Intelligent interview questionnaire method and system based on multi-round dialogue engine

Abstract

The present application relates to the technical field of intelligent interview questionnaire, and more particularly to an intelligent interview questionnaire method and system based on a multi-round dialogue engine. The method generates an initial question sequence by receiving interview purposes and parameter configurations, constructs questionnaire items containing structured question types and dialogue question types, extracts constraint features, performs consistency checking of semantic information and constraint features on the basis of answer contents obtained in dialogue interviews to generate contradiction marks, positions contradiction points by the dialogue engine according to the contradiction marks to generate follow-up questions, and combines complete dialogue records after receiving supplementary answers and updating contradiction marks. The method can effectively find and clarify contradictions in answers, and improve the accuracy and integrity of interview data.

Description

Technical Field

[0001] This invention relates to the field of intelligent interview questionnaire technology, and in particular to an intelligent interview questionnaire method and system based on a multi-turn dialogue engine. Background Technology

[0002] Traditional questionnaire surveys primarily rely on two methods: static questionnaires and human interviewers. Static questionnaires embed pre-designed questions, options, and transition logic in paper or electronic forms, with respondents answering in a fixed order. These questionnaires are then collected and processed by the survey team. Human interviews, on the other hand, involve interviewers asking questions in person or remotely, progressing step-by-step according to an interview outline. With the development of natural language processing technology, intelligent questionnaire systems have begun to appear in online interview scenarios. These systems present questions to respondents through voice or text interaction and record their answers. These systems are typically based on pre-programmed dialogue trees or intent recognition rules, mapping questionnaire questions to dialogue nodes and allowing respondents to jump to the corresponding branch questions based on their options or keywords.

[0003] Current practices often treat interviews and questionnaires as a one-time information collection process, focusing on question prioritization, branching path selection, and recording of responses. The core of design and execution is configuring questions before the interview and ensuring process stability during the interview. The system's processing of respondents' answers remains at the level of text collection and shallow semantic matching, such as identifying options, extracting keywords, or making simple sentiment judgments. Conventional practices lack the ability to dynamically verify and delve deeper into additional information or inconsistencies in respondents' answers. Respondents might choose an option in an objective question but give a contradictory explanation in an open-ended conversational question, or give contradictory answers on the same topic. Existing systems often mechanically record these instances as independent answers, failing to detect anomalies during the interview and guide respondents to clarify.

[0004] Another common drawback is the delayed data quality control. Because the interview process lacks real-time assessment of answer consistency, contradictory information can only be detected manually or by algorithms during the data processing stage, leading to high rework costs and reduced sample usability. Both of these limitations stem from insufficient ability to identify and respond to information conflicts during the interview process, making intelligent questionnaires less efficient at obtaining in-depth, authentic information. Summary of the Invention

[0005] This invention provides an intelligent interview questionnaire method and system based on a multi-turn dialogue engine, which can solve the problems in the prior art.

[0006] A first aspect of this invention provides an intelligent interview questionnaire method based on a multi-turn dialogue engine, comprising:

[0007] Receive the interview purpose description and interview parameter configuration, and call the language model interface to generate the initial interview question sequence;

[0008] Construct a questionnaire project that includes structured questions and conversational interview questions, obtain the respondents' answers to the structured questions, and extract constraint features from the answers;

[0009] In the conversational interview format, the interview questions in the initial interview question sequence are presented to the respondents. The respondents' answers to the interview questions are received, semantic information is extracted from the answers, the semantic information is checked for consistency with the constraint features, and contradiction markers are generated based on the consistency check results.

[0010] Input contradiction markers, semantic information, and constraint features into the dialogue engine. The dialogue engine locates the contradiction points between semantic information and constraint features based on the contradiction markers and generates follow-up questions for the contradiction points.

[0011] Present follow-up questions to the respondents, receive their supplementary answers to the follow-up questions, verify the consistency between the supplementary answers and the constraint features, update the contradiction markers based on the verification results, merge the supplementary answers and the answers to form a complete dialogue record and output it.

[0012] Receive the interview purpose description and interview parameter configuration, and call the language model interface to generate the initial interview question sequence, including:

[0013] Semantic analysis is performed on the description of the interview purpose to identify the binding and non-binding descriptive statements in the description of the interview purpose. Numerical range expressions and category enumeration expressions are extracted from the binding descriptive statements as binding elements, and opinion expressions and behavioral expressions are extracted from the non-binding descriptive statements as exploratory elements.

[0014] Construct closed-ended question templates with options based on the constraints and mark them as structured question templates; construct open-ended question templates without options based on the exploratory elements and mark them as conversational interview question templates.

[0015] The structured question template, the conversational interview question template, and the interview parameter configuration are combined into structured prompt text, and the structured prompt text is passed to the language model interface.

[0016] The language model interface generates structured questions for obtaining constraint features based on structured question templates, and generates conversational interview questions for accepting constraint feature verification based on conversational interview question templates, forming an initial interview question sequence.

[0017] A questionnaire project was constructed, including structured questions and conversational interview questions. Respondents' responses to the structured questions were obtained, and constraint features were extracted from the responses, including:

[0018] Structured question types are generated based on the structured question type template in the initial interview question sequence, and conversational interview question types are generated based on the conversational interview question type template in the initial interview question sequence. The structured question types and conversational interview question types are then combined to form questionnaire items.

[0019] Present the structured question types in the questionnaire to the respondents and receive their responses to the structured question types;

[0020] Syntactic dependency analysis is performed on the response results. Modification relations are identified from the syntactic dependency analysis results. The head words modified by numerals or quantifiers in the modification relations are extracted as numerical constraint units, and the head words modified by adjectives or adverbs in the modification relations are extracted as textual constraint units.

[0021] Based on the results of syntactic dependency analysis, it is determined that there is a direct dependency relationship between numerical constraint units and textual constraint units. Numerical constraint units and textual constraint units with direct dependency relationships are combined into constraint unit pairs.

[0022] Numerical constraint units, textual constraint units, and constraint unit pairs are combined to form constraint features.

[0023] The system receives respondents' answers to interview questions, extracts semantic information from the answers, performs consistency checks between the semantic information and constraint features, and generates contradiction markers based on the consistency check results, including:

[0024] Multidimensional semantic analysis is performed on the response content to identify event statements and state statements. The actions and frequency of action occurrence in the event statements are extracted as event semantic information, and the attributes and degree of attribute descriptions in the state statements are extracted as state semantic information. The event semantic information and state semantic information are combined to form semantic information.

[0025] Cross-validate the frequency of action occurrence in the semantic information of events with the numerical range represented by the numerical constraint units in the constraint features. When the frequency of action occurrence exceeds the numerical range, mark the frequency contradiction type.

[0026] The attribute degree description of state semantic information in semantic information is matched and verified with the semantic category represented by text-type constraint unit in constraint features. When the attribute degree description deviates from the semantic category, the degree contradiction type is labeled.

[0027] When labeling frequency contradiction type and degree contradiction type, record the corresponding contradiction fragment position. Based on the numerical constraint unit corresponding to the frequency contradiction type and the text constraint unit corresponding to the degree contradiction type, locate the structured question type from which the constraint feature originates. Combine the frequency contradiction type, degree contradiction type, structured question type and contradiction fragment position to form a contradiction marker.

[0028] The attribute degree description of state semantic information in semantic information is matched and verified with the semantic category represented by text-type constraint units in constraint features. When the attribute degree description deviates from the semantic category, the labeling degree contradiction types include:

[0029] Based on the constraint unit pairs in the constraint features, identify the text-type constraint units corresponding to the attributes in the state semantic information, extract the numerical constraint units that have a direct dependency relationship with the text-type constraint units in the constraint unit pairs, and convert the numerical range represented by the numerical constraint units into degree level intervals.

[0030] Map the attribute degree description in the state semantic information to the degree level interval, determine the degree level value corresponding to the attribute degree description according to the mapping position of the attribute degree description in the degree level interval, and match and verify the degree level value with the category boundary of the semantic category represented by the text-type constraint unit.

[0031] When the degree level value exceeds the category boundary of the semantic category, it is determined that the attribute degree description deviates from the semantic category. The degree contradiction type is marked, and the identifier of the constraint unit pair, the degree level value, and the category boundary are recorded in the degree contradiction type.

