Intelligent questionnaire generation method and system based on ai large model

Abstract

The application relates to the technical field of AI model questionnaire generation, in particular to an intelligent survey questionnaire generation method and system based on an AI large model, which comprises the following steps: obtaining a questionnaire theme, target respondent characteristics, historical questions, historical respondent characteristics and answer behavior data; matching and screening similar historical respondents, and binding their answer behaviors and historical questions; based on the binding answer completion degree, continuous willingness and rhythm stability, calculating a comprehensive acceptance performance value of the question; screening high acceptance questions, analyzing their understanding difficulty, interaction complexity and scene matching degree, determining the upper bound of understanding, the upper bound of answering and the lower bound of scene matching; calling AI to generate candidate questions, calculating the mismatch degree according to the three boundaries and performing directional optimization to generate a final questionnaire. The application solves the problem that AI questionnaires are difficult to understand and answer and have a high abandonment rate due to being separated from the characteristics of the respondents, and improves the quality of questionnaire research data.

Description

Technical Field

[0001] This application relates to the field of AI model questionnaire generation technology, specifically to a method and system for intelligent generation of survey questionnaires based on large AI models. Background Technology

[0002] In fields such as government services, market research, and medical follow-ups, questionnaires are a core tool for acquiring large-scale structured data. Traditional questionnaire design relies heavily on human experience, often facing bottlenecks such as long cycles, strong dependence, and unstable quality in complex and time-sensitive scenarios. With the evolution of AI technology, large-scale models have been gradually integrated into all stages of questionnaire generation. The core of these models is not to replace questionnaire methodologies, but to programmatically coordinate survey objectives, domain knowledge, question type rules, and format constraints, thereby significantly improving the efficiency, consistency, and configurability of questionnaire construction.

[0003] Existing solutions often generate relevant questions based solely on the survey topic, lacking a structured, quantitative, and constraining mechanism for the characteristics of the target population. As a result, while the questionnaires may appear superficially relevant, they fail to accurately match the respondents' knowledge background, language habits, and acceptance of terminology. Specifically, when targeting groups of different ages, occupations, or professional backgrounds, problems such as an imbalance in the difficulty of expression, detachment from reality, inappropriate question types, and unsuitable length often occur. This leads to misunderstandings, increased abandonment rates, and more invalid responses, ultimately resulting in a decline in the quality of the survey data. Summary of the Invention

[0004] To address the aforementioned technical issues, the purpose of this application is to provide a method and system for intelligently generating questionnaires based on AI large-scale models. The specific technical solution adopted is as follows: In a first aspect, embodiments of this application provide a method for intelligently generating questionnaires based on large AI models, the method comprising the following steps: Obtain the topics to be surveyed in the questionnaire to be generated, the characteristics of the target respondents, the historical questions in the historical questionnaires, and the characteristics and answering behavior data of the historical respondents; By comparing the characteristics of the target respondents with those of historical respondents, similar respondents to the target respondents are selected from the historical respondents, and the answer behavior data of the similar respondents are linked to the historical questions; For each historical question, the overall acceptance performance of the target respondents for each historical question was determined by analyzing the completion rate, willingness to continue answering, and stability of answering pace of similar respondents associated with it. Based on the overall acceptance performance value, high-acceptance questions were selected from all historical questions. By analyzing the difficulty of understanding the question, the complexity of the response interaction, and the degree of fit between the question and the real-life scenario, the upper bound of the target respondents' understanding, the upper bound of their response, and the lower bound of the scenario fit were determined. Based on the topic of the questionnaire to be generated, an AI model is called to generate candidate questions. The difficulty of understanding the question, the complexity of the answering interaction, and the degree of fit between the question and the real-life scenario are analyzed and compared with the upper bound of understanding, the upper bound of answering, and the lower bound of scenario fit, respectively, in order to optimize the generation of the questionnaire from the candidate questions.

[0005] Preferably, the answering behavior data includes: the number of historical respondents who entered each historical question, the number of historical respondents who completed the answering of each historical question in accordance with preset rules, the number of historical respondents who completed the answering of each historical question and entered the subsequent questions, and the median value of the answering time of all historical respondents for each historical question.

[0006] Preferably, the step of selecting similar respondents to the target respondent from historical respondents includes: Analyze the characteristics of the respondents and extract feature terms; Calculate the intersection of the feature word categories of each historical respondent and the feature word categories of the target respondent. If the proportion of this intersection in the total feature word categories of the target respondent is greater than the preset matching ratio threshold, then the corresponding historical respondent will be regarded as a similar respondent to the target respondent.

[0007] Preferably, the overall acceptance performance value is positively correlated with the completion rate of the answer, the willingness to continue answering, and the stability of the answering rhythm.

[0008] Preferably, the completion rate of each historical question is determined by the ratio of the number of similar respondents who completed the answer in accordance with the preset rules to the total number of similar respondents who entered the corresponding historical question.

[0009] Preferably, the willingness to continue answering each historical question is determined by the ratio of the number of similar respondents who have completed answering each historical question and moved on to the number of similar respondents who have completed answering according to preset rules.

[0010] Preferably, the stability of the answering rhythm for each historical question is determined by the deviation of the median answering time of all similar respondents for each historical question from the average answering time of all similar respondents for the same type of question.

[0011] Preferably, the process for determining the upper bound of the target respondent's understanding, the upper bound of their response, and the lower bound of the scenario fit is as follows: The median difficulty of understanding the question, the median complexity of the response interaction, and the median relevance of the question to real-life scenarios for all highly acceptable questions were used as the upper bound of the target respondents' understanding, the upper bound of their responses, and the lower bound of their scenario relevance, respectively.

