the model checks the options for each question based on parameters such as the range of option quantity and the uniformity of option distribution. It determines whether the number of options is within a reasonable range, whether the options cover all possible scenarios, and whether the options are clearly differentiated. For example, for a survey question about age groups, if the options only include "under 20" and "over 30," the 20-30 age group is omitted, making the option setting unreasonable.

[0061] Through the above multi-dimensional feature evaluation, the main analysis model obtains preliminary evaluation results, which include information on problem intent recognition, language expression quality detection, logical coherence analysis, and option setting rationality evaluation.

[0062] After obtaining the preliminary evaluation results, the system, based on a pre-set evaluation standard system and in conjunction with the preliminary evaluation results, conducts parallel analysis of the quality of each dimension of the questionnaire data through the master analysis model. The pre-set evaluation standard system is derived from extensive questionnaire data analysis and expert experience summarization, and it sets detailed scoring rules and question classification labels for each dimension. For example, in the question intent recognition dimension, if the model can accurately identify the core intent of the question and it perfectly matches the pre-set standard intent, a high score is given; if the identification result has some deviation but does not affect the overall understanding, a medium score is given; if the identification is incorrect or unrecognizable, a low score is given. Simultaneously, the question is labeled with corresponding classification labels based on the scoring, such as "clear intent," "relatively clear intent," and "vague intent." For the language expression quality detection dimension, scoring standards are established according to grammatical correctness, sentence fluency, and word accuracy, and corresponding scores and classification labels are given based on the model's detection results, such as "standard language," "relatively standard language," and "non-standard language." The logical coherence analysis and option setting rationality evaluation dimensions also adopt a similar approach, scoring and classifying based on the model's analysis results and pre-set standards. The main analytical model processes the questionnaire data in parallel across various dimensions, generating preliminary analytical results that include detailed scores and question classification labels.

[0063] In some embodiments, step S102 above, which involves distributing different types of problems to corresponding expert models for in-depth analysis based on the problem classification labels of the preliminary analysis results using an intelligent routing decision engine, specifically includes: Problem classification identifiers and feature information are extracted from the preliminary analysis results. The problem classification identifiers are input into the routing decision engine. The most suitable expert processing model is determined by the built-in model matching strategy, which is a multi-factor weighted evaluation algorithm that includes problem type, processing complexity, model load and response time requirements. Based on the output indication of the model matching strategy, the feature information is transmitted to the corresponding expert model service node. In each expert model service node, an expert model trained in a specific domain is invoked to perform specialized processing of the problem. The types of expert models include logical structure expert models, language expression expert models, option design expert models, and psychological effect expert models. The processing outputs of each expert model are integrated to obtain an expert evaluation result dataset.

[0064] In this embodiment, the preliminary analysis results are meticulously analyzed and stored in a structured data format, containing detailed information for each question. Using a specific data extraction algorithm, the system accurately extracts question classification identifiers and feature information from the preliminary analysis results. Question classification identifiers are labels indicating the type of question, such as "illogical," "vague language," "unreasonable option settings," or "psychologically suggestive." Feature information encompasses various aspects of the data, including the question's text content, the context of the questionnaire, and the question's position within the questionnaire.

[0065] After extracting the question classification identifier, the system inputs it into the intelligent routing decision engine. The intelligent routing decision engine incorporates a model matching strategy, a multi-factor weighted evaluation algorithm that includes question type, processing complexity, model load, and response time requirements. When processing the question type factor, the engine determines whether the question belongs to a logic, language, option, or psychological effect category based on the question classification identifier. Different types of questions correspond to different processing requirements. For example, logic-type questions require models with strong logical analysis capabilities. For processing complexity, the engine evaluates the question based on its text length, semantic complexity, and the knowledge domain involved; complex questions require more specialized and powerful models. The model load factor monitors the current load of each expert model service node in real time to avoid assigning questions to overloaded models, ensuring processing efficiency. The response time requirement factor considers the user's expected time for the questionnaire analysis results, prioritizing assigning questions to models with fast response times. The engine comprehensively evaluates each expert model using these multi-factor weighted evaluation algorithms to determine the most suitable expert model for handling the current question.

[0066] Based on the output instructions of the model matching strategy, the system transmits the extracted feature information to the corresponding expert model service nodes. The transmission process utilizes a pre-established secure and efficient data transmission channel to ensure the feature information arrives at the target service node accurately and promptly. During transmission, the system encrypts the feature information to protect data security and privacy. Simultaneously, the transmission process is monitored in real time. In the event of transmission anomalies, such as network interruptions or data loss, the system can promptly retransmit or take other remedial measures to ensure the feature information is transmitted to the expert model service nodes intact and without errors.

