Questionnaire personalized display method based on AI real-time analysis
By using AI to analyze questionnaire personalization methods in real time, constructing user profiles using semantic rules and multi-source data, and dynamically adjusting questionnaire content, the problems of complex configuration and insufficient flexibility of traditional questionnaires are solved, thus achieving efficient personalized surveys.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-01-26
- Publication Date
- 2026-05-12
AI Technical Summary
Traditional questionnaires are complex to configure and lack flexibility. They cannot dynamically adjust the survey content based on real-time user data and can only be selected based on the current question, thus failing to achieve personalized surveys.
It adopts AI-based semantic rule setting, real-time profiling construction of multi-source data and dynamic decision engine, and generates a questionnaire logic rule base by parsing requirements through the BERT-Tiny model. It integrates user behavior, basic information and third-party data, dynamically adjusts user profiles, and optimizes jump paths and question descriptions.
Significantly reduce the complexity of questionnaire configuration, improve flexibility, accurately adapt to users' real-time scenarios, deepen the personalization of questionnaires, and ensure that the survey is more in line with the actual situation of respondents.
Smart Images
Figure CN122022904A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the interdisciplinary field of big data, artificial intelligence and market research, and specifically relates to a method for personalized questionnaire display based on real-time AI analysis. Background Technology
[0002] Traditional questionnaires rely on predefined jump rules and question configurations, which are difficult to configure and lack flexibility when the logic is complex and the number of questions is large. They cannot dynamically adjust the survey content based on the user's real-time situation.
[0003] However, existing technologies have the following shortcomings: setting specific question jump rules one by one is difficult to configure in complex scenarios; jumping only based on the user's current question selection cannot integrate real-time data such as geographical location, device information, and answering time; and only optimizing the jump path cannot dynamically adjust the question description and options, thus failing to meet personalized survey needs.
[0004] Therefore, a new method is urgently needed. Summary of the Invention
[0005] The purpose of this invention is to provide a personalized questionnaire display method based on AI real-time analysis. This method significantly reduces the complexity of questionnaire configuration, improves flexibility, accurately adapts to users' real-time scenarios, and deepens the personalization of questionnaires by setting semantic rules, constructing real-time profiles from multi-source data, and optimizing dynamic decision engines, making the survey more relevant to the actual situation of respondents.
[0006] To achieve the above objectives, this invention provides a personalized questionnaire display method based on real-time AI analysis, comprising the following steps: S1. Edit basic questionnaire questions through the questionnaire setting module, set logical requirements using semantic matching degree, call the BERT-Tiny model to parse the requirements and generate adaptation rules, and build a questionnaire logic rule library. S2 collects user answering behavior, basic information, historical answers and third-party multi-source data in real time. After cleaning and anomaly data classification, the processed multi-source data is transmitted to S3. S3. Based on the multi-source data input from S2, integrate and analyze the data, dynamically adjust the user profile weights, generate standardized real-time user profile feature vectors, and pass the real-time user profile feature vectors to S4. S4. Based on the questionnaire logic rule base synchronized by S1 and the real-time user profile feature vector transmitted by S3, calculate the matching degree between the user profile and the questionnaire logic rules, and output personalized decision results. S5 receives the personalized decision results output by S4, calculates the page adaptation effect, and then displays the questionnaire according to the user's device and age characteristics.
[0007] Preferably, in S1, the BERT-Tiny model is constructed based on the teacher-student distillation technique, specifically as follows: The teacher model uses a pre-trained original BERT with 12 Transformer layers and 768 hidden layers; the student model is simplified to 4 / 6 Transformer layers, 312 hidden layers, and a linear projection layer, compressing the 312-dimensional features into 128-dimensional word vectors. Semantic knowledge is conveyed through a three-layer loss alignment of the embedding layer, Transformer layer, and prediction layer. The embedding layer uses mean squared error as the distillation loss function, while the prediction layer uses cross-entropy as the distillation loss function. The relationship between the embedding layer and the prediction layer is established through a fixed mapping, and the training is carried out in two stages: pre-training on a general corpus and fine-tuning on task data.
