Mental health assessment methods, equipment and storage media

Abstract

This application discloses a method, device, and storage medium for mental health assessment. Relating to the field of artificial intelligence technology, the method includes: collecting language expression samples from a user to be assessed; establishing a personalized expression anchor table for the user based on the language expression samples, wherein the personalized expression anchor table indicates the user's language expression anchor points at each option level in a preset rating field; conducting a scale assessment interaction with the user based on a preset mental health assessment form to obtain the user's natural language responses; calibrating the natural language responses according to the personalized expression anchor table to obtain the user's target option level for each item in the mental health assessment form, and generating assessment results based on each target option level. This application can improve the accuracy of mental health assessment results.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, and in particular to a method, device and storage medium for mental health assessment. Background Technology

[0002] As society pays increasing attention to mental health, the demand for mental health assessment and support services is growing in educational institutions, medical settings, and high-risk occupations, placing higher demands on the efficiency, accuracy, and interactive experience of mental health assessments.

[0003] Currently, mental health assessment technology has incorporated AI-assisted assessment solutions. Some systems use knowledge graphs and pre-defined question-and-answer trees to ask users questions and generate assessment results based on user feedback. However, the response and judgment logic of this type of assessment relies heavily on manually constructed knowledge bases and fixed question-and-answer rules. It can only execute standardized interactions according to preset processes, making it difficult to adapt to the complex, ambiguous, or personalized expressions of users in real-world applications. This leads to biases in the identification and judgment of users' psychological states, ultimately resulting in low accuracy of mental health assessment results.

[0004] Therefore, improving the accuracy of mental health assessment results is an urgent problem that needs to be solved. Summary of the Invention

[0005] The main purpose of this application is to provide a method, device and storage medium for mental health assessment, which aims to improve the accuracy of mental health assessment results.

[0006] To achieve the above objectives, this application provides a mental health assessment method, which includes: Collect language expression samples from users to be evaluated, and establish a personalized expression anchor point table for users to be evaluated based on the language expression samples. The personalized expression anchor point table is used to indicate the language expression anchor points of users to be evaluated at each option level of a preset scoring field. Based on a preset psychological assessment form, the scale assessment interaction is conducted with the user to be assessed to obtain the user's natural language response. The natural language responses are calibrated based on the personalized expression anchor table to obtain the target option level of the user to be evaluated for each item in the psychological assessment table, and assessment results are generated based on each target option level.

[0007] In one embodiment, the step of collecting language expression samples from the user to be evaluated includes: Engage in trust-aware interaction with the user to be evaluated to obtain language expression samples from the user to be evaluated; The language expression samples are used to characterize the personal expression scale of the user to be evaluated to preset rating fields, which include frequency, intensity, persistence and scope of influence fields.

[0008] In one embodiment, the step of establishing a personalized expression anchor point table for the user to be evaluated based on the language expression sample includes: The language expression samples are semantically encoded to obtain the corresponding expression vectors; According to the option levels of each of the preset rating fields, the expression vectors are grouped and aggregated to obtain multiple vector groups; Based on each of the vector groupings, the personalized expression anchor points of the user to be evaluated at each option level of each of the preset rating fields are determined, so as to construct the personalized expression anchor point table based on each of the personalized expression anchor points.

[0009] In one embodiment, the step of determining the personalized expression anchor point of the user to be evaluated at each option level of each preset rating field based on each of the vector groupings includes: For any first option level of each of the preset rating fields, determine the number of vectors in the first vector group corresponding to the first option level; If the number of vectors is lower than a first preset threshold, determine the second option level that is adjacent to the first option level among the option levels of the preset rating field; Based on the second vector grouping corresponding to the second option level and the first vector grouping, the personalized expression anchor point of the first option level is determined.

[0010] In one embodiment, the step of interacting with the user to be tested based on a preset psychological assessment form to obtain the user's natural language response includes: Transform the original questions in the pre-set psychological assessment scale into conversational questions; Each of the conversational questions is output sequentially to the user to be evaluated, and the user's response in natural language is obtained as the natural language answer.

[0011] In one embodiment, the step of calibrating the natural language response based on the personalized expression anchor table to obtain the target option level of the user to be assessed for each item in the psychological assessment table includes: For any item in the psychological assessment form, extract the first semantic vector of the natural language answer corresponding to the item, and the second semantic vector of the level of each candidate option of the item; Extract the third semantic vector related to the preset scoring field from the natural language answer corresponding to the question item, and obtain the anchor vector corresponding to each candidate option level of the question item in the personalized expression anchor table; Determine the first matching degree between the first semantic vector and the second semantic vector, and determine the second matching degree between the third semantic vector and the anchor vector; Candidate scores for each candidate option level are determined based on the first matching degree and the second matching degree; Based on each candidate score, the target option level for each question item is determined.