[0067] Once the feature information reaches each expert model service node, the system invokes a domain-specific trained expert model to process the question professionally. Expert models are diverse, including logical structure expert models, language expression expert models, option design expert models, and psychological effect expert models. Logical structure expert models focus on analyzing the logical relationships between questions in the questionnaire, checking the logical coherence and logical hierarchy between items. For example, for a market research questionnaire, it checks whether the questions are arranged in a reasonable order: first, understanding basic consumer information; then, inquiring about purchase intentions; and finally, exploring factors influencing purchases. Language expression expert models deeply evaluate the quality of the language used in the questions, checking for fluency, grammatical correctness, and accurate and appropriate word choice. They identify ambiguities and vague expressions in the questions and offer suggestions for improvement. Option design expert models primarily analyze the options provided, evaluating their comprehensiveness and rationality, and identifying any omissions of important options or insufficient differentiation between options. Psychological effect expert models analyze whether the questions contain psychological influencing factors and whether they will inappropriately affect the respondents' answers, thus ensuring the objectivity and authenticity of the questionnaire results.

[0068] After each expert model completes its specialized processing of the questions, the system integrates the outputs of each model. This integration process is not a simple data patching, but rather an organic fusion based on the importance and relevance of each expert model's results. The system assigns appropriate weights to each expert model's results, determined by factors such as the model's importance in the questionnaire analysis and the accuracy of its question processing. Through weighted fusion, the outputs of each expert model are integrated into a complete expert evaluation results dataset. This dataset contains detailed evaluation information and improvement suggestions for each question in the questionnaire in terms of logic, language, options, and psychological effects, providing strong support for generating a comprehensive and accurate questionnaire quality assessment report.

[0069] In some embodiments, step S103 above, which involves combining cultural feature data queried in real time from a multi-source cultural feature database to generate a cultural suitability score to supplement the analysis results of the expert model, specifically includes: Based on the expert evaluation results dataset, the relevant cultural dimensions that require cultural compatibility assessment are adapted. Based on relevant cultural dimensions, a real-time data query request is initiated to a multi-source cultural feature database to obtain multi-dimensional feature information related to the cultural background of the target region of the questionnaire. Multi-dimensional feature information is input into the cultural compatibility assessment model, and a cultural compatibility score and specific improvement suggestions are generated through multi-dimensional weighted calculation. The cultural fit score and specific improvement suggestions are organically integrated with the expert evaluation results dataset to generate a comprehensive evaluation result that includes cultural fit.

[0070] In this embodiment, the expert evaluation results dataset contains detailed evaluation information on each question in the questionnaire across multiple aspects, including logic, language, options, and psychological effects. Using natural language processing techniques and data mining algorithms, key clues are extracted from this evaluation information to determine which questions' analysis results may be influenced by cultural factors. For example, if the expert evaluation results show that the language expression of a question has some semantic ambiguity, and the question involves concepts or customs within a specific cultural context, the system determines that this question requires a cultural suitability assessment. Simultaneously, by combining the target region information of the questionnaire, cultural dimensions closely related to the target region's culture are further screened. These cultural dimensions may cover multiple aspects such as values, social etiquette, traditional customs, language habits, and aesthetic concepts.

[0071] After adapting the relevant cultural dimensions, the system initiates real-time data query requests to a multi-source cultural feature database based on these dimensions. This database is a vast and constantly updated knowledge base that integrates cultural feature data from various channels, including academic research, cultural institutions, and field surveys. Through predefined data interfaces and query protocols, the required cultural dimension information is accurately transmitted to the database. Upon receiving the query request, the database utilizes its efficient retrieval algorithm to quickly locate multi-dimensional feature information related to the cultural background of the target region within the massive dataset. This feature information includes not only descriptions of specific cultural phenomena but also detailed information such as the manifestations, importance, and trends of cultural features. For example, when querying about the cultural features of social etiquette in a certain region, the database returns specific norms and precautions regarding greetings, banquet etiquette, and gift-giving in that region.

[0072] After obtaining multi-dimensional feature information related to the cultural background of the target region of the questionnaire, this information is input into the cultural suitability assessment model. This model is an intelligent model trained and optimized with a large amount of data, capable of understanding the complex relationships between different cultural characteristics and questionnaire questions. Internally, each cultural dimension is assigned a corresponding weight, the determination of which comprehensively considers the importance of that cultural dimension within the target region's culture and its impact on the questionnaire questions. The model performs multi-dimensional weighted calculations on the input feature information, generating a cultural suitability score by comprehensively analyzing the matching degree between each cultural dimension and the questionnaire questions. The score range is typically set within a reasonable range, such as 0-100 points; a higher score indicates better suitability between the questionnaire questions and the target region's culture. Simultaneously, the model generates specific improvement suggestions based on the calculation results. These suggestions address the shortcomings of each question in terms of cultural suitability and are clearly actionable. For example, if a question uses inappropriate vocabulary from the target region's culture, the improvement suggestions will point out the specific inappropriate words and provide suitable alternatives.