[0032] Based on the contradiction markers, the system locates the contradiction points between semantic information and constraint features and generates follow-up questions targeting these contradiction points, including:

[0033] Based on the position of the contradictory fragment recorded in the contradictory marker, the contradictory semantic fragment is located in the semantic information; based on the structured question type identifier recorded in the contradictory marker, the contradictory constraint unit is located in the constraint feature; and the semantic expression corresponding to the contradictory semantic fragment and the constraint content corresponding to the contradictory constraint unit are extracted.

[0034] Semantic analysis is performed on semantic expressions and constraint content to extract core arguments and argument strength from semantic expressions, and constraint scope and constraint boundaries from constraint content.

[0035] Compare the semantic consistency between the core argument and the scope of constraints, and the numerical consistency between the argument strength and the constraint boundary. When the comparison results of semantic consistency or numerical consistency are inconsistent, the corresponding core argument, argument strength, constraint scope, and constraint boundary are identified as the contradiction points between semantic information and constraint features.

[0036] The type of conflict relationship is determined based on the direction of deviation of the core argument and the scope of constraints in the contradiction, and the degree of deviation of the argument intensity from the constraint boundary. The corresponding questioning template is then selected from the questioning template library based on the type of conflict relationship.

[0037] The core arguments and their intensity are combined and filled into the question subject slot of the follow-up question template. The constraint range and constraint boundary are combined and filled into the reference object slot of the follow-up question template. Follow-up questions targeting the contradictions are generated based on the content filled in the question subject slot and the content filled in the reference object slot.

[0038] Present follow-up questions to the respondents, receive their supplementary answers to the follow-up questions, perform consistency checks between the supplementary answers and the constraint features, update the contradiction markers based on the check results, merge the supplementary answers and the original answers to form a complete dialogue record, and output the following:

[0039] Based on the position of the contradictory fragment recorded in the contradictory marker, the original contradictory semantic fragment is located in the semantic information, and the supplementary answer is semantically parsed to extract the supplementary semantic fragment;

[0040] Calculate the semantic similarity between the supplementary semantic fragment and the original contradictory semantic fragment. When the semantic similarity is higher than the preset similarity threshold, it is determined that there is a correction relationship between the supplementary semantic fragment and the original contradictory semantic fragment. Then, the supplementary semantic fragment replaces the original contradictory semantic fragment to generate an updated semantic fragment.

[0041] When the semantic similarity is lower than the preset similarity threshold, it is determined that the supplementary semantic fragment and the original contradictory semantic fragment are in a supplementary relationship, and the supplementary semantic fragment and the original contradictory semantic fragment are merged to generate an updated semantic fragment;

[0042] Perform a consistency check between the updated semantic fragment and the contradictory constraint unit recorded in the contradiction marker. If the check result is consistent, delete the corresponding contradiction type in the contradiction marker. If the check result is inconsistent, retain the corresponding contradiction type in the contradiction marker and output the updated contradiction marker.

[0043] The updated semantic fragments will replace the positions corresponding to the original contradictory semantic fragments in the semantic information, and the supplementary answers will be inserted into the positions corresponding to the answers according to the time sequence of the follow-up questions, forming a complete dialogue record and outputting the complete dialogue record.

[0044] A second aspect of this invention provides an intelligent interview questionnaire system based on a multi-turn dialogue engine, comprising:

[0045] The question generation unit receives the interview purpose description and interview parameter configuration, and calls the language model interface to generate an initial interview question sequence.

[0046] The questionnaire construction unit is used to construct questionnaire items including structured questions and conversational interview questions, obtain the respondents' answers to structured questions, and extract constraint features from the answers.

[0047] The semantic verification unit is used in conversational interview questions to present the interview questions in the initial interview question sequence to the respondents, receive the respondents’ answers to the interview questions, extract semantic information from the answers, perform consistency verification between the semantic information and the constraint features, and generate contradiction markers based on the consistency verification results.

[0048] The contradiction follow-up unit is used to input contradiction markers, semantic information and constraint features into the dialogue engine. The dialogue engine locates the contradiction points between semantic information and constraint features based on the contradiction markers and generates follow-up questions for the contradiction points.

[0049] The dialogue output unit is used to present follow-up questions to the respondents, receive supplementary answers from the respondents to the follow-up questions, perform consistency checks between the supplementary answers and the constraint features, update the contradiction markers based on the check results, merge the supplementary answers and the answers to form a complete dialogue record and output it.

[0050] A third aspect of the present invention provides an electronic device, comprising:

[0051] processor;

[0052] Memory used to store processor-executable instructions;

[0053] The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0054] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0055] This method organically combines structured question-and-answer sessions with conversational interviews, utilizing the constraint features in the responses to verify consistency in subsequent dialogues, significantly improving the internal consistency quality of interview data. When respondents provide information in the dialogue that contradicts their earlier answers, the system can instantly generate a contradiction marker and accurately pinpoint the contradiction, avoiding the problem of overlooked contradictory information or difficulty in tracing it afterward, as is common in conventional interviews. This effectively reduces data cleaning costs and ensures the credibility and usability of the final output data. This invention utilizes a dialogue engine to automatically generate follow-up questions for contradictions, enabling the interview process to have dynamic error correction capabilities. Follow-up questions are not mechanical repetitions but rather targeted in-depth explorations around identified contradictions, guiding respondents to re-examine their statements and supplement more accurate and complete background information. This mechanism helps eliminate semantic ambiguity, reveal hidden attitudes or behavioral motivations, and simultaneously enhances the logicality and detail richness of the collected dialogue data, providing a more solid chain of evidence for subsequent analysis. The complete dialogue record merges supplementary answers with the original answers and updates the contradiction marker status simultaneously, ensuring that the final output not only covers a coherent semantic context but also includes clear and traceable consistency verification information. While ensuring data collection efficiency, it balances the structured advantages of quantitative data with the flexibility of qualitative data, making it suitable for various interview scenarios such as satisfaction surveys, user needs insights, and health screenings. It effectively reduces follow-up or re-interviews due to inconsistencies in information, improving the overall execution efficiency of research projects. Attached Figure Description

[0056] Figure 1 This is a flowchart illustrating the intelligent interview questionnaire method based on a multi-turn dialogue engine.

[0057] Figure 2 Flowchart for semantic contradiction resolution and labeling operations;

[0058] Figure 3 This is a diagram illustrating the comparison of conflict resolution rates.

[0059] Figure 4 This is a diagram illustrating the effectiveness of multiple rounds of follow-up questioning. Detailed Implementation

[0060] 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, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0061] The technical solution of the present invention will be described in detail below with reference to specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments.

[0062] Figure 1 This is a flowchart illustrating the intelligent interview questionnaire method based on a multi-turn dialogue engine, as described in an embodiment of the present invention.

[0063] Intelligent interview questionnaire methods based on multi-turn dialogue engines include:

[0064] Receive the interview purpose description and interview parameter configuration, and call the language model interface to generate the initial interview question sequence;

[0065] Construct a questionnaire project that includes structured questions and conversational interview questions, obtain the respondents' answers to the structured questions, and extract constraint features from the answers;

[0066] In the conversational interview format, the interview questions in the initial interview question sequence are presented to the respondents. The respondents' answers to the interview questions are received, semantic information is extracted from the answers, the semantic information is checked for consistency with the constraint features, and contradiction markers are generated based on the consistency check results.

[0067] Input contradiction markers, semantic information, and constraint features into the dialogue engine. The dialogue engine locates the contradiction points between semantic information and constraint features based on the contradiction markers and generates follow-up questions for the contradiction points.

[0068] Present follow-up questions to the respondents, receive their supplementary answers to the follow-up questions, verify the consistency between the supplementary answers and the constraint features, update the contradiction markers based on the verification results, merge the supplementary answers and the answers to form a complete dialogue record and output it.

[0069] Receive the interview purpose description and interview parameter configuration, and call the language model interface to generate the initial interview question sequence, including:

[0070] Semantic analysis is performed on the description of the interview purpose to identify the binding and non-binding descriptive statements in the description of the interview purpose. Numerical range expressions and category enumeration expressions are extracted from the binding descriptive statements as binding elements, and opinion expressions and behavioral expressions are extracted from the non-binding descriptive statements as exploratory elements.

[0071] Construct closed-ended question templates with options based on the constraints and mark them as structured question templates; construct open-ended question templates without options based on the exploratory elements and mark them as conversational interview question templates.

[0072] The structured question template, the conversational interview question template, and the interview parameter configuration are combined into structured prompt text, and the structured prompt text is passed to the language model interface.