[0012] In one embodiment, the step of determining the candidate score for each candidate option level based on the first matching degree and the second matching degree includes: For any candidate option level of the question item, determine the coverage confidence of the natural language answer corresponding to the question item to the preset scoring field, and determine the penalty confidence of the natural language answer corresponding to the question item, wherein the penalty confidence is evaluated from the dimensions of cross-round semantic contradiction and expression ambiguity; Based on the first matching degree, the second matching degree, the coverage confidence degree, and the penalty confidence degree corresponding to the candidate option level, the candidate score of the candidate option level is determined.

[0013] In one embodiment, the step of determining the target option level of the question item based on each of the candidate scores includes: Normalize each candidate score to obtain the candidate probability of each candidate option level; Determine the first candidate probability with the largest probability and the second candidate probability with the second largest probability among all the candidate probabilities, and determine the difference between the first candidate probability and the second candidate probability; If the difference is greater than or equal to the second preset threshold, the candidate option level corresponding to the first candidate probability is taken as the target option level of the question item; If the difference is less than a second preset threshold, at least one counterfactual answer is generated based on the natural language answer corresponding to the question item, forming a counterfactual set, and the target option level of the question item is determined from each of the candidate option levels based on the counterfactual set.

[0014] In one embodiment, the step of determining the target option level of the item from each of the candidate option levels based on the counterfactual set includes: Determine the counterfactual candidate probability for each counterfactual answer in the counterfactual set; For each counterfactual response, determine the third candidate probability, which is the largest among all counterfactual candidate probabilities corresponding to the counterfactual response, and obtain the counterfactual option level corresponding to the third candidate probability; The option level that appears most frequently among the various counterfactual option levels is taken as the target option level for the question item.

[0015] Furthermore, to achieve the above objectives, this application also provides a mental health assessment device, the mental health assessment device comprising: Anchor point construction module is used to collect language expression samples of users to be evaluated, and to build a personalized expression anchor point table for users to be evaluated based on the language expression samples. The personalized expression anchor point table is used to indicate the language expression anchor points of users to be evaluated at each option level of a preset scoring field. The interaction module is used to interact with the user to be tested based on a preset psychological assessment form to obtain the user's natural language response. The calibration module is used to calibrate the natural language response according to the personalized expression anchor table, obtain the target option level of the user to be evaluated for each item in the psychological assessment table, and generate assessment results based on each target option level.

[0016] In addition, to achieve the above objectives, this application also proposes an electronic device, the device comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, the computer program being configured to implement the steps of the mental health assessment method as described above.

[0017] In addition, to achieve the above objectives, this application also provides a storage medium, which is a computer-readable storage medium, on which a program for implementing a mental health assessment method is stored, and the program for implementing the mental health assessment method is executed by a processor to implement the steps of the mental health assessment method as described above.

[0018] In addition, to achieve the above objectives, this application also provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the mental health assessment method described above.

[0019] This application provides a method for assessing mental health. First, it collects language expression samples from the user to be assessed. Based on these samples, it establishes a personalized expression anchor table for the user, indicating the user's language expression anchor points at each option level in a preset rating field. Then, based on a preset psychological assessment form, it conducts a scale assessment interaction with the user to obtain their natural language responses. Finally, it calibrates the natural language responses according to the personalized expression anchor table to obtain the user's target option level for each item in the psychological assessment form, and generates assessment results based on these target option levels.

[0020] In summary, this application first collects language expression samples from users to be assessed and establishes a personalized expression anchor table for these users, achieving a precise characterization of their personal expression habits and providing a basis for subsequent identification of their true semantics. During the scale assessment interaction, this application calibrates the natural language responses of the users to be assessed based on the personalized expression anchor table, mapping the natural language responses to the target option levels of each item in the psychological assessment scale, and generating assessment results based on each target option level. Therefore, compared to the standardized matching method based on fixed question-and-answer rules and general semantic templates in traditional technologies, this application can adaptively understand the differences in personalized expressions of different users to preset scoring fields, calibrating the users' natural language responses with personal expression habits to standard option levels, making the mapping relationship between natural language responses and scale option levels more consistent with the user's true expression intentions, thereby improving the accuracy of psychological health assessment results. Attached Figure Description

[0021] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0022] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0023] Figure 1 This is a flowchart illustrating the first embodiment of the mental health assessment method of this application; Figure 2 This is a schematic diagram of the mental health assessment process involved in one embodiment of the mental health assessment method of this application; Figure 3 This is a schematic diagram of the system architecture involved in one embodiment of the mental health assessment method of this application; Figure 4This is a schematic diagram of the modular structure of the mental health assessment device in this application; Figure 5 This is a schematic diagram of the hardware operating environment involved in the mental health assessment method in this application embodiment.