[0073] The generated cultural fit scores and specific improvement suggestions are organically integrated with the expert evaluation dataset. This integration process is not a simple data patching, but rather fully considers the correlation and complementarity between the cultural fit assessment results and the expert evaluation results. The system re-examines and adjusts relevant questions in the expert evaluation results based on the cultural fit scores. For questions with poor cultural fit, the system focuses on optimizing them based on the improvement suggestions. For example, if the expert evaluation results consider a question's logical structure reasonable, but the cultural fit assessment indicates that the question might cause misunderstanding in the target region's culture, the system will modify the question's wording according to the improvement suggestions while preserving the question's logical structure, thereby improving the question's cultural adaptability. Through this organic integration, the system generates a comprehensive evaluation result that includes cultural adaptability. This result can more comprehensively and accurately reflect the quality and applicability of the questionnaire in different cultural contexts, providing strong support for further optimization and improvement of the questionnaire.

[0074] In some embodiments, in step S104 above, the analysis results based on multiple expert models are used to perform weighted fusion processing on the scores of each dimension using weight coefficients dynamically adjusted through a reinforcement learning mechanism to obtain a comprehensive evaluation score, specifically including: A dynamic set of weight coefficients is generated through reinforcement learning, and the dynamic set of weight coefficients is adaptively adjusted according to the characteristics of the current assessment task. The dynamic set of weight coefficients includes multiple sub-weight reorganizations that are applicable to different questionnaire types, application scenarios and cultural backgrounds. The analysis results of multiple expert models and the dynamic weight coefficient set are weighted and fused to generate a comprehensive evaluation score.

[0075] In this embodiment, a reinforcement learning mechanism is used to train the generation of a dynamic set of weight coefficients. The reinforcement learning mechanism is constructed in a simulated evaluation environment containing a large number of different types of questionnaire samples, diverse application scenarios, and rich cultural background information. In the initial training phase, a set of weight coefficients is randomly initialized for each questionnaire type, application scenario, and cultural background combination. These weight coefficients correspond to the degree of influence of different expert model analysis results in the overall evaluation. For example, for a market research questionnaire targeting young consumers in an e-commerce promotion scenario, the logical structure expert model, the language expression expert model, the option design expert model, and the psychological effect expert model would each be assigned initial random weight values.

[0076] In the simulated evaluation environment, the system uses these initial weighting coefficients to calculate a comprehensive evaluation score for the questionnaire samples. The calculated comprehensive evaluation score is compared with a pre-set ideal evaluation score, and a reward or penalty signal is generated based on the comparison result. If the calculated comprehensive evaluation score is close to the ideal evaluation score, the system will give a positive reward, indicating that the current weighting coefficient combination performs well in this specific questionnaire type, application scenario, and cultural context; conversely, if the difference is large, the system will give a negative penalty, prompting the need to adjust the weighting coefficients.

[0077] The reinforcement learning mechanism continuously adjusts the weight coefficients based on these reward and punishment signals. It employs an intelligent explore-exploitation strategy, trying different combinations of weight coefficients during the exploration phase to discover potentially better solutions; and selecting the best-performing weight coefficient combinations for further optimization based on past experience during the exploitation phase. After numerous training iterations, the system gradually converges to a relatively stable set of weight coefficients, forming a dynamic weight coefficient set. This set includes multiple sub-weight reconfigurations applicable to different questionnaire types, application scenarios, and cultural backgrounds. For example, there are corresponding sub-weight reconfigurations for different types of questionnaires, such as academic research questionnaires, commercial marketing questionnaires, and government public welfare questionnaires; there are also suitable sub-weight reconfigurations for different application scenarios, such as online questionnaires and offline questionnaires; and considering the cultural differences between different countries and regions, corresponding sub-weight reconfigurations are also prepared for each major cultural background.