[0073] The language model interface generates structured questions for obtaining constraint features based on structured question templates, and generates conversational interview questions for accepting constraint feature verification based on conversational interview question templates, forming an initial interview question sequence.

[0074] When performing semantic parsing on interview purpose descriptions, it is necessary to distinguish the nature and type of the descriptive statements. Interview purpose descriptions are typically input by researchers in natural language, containing two distinct types of descriptive statements: one type is restrictive descriptive statements with explicit limiting conditions, such as "the interviewee is between 25 and 45 years old," "monthly income is no less than 8,000 yuan," and "occupation category is teacher or doctor," etc.; the other type is non-restrictive descriptive statements without explicit numerical or category boundaries, such as "to understand the interviewee's views on remote work" and "to explore the interviewee's daily consumption decision-making behavior," etc. The semantic parsing process uses natural language processing technology to segment and classify the input text, labeling each statement as either restrictive or non-restrictive.

[0075] The elements extracted from the restrictive descriptive statements are called restrictive elements, which specifically include two forms of expression: numerical range expression and category enumeration expression. Numerical range expression corresponds to descriptions containing upper and lower bounds; for example, "25 to 45 years old" will be parsed as a numerical range constraint with a lower bound of 25 and an upper bound of 45, and the attribute being age. Category enumeration expression corresponds to descriptions listing a limited number of options; for example, "occupational category is teacher or doctor" will be parsed as an option set {teacher, doctor}, and the attribute being occupational enumeration constraint. These two types of restrictive elements will serve as the basis for constructing closed-ended question templates in subsequent steps, and also as a benchmark for verifying the consistency of respondents' answers.

[0076] Elements extracted from non-binding descriptive statements are called exploratory elements, which also include two forms of expression: opinion expression and behavior expression. Opinion expression corresponds to the part of the description involving subjective content such as attitudes, cognitions, and evaluations. For example, "views on remote work" is analyzed as an opinion element whose object of exploration is remote work and whose exploration dimension is subjective attitude. Behavioral expression corresponds to the part of the description involving objective content such as specific actions, habits, and decision-making processes. For example, "daily consumption decision-making behavior" is analyzed as a behavioral element whose object of exploration is consumption decision-making and whose exploration dimension is behavioral process. Exploratory elements do not contain numerical values ​​or category boundaries, therefore the corresponding questions do not have fixed options and need to guide respondents to express themselves freely in an open-ended manner.

[0077] When constructing closed-ended question templates based on constraint elements, for numerical range constraints, single-choice or multiple-choice question templates containing options within numerical ranges are generated. For example, the age constraint "25 to 45 years old" is transformed into a single-choice question framework containing several age range options. For category enumeration constraints, single-choice or multiple-choice question templates are generated with the enumerated categories as the option set. For example, the occupation constraint {teacher, doctor} is transformed into a single-choice question framework with these categories as options. These templates are marked as structured question templates, characterized by a fixed list of options for the questions, requiring respondents to choose only from preset options. The resulting responses are structured data, facilitating the subsequent extraction of constraint features.

[0078] When constructing open-ended question templates based on exploratory elements, for opinion-based elements, open-ended question templates are generated with guiding statements such as "How do you view…?" or "What is your evaluation of…?"; for behavior-based elements, open-ended question templates are generated with narrative statements such as "Please describe your specific experiences in…" or "How do you usually…?". These templates are labeled as conversational interview question templates, characterized by questions without fixed options, responses from respondents in free text format, and unstructured data requiring deep processing through semantic extraction and a dialogue engine.

[0079] Interview parameter configuration is a set of control parameters input by the interview designer in the system interface, typically including the maximum number of questions, interview language type, question difficulty level, and limits on follow-up question depth. When combining structured question templates, conversational interview question templates, and interview parameter configurations into structured prompt text, the three types of content are sequentially concatenated according to the prompt word format required by the language model interface. Specifically, the construction of structured prompt text follows this logic: first, the interview parameter configuration is converted into generation constraint instructions for the language model, such as "Please generate no more than..." A set of structured questions, each containing... "One option"; then embed the structured question template into the prompt text as an example or fill-in-the-blank; then embed the conversational interview question template in another block, explaining the exploration direction and expression style of the open-ended questions to be generated. This indicates the upper limit on the number of structured questions. This indicates the number of options for each structured question; both are derived from the interview parameter configuration.

[0080] After the structured prompt text is passed to the language model interface, the interface generates questions for two types of templates. Based on the structured question template, the language model interface generates structured questions for obtaining constraint features. These questions inherit the option framework of the closed-ended question template, and the generated question options are consistent with the numerical range or category enumeration in the constraint elements, ensuring that respondents can output structured data that can be directly used for constraint verification. Based on the conversational interview question template, the language model interface generates conversational interview questions for constraint feature verification. These questions guide respondents to elaborate using open-ended statements, while also semantically relating to the constraint elements, allowing respondents' answers to be processed by the subsequent consistency verification process.

[0081] The two types of questions are arranged in the order of structured questions first, followed by conversational interview questions, forming the initial interview question sequence. This arrangement has significant logical implications: placing structured questions first allows for the collection of respondents' constraint characteristics before the conversational interview begins, providing baseline data for subsequent consistency checks; placing conversational interview questions later allows for in-depth analysis of respondents' open-ended responses given the known constraint characteristics, effectively identifying contradictions between the response content and the constraint characteristics, and triggering the dialogue engine to generate targeted follow-up questions. The entire question sequence generation process organically combines the structured analysis of the interview purpose with the generative capabilities of the language model, achieving an automated transformation from natural language interview intent to an executable questionnaire structure.

[0082] A questionnaire project was constructed, including structured questions and conversational interview questions. Respondents' responses to the structured questions were obtained, and constraint features were extracted from the responses, including:

[0083] Structured question types are generated based on the structured question type template in the initial interview question sequence, and conversational interview question types are generated based on the conversational interview question type template in the initial interview question sequence. The structured question types and conversational interview question types are then combined to form questionnaire items.

[0084] Present the structured question types in the questionnaire to the respondents and receive their responses to the structured question types;

[0085] Syntactic dependency analysis is performed on the response results. Modification relations are identified from the syntactic dependency analysis results. The head words modified by numerals or quantifiers in the modification relations are extracted as numerical constraint units, and the head words modified by adjectives or adverbs in the modification relations are extracted as textual constraint units.

[0086] Based on the results of syntactic dependency analysis, it is determined that there is a direct dependency relationship between numerical constraint units and textual constraint units. Numerical constraint units and textual constraint units with direct dependency relationships are combined into constraint unit pairs.

[0087] Numerical constraint units, textual constraint units, and constraint unit pairs are combined to form constraint features.

[0088] When constructing the questionnaire, structured question types are generated one by one based on the pre-generated structured question template in the initial interview question sequence. These question types typically include single-choice questions, multiple-choice questions, rating scale questions, and fill-in-the-blank questions. Each question type carries a clear set of options or filling constraints, facilitating the subsequent structured analysis of the responses. Simultaneously, open-ended conversational interview questions are generated based on the conversational interview question template in the initial interview question sequence. These questions use natural language descriptions as the question carrier, do not have fixed options, and allow respondents to answer in a free-form manner. These two types of questions are combined in a predetermined logical order to form a complete questionnaire. The structured questions are usually placed at the beginning of the questionnaire to collect basic characteristic information of the respondents, while the conversational interview questions are placed at the end to delve deeper into the respondents' subjective cognition and attitudes.

[0089] After presenting the structured question types in the questionnaire to the respondents, the system receives their responses. The format of the responses depends on the question type: for multiple-choice questions, the responses are a set of option identifiers; for fill-in-the-blank questions, the responses are natural language text fragments entered by the respondents. For multiple-choice questions, the option identifiers need to be mapped back to the corresponding option text content, thus uniformly converting all responses into natural language text for subsequent unified syntactic analysis.

[0090] When performing syntactic dependency parsing on the response results, the response text is input into the syntactic dependency parser, which generates a dependency tree structure with words as nodes and dependency relations as directed edges. Syntactic dependency parsing can reveal the grammatical modification relations between words in a sentence, such as nummod, amod, and adverb modifiers. By traversing all nodes in the dependency tree, word pairs with modification relations are identified, that is, the set of word pairs consisting of the modifier (the starting point of the dependency arc) and the headword (the ending point of the dependency arc) is found.