[0024] The purpose, features, and advantages of this application will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0025] It should be understood that the specific embodiments described herein are merely illustrative of the technical solutions of this application and are not intended to limit this application.

[0026] To better understand the technical solution of this application, a detailed description will be provided below in conjunction with the accompanying drawings and specific implementation methods.

[0027] Currently, mental health assessment technology has incorporated AI-assisted assessment solutions. Some systems use knowledge graphs and pre-defined question-and-answer trees to ask users questions and generate assessment results based on user feedback. However, the response and judgment logic of this type of assessment relies heavily on manually constructed knowledge bases and fixed question-and-answer rules. It can only execute standardized interactions according to preset processes, making it difficult to adapt to the complex, ambiguous, or personalized expressions of users in real-world applications. This leads to biases in the identification and judgment of users' psychological states, ultimately resulting in low accuracy of mental health assessment results.

[0028] Therefore, improving the accuracy of mental health assessment results is an urgent problem that needs to be solved.

[0029] The main solution of this application is as follows: Collect language expression samples from users to be assessed; establish a personalized expression anchor table for the users based on the language expression samples, wherein the personalized expression anchor table is used to indicate the language expression anchor points of the users at each option level in a preset rating field; based on a preset psychological assessment table, conduct a scale assessment interaction with the users to be assessed to obtain their natural language responses; calibrate the natural language responses according to the personalized expression anchor table to obtain the target option levels of the users for each item in the psychological assessment table, and generate assessment results based on each target option level.

[0030] This application first collects language expression samples from users to be assessed and establishes a personalized expression anchor table for these users, achieving a precise characterization of their personal expression habits and providing a basis for subsequent identification of their true semantics. During the scale assessment interaction, this application calibrates the natural language responses of the users to be assessed based on the personalized expression anchor table, mapping the natural language responses to the target option levels of each item in the psychological assessment scale, and generating assessment results based on each target option level. Therefore, compared to the standardized matching method based on fixed question-and-answer rules and general semantic templates in traditional technologies, this application can adaptively understand the personalized expression differences of different users on preset scoring fields, calibrating the users' natural language responses with personal expression habits to standard option levels, making the mapping relationship between natural language responses and scale option levels more consistent with the user's true expressive intent, thereby improving the accuracy of psychological health assessment results.

[0031] It should be noted that the executing entity of the method in each embodiment of the mental health assessment method of this application can be a mental health assessment system, or a computing service device with data processing, network communication, and program execution functions, such as a tablet computer, personal computer, or mobile phone, or a mental health assessment device capable of the above functions. This embodiment does not specifically limit this. The following uses a mental health assessment system as the executing entity as an example to describe this embodiment and the following embodiments.

[0032] Based on this, this application proposes a mental health assessment method according to the first embodiment, please refer to... Figure 1 The mental health assessment method includes steps S10 to S30: Step S10: Collect language expression samples of the user to be evaluated, and establish a personalized expression anchor point table for the user to be evaluated based on the language expression samples. The personalized expression anchor point table is used to indicate the language expression anchor point of the user to be evaluated at each option level of the preset scoring field.

[0033] Language expression samples refer to text or speech-to-text samples provided by the user during natural language interaction that reflect their personal expression habits. Specifically, these can include the user's everyday descriptions of their emotional state, frequency of behavior, intensity of feelings, duration, and scope of impact. A personalized expression anchor table is a data structure used to store the user's language expression anchors at each option level of a preset rating field. The preset rating fields include at least frequency, intensity, duration, and scope of impact fields. Each option level represents multiple preset scales of the psychological assessment scale under each field. Language expression anchors characterize the user's personal expression scale at the corresponding field and level; for example, the typical expression corresponding to level 3 ("almost every day") in the frequency field.

[0034] First, language expression samples are collected from the users to be evaluated; then, the collected language expression samples are analyzed to extract the users' expression habits at each option level of each preset scoring field; finally, a personalized expression anchor point table is established based on the analysis results.