[0078] After training the dynamic weight coefficient set, the system adaptively adjusts it based on the characteristics of the current evaluation task. The system first performs a detailed analysis of the current evaluation task, extracting key feature information, including questionnaire type, expected application scenario, and the cultural background of the target audience. For example, if the current evaluation task is a health survey questionnaire targeting the elderly population in a European country, the system will identify the questionnaire type as a health survey, the application scenario as an offline field survey, and the cultural background as a specific European culture. Then, the system searches the dynamic weight coefficient set for the sub-weight recombination that best matches these features. If no perfectly matching sub-weight recombination is found, the system uses a similarity-based algorithm to select the closest sub-weight recombinations from the existing ones for interpolation calculation, generating an adaptive subset of weight coefficients suitable for the current evaluation task. For example, if there is no sub-weight recombination directly targeting the health survey of the elderly population in a European country, the system will refer to sub-weight recombinations from other European countries' elderly population surveys and general health surveys, calculating their similarity and performing weighted interpolation to obtain weight coefficients suitable for the current task.

[0079] After obtaining the adaptively adjusted set of dynamic weight coefficients, the system performs a weighted fusion calculation of the analysis results from multiple expert models and the dynamic weight coefficient set. Each expert model provides corresponding analysis results for each question in the questionnaire, presented as dimensional scores. For example, the logical structure expert model provides a score for the logicality of the question, while the language expression expert model provides scores for the accuracy and fluency of the language. The system weights the dimensional scores of each expert model based on the adaptively adjusted weight coefficients. Specifically, for each question, the dimensional scores of each expert model are multiplied by their corresponding weight coefficients, and then these weighted scores are summed to obtain the comprehensive evaluation score for that question. Finally, the comprehensive evaluation scores of all questions in the questionnaire are aggregated and further processed, such as by averaging or weighted averaging based on question importance, to generate the comprehensive evaluation score for the entire questionnaire. This comprehensive evaluation score more comprehensively and accurately reflects the questionnaire's performance across different dimensions, while fully considering the influence of factors such as questionnaire type, application scenario, and cultural background, providing a reliable basis for questionnaire quality assessment and optimization.

[0080] Furthermore, the step of training and generating a dynamic set of weight coefficients through a reinforcement learning mechanism specifically includes: Extract multidimensional feature vectors from the current assessment task, including questionnaire type features, application scenario features, and cultural background features; The multidimensional feature vector is input into the policy network, and a preliminary weight adjustment scheme is generated through forward propagation. A reward function is constructed based on historical evaluation data. The effect of the initial weight adjustment plan is quantitatively evaluated through the reward function, and a reward signal is generated. The reward signal is used to update the parameters of the policy network, and the policy gradient method is used to optimize the network weights, generating a dynamic set of weight coefficients.

[0081] In this embodiment, the current assessment task is analyzed to obtain various key information related to the questionnaire. For questionnaire type characteristics, the system deeply analyzes the questionnaire's theme, purpose, and structure. Application scenario characteristics cover factors such as the questionnaire distribution method and the environment in which the respondents are located. Cultural background characteristics involve the cultural region, cultural traditions, and values ​​of the target audience. The system integrates and quantifies these different characteristics to form a multi-dimensional feature vector containing questionnaire type characteristics, application scenario characteristics, and cultural background characteristics.

[0082] The extracted multidimensional feature vectors are input into the policy network. The policy network is a carefully designed and trained neural network model capable of performing complex analysis and processing of input features. Once the multidimensional feature vectors enter the policy network, they propagate forward through each layer of the network sequentially. At different layers, neurons perform operations such as weighted summation and non-linear activation on the input features, gradually extracting and transforming feature information. Through this series of forward propagation processes, the policy network can calculate and generate a preliminary weight adjustment scheme based on the input multidimensional feature vectors. This preliminary weight adjustment scheme is a suggestion to adjust the weight of each expert model analysis result in the overall evaluation of the current assessment task. For example, if the policy network determines, based on the input features, that the accuracy of language expression has a significant impact on the quality of the questionnaire within its cultural context, then the preliminary weight adjustment scheme might increase the weight corresponding to the language expression expert model analysis result.

[0083] After obtaining the initial weight adjustment plan, a reward function is constructed based on historical evaluation data. This historical evaluation data is a detailed record of numerous past questionnaire evaluation tasks, including evaluation results and corresponding overall evaluation effects across different questionnaire types, application scenarios, and cultural backgrounds. The reward function aims to quantify and generate reward signals based on the potential effects of implementing the initial weight adjustment plan. For example, if adjusting the weights according to the initial plan significantly narrows the gap between the new overall evaluation score and the ideal score, it indicates that the adjustment plan is effective, and the reward function will generate a large positive reward signal. Conversely, if the overall evaluation effect deteriorates after the adjustment, the reward function will generate a negative penalty signal. The reward function comprehensively considers multiple factors, such as changes in the overall evaluation score and the balance of scores across different dimensions of the questionnaire, to ensure a comprehensive and accurate evaluation of the initial weight adjustment plan.