[0091] In the identified modification relationships, word pairs with modifiers whose parts of speech are numerals or classifiers are extracted, and the headword in the word pair is marked as a numerical constraint unit. A numerical constraint unit represents a core concept in the response that carries a specific quantitative meaning. For example, in "worked for 5 years," "year" is modified by the numeral "5," so "year" is extracted as a numerical constraint unit, and its corresponding numerical modifier "5" is recorded as the numerical attribute of this constraint unit. Word pairs with modifiers whose parts of speech are adjectives or adverbs are extracted, and the headword in the word pair is marked as a textual constraint unit. Textual constraint units represent core concepts in the response that carry a qualitative descriptive meaning. For example, in "very satisfied," "satisfied" is modified by the adverb "very," so "satisfied" is extracted as a textual constraint unit, and its corresponding modifier "very" is recorded as the degree attribute of this constraint unit. Through the above extraction process, the quantitative and qualitative information implied in the answer results can be structured into sets of numerical constraint units and sets of textual constraint units, respectively.

[0092] After extracting numerical and textual constraint units, the syntactic relationships between the two types of constraint units are further analyzed. Specifically, in the dependency tree, it is determined whether a direct dependency edge exists between a numerical constraint unit node and a textual constraint unit node, i.e., whether there is a one-step reachable directed dependency relationship between the two nodes, rather than an indirect path requiring intermediate nodes. If the numerical constraint unit... Text-type constraint unit If a direct dependency exists between them, then the two will be combined into a constraint element pair. This records the semantic binding relationship between the two constraint units in the pair. For example, in the statement "monthly income of about 20,000 yuan and relatively high quality of life", if "10,000 yuan" (numerical constraint unit, numerical attribute "two") and "quality" (textual constraint unit, degree attribute "relatively high") have a direct dependency edge in the dependency tree, then the two are combined into a constraint unit pair so as to verify the consistency of the respondents' statements on the relationship between income and quality of life in the subsequent conversational interview stage.

[0093] It should be noted that not all numerical constraint units and textual constraint units can form constraint unit pairs. Only two constraint units with a direct dependency relationship in the dependency tree are combined. Numerical constraint units that cannot form constraint unit pairs with any textual constraint units, and textual constraint units that cannot form constraint unit pairs with any numerical constraint units, are retained as independent constraint units in their respective sets and participate in the subsequent consistency verification process. This design ensures the integrity of constraint features and avoids the loss of valid constraint information due to forced pairing.

[0094] The constraint features are formed by merging the sets of numerical constraint units, text constraint units, and constraint unit pairs. These constraint features are stored in a structured data format. Each numerical constraint unit record includes a central term, a numerical attribute, and its source question number; each text constraint unit record includes a central term, a degree attribute, and its source question number; and each constraint unit pair record includes the binding relationship between the numerical and text constraint unit identifiers. This structured storage method facilitates rapid retrieval and recall during the conversational interview phase, and supports the dialogue engine in accurately identifying potential contradictions between the respondent's answers and the structured response results when generating follow-up questions.

[0095] The core of the entire constraint feature extraction process lies in transforming unstructured natural language responses into computable and comparable structured constraint representations, thereby providing reliable benchmark data for subsequent consistency verification. Syntactic dependency parsing, as the technical foundation of this process, can automatically identify quantitative and qualitative descriptive relationships from any natural language text without relying on a pre-set knowledge base. It has strong domain adaptability and can be applied to questionnaire data processing needs in various interview scenarios such as healthcare, education, consumer behavior, and social surveys.

[0096] like Figure 2 As shown, Figure 2 This is a flowchart of semantic contradiction resolution and labeling operations.

[0097] The system receives respondents' answers to interview questions, extracts semantic information from the answers, performs consistency checks between the semantic information and constraint features, and generates contradiction markers based on the consistency check results, including:

[0098] Multidimensional semantic analysis is performed on the response content to identify event statements and state statements. The actions and frequency of action occurrence in the event statements are extracted as event semantic information, and the attributes and degree of attribute descriptions in the state statements are extracted as state semantic information. The event semantic information and state semantic information are combined to form semantic information.

[0099] Cross-validate the frequency of action occurrence in the semantic information of events with the numerical range represented by the numerical constraint units in the constraint features. When the frequency of action occurrence exceeds the numerical range, mark the frequency contradiction type.

[0100] The attribute degree description of state semantic information in semantic information is matched and verified with the semantic category represented by text-type constraint unit in constraint features. When the attribute degree description deviates from the semantic category, the degree contradiction type is labeled.

[0101] When labeling frequency contradiction type and degree contradiction type, record the corresponding contradiction fragment position. Based on the numerical constraint unit corresponding to the frequency contradiction type and the text constraint unit corresponding to the degree contradiction type, locate the structured question type from which the constraint feature originates. Combine the frequency contradiction type, degree contradiction type, structured question type and contradiction fragment position to form a contradiction marker.

[0102] When performing multidimensional semantic analysis on the responses, it is necessary to first perform syntactic analysis and semantic role labeling on the natural language text input by the respondents, breaking down the responses into several semantic units. Descriptions involving a subject performing an action, an event occurring, or a behavior appearing are categorized as event statements; expressions describing a subject's state, characteristics, or attributes are categorized as state statements. This classification process relies on a pre-trained language model's deep understanding of sentence structure, enabling it to distinguish between different expressions such as "I go to the gym three times a week" (event statement) and "My physical condition has always been very good" (state statement).

[0103] For event statements, the focus is on extracting action words and their associated frequency expressions. Action words are obtained through predicate recognition, while frequency expressions include explicit quantifiers (such as "three times," "twice a week") and implicit frequency descriptions (such as "occasionally," "frequently," "never"). For implicit frequency descriptions, they need to be mapped to standardized frequency ranges; for example, "occasionally" corresponds to a low-frequency range, and "frequently" corresponds to a high-frequency range, thus forming quantifiable and comparable event semantic information. For state statements, the focus is on extracting attribute words and their degree modifiers. Attribute words describe a trait or state of the respondent, while degree modifiers include degree adverbs (such as "very," "slightly," "relatively") and comparative structures (such as "much better than before," "not as serious as before"). Combining attribute words and degree modifiers forms state semantic information with intensity annotations. Finally, the event semantic information and state semantic information are merged to form a complete semantic information representation, providing input for subsequent consistency verification.

[0104] The first dimension of consistency verification is to cross-validate the frequency of actions in the event semantic information with the numerical range represented by the numerical constraint units in the constraint features. Numerical constraint units The data comes from respondents' answers to structured questions. For example, if a respondent selects "1 to 2 times" in a single-choice question about "how many times do you exercise per week," then a numerical range is formed. ,in This is the lower bound of the range. The upper bound is the frequency of actions extracted from the response content. satisfy or When a frequency discrepancy is detected, it is identified and the type of frequency discrepancy is labeled. The labeling of the frequency discrepancy type not only records the existence of the discrepancy, but also records the direction of the deviation, i.e., whether the frequency is too low or too high, so that the dialogue engine can guide the respondents to clarify the source of the difference in a targeted manner when generating follow-up questions.

[0105] The second dimension of consistency verification is to combine the attribute-level descriptions in the state semantic information with the textual constraint units in the constraint features. The semantic categories represented are matched and verified. Textual constraint units originate from the text options selected by respondents in structured questions. For example, selecting "good" in the question "Your overall health status" corresponds to a semantic category centered on "good." This category is expressed through the semantic embedding space of a pre-trained language model, covering a descriptive range semantically similar to "good." The attribute degree descriptions extracted from the state semantic information are also mapped to the semantic embedding space, and their semantic distance to the semantic category center of the textual constraint unit is calculated. .when Exceeding the preset threshold When an attribute degree description deviates from its semantic scope, it is marked as a degree contradiction type. The degree contradiction type also records the direction of deviation, that is, whether the attribute degree description, relative to the structured question type answer, leans towards a more positive or more negative semantic range, providing directional basis for generating follow-up questions.

[0106] While labeling the frequency and degree of contradiction, it is also necessary to accurately record the location information of contradictory fragments within the original answer content. The location of contradictory fragments is expressed as character offsets. It means that, among them The index of the starting character of the contradictory fragment in the response text. This is the end character index. Recording this positional information allows the subsequent dialogue engine to directly locate the specific statements in the respondent's statement that cause contradictions, rather than broadly questioning the entire answer, thereby improving the accuracy of follow-up questions and the efficiency of the dialogue.