[0012] Preferably, the optimized candidate questions for generating the questionnaire include: By analyzing the degree of mismatch between the difficulty of understanding the question, the complexity of the answering interaction, and the degree of fit between the question and the real-life scenario of the candidate questions and the upper bound of understanding, the upper bound of answering, and the lower bound of scenario fit, respectively, the degree of mismatch of the difficulty of understanding the question, the degree of mismatch of the answering interaction, and the degree of mismatch of the real-life scenario of the candidate questions are constructed, and the mismatch source sequence of the candidate questions is formed in descending order. Extract the core fragments of candidate questions, and perform a preset directional reconstruction action on the first element in the mismatch source sequence to modify the text outside the core fragments of the candidate questions. Update the mismatch source sequence based on the text corresponding to the modified candidate questions until the preset convergence condition is met to obtain the final modified text of the candidate questions, and generate a questionnaire.

[0013] Secondly, embodiments of this application also provide an intelligent questionnaire generation system based on an AI large model, including a memory, a processor, and a computer program stored in the memory and running on the processor. When the processor executes the computer program, it implements the steps of any of the above-described intelligent questionnaire generation methods based on an AI large model.

[0014] This application has at least the following beneficial effects: This application selects similar respondents by matching the target with the multidimensional features of historical respondents, and binds the probability distribution of similar respondents' answers to historical questions to build a predictive model. This approach uses the real behavior of highly similar groups as a reliable reference, which avoids the problem of no samples being available due to absolute matching, and filters out low-overlap mismatched samples. This transforms historical questions into predictive labels with the target respondents' answering tendencies, providing an objective quantitative basis for subsequent evaluation of question acceptance. Furthermore, this application calculates the overall acceptance performance value by quantifying and averaging the completion rate, willingness to continue answering, and stability of answering rhythm of similar respondents on historical questions. This method transforms the implicit objective burden of respondents' "inability to understand, answer, and feel" into quantifiable continuous behavioral signals, accurately depicting the true fit of the questions to the target respondents in a multi-dimensional and integrated manner, and providing an objective and reliable quantitative benchmark for subsequent reverse analysis of the boundaries of highly acceptable questions. Furthermore, this application selects highly acceptable questions based on comprehensive acceptance performance values ​​and reverse-analyzes their common characteristics in text expression, interaction structure, and contextual immersion. The median of the difficulty of understanding the question, the complexity of the response interaction, and the degree of scenario fit are respectively marked as the upper bound of understanding, the upper bound of response, and the lower bound of scenario fit. This method transforms the implicit "good question genes" into explicit quantitative boundary red lines, giving the vague acceptance threshold of the target respondents a calculable engineering constraint, and providing a precise and objective comparison benchmark for the mismatch judgment and targeted reconstruction of candidate questions generated by the subsequent AI large model. Ultimately, this application generates candidate questions and constructs a pool of alternatives by calling a large AI model. It uses directional asymmetric logic to calculate the mismatch degree of each dimension relative to the acceptance boundary to generate a sequence of mismatch sources. Then, under the premise of locking the core measurement segment, it performs directional reconstruction on the first mismatch source until the convergence conditions of zero mismatch, reaching the top of the round, or complete blockade are met. Finally, it combines the cumulative load curve to achieve the rhythm convergence of the entire questionnaire. This method transforms questionnaire generation from unconstrained topic divergence to precise constraints around the acceptance boundary. It avoids the drift of measurement intent caused by rewriting the entire question through single-dimensional directional reconstruction and eliminates continuous answering fatigue through load rearrangement of the entire questionnaire, thereby improving the quality of questionnaire survey data. Attached Figure Description

[0015] Figure 1 A flowchart illustrating the steps of an AI-based large-scale model-based intelligent questionnaire generation method provided in one embodiment of this application; Figure 2 This is a schematic diagram of a high-acceptance topic screening process provided in one embodiment of this application. Detailed Implementation

[0016] The following, in conjunction with the accompanying drawings, details the specific scheme of the AI-based large-scale model-based intelligent questionnaire generation method and system provided in this application.

[0017] Please see Figure 1 The diagram illustrates a flowchart of a method for intelligently generating questionnaires based on an AI large model, according to an embodiment of this application. The method includes the following steps: Step S1: Obtain the topic to be investigated in the questionnaire to be generated, the characteristics of the target respondents, the historical questions in the historical questionnaire, and the characteristics and answering behavior data of the historical respondents.

[0018] This embodiment selects community-based follow-up of elderly people with chronic diseases as a representative scenario. In this scenario, elderly respondents generally exhibit characteristics such as limited acceptance of terminology, low tolerance for continuous reading, insufficient stability of mobile device operation, and weak empathy for abstract expressions. Furthermore, questions generated by existing general-purpose models are prone to problems such as dense terminology and insufficient anchoring to real-life situations, leading to skipped answers, interruptions, and data distortion. It should be noted that this scenario is for illustrative purposes only; the following steps can be applied to other questionnaire generation scenarios targeting specific audiences. The specific data acquisition and preprocessing process is as follows: Obtain the subject text of the questionnaire to be generated and the feature description of the target respondents; wherein, in this embodiment, the feature description of the target respondents includes multi-dimensional feature information such as the target respondents' age group, management status stage (e.g., initial diagnosis period, stable maintenance period, complication period), terminal type (e.g., smartphone, tablet, PC) and scenario (e.g., residential scenario, medical treatment scenario); The system synchronously acquires historical questionnaires, historical respondents' characteristic descriptions, and answer behavior data. In this embodiment, the characteristic descriptions of historical respondents include multi-dimensional characteristic information such as age group, management status (e.g., living alone, children accompanying), terminal type (e.g., smartphone, tablet, PC), and scenario (e.g., "community hospital", "home rehabilitation"). The answer behavior data includes: the number of historical respondents who entered each historical question, the number of historical respondents who completed each historical question according to preset rules, the number of historical respondents who completed each historical question and entered subsequent questions, and the median time spent by all historical respondents on each historical question.