[0084] The generated reward signal is used to update the parameters of the policy network. The policy gradient method is a commonly used approach to optimize neural network weights. It adjusts the network parameters based on the magnitude and direction of the reward signal, enabling the network to generate better weight adjustment schemes for subsequent inputs. Specifically, the system calculates the gradient of the reward signal with respect to the policy network parameters and then updates the parameters in the opposite direction of the gradient. Through continuous iteration of this process, the parameters of the policy network are gradually optimized, and its ability to generate preliminary weight adjustment schemes continuously improves. After multiple training and parameter updates, the policy network can ultimately generate a dynamic set of weight coefficients adapted to different questionnaire types, application scenarios, and cultural backgrounds. This dynamic set of weight coefficients can be flexibly adjusted according to the characteristics of the actual evaluation task, providing accurate and reasonable weight basis for subsequent comprehensive evaluation based on the analysis results of multiple expert models.

[0085] In some embodiments, step S105 above, which involves collecting feedback data after the user completes the interaction, specifically includes: Real-time monitoring of user interactions with the user interface, collection of feedback data including rating modification records, suggestion adoption decisions, and free text feedback; Spatiotemporal feature extraction and semantic analysis are performed on the feedback data, and a training sample set with confidence labels is generated through an importance-weighted algorithm; Based on the training sample set, a differentiated parameter update algorithm is used to adjust the general parameters of the main model and the specific parameters of the expert model in a targeted manner. Based on the training sample set, the configuration parameters and content structure of the evaluation parameter settings are dynamically optimized using the policy gradient method.

[0086] In this embodiment, a real-time monitoring mechanism closely tracks every interaction between the user and the user interface. From the moment the user opens the user interface, the system begins recording various user behaviors. When a user rates a question in a questionnaire, the system precisely captures every detail of the rating modification, including the initial rating, the modified rating, and the time of modification. For example, if a user initially rates a product's appearance design as 3 points, and then changes it to 4 points after reflection, the system will fully record this rating modification record, along with the specific time the modification occurred.

[0087] When users encounter suggested options in the questionnaire, the system records their decision to adopt those suggestions. If the questionnaire provides suggestions for improving the questions, such as "Do you think it would be better to rephrase the questions in a more easily understandable way?", the system records whether the user chooses to adopt the suggestion, and the time it takes to make that decision. This helps understand the user's acceptance of suggestions for improving the questionnaire and provides a basis for optimizing the questionnaire design.

[0088] In addition, the system provides users with an entry point for free text feedback, encouraging them to express any thoughts and opinions they have about the questionnaire. As users enter content into the free text boxes, the system collects this feedback information in real time, including the text content and the duration of their input. This free text feedback often contains users' most authentic and personalized thoughts, which is of great value for gaining a deeper understanding of user needs and identifying problems with the questionnaire.

[0089] After collecting the feedback data, comprehensive spatiotemporal feature extraction and in-depth semantic analysis are performed. For rating modification records, spatiotemporal feature extraction focuses on temporal features such as the time interval and frequency of rating modifications, as well as spatial features such as the user's position in the questionnaire at the time of modification. For example, if a user frequently modifies their rating at the beginning of the questionnaire, it may indicate that the user has difficulty understanding the questionnaire as a whole; while modifying the rating at the end of the questionnaire may be related to the user's overall feeling about the questionnaire.

[0090] Semantic analysis is conducted on suggestion adoption decisions and free-text feedback. For suggestion adoption decisions, the logical relationship between the content of the suggestion and the user's decision is analyzed to determine the user's considerations in making the decision to adopt or not adopt it. For free-text feedback, natural language processing techniques are used to identify keywords, themes, and sentiment in the text. For example, keywords such as "vague questioning" and "insufficient options" are extracted from users' free-text feedback to determine that the theme of the user's feedback is questioning the questionnaire questions, while the sentiment in the text is analyzed to determine whether it is positive, negative, or neutral.

[0091] After completing spatiotemporal feature extraction and semantic analysis, an importance-weighted algorithm is used to generate confidence labels for each feedback data sample. This algorithm comprehensively considers multiple factors related to the feedback data, such as the level of detail and the user's historical behavior. For example, if a user provides detailed and well-organized free-text feedback, and has a history of providing high-quality feedback, the system will assign a higher confidence label to this feedback data sample. In this way, a training sample set with confidence labels is generated, providing high-quality data support for subsequent model training.