[0107] In the final stage of combining contradiction markers, it is necessary to associate the frequency contradiction type and the degree contradiction type with their corresponding constraint feature sources. Specifically, for each frequency contradiction, its corresponding numerical constraint unit is traced. Furthermore, it determines which structured question the constraint unit originates from; for each degree of contradiction, it traces back to its corresponding text-based constraint unit. Similarly, the structured question type from which the contradiction originates is identified. Four types of information—frequency of contradiction type, degree of contradiction type, corresponding structured question type number, and location of the contradiction fragment—are combined to form a structured contradiction marker object. This contradiction marker object is stored in key-value pairs, where the key is the contradiction type identifier, and the value contains contradiction details, constraint source question type identifier, and contradiction fragment location range.

[0108] When multiple contradictions exist within a single response, the contradiction markers are organized as a list, with each contradiction corresponding to an independent record. This structured storage allows the dialogue engine to prioritize contradictions based on their severity or type during subsequent processing, determining which contradiction to address first and avoiding overwhelming the respondent with multiple follow-up questions in a single round of dialogue. The entire contradiction marker generation process forms a closed loop, ensuring that the differences between semantic information and constraint features are fully captured and transmitted in an operable, structured form to subsequent dialogue flows.

[0109] The attribute degree description of state semantic information in semantic information is matched and verified with the semantic category represented by text-type constraint units in constraint features. When the attribute degree description deviates from the semantic category, the labeling degree contradiction types include:

[0110] Based on the constraint unit pairs in the constraint features, identify the text-type constraint units corresponding to the attributes in the state semantic information, extract the numerical constraint units that have a direct dependency relationship with the text-type constraint units in the constraint unit pairs, and convert the numerical range represented by the numerical constraint units into degree level intervals.

[0111] Map the attribute degree description in the state semantic information to the degree level interval, determine the degree level value corresponding to the attribute degree description according to the mapping position of the attribute degree description in the degree level interval, and match and verify the degree level value with the category boundary of the semantic category represented by the text-type constraint unit.

[0112] When the degree level value exceeds the category boundary of the semantic category, it is determined that the attribute degree description deviates from the semantic category. The degree contradiction type is marked, and the identifier of the constraint unit pair, the degree level value, and the category boundary are recorded in the degree contradiction type.

[0113] In the consistency verification process of conversational interview questions, when extracting state semantic information from the respondents' answers, it is necessary to match and verify the semantic categories represented by the attribute degree descriptions and textual constraint units in the constraint features. State semantic information refers to the degree-modified expressions used by respondents when describing their own state, habits, or behavioral characteristics, such as "occasionally," "often," "very much like," and "almost never." These expressions contain two dimensions of information: the attribute itself (such as "exercise frequency" and "dietary preferences") and the degree modification of that attribute (such as "occasionally" and "very much"). The accuracy of the attribute degree description directly affects whether the subsequent dialogue engine can accurately locate contradictions; therefore, a structured degree contradiction labeling mechanism needs to be established.

[0114] The constraint units in the constraint features are organized in the form of constraint unit pairs. Each constraint unit pair consists of textual constraint units and numerical constraint units, which are directly dependent on each other. Textual constraint units record the semantic category labels selected by the respondents in the structured question types, such as "basically no exercise," "light exercise," "moderate exercise," and "high-intensity exercise." Numerical constraint units record the quantitative data directly corresponding to the semantic category label, such as the specific numerical range of weekly exercise frequency. When identifying textual constraint units corresponding to attributes in the state semantic information, semantic matching is performed between the attribute name and the attribute identifier in the constraint unit pair to locate the target constraint unit pair. Based on this, the numerical constraint units that have a direct dependency relationship with the textual constraint unit are extracted, and the numerical range they represent is obtained, providing a quantitative basis for subsequent degree level conversion.

[0115] Converting the numerical range represented by numerical constraint units into severity level intervals is a crucial step in achieving cross-modal consistency verification. Numerical ranges are typically stored as intervals. The conversion process evenly divides this interval into several severity levels, forming a discretized sequence of severity levels. The granularity of the severity level interval division is determined by the precision settings in the interview parameter configuration. By default, a five-level division is used, corresponding to five severity levels: "very low," "relatively low," "moderate," "relatively high," and "very high." Each severity level corresponds to a sub-interval within the numerical range, and the severity level value... It is an integer, and its value range is 1. ,in This represents the total number of degree levels. The lower bound of the numerical range. Corresponding level value Upper Realm Corresponding level value The intermediate levels are allocated by linear mapping in equal proportion.

[0116] When mapping attribute degree descriptions in state semantic information to degree level intervals, the language model interface is invoked to perform semantic quantization of the attribute degree descriptions, converting natural language degree words into normalized degree scores. The range of values ​​is Normalization score The mapping relationship with the degree level range is as follows: ,in This indicates the rounding operation. For example, "occasionally" is quantized by the language model as... In the five-level severity level range, it is mapped to severity level values. This corresponds to a "lower" level. "Frequently" is quantified as... Mapped to a degree level value This corresponds to a "higher" level. In this way, degree modifiers in natural language are converted into structured degree level values ​​that can be numerically compared with constraint features.

[0117] The semantic categories represented by text-based constraint units have clearly defined category boundaries, which are stored in the form of degree level values, including the lower bound level value of the semantic category. and upper limit level value ,satisfy Describe the degree level value corresponding to the attribute degree description. When performing matching verification with category boundaries, the judgment condition is: or When the degree level value Falling on the boundary of the category In other cases, the description of the attribute degree deviates from the semantic scope, triggering the degree contradiction type annotation process.

[0118] The labeling results of the degree of contradiction type are written into the contradiction label structure as part of the contradiction label, which records the following information: the identifier of the constraint element pair. (Used to accurately locate the constraint unit pairs involving contradictions in the generation of subsequent follow-up questions), the degree level value corresponding to the attribute degree description. (Reflecting the degree and intensity of the actual expression made by the respondents in their responses) and the category boundaries of semantic categories. (This reflects the expected range of degree established by the answers to structured questions). These three pieces of information together constitute a complete description of the degree of contradiction type, enabling the dialogue engine to accurately understand the nature of the contradiction and the direction of the degree deviation.

[0119] To illustrate the above process, consider a specific application scenario: A respondent selected "basically no exercise" as the descriptive label for their exercise habits in a structured question format. The semantic category boundary of the corresponding text-type constraint unit is the degree level. (i.e., the range from "very low" to "relatively low"); the numerical constraint unit directly dependent on this text-type constraint unit records the number of exercise sessions per week as ranging from 0 to 1. After converting this to a five-level severity interval, level 1 corresponds to 0 times and level 2 corresponds to 1 time. However, in the conversational interview question, the respondent answered "I occasionally go to the gym, sometimes two or three times a week." The extracted state semantic information includes the attribute "exercise frequency" and the attribute severity description "occasionally, sometimes two or three times a week." The language model quantifies this composite severity description as... Mapped to a degree level value This corresponds to the "medium" level. Because... Beyond the upper limit of the category boundary Determine if the attribute description deviates from the semantic category, label the degree of contradiction type, and record the constraint unit's effect on the identifier in the contradiction marker. Degree level value and the boundaries of the scope Upon receiving the contradictory marker, the dialogue engine generates targeted follow-up questions to guide the respondent to clarify their actual movement frequency, thereby ensuring the internal consistency of the final dialogue record.

[0120] Based on the contradiction markers, the system locates the contradiction points between semantic information and constraint features and generates follow-up questions targeting these contradiction points, including:

[0121] Based on the position of the contradictory fragment recorded in the contradictory marker, the contradictory semantic fragment is located in the semantic information; based on the structured question type identifier recorded in the contradictory marker, the contradictory constraint unit is located in the constraint feature; and the semantic expression corresponding to the contradictory semantic fragment and the constraint content corresponding to the contradictory constraint unit are extracted.

[0122] Semantic analysis is performed on semantic expressions and constraint content to extract core arguments and argument strength from semantic expressions, and constraint scope and constraint boundaries from constraint content.

[0123] Compare the semantic consistency between the core argument and the scope of constraints, and the numerical consistency between the argument strength and the constraint boundary. When the comparison results of semantic consistency or numerical consistency are inconsistent, the corresponding core argument, argument strength, constraint scope, and constraint boundary are identified as the contradiction points between semantic information and constraint features.