[0019] It should be noted that the preset rules are age-appropriate verification standards used to identify the quality of elderly respondents' answers and their cognitive state: First, a basic time limit for behavior is set. Given the decline in eyesight and reaction speed among the elderly, the answering time for a single question must not be less than a reasonable reading limit (e.g., 5 seconds), eliminating the behavior of blindly selecting answers without fully understanding the question, and removing samples with extremely long answering times; Second, the absolute self-consistency of medication logic must be strictly controlled, and cross-validation is set for commonly used medications for chronic diseases (e.g., if "taking hypoglycemic drugs" is selected in the previous question, but "never diagnosed with diabetes" is selected later, or two drugs that are pharmacologically prohibited from being taken together are selected, it is judged as a logical contradiction); Third, it is mandatory to answer all questions on core medication points completely, and missing medication records and skipping questions are strictly prohibited; In addition, abnormal patterns of answering are specially monitored (e.g., selecting the same option for more than 10 consecutive questions, which is common when elderly respondents lose patience or do not understand the operation and keep clicking "next question"); Only when the respondent's answering time, pathological logic, and behavioral trajectory all meet the above constraints are they included in the valid sample.

[0020] It should be noted that in this embodiment, the historical time period corresponding to the historical questions and historical respondents is six months backward from the current moment. In actual application, when facing other types of questionnaires, the historical time period needs to be set according to the specific scenario. This embodiment does not impose any special restrictions.

[0021] Step S2: Compare the characteristics of the target respondents with those of historical respondents, select similar respondents from the historical respondents, and link the answer behavior data of similar respondents with historical questions.

[0022] Since the target respondents are entirely new individuals who have not yet answered the questions, their actual answering behavior is uncertain, and blindly predicting it directly often lacks a basis. Therefore, the core solution lies in finding a "reference point" in historical data—based on the underlying logic that "the more similar the characteristics of a group, the more consistent their answering responses and behavioral patterns will be when faced with the same questions," it is necessary to compare the characteristics of the target respondents with those of historical respondents, select similar respondents from the historical respondents, and link the answering behavior data of similar respondents with historical questions. This serves as a reliable basis for predicting the target respondents' answering behavior. The specific process is as follows: In this embodiment, the characteristics of the respondents are analyzed and feature terms are extracted. Specifically, in the process of extracting feature terms, the feature description of the target respondents is first segmented and filtered for stop words. Then, based on the preset population feature ontology library, the remaining terms are classified and mapped to four dimensions: age (e.g., "advanced age", "early aging"), scenario (e.g., "community hospital", "home rehabilitation"), terminal (e.g., "touchscreen mobile phone", "voice assistance"), and management status (e.g., "living alone", "children accompanying"), thereby transforming unstructured text into structured feature terms.

[0023] Furthermore, the intersection of the feature word categories of each historical respondent and the feature word categories of the target respondent is calculated. If the proportion of this intersection in the total feature word categories of the target respondent is greater than the preset matching ratio threshold, then the corresponding historical respondent is regarded as a similar respondent to the target respondent.

[0024] It should be noted that the preset matching ratio threshold is set at 75%. This is because this threshold can ensure that historical respondents are highly similar to target respondents in terms of core feature dimensions, avoiding the introduction of mismatched samples with extremely low feature overlap due to a threshold that is too low (e.g., 50%). At the same time, it can accommodate the objective situation in actual business where it is difficult to achieve 100% complete coverage of historical data labels, preventing the matching conditions from being too strict due to a threshold that is too high (e.g., 100%), which would fail to select enough similar respondents as a reference for predicting answer behavior. In actual application, as another implementation method, implementers can also set the preset matching ratio threshold according to specific circumstances. This embodiment does not impose any special restrictions.

[0025] It should be noted that the reason for adopting the above matching logic is that the respondents' answering habits in the elderly medication survey scenario are deeply dependent on their own characteristics: scenario-related and age-related terms determine the elderly's cognitive level and comprehension ability, management status terms (such as living alone or being cared for by children) reflect the differences in their channels for obtaining medication information, and terminal usage terms directly affect their operational proficiency when answering questions; when no perfectly matching historical records can be found, absolute matching is abandoned and coverage ratio is used instead, which avoids the dilemma of having no samples available due to overly detailed feature dimension division, effectively filters out marginal samples with too low feature overlap, and thus ensures that the selected similar groups have sufficient reference value and representativeness in terms of core features.

[0026] Furthermore, after identifying similar respondents, a "feature aggregation mapping" approach is used for binding: First, the actual answer records of all similar respondents in historical questionnaires are extracted, and for each historical question, the distribution ratio of the number of people in this group choosing each option is calculated; then, this group's answering tendency with probability distribution is mapped and bound to the corresponding historical question as a whole, forming a binding relationship model of "historical question - similar group option distribution", so that each historical question is no longer just dry text, but comes with a predictive label of "how likely a person like the target respondent is to choose a certain option".

[0027] Thus, this embodiment selects similar respondents by matching the target with the multidimensional features of historical respondents, and binds the probability distribution of similar respondents' answers to historical questions to build a predictive model. This approach uses the real behavior of highly similar groups as a reliable reference, which avoids the problem of no samples being available due to absolute matching, and filters out samples with low overlap and mismatch. This transforms historical questions into predictive labels with the target respondents' answering tendencies, providing an objective quantitative basis for subsequent evaluation of question acceptance.

[0028] Step S3: For each historical question, determine the target respondents' overall acceptance performance value for each historical question by analyzing the completion rate, willingness to continue answering, and stability of answering pace of similar respondents associated with it.

[0029] Community-based elderly chronic disease follow-up subjects are limited by a combination of factors when reading and answering questions: vocabulary familiarity, sentence meaning construction ability, and terminal operation tolerance. Too many professional expressions increase the burden of word recognition; multiple conditions and transitions consume limited sentence meaning retention capacity; matrix questions or multiple-choice questions with a large number of options amplify the amount of clicking and comparison operations; and when there is a lack of life management context such as "taking medicine, measuring, and follow-up examination", the questions are not easy to be quickly applied even if they are related to the topic. The above-mentioned objective burdens are not just at the level of subjective feelings, but will inevitably leave quantifiable and stable behavioral manifestations in the answer log. For example, when encountering questions that they do not understand, respondents often give up answering directly; when encountering questions with complicated operations, they barely finish answering but are unwilling to continue to the next question; and when encountering questions that lack a sense of identification, they hesitate repeatedly, resulting in abnormal dwell time.