[0008] Preferably, in S1, the formula for calculating the semantic matching degree is: ; In the formula, For semantic logic requirements With question tags Overall matching degree between them; Semantic logic requirements for user input; This is a set of core tags for the questionnaire items; For semantic logic requirements 128-dimensional lightweight word vectors output by the BERT-Tiny model; Tags for the question 128-dimensional lightweight word vectors output by the BERT-Tiny model; Scoring based on exact keyword matching; These are the weighting coefficients. ; Cosine similarity between two lightweight word vectors; For semantic requirements With question tags Keyword exact match score.
[0009] Preferably, in S2, the formula for classifying and processing abnormal data is: ; In the formula, The anomaly level is defined as follows: the number of anomaly fields is the number of fields in the collected data that contain errors, are missing, or are illogical; the total number of fields is the total number of all fields contained in the collected data. Repairable exceptions with less than 30% of abnormal fields are filled using neighborhood padding; invalid exceptions with ≥30% of abnormal fields are stored separately.
[0010] Preferably, in S3, the user profile weights are dynamically adjusted using the following formula: ; In the formula, These are the initial weights; Contribute incremental value to behavior; This is the time decay coefficient; For the first Each user profile sub-dimension in Real-time weight of each moment; The feature vector of the portrait is normalized to a magnitude of 1.
[0011] Preferably, in S4, the formula for calculating the matching degree between the user profile and the questionnaire logic rules is: ; In the formula, For the first The feature vectors of each rule are 1 if they conform to the rule, and 0 if they do not. Score the user's feature vector; For dimension weights; For the first Feature vectors of a rule The Middle The values that can be taken for each sub-dimension.
[0012] Preferably, in S5, the evaluation formula for page adaptation effect is: ; In the formula, Score for layout adaptation; Standardized values for page load speed; Score for font adaptation.
[0013] Therefore, the present invention employs the above-mentioned AI-based real-time analysis-based personalized questionnaire display method, and compared with the prior art, the technical solution of the present invention has the following beneficial effects: (1) By using semantic demand description to replace specific rules and the system to automatically generate adaptation rules, the technical problem of traditional questionnaires requiring specific jump rules to be set one by one, complex logic and large number of questions is difficult to configure, thereby achieving the technical effect of reducing the complexity of questionnaire logic configuration and improving configuration flexibility. (2) By using the technical means of integrating behavior, basic information, historical answers and third-party multi-source data to build real-time user profiles, the existing technology can overcome the technical problems that it cannot comprehensively integrate the real-time situation of users, only selects to jump based on the current question and lacks scene adaptability, thereby achieving the technical effect of accurately adapting to the current scene of users and making the questionnaire display fit the real-time status. (3) By using a dynamic decision engine to optimize the jump path, dynamically modify the question description and options, and shorten the questionnaire to a limited extent, the technical problems of existing technologies that can only optimize the path, cannot adjust the question content and have insufficient personalization are overcome. This achieves the technical effect of deepening the personalization of the questionnaire and making the survey more accurate and in line with the actual situation of the respondents.
[0014] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description
[0015] Figure 1 This is a flowchart illustrating an embodiment of a questionnaire personalized display method based on real-time AI analysis according to the present invention. Figure 2 This is a flowchart illustrating the dynamic adaptation of a questionnaire based on a real-time AI analysis-based questionnaire personalized display method according to an embodiment of the present invention. Detailed Implementation
[0016] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. All other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention. Unless otherwise defined, the technical or scientific terms used in the present invention should have the ordinary meaning understood by those skilled in the art.
[0017] Example 1 like Figures 1-2 As shown, this embodiment provides a personalized questionnaire display method based on real-time AI analysis. It should be understood that the specific parameters, models and protocols mentioned in this embodiment are merely examples to help those skilled in the art understand the present invention, and are not intended to limit the present invention.