[0124] The type of conflict relationship is determined based on the direction of deviation of the core argument and the scope of constraints in the contradiction, and the degree of deviation of the argument intensity from the constraint boundary. The corresponding questioning template is then selected from the questioning template library based on the type of conflict relationship.

[0125] The core arguments and their intensity are combined and filled into the question subject slot of the follow-up question template. The constraint range and constraint boundary are combined and filled into the reference object slot of the follow-up question template. Follow-up questions targeting the contradictions are generated based on the content filled in the question subject slot and the content filled in the reference object slot.

[0126] After the dialogue engine receives the contradiction marker, semantic information, and constraint features, it first precisely locates the contradictory semantic fragment within the semantic information based on the contradiction fragment location information recorded in the contradiction marker. The contradiction marker stores the character index range written during the previous consistency verification stage. Based on this range, the specific text fragment that conflicts with the constraint features can be extracted from the respondent's complete answer text; this fragment is the contradictory semantic fragment. Simultaneously, the contradiction marker also records the structured question type identifier that triggers the contradiction. Using this identifier, the corresponding constraint unit can be retrieved from the constraint feature set, thus obtaining the contradictory constraint unit. The semantic expression corresponding to the contradictory semantic fragment and the constraint content corresponding to the contradictory constraint unit are extracted simultaneously and used as input for subsequent semantic parsing.

[0127] Semantic parsing is performed on both semantic expressions and constraint content. For semantic expressions, a language model interface is used to structurally decompose contradictory semantic fragments, extracting core arguments and argument strength. Core arguments refer to the main viewpoints or behavioral facts stated by the respondents in the fragment, such as the frequency of a certain type of activity or a specific description of a certain attribute. Argument strength is a numerical representation of the quantitative degree or emotional tendency contained in the core arguments, reflecting the intensity or magnitude of the respondents' statements. Similarly, for constraint content, a language model interface is used to parse contradictory constraint units, extracting constraint scope and constraint boundaries. Constraint scope represents the legal semantic range defined by the constraint unit, such as whether a behavior falls into the category of "occasionally" or "frequently." Constraint boundaries are the quantitative boundary values ​​of the constraint scope, used to support subsequent numerical consistency comparisons.

[0128] After semantic parsing, a semantic consistency comparison is performed between the core argument and the constraint scope, and a numerical consistency comparison is performed between the argument strength and the constraint boundary. Semantic consistency comparison determines whether they belong to the same semantic category by calculating the similarity between the semantic vector of the core argument and the semantic center vector of the constraint scope. If the similarity is below a preset threshold, the semantic consistency comparison result is considered inconsistent. Numerical consistency comparison directly determines the interval assignment between the quantized value corresponding to the argument strength and the upper and lower bounds of the constraint boundary. If the quantized value corresponding to the argument strength falls outside the legal interval defined by the constraint boundary, the numerical consistency comparison result is considered inconsistent. When the semantic consistency or numerical consistency comparison results are inconsistent, the corresponding core argument, argument strength, constraint scope, and constraint boundary are collectively identified as the point of contradiction between semantic information and constraint features, and specific inconsistency type markers are recorded for subsequent conflict relationship classification.

[0129] The type of conflict relationship is determined based on the direction of deviation between the core argument and the scope of constraints, as well as the degree of deviation between the argument's intensity and the constraint boundary. The direction of deviation refers to whether the semantic tendency of the core argument is higher or lower than the semantic center of the constraint scope. For example, if a respondent selects "never" a certain behavior in a structured question but describes specific details of that behavior in a conversational interview, the direction of deviation is "high deviation." The degree of deviation refers to the absolute value of the difference between the quantitative value of the argument's intensity and the nearest boundary value of the constraint boundary, reflecting the severity of the conflict. Combining the direction and degree of deviation, conflict relationship types are divided into several predefined categories, such as "slightly high deviation," "significantly high deviation," "slightly low deviation," "significantly low deviation," and "semantic category transgression." Different conflict relationship types correspond to different probing strategies; therefore, it is necessary to select probing templates that match the conflict relationship type from a probing template library. The follow-up question template library pre-stores template sets for various conflict relationship types, with multiple candidate templates for each type of conflict relationship, in order to improve the linguistic diversity of follow-up questions and avoid respondents developing resistance due to repetitive follow-up questions.

[0130] The structure of the follow-up question template includes two core filling areas: the question subject slot and the reference object slot. These two slots are connected by a fixed question sentence framework, forming a complete follow-up question statement skeleton. The core argument and its intensity are combined and filled into the question subject slot. The combination method is to concatenate the textual description of the core argument with the quantitative expression of the argument intensity into a complete phrase. For example, combining the core argument "exercises daily" with the argument intensity "about 60 minutes" results in "You mentioned exercising for about 60 minutes daily," which is the content filled into the question subject slot. The constraint range and constraint boundary are combined and filled into the reference object slot. The combination method is to concatenate the semantic description of the constraint range with the quantitative range of the constraint boundary into a complete reference phrase. For example, combining the constraint range "exercises occasionally" with the constraint boundary "no more than twice a week" results in "You previously selected 'exercises occasionally (no more than twice a week)'," which is the content filled into the reference object slot.

[0131] After the question subject slot and reference object slot are filled, their contents are substituted into the question sentence framework of the follow-up question template to generate complete follow-up question text. During generation, the follow-up questions undergo a fluency check. A language model interface is used to evaluate the semantic coherence of the generated follow-up questions. If the evaluation result is below the fluency threshold, a regeneration process is triggered. Another template is selected from the candidate templates corresponding to the same type of conflict relationship and re-filled until a follow-up question that meets the fluency requirements is generated. The final generated follow-up questions precisely target the points of conflict, including both the respondent's specific statements in the conversational interview and their choices in the structured question type. This allows the respondent to clearly understand the basis for the follow-up questions, thus providing more targeted supplementary answers and high-quality input data for subsequent consistency checks and conflict marker updates.

[0132] Present follow-up questions to the respondents, receive their supplementary answers to the follow-up questions, perform consistency checks between the supplementary answers and the constraint features, update the contradiction markers based on the check results, merge the supplementary answers and the original answers to form a complete dialogue record, and output the following:

[0133] Based on the position of the contradictory fragment recorded in the contradictory marker, the original contradictory semantic fragment is located in the semantic information, and the supplementary answer is semantically parsed to extract the supplementary semantic fragment;

[0134] Calculate the semantic similarity between the supplementary semantic fragment and the original contradictory semantic fragment. When the semantic similarity is higher than the preset similarity threshold, it is determined that there is a correction relationship between the supplementary semantic fragment and the original contradictory semantic fragment. Then, the supplementary semantic fragment replaces the original contradictory semantic fragment to generate an updated semantic fragment.

[0135] When the semantic similarity is lower than the preset similarity threshold, it is determined that the supplementary semantic fragment and the original contradictory semantic fragment are in a supplementary relationship, and the supplementary semantic fragment and the original contradictory semantic fragment are merged to generate an updated semantic fragment;

[0136] Perform a consistency check between the updated semantic fragment and the contradictory constraint unit recorded in the contradiction marker. If the check result is consistent, delete the corresponding contradiction type in the contradiction marker. If the check result is inconsistent, retain the corresponding contradiction type in the contradiction marker and output the updated contradiction marker.

[0137] The updated semantic fragments will replace the positions corresponding to the original contradictory semantic fragments in the semantic information, and the supplementary answers will be inserted into the positions corresponding to the answers according to the time sequence of the follow-up questions, forming a complete dialogue record and outputting the complete dialogue record.

[0138] After receiving the respondent's supplementary answer to follow-up questions, this supplementary answer needs to be correlated with previously recorded contradictory information to determine whether the respondent's supplementary explanation has truly resolved the original contradiction. The contradiction markers record the start and end character indices of the contradictory fragment in the answer text. Using these two positional information, the original contradictory semantic fragment can be accurately located from the semantic information. The original contradictory semantic fragment refers to the specific semantic content that was determined to contradict the constraint features during the consistency verification phase, such as a respondent's description of the frequency of a certain behavior or a description of the degree of a certain attribute. After location, semantic parsing is performed on the supplementary answer to extract its core semantic units, forming a supplementary semantic fragment. During semantic parsing, key semantic components such as themes, attribute descriptive words, and frequency descriptive words involved in the supplementary answer are identified to ensure that the supplementary semantic fragment can fully carry the substantive information of the respondent's supplementary explanation.