[0030] Given the deep causal relationship between the aforementioned objective burden and behavioral representation, to accurately assess the true suitability of historical questions for target respondents, it is necessary to move beyond single-dimensional judgment and instead capture exit signals corresponding to "not understanding, unable to answer, and lack of engagement" from the respondents' continuous behavioral trajectories. Therefore, this embodiment determines the target respondents' overall acceptance performance value for each historical question by analyzing the completion rate, willingness to continue answering, and stability of answering rhythm of similar respondents associated with it. The specific process is as follows: In this embodiment, the overall acceptance performance value is positively correlated with the completion rate of the answer, the willingness to continue answering, and the stability of the answering rhythm.

[0031] It should be understood that a positive correlation means that the dependent variable increases as the independent variable increases, and the dependent variable decreases as the independent variable decreases. The specific relationship can be additive or multiplicative, etc., and is determined by the actual application. This application does not impose any special restrictions.

[0032] Preferably, in this embodiment, as one implementation method, the average of the quantitative results of the completion rate of each question, the willingness to continue answering, and the stability of the answering rhythm is used as the overall acceptance value of the target respondents for each question.

[0033] To clarify, the quantification process for the above-mentioned completion rate, willingness to continue answering, and stability of answering rhythm is as follows: (1) The process for determining the completion rate of each history question is as follows: Among all similar respondents, the ratio of the number of similar respondents who completed the answer according to the preset rules to the total number of similar respondents who entered the corresponding historical question is used as the answer completion rate of each historical question; in particular, when the total number of historical respondents who entered the corresponding historical question is 0, the answer completion rate of each historical question is set to 0.

[0034] Based on the completion rate, it can be understood that the completion rate is used to characterize the effective conversion rate of the question, reflecting the degree of impact of the question text on the cognitive load of the respondents in terms of vocabulary density and logical structure. The higher the completion rate, the higher the proportion of respondents who understand and answer the question smoothly, reflecting that the question is in line with the audience's comprehension ability and does not trigger reading obstacles. Conversely, the lower the completion rate, the more respondents will give up answering the question directly due to unfamiliar terminology, obscure sentence meaning, or convoluted logic. This will lead to a sharp reduction in the effective sample size of the question, a lack of statistical support for subsequent analysis, and may even introduce selection bias due to difficulty.

[0035] (2) The process of quantifying the continued willingness to answer each history question is as follows: The ratio of the number of similar respondents who completed all historical questions and moved on to the number of similar respondents who completed the questions in accordance with the preset rules is used as the quantitative result of the continued willingness to answer each historical question; in particular, when the number of respondents who completed the questions in accordance with the preset rules is 0, the continued willingness to answer each historical question is set to 0.

[0036] Based on the willingness to continue answering, it can be understood that the willingness to continue answering is used to characterize the degree of risk of interruption in answering questions. It reflects the degree to which the cumbersome operation of previous questions and the interaction friction consume the patience of respondents. The greater the willingness to continue answering, the more smoothly respondents can answer questions after completing the corresponding questions, reflecting that the operational and emotional burden of the corresponding questions is within a tolerable range. Conversely, the lower the willingness to continue answering, the more likely respondents are to have exhausted their patience by the comparison of a large number of options or the lengthy and tedious interaction process, which directly affects the arrival rate of subsequent questions in the entire questionnaire and is very likely to cause a large number of people to abandon the answering halfway.

[0037] (3) The quantitative process of determining the stability of the answering rhythm for each history question is as follows: Quantitative results of the stability of the answering rhythm of historical question i The expression is: In the formula, This represents the median time spent by all similar respondents on historical question i. This represents the average time all similar respondents spent answering the same type of question in historical question i; The interquartile range represents the time spent by all similar respondents on the same type of question in historical question i. This represents a preset constant greater than 0, used to prevent the denominator from being 0. In this embodiment... The value of is 0.01. Provided that the denominator is not zero and does not excessively affect the calculation result, the implementer may also set it according to the specific situation. This embodiment does not impose any special restrictions.

[0038] The calculation process of the interquartile range is a well-known technique and will not be elaborated further.

[0039] Based on the quantitative results of the answer rhythm stability, it can be understood that the answer rhythm stability is used to characterize the coherence and confidence of the answering behavior, reflecting whether the context of the question causes the respondent to hesitate, be confused or need to repeatedly review the question. The greater the answer rhythm stability, the more smoothly the respondent reads and answers the question, reflecting that the question is straightforward and has a strong sense of immersion, and the respondent does not need to pause to think. Conversely, the smaller the answer rhythm stability, the more the dwell time deviates significantly from the natural benchmark of similar question types, reflecting that the respondent is caught in a long period of comprehension hesitation, recall difficulties or operational confusion. This not only reduces the efficiency of a single question, but also makes it very easy for hesitation to turn into abandonment of answering.

[0040] Based on the overall acceptance performance score, it can be understood that the overall acceptance performance score is used to characterize the comprehensive fit quality of the question under the multi-dimensional constraints of "understanding, completing the answer, and responding smoothly". It reflects whether the superimposed effect of the text comprehension load, interactive operation load and situational immersion of the question has exceeded the acceptance threshold of the target audience. The larger the overall acceptance performance score, the better the corresponding historical question performs in terms of completion, continued willingness and rhythm stability. It reflects that this is a high-quality benchmark question that is highly consistent with the cognitive habits and operational abilities of the target respondents. Conversely, the smaller the overall acceptance performance score, the more serious the corresponding historical question is in terms of understanding, operation or immersion. If such questions are used to construct questionnaires, it will directly lead to data distortion and a collapse in the response rate.