[0035] In this embodiment, the step of "collecting language expression samples from the user to be evaluated" in step S10 may include: Step S101: Conduct trust perception interaction with the user to be evaluated to obtain language expression samples of the user to be evaluated; wherein, the language expression samples are used to characterize the user's personal expression scale for preset rating fields, the preset rating fields include frequency field, intensity field, persistence field and scope of influence field.

[0036] Trust-aware interaction refers to a natural language dialogue process between the system and the user being evaluated, aimed at establishing a trust relationship and initially perceiving the user's situation. This interaction process guides the user to express themselves in a low-pressure atmosphere using non-judgmental and empathetic dialogue strategies, which differs from traditional standardized questionnaires. The frequency field is used to characterize the number of times or how often the user's described state occurs; the intensity field is used to characterize the subjective severity or emotional intensity of the user's described state; the duration field is used to characterize the duration or continuity of the related state; and the scope of influence field is used to characterize the degree of impact of the state on functional areas such as learning, work, life, sleep, and social interaction.

[0037] The system first initiates a trust-aware interaction with the user to be assessed, guiding the user to express themselves naturally through a non-judgmental and empathetic dialogue. During the interaction, the system collects language expression samples provided by the user in natural language dialogue that reflect their personal expression scale for preset scoring fields such as frequency, intensity, duration, and scope of influence. These samples serve as the data basis for subsequently building a personalized expression anchor table. This collection process does not generate psychological assessment conclusions; it is only used to determine the user's personal language scale for each preset scoring field.

[0038] In one feasible implementation, the system outputs an empathetic opening and open-ended situational questions to the user through a first agent (i.e., a trust-initiating and context-aware agent), such as "Hello, this is your private psychological support space. You can feel safe to talk about your recent feelings," and "Has anything been particularly stressful lately?" After the user replies in natural language, the first agent extracts expression samples related to frequency, intensity, persistence, and scope of impact from the reply. For example, if the user answers, "I haven't been able to sleep almost every day for the past week, and it's been very difficult," the system extracts "almost every day" (frequency field expression sample) and "very difficult" (intensity field expression sample) as the user's language expression samples. The first agent, responsible for the initial interaction, can utilize the empathic capabilities of a large language model to create a safe, non-judgmental dialogue atmosphere, guiding the user to accept and be willing to begin communicating. Through open-ended questions, without intrusive evaluation, the system perceives the user's current emotional state, environment, and main concerns, while simultaneously collecting basic user information to prepare for the next stage of assessment.

[0039] Step S20: Based on a preset psychological assessment form, conduct a scale assessment interaction with the user to be assessed to obtain the user's natural language response.

[0040] The pre-set psychological assessment scale is a standard questionnaire tool validated in the field of psychology, containing multiple items, corresponding candidate option levels for each item, and their scoring rules. Scale assessment interaction refers to the interactive process where the system presents scale items to the user in a conversational manner, and the user responds freely in natural language. Natural language responses refer to the free text content expressed in everyday conversational style that the user inputs during the scale assessment interaction.

[0041] The system acquires a pre-set psychological assessment scale, presents each item in the scale to the user in the form of a natural language dialogue, and acquires the user's answers in natural language in each round of interaction until all items have been collected.

[0042] In this embodiment, step S20 may include: Step S201: Convert the original questions in the preset psychological assessment scale into conversational questions; Step S202: Output each of the conversational questions to the user to be evaluated in sequence, and obtain the responses input by the user in natural language form as the natural language answers.

[0043] The pre-set psychological assessment scale is a standard questionnaire tool validated in the field of psychology. It includes multiple original questions, corresponding candidate option levels for each original question, and scoring rules. The original questions are usually presented in a written, professional style, such as the original question in PHQ-9: "Have you felt listless or lacking interest in doing anything in the past two weeks?". Conversational questions transform the written, professional expressions of the original questions into conversational, natural, and interactive question forms suitable for dialogue scenarios. For example, the original question above could be transformed into "Have you felt listless and lacking interest in doing anything in the past two weeks?". Conversational questions retain the core semantics and scoring orientation of the original questions but use expressions closer to everyday communication habits. Natural language responses refer to the text content that users freely input in a conversational style after receiving the conversational questions. This response is not limited by the pre-set options; users can express themselves using their own language habits.

[0044] The system first reads a preset psychological assessment scale and obtains the original questions in the scale; then it converts each original question into a conversational question while retaining the core semantics and scoring direction of the original question; then it outputs each conversational question to the user in turn and obtains the user's response in natural language in each round of interaction; finally, all the obtained natural language responses are used as the natural language answers for subsequent calibration.