[0139] After extracting the supplementary semantic fragments, the semantic similarity between the supplementary semantic fragments and the original contradictory semantic fragments is calculated. The semantic similarity calculation is based on the vectorized encoding of the two text segments using a language model, and similarity values ​​are obtained through methods such as cosine similarity. ,in A higher value indicates that the two pieces of semantic content are more similar. A preset similarity threshold is used. Used to distinguish two different semantic relations: when At that time, it was considered that the supplementary semantic fragment was highly consistent with the original contradictory semantic fragment in terms of semantic theme, which constituted a direct correction to the original contradictory content, that is, the respondents intended to correct or update their previous statements; when At that time, it was believed that the supplementary semantic fragment and the original contradictory semantic fragment had certain semantic differences and belonged to a supplementary relationship. That is, the respondents provided additional information beyond the original statement, and the two together constituted a complete answer to the question.

[0140] Under the correction relation, the original contradictory semantic fragment is replaced with a supplementary semantic fragment to generate an updated semantic fragment. The replacement operation means that content in the original contradictory semantic fragment deemed inconsistent with the constraint features is replaced by the new expression, and the updated semantic fragment serves as input for subsequent consistency checks. Under the supplementary relation, the supplementary semantic fragment is merged with the original contradictory semantic fragment. During the merging, the main structure of the original contradictory semantic fragment is preserved, and the newly added information from the supplementary semantic fragment is appended or integrated into the corresponding semantic position to form the updated semantic fragment. The merging operation ensures that the contextual information of the original expression is not lost, while incorporating the respondent's supplementary explanations into the complete semantic expression. Regardless of whether it is a correction or supplementary relation, the updated semantic fragment serves as a unified input for subsequent processing.

[0141] The updated semantic fragment is compared with the contradictory constraint units recorded in the contradiction markers for consistency verification. A contradictory constraint unit refers to a constraint feature unit identified as contradicting the semantic information during the initial consistency verification phase; this may be a numerical or textual constraint unit. For numerical constraint units, the frequency or quantity information extracted from the updated semantic fragment is verified to fall within the constraint range. For textual constraint units, the attribute degree description in the updated semantic fragment is verified to meet the requirements of the semantic category. When the verification result is consistent, it indicates that the respondent's supplementary response has effectively resolved the contradiction, and the corresponding contradiction type record is deleted from the contradiction marker, indicating that the contradiction has been resolved. When the verification result is still inconsistent, it indicates that the supplementary response has failed to resolve the contradiction, and the corresponding contradiction type record in the contradiction marker is retained for further processing in subsequent steps or marked as an unresolved contradiction. After the above operations, the updated contradiction marker is output, reflecting the contradiction status that still exists after the current dialogue round ends.

[0142] After updating the contradiction markers, the semantic information and response content are updated synchronously to form a complete dialogue record. At the semantic information level, updated semantic fragments replace the content at the corresponding positions of the original contradictory semantic fragments in the semantic information, ensuring that the semantic information remains up-to-date and complete. If a new round of follow-up questions is triggered, the dialogue engine will perform contradiction analysis based on the updated semantic information, rather than reusing the replaced old semantic content. At the response content level, supplementary responses are inserted into the corresponding positions of the responses in chronological order of the follow-up questions, i.e., after the original responses to the follow-up questions, ensuring the sequentiality and traceability of the dialogue record in the temporal dimension. This insertion method preserves the natural flow of the dialogue, allowing the complete dialogue record to fully reconstruct the evolution of the respondents' expressions in different rounds of questioning and answering.

[0143] The complete dialogue transcript includes the original interview questions, the respondent's initial answers, follow-up questions, and corresponding supplementary answers, arranged chronologically. When outputting the complete transcript, updated conflict markers are simultaneously output; both serve as the final output of the interview results. If unresolved conflicts remain in the updated conflict markers, a new round of follow-up questions can be initiated, or the conflict status can be marked in the complete transcript for subsequent data analysis to assess the quality of the respondent's answers. This mechanism of gradually resolving conflicts and continuously updating semantic information effectively improves the internal consistency of the interview data, reduces data quality issues caused by unclear or contradictory statements from respondents, and provides a reliable structured semantic foundation for in-depth analysis of the interview results.

[0144] like Figure 3 As shown, Figure 3 This diagram illustrates the comparison of conflict resolution rates, showing the difference in resolution between frequency-related conflicts and degree-related conflicts after follow-up questioning. Frequency-related conflicts, involving specific numerical values ​​(such as frequency per week), are easily calibrated with supplementary answers, achieving a resolution rate of 91%. In contrast, degree-related conflicts, involving subjective feelings (such as satisfaction levels), have a resolution rate of only 76%, requiring multiple rounds of guidance. This indicates that the system is more sensitive to objective data verification but needs to enhance its semantic understanding of subjective expressions.

[0145] like Figure 4 As shown, Figure 4 This diagram illustrates the effectiveness of multi-round follow-up questioning, showing that the cumulative conflict resolution rate increases logarithmically with each round of dialogue. The first round of follow-up questioning resolves approximately 60% of conflicts, the second round reaches 82%, and the third round reaches 94%. This demonstrates that the iterative follow-up questioning mechanism of the multi-round dialogue engine can effectively uncover deeper information, but the marginal benefit diminishes, with the improvement slowing down after the third round. This provides data support for setting a maximum number of follow-up questioning rounds (e.g., 3 rounds) in the system.

[0146] A second aspect of this invention provides an intelligent interview questionnaire system based on a multi-turn dialogue engine, comprising:

[0147] The question generation unit receives the interview purpose description and interview parameter configuration, and calls the language model interface to generate an initial interview question sequence.

[0148] The questionnaire construction unit is used to construct questionnaire items including structured questions and conversational interview questions, obtain the respondents' answers to structured questions, and extract constraint features from the answers.

[0149] The semantic verification unit is used in conversational interview questions to present the interview questions in the initial interview question sequence to the respondents, receive the respondents’ answers to the interview questions, extract semantic information from the answers, perform consistency verification between the semantic information and the constraint features, and generate contradiction markers based on the consistency verification results.

[0150] The contradiction follow-up unit is used to input contradiction markers, semantic information and constraint features into the dialogue engine. The dialogue engine locates the contradiction points between semantic information and constraint features based on the contradiction markers and generates follow-up questions for the contradiction points.

[0151] The dialogue output unit is used to present follow-up questions to the respondents, receive supplementary answers from the respondents to the follow-up questions, perform consistency checks between the supplementary answers and the constraint features, update the contradiction markers based on the check results, merge the supplementary answers and the answers to form a complete dialogue record and output it.

[0152] A third aspect of the present invention provides an electronic device, comprising:

[0153] processor;

[0154] Memory used to store processor-executable instructions;

[0155] The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0156] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0157] This invention can be a method, apparatus, system, and / or computer program product. The computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for performing various aspects of the invention.

[0158] 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 the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. An intelligent interview questionnaire method based on a multi-turn dialogue engine, characterized in that, include: Receive the interview purpose description and interview parameter configuration, and call the language model interface to generate the initial interview question sequence; Construct a questionnaire project that includes structured questions and conversational interview questions, obtain the respondents' answers to the structured questions, and extract constraint features from the answers; In the conversational interview format, the interview questions in the initial interview question sequence are presented to the respondents. The respondents' answers to the interview questions are received, semantic information is extracted from the answers, the semantic information is checked for consistency with the constraint features, and contradiction markers are generated based on the consistency check results. Input contradiction markers, semantic information, and constraint features into the dialogue engine. The dialogue engine locates the contradiction points between semantic information and constraint features based on the contradiction markers and generates follow-up questions for the contradiction points. Present follow-up questions to the respondents, receive their supplementary answers to the follow-up questions, verify the consistency between the supplementary answers and the constraint features, update the contradiction markers based on the verification results, merge the supplementary answers and the answers to form a complete dialogue record and output it.