[0041] Thus, this embodiment calculates the overall acceptance performance value by quantifying and averaging the completion rate, willingness to continue answering, and rhythm stability of similar respondents on historical questions. This method transforms the implicit objective burden of respondents' "inability to understand, answer, and feel" into quantifiable continuous behavioral signals, accurately depicting the true suitability of the questions to the target respondents in a multi-dimensional way, and providing an objective and reliable quantitative benchmark for subsequent reverse analysis of the boundaries of highly acceptable questions.

[0042] Step S4: Based on the overall acceptance performance value, select high-acceptance questions from all historical questions. By analyzing the difficulty of understanding the question, the complexity of the answering interaction, and the degree of fit between the question and the life scenario of each high-acceptance question, determine the upper limit of the target respondents' understanding, the upper limit of their answer, and the lower limit of the scenario fit.

[0043] The overall acceptance performance score has quantified the real feedback of historical questions in terms of "completion rate, continued willingness, and rhythm stability," thus providing an objective benchmark for distinguishing between "good" and "bad" questions. However, simply selecting highly acceptable questions is insufficient to guide the generation of new questionnaires. It is necessary to reverse-analyze these highly acceptable questions, breaking down their common characteristics in text expression, interaction structure, and contextualization, in order to transform the "genes of good questions" into the boundary lines that constrain the generation of large AI models. Therefore, this embodiment selects highly acceptable questions from all historical questions based on the overall acceptance performance score. By analyzing the difficulty of understanding the meaning of each highly acceptable question, the complexity of the answering interaction, and the degree of fit between the question and real-life scenarios, the upper bound of the target respondents' understanding, the upper bound of their answers, and the lower bound of scenario fit are determined. The specific calculation process is as follows: The overall acceptance score of the target respondents for all historical questions is used as the input of the threshold segmentation algorithm, and the output is the segmentation threshold. Historical questions with an overall acceptance score greater than the segmentation threshold are regarded as high acceptance questions. In this embodiment, the threshold segmentation algorithm adopts the Otsu threshold segmentation algorithm. This embodiment does not impose any special restrictions on the selection of segmentation algorithms.

[0044] The process of obtaining the segmentation threshold using the Otsu threshold segmentation algorithm is a well-known technique and will not be described in detail here.

[0045] Preferably, the schematic diagram of the high-acceptability question screening process provided in this embodiment is as follows: Figure 2 As shown.

[0046] Furthermore, in this embodiment, the median of the difficulty of understanding the question, the median of the complexity of the response interaction, and the median of the degree of fit between the question and the life scenario of all highly acceptable questions are used as the upper bound of the target respondents' understanding, the upper bound of their response, and the lower bound of the scenario fit, respectively.

[0047] Additionally, in this embodiment, the quantification process for the difficulty of understanding the question, the complexity of the answering interaction, and the degree of relevance of the question to real-life scenarios for each highly acceptable question is as follows: (1) The quantification process of the difficulty of understanding the question meaning of each highly acceptable question is as follows: In this embodiment, for each highly acceptable question, the number of professional expressions and abbreviations, the number of conditional, adversative, and causal logical connectors, and the number of negation and comparison words are extracted respectively. In this embodiment, the sum of the number of technical terms and abbreviations, twice the number of conditional, adversative, and causal logical connectors, and the number of negation and comparison words is used as the difficulty of understanding the meaning of each highly acceptable question.

[0048] Based on the difficulty of understanding the question, it can be understood that the difficulty of understanding the question is used to represent the reading comprehension threshold of the text. It reflects the degree to which the density of professional terminology, complex logical nesting, and negative comparison sentences occupy the brain's information processing channels. The greater the difficulty of understanding the question, the more background knowledge the respondents need to mobilize to decipher unfamiliar words and dissect multiple conditions. This reflects that the corresponding question is very likely to cause misinterpretation or directly discourage the respondents. Conversely, the lower the difficulty of understanding the question, the more common the words in the question stem and the more straightforward the sentence structure. The respondents can grasp the measurement intention instantly, which directly determines the effective response rate and the length of the response time.

[0049] To clarify, professional expressions refer to medical or academic terms beyond everyday understanding, such as "target organ damage" and "two-hour postprandial blood glucose"; abbreviations refer to terms composed of initials or abbreviations, such as "CAD" and "BP"; conditional, adversative, and causal logical connectors refer to conjunctions that require semantic connection across clauses, such as "if...then...", "although...but...", and "therefore"; negation words refer to words representing opposites or exclusions, such as "not" and "not"; and comparison words refer to words representing degree or state contrasts, such as "not as good as" and "equivalent to". The extraction process is as follows: First, based on the word segmentation and part-of-speech tagging results of the question stem, terms marked as nouns and falling into the pre-constructed chronic disease professional terminology database are identified as professional expressions. Terms containing uppercase letters or conforming to abbreviation rules and within the terminology database are identified as abbreviations. Then, all words are traversed in sequence. Terms marked as conjunctions and matching the predefined logical connector word list are identified as logical connectors. Terms marked as adverbs and matching the negation word list are identified as negation words. Terms marked as adverbs or prepositions and matching the comparison word list are identified as comparison words.

[0050] (2) The quantification process of the interactive complexity of answering each highly acceptable question is as follows: Interaction complexity of answering highly acceptable question k The expression is: In the formula, , , representing the number of actions and the mental comparison quantity for the highly acceptable question k, respectively.