[0045] In one feasible implementation, taking the PHQ-9 depression scale as an example, the second agent (i.e., the dynamic assessment agent) reads the original question in the scale, "Have you felt listless or uninterested in doing things in the past two weeks?", and converts the original question into a conversational question using a preset conversion template, "I would like to ask, in the past two weeks, have you felt listless when doing things, or have you lost interest in things you used to enjoy?" After receiving the question, the user enters "Yes, I've felt like this for several days" in the dialog box. The second agent saves this response as the natural language response for that item, and then continues to process the next original question in the scale in the same way.

[0046] Furthermore, before converting the original questions of the standard psychological assessment scale into conversational questions, the system first performs structured processing on each scale item. Each standard scale item is pre-constructed as an item object, which includes at least the item number, the dimension it belongs to, the item semantics, the time range field, the frequency field, the intensity field, the persistence field, the scope of influence field, the candidate option level set, and the scoring rules. The time range field defines the time period covered by the user's response, such as "the most recent week" or "the past month"; the frequency field describes the number of times the user's described state occurs; the intensity field describes the subjective severity or emotional intensity of the user's described state; the persistence field describes the duration or continuity of the relevant state; the scope of influence field describes the degree of impact of the state on functional areas such as learning, work, life, sleep, and social interaction; the scoring rules include positive scoring, negative scoring, missing value handling, dimension summarization rules, and report generation rules. This item object serves as the data foundation for subsequent user response analysis, response anchor comparison, candidate score generation, and stability gating scoring.

[0047] In one implementation, the aforementioned item objects are obtained through a combination of "scale rule parsing, language model-assisted extraction, and expert confirmation." The system first reads the original item text, option levels, scoring rules, and dimension attribution from the standard psychological scale text. Then, the item structuring module automatically extracts scoring fields such as time range, frequency, intensity, duration, and scope of influence from the item stem. For items with ambiguity or missing explicit fields, the system uses a large language model to generate candidate field annotation results, which are then confirmed by a preset rule or expert review module. Only fields that pass rule verification or expert confirmation are written into the item object for that item.

[0048] Step S30: The natural language response is calibrated according to the personalized expression anchor table to obtain the target option level of the user to be evaluated for each item in the psychological assessment table, and the assessment result is generated based on each target option level.

[0049] The target option level refers to the standard option level determined by the system for each item after calibration. This level is used for subsequent scoring according to the scoring rules of the psychological assessment scale. The assessment result refers to the quantitative evaluation report generated by the system based on the target option level of each item, according to the scoring rules of the psychological assessment scale.

[0050] For each item in the psychological assessment scale, the corresponding natural language response and the level of each candidate option for that item are obtained. Based on the user's personalized expression anchor table, the user's natural language response is matched with the level of each candidate option to determine the target option level for that item. The above calibration process is performed for each item. After obtaining the target option level for each item, the scores are calculated and summarized according to the scoring rules of the psychological assessment scale to generate the final assessment result.

[0051] For example, such as Figure 2 The diagram illustrates the mental health assessment process. First, a trust-building interaction is established with the user to obtain a sample of their verbal expression. A personalized expression anchor point table is then created based on this sample. Next, a scale assessment interaction is conducted with the user based on a pre-set psychological assessment form, yielding their natural language responses. These responses are then calibrated according to the personalized expression anchor point table to obtain the user's true choices (i.e., target option levels). Finally, assessment results are generated based on the target option levels for each question.

[0052] Thus, this embodiment of the application, by collecting language expression samples from users to be assessed and establishing a personalized expression anchor table for those users, achieves a precise characterization of users' personal expression habits, providing a basis for subsequent identification of users' true semantics. During the scale assessment interaction, this embodiment of the application calibrates the natural language responses of the users to be assessed according to the personalized expression anchor table, mapping the natural language responses to the target option levels of each item in the psychological assessment scale, and generating assessment results based on each target option level. Therefore, compared with the standardized matching method based on fixed question-and-answer rules and general semantic templates in traditional technologies, this embodiment of the application can adaptively understand the differences in personalized expressions of different users to preset scoring fields, calibrate the natural language responses with personal expression habits of users to standard option levels, and make the mapping relationship between natural language responses and scale option levels more consistent with the user's true expression intention, thereby improving the accuracy of psychological health assessment results.

[0053] Based on the first embodiment of this application, in the second embodiment of this application, the content that is the same as or similar to that in the first embodiment described above can be referred to the above description and will not be repeated hereafter. Based on this, the step S10 of "establishing a personalized expression anchor point table for the user to be evaluated based on the language expression sample" may include: Step A10: Semantically encode the language expression sample to obtain the corresponding expression vector.