[0092] Based on the generated training sample set with confidence levels, a differentiated parameter update algorithm is used to adjust the general parameters of the main model and the specialized parameters of the expert model. The main model is responsible for handling the overall evaluation of the questionnaire, and its general parameters affect the model's comprehensive judgment ability on various aspects of the questionnaire. For updating the general parameters of the main model, the system considers the overall characteristics and trends of all training samples and adopts a more macro-level update strategy. For example, if most users in the training sample set have suggested improvements to the logical structure of the questionnaire, the general parameters related to the logical structure evaluation in the main model are adjusted accordingly, so that the main model pays more attention to the logical rationality of the questionnaire in subsequent evaluations.

[0093] Expert models focus on evaluating specific dimensions of the questionnaire, such as language expression and option design. Their specialized parameters determine the model's evaluation accuracy in those specific dimensions. Updates to these specialized parameters are made through fine-tuning based on specific feedback from the training samples related to the expert model. For example, if an expert model is responsible for evaluating the language expression of a questionnaire, and many users in the training samples point out grammatical errors and inappropriate word choice, the specialized parameters for the language expression expert model are updated to address these issues, improving the model's ability to identify and evaluate language expression problems. This differentiated parameter update algorithm allows both the main model and expert models to be optimized at different levels, improving the overall performance of the evaluation system.

[0094] Based on the training sample set, the policy gradient method is used to dynamically optimize the configuration parameters and content structure of the evaluation parameters. The evaluation parameters include various parameters involved in the questionnaire evaluation process, such as scoring criteria and weight allocation. The configuration parameters and content structure directly affect the accuracy and reasonableness of the evaluation results.

[0095] The policy gradient method calculates the reward value for the current evaluation parameter settings based on feedback information from the training sample set. The reward value calculation comprehensively considers multiple factors, such as the degree of match between the evaluation result and the user's actual feedback, and the efficiency of the evaluation process. If the current evaluation parameter settings allow the evaluation result to better reflect the user's opinions and needs, the system will award a higher reward value; conversely, if the evaluation result deviates significantly from the user feedback, the reward value will be lower.

[0096] Based on the calculated reward value, the strategy gradient method adjusts the configuration parameters and content structure of the evaluation parameters along the gradient ascending direction. For example, if it finds that the current scoring criteria are too lenient, resulting in generally high evaluation results that do not match actual user feedback, the strategy gradient method will adjust the relevant configuration parameters of the scoring criteria to make them more stringent. Through continuous iteration of this process, the system can dynamically optimize the evaluation parameter settings, making the evaluation process more scientific and reasonable, thereby improving the quality and reliability of the entire questionnaire evaluation system.

[0097] Reference Figure 2 An embodiment of the present invention provides a processing system 2 for intelligent questionnaire surveys, the system 2 specifically comprising: The first processing module 201 is used to acquire the questionnaire data and evaluation parameter settings input by the user, and to call the main model through the Dify workflow engine to perform general dimension checks on the questionnaire and obtain preliminary analysis results. The second processing module 202 is used to distribute different types of problems to the corresponding expert models for in-depth analysis based on the problem classification labels of the preliminary analysis results and through the intelligent routing decision engine. The third processing module 203 is used to combine cultural feature data queried in real time from a multi-source cultural feature database to generate a cultural suitability score to supplement the analysis results of the expert model. The fourth processing module 204 is used to perform weighted fusion processing on the scores of each dimension based on the analysis results of multiple expert models and the weight coefficients dynamically adjusted through the reinforcement learning mechanism to obtain a comprehensive evaluation score. The fifth processing module 205 is used to return the comprehensive evaluation score to the user interface and collect feedback data after the user completes the interaction. The feedback data is used to iteratively update the model parameters and evaluation parameter settings.

[0098] It is understandable that, such as Figure 1 The content of the processing method embodiments for intelligent questionnaire surveys shown herein is applicable to the processing system embodiments for intelligent questionnaire surveys. The specific functions implemented by the processing system embodiments for intelligent questionnaire surveys are as follows: Figure 1 The processing method for intelligent questionnaire surveys shown is the same as that implemented in this example, and the beneficial effects achieved are the same as those described above. Figure 1 The beneficial effects achieved by the processing method embodiment for intelligent questionnaire surveys shown are also the same.

[0099] It should be noted that the information interaction and execution process between the above systems are based on the same concept as the method embodiments of the present invention. For details on their specific functions and technical effects, please refer to the method embodiments section, which will not be repeated here.