[0051] When the highly acceptable question k is a single-choice question, respondents only need to perform one click action after reading all the options. Choose 1; since single-choice questions only require one-way selection and do not require cross-category comparison, therefore... Set to 0; When the highly acceptable question k is a multiple-choice question, respondents need to independently check each option. The total number of options needs to be calculated, and multiple options need to be scanned and weighed for trade-offs. Count the total number of options; When the highly acceptable question k is a matrix question (such as a 3x5 scale), respondents need to click one cell in each row, requiring a total of "row number" clicks. The number of rows is taken, but during the judgment, the left-side question stem and the top 5 column headers need to be repeatedly aligned and compared. Get the number of columns; When the highly acceptable question k is an open-ended question, respondents only need to activate the input box once. Taking 1, there is no control load, therefore Take 0.

[0052] Based on the complexity of the interaction for each highly acceptable question, it can be understood that the complexity of the interaction is used to characterize the operational friction of the question's interactive design, reflecting the degree to which the question structure and option configuration consume the audience's motor precision and working memory. The greater the complexity of the interaction, the more buttons the respondents need to click or repeatedly align and compare in matrix tables and long option lists, reflecting that the corresponding highly acceptable questions are very likely to cause hand fatigue and mental distraction. Conversely, the less complex the interaction, the shorter the operation path and the fewer comparison objects, allowing respondents to quickly select based on intuition, directly determining the level of willingness to continue answering and the occurrence of midway abandonment.

[0053] (3) The quantification process for the degree of fit between each highly acceptable topic and real-life scenarios is as follows: First, the feedback text of historical questions is obtained. This feedback text includes the open-ended answers submitted by respondents to historical questions, as well as supplementary contextual explanations attached to historical questionnaires. The feedback text of high-acceptance question k is segmented into words, and the length distribution of all segmented word segments is statistically analyzed. The median length of this distribution is taken as the N value of the N-gram sliding window. This window is used to slide segment the feedback text, and the resulting segments are recorded as candidate segments. Then, the consistency of part-of-speech tags and the continuity of the original text between adjacent candidate segments are checked to literally merge them to restore complete natural language blocks. Next, the frequency of each candidate segment in the feedback text of all historical questions is counted. Only candidate segments whose two statistical values ​​are both higher than their respective medians are retained as scene segments that represent expressions familiar to the target respondents. Finally, the ratio of the number of scene segments in the feedback text of high-acceptance question k to the total number of candidate segments is used as the quantitative result of the degree of fit between high-acceptance question k and real-life scenarios.

[0054] It should be noted that the value of N is set to 4 in this embodiment because the core objective of this scenario is to extract familiar life management segments from the target respondents (such as elderly patients with chronic diseases in the community). If the value of N is set too small, it degenerates into single-word or word segmentation, and cannot capture continuous actions and life phrases with complete contextual meaning, such as "taking medication on time" and "regularly measuring blood pressure", making it difficult to form effective scene anchoring. If the value of N is set too large, the terms selected by the sliding window will be too long, which is very easy to cross the natural semantic boundary and forcibly bind irrelevant modifiers or multiple independent events, resulting in lengthy and fragmented extracted segments, and the matching threshold is too high, thus losing the ability to generalize and cover. In actual application, as other implementation methods, implementers can also set their own values ​​according to specific circumstances. This embodiment does not impose any special restrictions.

[0055] The quantitative results of the fit between each highly acceptable question and real-life scenario indicate that the fit between the question and the real-life scenario represents the sense of contextual relatability of the question, reflecting the close connection between the abstract survey intention and the respondents' concrete life experience. The greater the fit between the question and the real-life scenario, the more the corresponding highly acceptable question can evoke familiar memories of daily scenarios such as taking medication and follow-up visits, reflecting that the respondents can quickly enter the answering state based on intuition and experience. Conversely, the smaller the fit between the question and the real-life scenario, the more the corresponding highly acceptable question is expressed stiffly and superficially, detached from the audience's real-life context, causing respondents to have difficulty understanding and resist answering due to a lack of resonance, which seriously affects the authenticity and validity of the measurement.

[0056] Thus, this embodiment selects highly acceptable questions based on comprehensive acceptance performance values ​​and reverse-analyzes their common characteristics in text expression, interaction structure, and contextual immersion. The median of the difficulty of understanding the question, the complexity of the answer interaction, and the degree of scenario fit are respectively marked as the upper bound of understanding, the upper bound of answering, and the lower bound of scenario fit. This method transforms the implicit "good question genes" into explicit quantitative boundary red lines, giving the vague acceptance threshold of the target respondents a calculable engineering constraint, and providing a precise and objective comparison benchmark for the subsequent mismatch judgment and targeted reconstruction of candidate questions generated by the AI ​​large model.

[0057] Step S5: Based on the topic of the questionnaire to be generated, call the AI ​​big model to generate candidate questions. Analyze the difficulty of understanding the question, the complexity of the answering interaction, and the degree of fit between the question and the life scenario compared with the upper bound of understanding, the upper bound of answering, and the lower bound of scenario fit, respectively, to optimize the generation of the questionnaire from the candidate questions.

[0058] After accurately identifying the upper bound of the target respondents' understanding, upper bound of their responses, and lower bound of scenario fit using historical data, the generation logic of the new questionnaire must shift from the traditional "divergent thinking around the topic" to "constraints around the acceptance boundary." Although calling an AI model can quickly generate candidate questions that meet content relevance based on the topic of the questionnaire to be generated, it cannot determine whether the generated questions deviate from the target respondents' acceptance boundary due to excessive understanding load, excessive operational load, or insufficient scenario anchoring, nor can it provide a clear basis for subsequent revisions. Therefore, based on the topic of the questionnaire to be generated, an AI model is used to generate candidate questions. By analyzing the mismatch between the difficulty of understanding the candidate questions, the complexity of the response interaction, and the degree of fit between the questions and real-life scenarios compared to the aforementioned upper bound of understanding, upper bound of responses, and lower bound of scenario fit, the candidate questions are optimized to generate the questionnaire. The specific process is as follows: First, based on the theme of the questionnaire to be generated, it is decomposed into multiple specific measurement dimensions. The survey objectives of each dimension, the characteristics of the target respondents (such as elderly patients with chronic diseases in the community), and the pre-set question structure constraints are encapsulated into structured prompt words. Then, the AI ​​model is called to control it to generate at least two semantically parallel but differently expressed initial questions for the same measurement dimension. During the generation process, the model is forced to retain clear time limits, actions, and evaluation objects through instructions, thereby constructing a candidate question pool that meets the theme relevance and has a margin of choice.