[0054] Semantic encoding refers to the process of converting user-provided language expressions in natural language into numerical vector representations that can be processed by computers. Through semantic encoding, semantically similar expressions have similar distribution positions in the vector space. For example, "can't sleep every day" and "can't sleep well all night" should have high vector similarity after encoding. The expression vector is a numerical vector obtained after semantic encoding, used to represent the semantic information of the language expression sample.

[0055] Step A20: Group and aggregate the expression vectors according to the option levels of each preset rating field to obtain multiple vector groups.

[0056] Vector grouping refers to the process of classifying and grouping expression vectors according to preset scoring fields and their respective option levels. For example, all expression vectors belonging to the frequency field and corresponding to level 3 are grouped into the same group.

[0057] It should be noted that, in one feasible implementation, grouping and aggregation are based on the rating field labels and option level labels already associated with each language expression sample during collection. Specifically, when collecting language expression samples, each collection question is pre-associated with a corresponding rating field and option level. The language expression sample provided by the user for that collection question automatically inherits these rating field and option level as its labels. During grouping and aggregation, the system reads the rating field labels and option level labels already carried by each language expression sample, and groups expression vectors with the same rating field labels and the same option level labels into the same vector group.

[0058] The rating field labels and option level labels of the language expression samples are determined by at least one of the following methods: The first type involves pre-set calibration questions. During the design and configuration phase, scale experts or designers define the option levels of the psychological assessment scales from which the questions originate. Each calibration question is assigned a corresponding scoring field and option level. User responses to these calibration questions automatically acquire the label associated with that question. For example, if the system's pre-set calibration question is "How many days in the past week have you had trouble sleeping?", this question is configured as level 3 of the corresponding frequency field during the design phase. Any natural language response from the user to this question is labeled as "Frequency Field - Level 3" during data collection, without requiring the system to infer the level from the response content.

[0059] The second category consists of user answer samples that are confirmed as "official scores" during the formal evaluation process through stability gating. The system will then incorporate these answer samples, along with their corresponding scoring fields and option levels, into the anchor table as the user's expression sample.

[0060] The third category consists of historical answer samples whose scoring fields and option levels have been reviewed and confirmed by experts. After obtaining the appropriate authorization and completing the anonymization process, the system will incorporate these historical answer samples and the fields and level labels confirmed by experts into the anchor table.

[0061] In this way, when performing grouping and aggregation, the system directly classifies each expression sample based on the rating field labels and option level labels already carried by it, without relying on real-time semantic inference of the answer content to determine its level, thus ensuring the certainty and consistency of grouping.

[0062] Step A30: Based on each of the vector groups, determine the personalized expression anchor points of the user to be evaluated at each option level of each of the preset rating fields, and construct the personalized expression anchor point table based on each of the personalized expression anchor points.

[0063] Personalized expression anchors refer to a comprehensive vector representation determined by the system based on multiple expression vectors within the same group. This comprehensive vector characterizes the user's personal expression scale at a specific option level in a specific preset rating field. For example, a user's language expression anchor at level 3 in the frequency field can be determined by the average of expression vectors such as "can't sleep every day," "this happens seven days a week," and "almost every day has insomnia" within that group. This anchor represents the user's personal expression habit at the "almost every day" level. The personalized expression anchor table is a data structure composed of the aforementioned preset rating fields, option levels, and their corresponding personalized expression anchors.

[0064] The system first performs semantic encoding on the collected language expression samples, converting each sample into a corresponding expression vector; then, it groups and aggregates the expression vectors according to the option levels of each preset rating field, resulting in multiple vector groups, each vector group corresponding to an option level of a preset rating field; then, based on each vector group, it determines the user's personalized expression anchor points at each option level of each preset rating field; finally, it constructs a personalized expression anchor point table based on each personalized expression anchor point.

[0065] In one feasible implementation, for user u, a personalized expression anchor table is established, represented as follows:

[0066] Where M represents the set of scale rating fields, including fields such as frequency, intensity, duration, and scope of influence (i.e., preset rating fields); k represents the option level; Indicates the highest option level for field m; This represents the language expression anchor (i.e., the personalized expression anchor) for user u at the corresponding option level k in field m. This language expression anchor can consist of the user's historical response text, calibration interaction text, or expert-verified samples, and can be represented as a set of text fragments or a vector representation.