[0100] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is merely an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the system can be divided into different functional units or modules to complete all or part of the functions described above. The functional units and modules in the embodiments can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit. Furthermore, the specific names of the functional units and modules are only for easy differentiation and are not intended to limit the scope of protection of this application. The specific working process of the units and modules in the above system can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0101] Reference Figure 3 The present invention also provides a computer device 3, including: a memory 302 and a processor 301, and a computer program 303 stored in the memory 302. When the computer program 303 is executed on the processor 301, it implements the processing method for intelligent questionnaire survey as described in any of the above methods.

[0102] The computer device 3 may be a desktop computer, laptop, handheld computer, or cloud server, etc. The computer device 3 may include, but is not limited to, a processor 301 and a memory 302. Those skilled in the art will understand that... Figure 3 The computer device 3 is merely an example and does not constitute a limitation on the computer device 3. It may include more or fewer components than shown in the figure, or combine certain components, or different components, such as input / output devices, network access devices, etc.

[0103] The processor 301 may be a Central Processing Unit (CPU), or it may be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor may be a microprocessor or any conventional processor.

[0104] In some embodiments, the memory 302 may be an internal storage unit of the computer device 3, such as a hard disk or memory of the computer device 3. In other embodiments, the memory 302 may be an external storage device of the computer device 3, such as a plug-in hard disk, smart media card (SMC), secure digital (SD) card, flash card, etc., equipped on the computer device 3. Furthermore, the memory 302 may include both internal and external storage units of the computer device 3. The memory 302 is used to store the operating system, applications, boot loader, data, and other programs, such as the program code of the computer program. The memory 302 can also be used to temporarily store data that has been output or will be output.

[0105] This invention also provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the processing method for intelligent questionnaire surveys as described in any of the above methods.

[0106] In this embodiment, if the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can include at least: any entity or device capable of carrying computer program code to a photographing device / terminal device, a recording medium, a computer memory, a read-only memory (ROM), a random access memory (RAM), an electrical carrier signal, a telecommunication signal, and a software distribution medium. Examples include USB flash drives, portable hard drives, magnetic disks, or optical disks. In some jurisdictions, according to legislation and patent practice, computer-readable media cannot be electrical carrier signals or telecommunication signals.

[0107] In the above embodiments, the descriptions of each embodiment have different focuses. For parts that are not described in detail or recorded in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0108] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0109] In the embodiments disclosed in this application, it should be understood that the disclosed devices / terminal equipment and methods can be implemented in other ways. For example, the device / terminal equipment embodiments described above are merely illustrative. For instance, the division of modules or units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the displayed or discussed mutual coupling or direct coupling or communication connection may be through some interfaces; the indirect coupling or communication connection between devices or units may be electrical, mechanical, or other forms.

[0110] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

Claims

1. A processing method for intelligent questionnaire surveys, characterized in that, The method specifically includes: Obtain user-inputted questionnaire data and evaluation parameter settings, and use the Dify workflow engine to call the main model to perform general dimension checks on the questionnaire and obtain preliminary analysis results; Based on the problem classification labels from the preliminary analysis results, the intelligent routing decision engine distributes different types of problems to the corresponding expert models for in-depth analysis. By combining cultural feature data queried in real time from a multi-source cultural feature database, a cultural suitability score is generated to supplement the analysis results of the expert model; Based on the analysis results of multiple expert models, the scores of each dimension are weighted and fused using weight coefficients that are dynamically adjusted through a reinforcement learning mechanism to obtain a comprehensive evaluation score. The comprehensive evaluation score is returned to the user interface, and feedback data is collected after the user completes the interaction. The feedback data is used to iteratively update the model parameters and evaluation parameter settings.

2. The method according to claim 1, characterized in that, The process of using the Dify workflow engine to call the main model to perform general dimension checks on the questionnaire and obtain preliminary analysis results specifically includes: Input the questionnaire data and evaluation parameter settings into the Dify workflow engine; In the workflow environment of the Dify workflow engine, the master analysis model is used to evaluate the questionnaire data in multiple dimensions to obtain preliminary evaluation results. The preliminary evaluation results include question intent recognition, language expression quality detection, logical coherence analysis, and option setting rationality evaluation. Based on the pre-set evaluation criteria system and combined with the preliminary evaluation results, the quality of each dimension of the questionnaire data is analyzed in parallel through the master analysis model, generating preliminary analysis results containing detailed scores and question classification labels.