[0059] Furthermore, by analyzing the mismatch between the difficulty of understanding the question, the complexity of the answering interaction, and the degree of fit between the question and real-life scenarios of the candidate questions compared to the upper bound of understanding, the upper bound of answering, and the lower bound of scenario fit, respectively, a mismatch degree of the difficulty of understanding the question, the mismatch degree of the answering interaction complexity, and the mismatch degree of fit with real-life scenarios of the candidate questions are constructed, and these are then arranged in descending order to form a mismatch source sequence of candidate questions. Specifically: In this embodiment, the difficulty of understanding the question, the complexity of the answering interaction, and the degree of fit between the question and real-life scenarios are extracted from the candidate questions and the historical questions, which are calculated using the same rules. A directional asymmetric computation logic is adopted, and mismatch is triggered only when the load exceeds the upper limit or the anchoring is insufficient to the lower limit. The mismatches of the difficulty of understanding the question, the complexity of the answering interaction, and the degree of fit with real-life scenarios for the candidate questions are constructed. The specific process is as follows: ; ; ; In the formula, , , These represent the mismatch in difficulty of understanding the question, the mismatch in complexity of answering interaction, and the mismatch in relevance to the real-life scenario for candidate question j, respectively. , , These represent the quantitative results of the difficulty in understanding the question, the complexity of the answering interaction, and the degree of relevance of the question to real-life scenarios for candidate question j, respectively. , , These represent the upper bound of the target respondents' understanding, the upper bound of their responses, and the lower bound of the scenario fit, respectively. , , These represent the interquartile ranges of the quantitative results for the difficulty of understanding the question, the interquartile range of the quantitative results for the complexity of the response interaction, and the interquartile range of the quantitative results for the relevance of the question to real-life scenarios, respectively; max() represents the maximum value function. These represent constants that are preset to be greater than 0, used to prevent the denominator from being 0. In this embodiment, the value of is 0.01. Under the premise of ensuring that the denominator is not 0 and does not excessively affect the calculation result, the implementer can also set it according to the specific situation. This embodiment does not impose any special restrictions.

[0060] Furthermore, the candidate questions are sorted in descending order of the degree of mismatch between the difficulty of understanding the question, the degree of mismatch between the interaction complexity of answering, and the degree of mismatch between the candidate questions and the real-life scenarios, forming a sequence of mismatch sources for the candidate questions.

[0061] Furthermore, core fragments consisting of subject matter words, action words, evaluation words, and time-limited words are extracted from the candidate questions. For the first element in the mismatch source sequence, a pre-defined targeted reconstruction action is performed to modify the text outside the core fragments of the candidate questions. Specifically: if the first mismatch is due to a mismatch in the difficulty of understanding the question, professional terminology is replaced with easily accepted everyday expressions, and redundant logical fragments are pruned and removed according to a pre-defined core retention ratio (e.g., set to 60% of the original length). These redundant logical fragments are identified through dependency parsing tree analysis, specifically extracting and removing continuous modifying clauses or non-restrictive relative clauses that further modify the core meaning group. If the first mismatch is due to a mismatch in the complexity of the answer interaction, matrix questions and high-option multiple-choice questions are removed. The process is broken down into step-by-step single-item questions. If the first mismatch is due to a mismatch in the degree of relevance to the life scenario, then an internal full-text search is performed based on the pre-built historical high-rated questionnaire question bank to locate the scenario. The natural word fragments of the scenario that are most frequently associated with the core fragment within the same short sentence interval are extracted and inserted as prefixes. After each round of targeted reconstruction, the mismatch source sequence is updated based on the text corresponding to the modified candidate question. If the original first mismatch source has been corrected, the new first mismatch source is processed in sequence. If the reconstruction results in the loss of core fragment elements, the process is backtracked and jumps to the next mismatch source. If all mismatch sources cannot be deloaded, parallel questions are called for replacement until the preset convergence condition of all mismatch sources being processed or the mismatch degree being zero is met, and the final modified text of the candidate question is obtained. It should be noted that the preset convergence condition is the core mechanism for controlling the termination of the targeted reconstruction of the question and preventing excessive rewriting from causing the measurement intent to drift. The specific judgment rules include three scenarios: First, the mismatch is reduced to zero, that is, the mismatch degree of all dimensions in the updated mismatch source sequence is reduced to 0, indicating that the candidate question has fully adapted to the acceptance boundary of the target respondents; Second, the round reaches the limit, since each round of reconstruction only processes the current first mismatch source and the mismatch sequence may be dynamically rotated after reconstruction, the maximum number of reconstruction rounds is set to be equal to the number of mismatch source types that can be identified at the beginning of the question. When all types of mismatch sources have completed one targeted processing, the reconstruction is forcibly terminated; Third, the entire line is blocked, that is, in a certain round of reconstruction, the risk of loss of core fragment elements is triggered to trigger a rollback, which causes the current first mismatch source to be marked as undownloadable, and the reconstruction is terminated when all subsequent mismatch sources are in an undownloadable state. At this time, it is determined that the candidate question has reached the adaptation limit under the current structure, and a parallel question will be directly called from the candidate pool to replace the entire question, so as to ensure that the reconstruction process achieves a strict balance between adaptation improvement and measurement intent preservation.

[0062] Finally, the cumulative load of the reconstructed candidate items, calculated by adding the understanding and operational loads according to the questionnaire order, is compared with the historical natural cumulative load curve. High-load segments are rearranged or replaced with low-load, high-fit parallel items without disrupting the dimensional order, thus eliminating the cumulative fatigue of continuous answering and generating the questionnaire.