[0067] The system encodes user calibration responses into a vector e(x) and groups them according to user, rating field, and option level. Let... This represents the set of language expression samples for user u under the rating field m and option level k. (Language expression sample set) The sample can consist of at least one of the following: The first type consists of user response samples with clearly defined scoring fields and option level labels in the system's preset calibration questions. The scoring fields and option level labels for this type of sample originate from the manual configuration of the calibration questions during the design phase. That is, each calibration question is pre-specified by scale experts or designers based on the option level definition of its corresponding psychological assessment scale. User responses to these calibration questions automatically inherit the labels of that calibration question, eliminating the need for the system to infer levels from the responses through semantic analysis after collection. The second type consists of user response samples that have been confirmed as "official scores" through stability gating during the formal assessment process. The third type consists of historical response samples whose scoring fields and option levels have been verified by experts. For the second and third types of samples, the system only allows their inclusion in the user response anchor table after obtaining the appropriate authorization and completing the anonymization process. The response anchor vector for user u on field m and level k is shown below. It can be represented as:

[0068] Where e(x) represents the vector obtained after semantically encoding the language expression sample x. Represents the sample set The formula represents the number of samples in the dataset. It describes the vector average of multiple expression samples from the same user, within the same rating field, and at the same option level, to obtain the user's personalized expression anchor point for that field and level. The system performs the above calculation on each option level k of each preset rating field m to obtain the... , with each Construct a personalized expression anchor table.

[0069] In this embodiment, step A30 may include: Step A301: For any first option level of each of the preset rating fields, determine the number of vectors in the first vector group corresponding to the first option level; Step A302: If the number of vectors is lower than a first preset threshold, determine the second option level that is adjacent to the first option level among the option levels of the preset rating field; Step A303: Based on the second vector grouping corresponding to the second option level and the first vector grouping, determine the personalized expression anchor point of the first option level.

[0070] The first option level refers to any option level for which anchor point determination is currently needed. For example, in the frequency field, level 1, which the system is currently processing, is the first option level. The first vector group refers to the vector group corresponding to the first option level, i.e., the set of all expression vectors under that level. The first preset threshold is a threshold used to determine whether the number of samples in the vector group is sufficient. When the number of vectors is lower than this threshold, it indicates that the samples at that level are insufficient to independently form a reliable anchor point, and estimation needs to be performed using information from adjacent levels. The second option level refers to the option level adjacent to the first option level in the same preset scoring field. For example, when the first option level is level 1, the second option level is level 0 or level 2.

[0071] For any first option level in each preset rating field, the system first obtains the first vector group corresponding to that first option level and counts the number of vectors within it. Then, it compares the counted vector count with a first preset threshold to determine if the sample size in the first vector group is sufficient. If the number of vectors is lower than the first preset threshold, it determines the second option level adjacent to the first option level in the preset rating field. Finally, based on the second vector group and the first vector group corresponding to the second option level, it determines the personalized expression anchor point for the first option level. When the number of vectors in the first vector group is not lower than the first preset threshold, the system does not perform smoothing estimation but directly determines the personalized expression anchor point for that option level based on the aggregation result of each expression vector in the first vector group, for example, by directly calculating the average value of each expression vector in the group as the anchor point for that level. The system performs the above judgment and processing for each option level of each preset rating field until all levels have obtained corresponding personalized expression anchor points, ultimately constructing a complete personalized expression anchor point table.

[0072] In one feasible implementation, when the number of samples at a certain level k is lower than a preset threshold, the system performs a smooth estimation based on the response anchor vectors of levels adjacent to that level, as shown in the following formula:

[0073] in, The smoothed response anchor vector (i.e., the smoothed expression vector) is represented by λ, which is the smoothing weight, ranging from 0 to 1. If the current level k is the lowest or highest level, only its neighboring levels are used for smoothing estimation. Furthermore, when there are no calibrated expression samples at the current level k, the system records the initial anchor vector of the current level as null and sets the smoothing weight λ=0, estimating only based on the response anchor vectors of neighboring levels. When there are no available response anchor vectors at neighboring levels, the system uses a general response anchor preset by experts or obtained from group sample statistics as a temporary anchor and marks this temporary anchor as a low-confidence anchor. For candidate scoring results generated using low-confidence anchors, the system increases the priority of clarification or review in subsequent stability gating to avoid directly incorporating them into the formal evaluation results.

[0074] Thus, the embodiments of this application can accurately characterize the user's personal expression scale in a vectorized form at each option level of preset rating fields such as frequency, intensity, duration and scope of influence. This allows the user's natural language response to be compared with the user's own expression anchor point during the subsequent calibration process, avoiding misjudgment of expression scale caused by the use of general semantic templates to unify the processing of different users in traditional technologies.