3. The method according to claim 1, characterized in that, Based on the preliminary analysis results, the problem classification labels are used to distribute different types of problems to corresponding expert models for in-depth analysis via an intelligent routing decision engine. Specifically, this includes: Problem classification identifiers and feature information are extracted from the preliminary analysis results. The problem classification identifiers are input into the routing decision engine. The most suitable expert processing model is determined by the built-in model matching strategy, which is a multi-factor weighted evaluation algorithm that includes problem type, processing complexity, model load and response time requirements. Based on the output indication of the model matching strategy, the feature information is transmitted to the corresponding expert model service node. In each expert model service node, an expert model trained in a specific domain is invoked to perform specialized processing of the problem. The types of expert models include logical structure expert models, language expression expert models, option design expert models, and psychological effect expert models. The processing outputs of each expert model are integrated to obtain an expert evaluation result dataset.

4. The method according to claim 3, characterized in that, The process of combining cultural feature data queried in real time from a multi-source cultural feature database to generate a cultural suitability score to supplement the analysis results of the expert model specifically includes: Based on the expert evaluation results dataset, the relevant cultural dimensions that require cultural compatibility assessment are adapted. Based on relevant cultural dimensions, a real-time data query request is initiated to a multi-source cultural feature database to obtain multi-dimensional feature information related to the cultural background of the target region of the questionnaire. Multi-dimensional feature information is input into the cultural compatibility assessment model, and a cultural compatibility score and specific improvement suggestions are generated through multi-dimensional weighted calculation. The cultural fit score and specific improvement suggestions are organically integrated with the expert evaluation results dataset to generate a comprehensive evaluation result that includes cultural fit.

5. The method according to claim 1, characterized in that, The analysis results based on multiple expert models are weighted and fused using weight coefficients dynamically adjusted through a reinforcement learning mechanism to obtain a comprehensive evaluation score, specifically including: A dynamic set of weight coefficients is generated through reinforcement learning, and the dynamic set of weight coefficients is adaptively adjusted according to the characteristics of the current assessment task. The dynamic set of weight coefficients includes multiple sub-weight reorganizations that are applicable to different questionnaire types, application scenarios and cultural backgrounds. The analysis results of multiple expert models and the dynamic weight coefficient set are weighted and fused to generate a comprehensive evaluation score.

6. The method according to claim 5, characterized in that, The process of training and generating a dynamic set of weight coefficients through a reinforcement learning mechanism specifically includes: Extract multidimensional feature vectors from the current assessment task, including questionnaire type features, application scenario features, and cultural background features; The multidimensional feature vector is input into the policy network, and a preliminary weight adjustment scheme is generated through forward propagation. A reward function is constructed based on historical evaluation data. The effect of the initial weight adjustment plan is quantitatively evaluated through the reward function, and a reward signal is generated. The reward signal is used to update the parameters of the policy network, and the policy gradient method is used to optimize the network weights, generating a dynamic set of weight coefficients.

7. The method according to any one of claims 1 to 6, characterized in that, The collection of feedback data after the user completes the interaction specifically includes: Real-time monitoring of user interactions with the user interface, collection of feedback data including rating modification records, suggestion adoption decisions, and free text feedback; Spatiotemporal feature extraction and semantic analysis are performed on the feedback data, and a training sample set with confidence labels is generated through an importance-weighted algorithm; Based on the training sample set, a differentiated parameter update algorithm is used to adjust the general parameters of the main model and the specific parameters of the expert model in a targeted manner. Based on the training sample set, the configuration parameters and content structure of the evaluation parameter settings are dynamically optimized using the policy gradient method.

8. A processing system for intelligent questionnaire surveys, characterized in that, The system specifically includes: The first processing module is used to acquire the questionnaire data and evaluation parameter settings input by the user, and to call the main model through the Dify workflow engine to perform general dimension checks on the questionnaire and obtain preliminary analysis results. The second processing module is used to distribute different types of problems to the corresponding expert models for in-depth analysis based on the problem classification labels of the preliminary analysis results and through the intelligent routing decision engine. The third processing module is used to combine cultural feature data queried in real time from a multi-source cultural feature database to generate a cultural suitability score to supplement the analysis results of the expert model. The fourth processing module is used to perform weighted fusion processing on the scores of each dimension based on the analysis results of multiple expert models, using weight coefficients dynamically adjusted through reinforcement learning mechanism, to obtain a comprehensive evaluation score. The fifth processing module is used to return the comprehensive evaluation score to the user interface and collect feedback data after the user completes the interaction. The feedback data is used to iteratively update the model parameters and evaluation parameter settings.

9. A computer device, characterized in that, include: The memory and processor, and the computer program stored in the memory, when the computer program is executed on the processor, implement the processing method for intelligent questionnaire survey as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, It stores a computer program, which, when executed by a processor, implements the processing method for intelligent questionnaire surveys as described in any one of claims 1 to 7.

Images