[0018] The present invention provides a personalized questionnaire display method based on real-time AI analysis, comprising the following steps: S1. Edit basic questionnaire questions through the questionnaire setting module, set logical requirements using semantic matching degree, call the BERT-Tiny model to parse the requirements and generate adaptation rules, and build a questionnaire logic rule library. In this step, the system has built-in industry questionnaire templates such as consumer surveys and user experience surveys, which users can directly reuse to generate basic questionnaires; at the same time, the TextRank algorithm is used to automatically extract the core keywords of the questions as tags, avoiding the repetitive work of manually entering tags; The BERT-Tiny model was constructed based on the teacher-student distillation technique. The teacher model used a pre-trained original BERT with 12 Transformer layers and 768 hidden layers. The student model was simplified to 4 / 6 Transformer layers and 312 hidden layers, with the addition of a linear projection layer containing 39,936 parameters, which compressed the 312-dimensional features into 128-dimensional word vectors. Semantic knowledge is conveyed through a three-layer loss alignment of the embedding layer, Transformer layer, and prediction layer. The embedding layer uses mean squared error as the distillation loss function, while the prediction layer uses cross-entropy. Since the number of Transformer layers differs between the teacher and student models, a fixed mapping is used to establish the relationship between the embedding layer and the prediction layer. The intermediate Transformer layers are mapped according to the ratio of "student layer number / teacher layer number" (e.g., student layer 1 corresponds to teacher layer 3) to ensure that students accurately learn the key knowledge of the teacher. First, pre-train on a general corpus to learn general semantics and optimize the projection layer; then fine-tune using a questionnaire semantic parsing task dataset to ensure that the 128-dimensional word vectors accurately capture the core elements. The user inputs a natural language logical requirement (e.g., "Mobile phone users under 3000 yuan answer experience questions"), and the BERT-Tiny model is invoked to parse the requirement, extracting core elements such as "condition dimension (e.g., price)" and "target dimension (e.g., question type)". The calculation formula is as follows: ; In the formula, For semantic logic requirements With question tags Overall matching degree between them; Semantic logic requirements for user input; This is a set of core tags for the questionnaire items; For semantic logic requirements 128-dimensional lightweight word vectors output by the BERT-Tiny model; Tags for the question 128-dimensional lightweight word vectors output by the BERT-Tiny model; Scoring based on exact keyword matching; These are the weighting coefficients. ; Cosine similarity between two lightweight word vectors; For semantic requirements With question tags Keyword exact match score; A match is considered valid when the match result is ≥0.65, and the recall rate remains above 95%. The parsed elements are associated with the question tag library to generate structured rules containing "condition threshold, matching tag, and jump priority". High-frequency rules are cached in the Redis database, while core rules are loaded directly into memory to improve the efficiency of subsequent calls. Automatically detects conflicts between rules (such as multiple contradictory questions corresponding to the same condition), supports users to confirm and fix with one click, and provides batch verification function to adapt to configuration scenarios of large-scale questionnaires. Obtain the questionnaire logic rule base and synchronize it to the dynamic decision engine; S2. Collect user answer behavior data in real time; collect basic user information (information filled in by the user on the page, geographical location, device UA, IP, browser information, etc. obtained by the system authorization); retrieve user historical answer data and optional third-party matching data sources; after cleaning and abnormal data classification, transfer the processed multi-source data to S3; In this step, user answering behavior data (such as answering time, click frequency, and option dwell time) is collected in real time through front-end tracking. If a network interruption occurs, the data will be temporarily stored on the front-end and reported in batches after the network is restored. At the same time, non-core behavior data is sampled at a ratio of 1:10, while core data is collected in full to reduce transmission pressure. The browser's User Agent (UA) string is parsed to obtain device information. If parsing fails, screen size and system identifier are used as fallback options. Geographic information is obtained primarily through GPS. If GPS is empty, a three-level verification is performed using IP address and the user's historical delivery address to improve the accuracy of geographic information. The collected data is cleaned and completed; at the same time, abnormal data is classified and processed, and the calculation formula is as follows: ; In the formula, The anomaly level is defined as follows: the number of anomaly fields is the number of fields in the collected data that contain errors, are missing, or are illogical; the total number of fields is the total number of all fields contained in the collected data. Repairable exceptions