2. The method according to claim 1, characterized in that, Receive the interview purpose description and interview parameter configuration, and call the language model interface to generate the initial interview question sequence, including: Semantic analysis is performed on the description of the interview purpose to identify the binding and non-binding descriptive statements in the description of the interview purpose. Numerical range expressions and category enumeration expressions are extracted from the binding descriptive statements as binding elements, and opinion expressions and behavioral expressions are extracted from the non-binding descriptive statements as exploratory elements. Construct closed-ended question templates with options based on the constraints and mark them as structured question templates; construct open-ended question templates without options based on the exploratory elements and mark them as conversational interview question templates. The structured question template, the conversational interview question template, and the interview parameter configuration are combined into structured prompt text, and the structured prompt text is passed to the language model interface. The language model interface generates structured questions for obtaining constraint features based on structured question templates, and generates conversational interview questions for accepting constraint feature verification based on conversational interview question templates, forming an initial interview question sequence.

3. The method according to claim 1, characterized in that, A questionnaire project was constructed, including structured questions and conversational interview questions. Respondents' responses to the structured questions were obtained, and constraint features were extracted from the responses, including: Structured question types are generated based on the structured question type template in the initial interview question sequence, and conversational interview question types are generated based on the conversational interview question type template in the initial interview question sequence. The structured question types and conversational interview question types are then combined to form questionnaire items. Present the structured question types in the questionnaire to the respondents and receive their responses to the structured question types; Syntactic dependency analysis is performed on the response results. Modification relations are identified from the syntactic dependency analysis results. The head words modified by numerals or quantifiers in the modification relations are extracted as numerical constraint units, and the head words modified by adjectives or adverbs in the modification relations are extracted as textual constraint units. Based on the results of syntactic dependency analysis, it is determined that there is a direct dependency relationship between numerical constraint units and textual constraint units. Numerical constraint units and textual constraint units with direct dependency relationships are combined into constraint unit pairs. Numerical constraint units, textual constraint units, and constraint unit pairs are combined to form constraint features.

4. The method according to claim 1, characterized in that, The system receives respondents' answers to interview questions, extracts semantic information from the answers, performs consistency checks between the semantic information and constraint features, and generates contradiction markers based on the consistency check results, including: Multidimensional semantic analysis is performed on the response content to identify event statements and state statements. The actions and frequency of action occurrence in the event statements are extracted as event semantic information, and the attributes and degree of attribute descriptions in the state statements are extracted as state semantic information. The event semantic information and state semantic information are combined to form semantic information. Cross-validate the frequency of action occurrence in the semantic information of events with the numerical range represented by the numerical constraint units in the constraint features. When the frequency of action occurrence exceeds the numerical range, mark the frequency contradiction type. The attribute degree description of state semantic information in semantic information is matched and verified with the semantic category represented by text-type constraint unit in constraint features. When the attribute degree description deviates from the semantic category, the degree contradiction type is labeled. When labeling frequency contradiction type and degree contradiction type, record the corresponding contradiction fragment position. Based on the numerical constraint unit corresponding to the frequency contradiction type and the text constraint unit corresponding to the degree contradiction type, locate the structured question type from which the constraint feature originates. Combine the frequency contradiction type, degree contradiction type, structured question type and contradiction fragment position to form a contradiction marker.

5. The method according to claim 4, characterized in that, The attribute degree description of state semantic information in semantic information is matched and verified with the semantic category represented by text-type constraint units in constraint features. When the attribute degree description deviates from the semantic category, the labeling degree contradiction types include: Based on the constraint unit pairs in the constraint features, identify the text-type constraint units corresponding to the attributes in the state semantic information, extract the numerical constraint units that have a direct dependency relationship with the text-type constraint units in the constraint unit pairs, and convert the numerical range represented by the numerical constraint units into degree level intervals. Map the attribute degree description in the state semantic information to the degree level interval, determine the degree level value corresponding to the attribute degree description according to the mapping position of the attribute degree description in the degree level interval, and match and verify the degree level value with the category boundary of the semantic category represented by the text-type constraint unit. When the degree level value exceeds the category boundary of the semantic category, it is determined that the attribute degree description deviates from the semantic category. The degree contradiction type is marked, and the identifier of the constraint unit pair, the degree level value, and the category boundary are recorded in the degree contradiction type.

6. The method according to claim 1, characterized in that, Based on the contradiction markers, the system locates the contradiction points between semantic information and constraint features and generates follow-up questions targeting these contradiction points, including: Based on the position of the contradictory fragment recorded in the contradictory marker, the contradictory semantic fragment is located in the semantic information; based on the structured question type identifier recorded in the contradictory marker, the contradictory constraint unit is located in the constraint feature; and the semantic expression corresponding to the contradictory semantic fragment and the constraint content corresponding to the contradictory constraint unit are extracted. Semantic analysis is performed on semantic expressions and constraint content to extract core arguments and argument strength from semantic expressions, and constraint scope and constraint boundaries from constraint content. Compare the semantic consistency between the core argument and the scope of constraints, and the numerical consistency between the argument strength and the constraint boundary. When the comparison results of semantic consistency or numerical consistency are inconsistent, the corresponding core argument, argument strength, constraint scope, and constraint boundary are identified as the contradiction points between semantic information and constraint features. The type of conflict relationship is determined based on the direction of deviation of the core argument and the scope of constraints in the contradiction, and the degree of deviation of the argument intensity from the constraint boundary. The corresponding questioning template is then selected from the questioning template library based on the type of conflict relationship. The core arguments and their intensity are combined and filled into the question subject slot of the follow-up question template. The constraint range and constraint boundary are combined and filled into the reference object slot of the follow-up question template. Follow-up questions targeting the contradictions are generated based on the content filled in the question subject slot and the content filled in the reference object slot.

7. The method according to claim 1, characterized in that, Present follow-up questions to the respondents, receive their supplementary answers to the follow-up questions, perform consistency checks between the supplementary answers and the constraint features, update the contradiction markers based on the check results, merge the supplementary answers and the original answers to form a complete dialogue record, and output the following: Based on the position of the contradictory fragment recorded in the contradictory marker, the original contradictory semantic fragment is located in the semantic information, and the supplementary answer is semantically parsed to extract the supplementary semantic fragment; Calculate the semantic similarity between the supplementary semantic fragment and the original contradictory semantic fragment. When the semantic similarity is higher than the preset similarity threshold, it is determined that there is a correction relationship between the supplementary semantic fragment and the original contradictory semantic fragment. Then, the supplementary semantic fragment replaces the original contradictory semantic fragment to generate an updated semantic fragment. When the semantic similarity is lower than the preset similarity threshold, it is determined that the supplementary semantic fragment and the original contradictory semantic fragment are in a supplementary relationship, and the supplementary semantic fragment and the original contradictory semantic fragment are merged to generate an updated semantic fragment; Perform a consistency check between the updated semantic fragment and the contradictory constraint unit recorded in the contradiction marker. If the check result is consistent, delete the corresponding contradiction type in the contradiction marker. If the check result is inconsistent, retain the corresponding contradiction type in the contradiction marker and output the updated contradiction marker. The updated semantic fragments will replace the positions corresponding to the original contradictory semantic fragments in the semantic information, and the supplementary answers will be inserted into the positions corresponding to the answers according to the time sequence of the follow-up questions, forming a complete dialogue record and outputting the complete dialogue record.

8. An intelligent interview questionnaire system based on a multi-turn dialogue engine, used to implement the method as described in any one of claims 1-7, characterized in that, include: The question generation unit receives the interview purpose description and interview parameter configuration, and calls the language model interface to generate an initial interview question sequence. The questionnaire construction unit is used to construct questionnaire items including structured questions and conversational interview questions, obtain the respondents' answers to structured questions, and extract constraint features from the answers. The semantic verification unit is used in conversational interview questions to present the interview questions in the initial interview question sequence to the respondents, receive the respondents’ answers to the interview questions, extract semantic information from the answers, perform consistency verification between the semantic information and the constraint features, and generate contradiction markers based on the consistency verification results. The contradiction follow-up unit is used to input contradiction markers, semantic information and constraint features into the dialogue engine. The dialogue engine locates the contradiction points between semantic information and constraint features based on the contradiction markers and generates follow-up questions for the contradiction points. The dialogue output unit is used to present follow-up questions to the respondents, receive supplementary answers from the respondents to the follow-up questions, perform consistency checks between the supplementary answers and the constraint features, update the contradiction markers based on the check results, merge the supplementary answers and the answers to form a complete dialogue record and output it.

9. An electronic device, characterized in that, include: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the method according to any one of claims 1 to 7.

10. A computer-readable storage medium having computer program instructions stored thereon, characterized in that, When the computer program instructions are executed by the processor, they implement the method described in any one of claims 1 to 7.

Images