[0063] In this embodiment, candidate questions are generated and a pool of alternative questions are constructed by calling a large AI model. Directional asymmetric logic is used to calculate the mismatch degree of each dimension relative to the acceptance boundary to generate a sequence of mismatch sources. Then, under the premise of locking the core measurement segment, directional reconstruction is performed on the first mismatch source until the convergence conditions of zero mismatch, reaching the top of the round, or complete blockade are met. Finally, the overall questionnaire rhythm convergence is achieved by combining the cumulative load curve. This method transforms questionnaire generation from unconstrained topic divergence to precise constraints around the acceptance boundary. It avoids the measurement intent drift caused by rewriting the entire question through single-dimensional directional reconstruction and eliminates continuous answering fatigue through load rearrangement, thereby improving the quality of questionnaire survey data.

[0064] Based on the same inventive concept as the above methods, this application also provides an intelligent questionnaire generation system based on an AI large model, including a memory, a processor, and a computer program stored in the memory and running on the processor. When the processor executes the computer program, it implements the steps of any one of the above-described intelligent questionnaire generation methods based on an AI large model.

[0065] The above description is only a preferred embodiment of this application and is not intended to limit this application. Any modifications, equivalent substitutions, improvements, etc., made within the principles of this application should be included within the protection scope of this application.

Claims

1. An AI large model-based questionnaire intelligent generation method, characterized in that, The method includes the following steps: Obtain the topics to be surveyed in the questionnaire to be generated, the characteristics of the target respondents, the historical questions in the historical questionnaires, and the characteristics and answering behavior data of the historical respondents; By comparing the characteristics of the target respondents with those of historical respondents, similar respondents to the target respondents are selected from the historical respondents, and the answer behavior data of the similar respondents are linked to the historical questions; For each historical question, the overall acceptance performance of the target respondents for each historical question was determined by analyzing the completion rate, willingness to continue answering, and stability of answering pace of similar respondents associated with it. Based on the overall acceptance performance value, high-acceptance questions were selected from all historical questions. By analyzing the difficulty of understanding the question, the complexity of the response interaction, and the degree of fit between the question and the real-life scenario, the upper bound of the target respondents' understanding, the upper bound of their response, and the lower bound of the scenario fit were determined. Based on the topic of the questionnaire to be generated, an AI model is called to generate candidate questions. The difficulty of understanding the question, the complexity of the answering interaction, and the degree of fit between the question and the real-life scenario are analyzed and compared with the upper bound of understanding, the upper bound of answering, and the lower bound of scenario fit, respectively, in order to optimize the generation of the questionnaire from the candidate questions.

2. The intelligent questionnaire generation method based on AI large model as described in claim 1, characterized in that, The answering behavior data includes: the number of historical respondents who entered each historical question, the number of historical respondents who completed each historical question in accordance with preset rules, the number of historical respondents who completed each historical question and entered subsequent questions, and the median time spent by all historical respondents on each historical question.

3. The intelligent questionnaire generation method based on AI large model as described in claim 1, characterized in that, The similar respondents selected from historical respondents to identify the target respondent include: Analyze the characteristics of the respondents and extract feature terms; Calculate the intersection of the feature word categories of each historical respondent and the feature word categories of the target respondent. If the proportion of this intersection in the total feature word categories of the target respondent is greater than the preset matching ratio threshold, then the corresponding historical respondent will be regarded as a similar respondent to the target respondent.

4. The intelligent questionnaire generation method based on AI large model as described in claim 2, characterized in that, The overall acceptance performance value is positively correlated with the completion rate of the response, the willingness to continue responding, and the stability of the response rhythm.

5. The intelligent questionnaire generation method based on AI large model as described in claim 4, characterized in that, The completion rate of each history question is determined by the ratio of the number of similar respondents who completed the question in accordance with the preset rules to the total number of similar respondents who entered the corresponding history question.

6. The intelligent questionnaire generation method based on AI large model as described in claim 4, characterized in that, The willingness to continue answering each history question is determined by the ratio of the number of similar respondents who have completed each history question and moved on to the number of similar respondents who have completed the questions in accordance with the preset rules.

7. The intelligent questionnaire generation method based on AI large model as described in claim 4, characterized in that, The stability of the answering pace for each history question is determined by the deviation of the median answering time for each history question from the average answering time for the same type of question among all similar respondents.

8. The intelligent questionnaire generation method based on AI large model as described in claim 1, characterized in that, The process for determining the upper bound of the target respondents' understanding, the upper bound of their responses, and the lower bound of the scenario fit is as follows: The median difficulty of understanding the question, the median complexity of the response interaction, and the median relevance of the question to real-life scenarios for all highly acceptable questions were used as the upper bound of the target respondents' understanding, the upper bound of their responses, and the lower bound of their scenario relevance, respectively.

9. The intelligent questionnaire generation method based on AI large model as described in claim 1, characterized in that, The optimized candidate questions are used to generate a questionnaire, including: By analyzing the degree of mismatch between the difficulty of understanding the question, the complexity of the answering interaction, and the degree of fit between the question and the real-life scenario of the candidate questions and the upper bound of understanding, the upper bound of answering, and the lower bound of scenario fit, respectively, the degree of mismatch of the difficulty of understanding the question, the degree of mismatch of the answering interaction, and the degree of mismatch of the real-life scenario of the candidate questions are constructed, and the mismatch source sequence of the candidate questions is formed in descending order. Extract the core fragments of candidate questions, and perform a preset directional reconstruction action on the first element in the mismatch source sequence to modify the text outside the core fragments of the candidate questions. Update the mismatch source sequence based on the text corresponding to the modified candidate questions until the preset convergence condition is met to obtain the final modified text of the candidate questions, and generate a questionnaire.

10. A questionnaire intelligent generation system based on an AI large model, comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the intelligent generation method for questionnaires based on AI large models as described in any one of claims 1-9.

Images