with less than 30% of abnormal fields are filled using neighborhood padding; invalid exceptions with ≥30% of abnormal fields are stored separately to avoid data waste. The processed multi-source data is then input into the real-time user profiling engine. S3. Based on the multi-source data input from S2, integrate and analyze the data, dynamically adjust the user profile weights, generate standardized real-time user profile feature vectors, and pass the real-time user profile feature vectors to S4. In this step, after determining the initial weights of each dimension using the analytic hierarchy process (AHP), if a user continuously focuses on a certain dimension (e.g., repeatedly selecting phone-related options), the weight of that dimension will increase in real time (e.g., the weight of device price increases from 0.2 to 0.35). If a certain dimension has no relevant behavior for a long period, its weight will decrease linearly to ensure the timeliness of the user profile. The calculation formula is as follows: ; In the formula, These are the initial weights; Contribute incremental value to behavior; This is the time decay coefficient; For the first Each user profile sub-dimension in Real-time weight of each moment; The profile score is calculated only for newly added data, instead of recalculating the entire data, which shortens the calculation time per operation; at the same time, the profile data of highly active users is stored in memory, while the profile data of ordinary users is stored in Redis to reduce memory usage. The scores for each dimension are arranged in a fixed order, and then L2 normalization is used to eliminate the bias caused by excessively high scores in a single dimension, resulting in a standardized user profile feature vector. The normalized vector has a magnitude of 1, and the calculation formula is as follows: ; In the formula, For the original 3D user profile feature vector; The first to the second of the original feature vector Components of each dimension; The user profile feature vector after L2 normalization; The total number of dimensions of the feature vectors; S4. Based on the questionnaire logic rule base synchronized by S1 and the real-time user profile feature vector transmitted by S3, calculate the matching degree between the user profile and the questionnaire logic rules, and output personalized decision results. In this step, the matching degree between the user profile feature vector and the rule feature vector in the questionnaire logic rule base is calculated in batches, and the top 3 rules with the highest matching degree are selected. This balances accuracy and efficiency. The calculation formula is as follows: ; In the formula, For the first The feature vectors of each rule are 1 if they conform to the rule, and 0 if they do not. Score the user's feature vector; For dimension weights; For the first Feature vectors of a rule The Middle The values that can be taken for each sub-dimension; By combining matching rules with the user's current answering progress, the redirection path is determined to avoid repeating questions or missing core questions; Extract scenario information (such as region and device model) from user profiles and replace generalized expressions in questions (e.g., replace "local supermarket" with "Carrefour supermarket in Chaoyang District, Beijing"); at the same time, eliminate options with a matching degree of less than 0.4, and adjust the questionnaire threshold according to user age (0.7 for elderly users and 0.85 for young users) to reduce redundancy in answers; S5. Receive the personalized decision results output by S4, calculate the page adaptation effect, and then adapt the questionnaire to the user's device and age characteristics. In this step, after receiving the results from the dynamic decision engine, the page layout is adjusted according to the user's device type (single-column layout for mobile devices, double-column layout for PC devices). The calculation formula is as follows: ; In the formula, Score for layout adaptation; Standardized values for page load speed; Font adaptation score; when the adaptation score is ≥0.85, it is judged as suitable. It also automatically recognizes elderly users, enlarges the font size, and improves the reading experience; If the decision result is empty (no matching rule), the preset default question path is displayed; after the user submits the current question, feedback data such as "submission time, whether the answer is modified, and whether the question is relevant" are collected and fed back to S3 to optimize the subsequent profile weight; After a user submits the current question, the system will trigger the S2 multi-source data collection process again, entering the next cycle of "answering questions - collecting data - making decisions - displaying data", achieving real-time dynamic adaptation of the entire process.
[0019] Therefore, the present invention adopts the above-mentioned personalized questionnaire display method based on AI real-time analysis. This method significantly reduces the complexity of questionnaire configuration, improves flexibility, accurately adapts to the user's real-time scenario, and deepens the personalization of the questionnaire by setting semantic rules, constructing real-time profiles from multi-source data, and optimizing the dynamic decision engine, making the survey more in line with the actual situation of the respondents.
[0020] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.
[0021] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.
Claims
1. A method for personalized questionnaire display based on real-time AI analysis, characterized in that, Includes the following steps: S1. Edit basic questionnaire questions through the questionnaire setting module, set logical requirements using semantic matching degree, call the lightweight model to parse the requirements and generate adaptation rules, and build a questionnaire logic rule library. S2 collects user answering behavior, basic information, historical answers and third-party multi-source data in real time. After cleaning and anomaly data classification, the processed multi-source data is transmitted to S3. S3. Based on the multi-source data input from S2, integrate and analyze the data, dynamically adjust the user profile weights, generate standardized real-time user profile feature vectors, and pass the real-time user profile feature vectors to S4. S4. Based on the questionnaire logic rule base synchronized by S1 and the user real-time profile feature vector transmitted by S3, calculate the matching degree between the user profile and the questionnaire logic rules, and output personalized decision results. S5 receives the personalized decision results output by S4, calculates the page adaptation effect, and then displays the questionnaire according to the user's device and age characteristics.
2. The personalized questionnaire display method based on AI real-time analysis according to claim 1, characterized in that, In S1, the lightweight model is the BERT-Tiny model, constructed based on the teacher-student distillation technique, specifically as follows: The teacher model uses a pre-trained original BERT with 12 Transformer layers and 768 hidden layers; the student model is simplified to 4 / 6 Transformer layers, 312 hidden layers, and a linear projection layer, compressing the 312-dimensional features into 128-dimensional word vectors. Semantic knowledge is conveyed through a three-layer loss alignment of the embedding layer, Transformer layer, and prediction layer. The embedding layer uses mean squared error as the distillation loss function, while the prediction layer uses cross-entropy as the distillation loss function. The relationship between the embedding layer and the prediction layer is established through a fixed mapping, and the training is carried out in two stages: pre-training on a general corpus and fine-tuning on task data.
3. The personalized questionnaire display method based on real-time AI analysis according to claim 2, characterized in that, In S1, the formula for calculating the semantic matching degree is: ； In the formula, For semantic logic requirements With the title tag Overall matching degree between them; Semantic logic requirements for user input; This is a set of core tags for the questionnaire items; For semantic logic requirements 128-dimensional lightweight word vectors output by the BERT-Tiny model; Tags for the question 128-dimensional lightweight word vectors output by the BERT-Tiny model; Scoring based on exact keyword matching; These are the weighting coefficients. ; Cosine similarity between two lightweight word vectors; For semantic requirements With the title tag Keyword exact match score.
4. The personalized questionnaire display method based on real-time AI analysis according to claim 3, characterized in that, In S2, the formula for classifying and processing abnormal data is: ； In the formula, The anomaly level of a single data entry; The number of abnormal fields is the number of fields in this data collection that contain errors, are missing, or do not conform to logic. The total number of fields is the total number of all fields included in this data collection record; Repairable exceptions with less than 30% of abnormal fields are filled using neighborhood padding; invalid exceptions with ≥30% of abnormal fields are stored separately.
5. The personalized questionnaire display method based on AI real-time analysis according to claim 4, characterized in that, In S3, the user profile weights are dynamically adjusted using the following formula: ； In the formula, These are the initial weights; Contribute incremental value to behavior; This is the time decay coefficient; For the first Each user profile sub-dimension Real-time weight of each moment; The feature vector of the portrait is normalized to a magnitude of 1.
6. The personalized questionnaire display method based on AI real-time analysis according to claim 5, characterized in that, In S4, the formula for calculating the matching degree between user profiles and questionnaire logic rules is: ； In the formula, For the first The feature vectors of each rule are 1 if they conform to the rule, and 0 if they do not. Score the user's feature vector; For dimension weights; For the first Feature vectors of a rule The Middle The values that can be taken for each sub-dimension.
7. The personalized questionnaire display method based on real-time AI analysis according to claim 6, characterized in that, In S5, the formula for evaluating page adaptation effectiveness is: ； In the formula, Score for layout adaptation; Standardized values for page load speed; Score for font adaptation.
8. A computer device, characterized in that, include: A processor configured to be coupled to a memory, read and execute instructions and / or program code in the memory to perform the method as described in any one of claims 1-7.
9. A computer-readable medium, characterized in that, The computer-readable medium stores computer program code that, when executed on a computer, causes the computer to perform the method as described in any one of claims 1-7.