[0075] Based on the first and second embodiments of this application, in the third embodiment of this application, the content that is the same as or similar to the first and second embodiments described above can be referred to the above description and will not be repeated hereafter. Based on this, step S30 may include: Step S301: For any item in the psychological assessment form, extract the first semantic vector of the natural language answer corresponding to the item, and the second semantic vector of the level of each candidate option of the item.

[0076] The first semantic vector is a vector representation obtained by semantically encoding the user's natural language answer to a question item as a whole, used to represent the overall semantic information of the answer. The second semantic vector is a vector representation obtained by encoding the standard semantic descriptions of the levels of each candidate option for that question item, used to represent the semantic information of each candidate option level.

[0077] Step S302: Extract the third semantic vector related to the preset scoring field from the natural language answer corresponding to the question item, and obtain the anchor vector corresponding to the candidate option level of the question item in the personalized expression anchor table.

[0078] The third semantic vector refers to the vector representation of semantic fragments extracted from the user's natural language response that are related to the preset rating fields (including frequency, intensity, duration, and scope of influence). For example, from the response "I haven't been able to sleep almost every day for the past week, and I feel terrible," the semantic vectors related to frequency ("almost every day") and intensity ("feel terrible") can be extracted. Anchor vectors refer to the vector representations of individualized language expression anchors in the personalized expression anchor table that correspond to the option levels of the preset rating fields. For example, the vector representation of the user's language expression anchor at level 3 in the frequency field.

[0079] Step S303: Determine the first matching degree between the first semantic vector and the second semantic vector, and determine the second matching degree between the third semantic vector and the anchor vector.

[0080] The first matching degree refers to the similarity between the first and second semantic vectors, used to measure the degree of matching between the overall semantics of a user's answer and the standard semantics of a certain candidate option level. The second matching degree refers to the similarity between the third semantic vector and the anchor vector, used to measure the degree of matching between the user's expression habits in the preset scoring field and the user's personal expression anchors in that field and at that level.

[0081] Step S304: Determine the candidate score for each of the candidate option levels based on the first matching degree and the second matching degree.

[0082] Candidate score refers to the comprehensive score obtained by combining the first and second matching degrees, used to measure whether a user's answer belongs to a certain level of candidate options.

[0083] Step S305: Based on each candidate score, determine the target option level for the question item.

[0084] For any item in the psychological assessment scale, the system first performs overall semantic encoding on the user's natural language response to obtain a first semantic vector, and then encodes the standard semantics of each candidate option level for that item to obtain second semantic vectors. Simultaneously, it extracts semantic fragments related to preset scoring fields from the user's natural language response and encodes them to obtain a third semantic vector. It also retrieves anchor vectors corresponding to each candidate option level for that item from a personalized expression anchor table. Next, it calculates the first matching degree for each candidate option level based on the first and second semantic vectors, and the second matching degree based on the third semantic vector and anchor vectors. Then, for each candidate option level, it determines the candidate score for that level based on the first and second matching degrees. Finally, based on the candidate scores for each candidate option level, it determines the target option level for that item from among the candidate option levels.

[0085] Thus, this embodiment of the application extracts the overall semantic vector of the user's natural language answer and the local semantic vector related to the preset scoring field, and matches them with the standard semantic vector of each candidate option level and the anchor vector in the personalized expression anchor table, respectively. This achieves a comprehensive evaluation of the user's answer from two dimensions: "global semantics" and "personal expression scale." Global semantic matching ensures the basic semantic correspondence between the answer and the option level, while personal expression scale matching effectively corrects scoring biases caused by differences in personal expression habits by comparing expressions related to frequency, intensity, persistence, and scope of influence in the answer with the user's own language expression anchors. This ensures that subsequent candidate scores reflect both the overall meaning of the answer and the user's personal expression scale.

[0086] In this embodiment, step S304 may include: Step B10: For any candidate option level of the question item, determine the coverage confidence of the natural language answer corresponding to the question item to the preset scoring field, and determine the penalty confidence of the natural language answer corresponding to the question item, wherein the penalty confidence is evaluated from the dimensions of cross-round semantic contradiction and expression ambiguity. Step B20: Determine the candidate score for the candidate option level based on the first matching degree, the second matching degree, the coverage confidence degree, and the penalty confidence degree corresponding to the candidate option level.

[0087] Coverage confidence refers to whether a user's natural language response contains the necessary information for each of the pre-defined scoring fields required by the question, and the measure of the clarity of this information. It is used to assess the sufficiency of the response's coverage of the required scoring fields. For example, if a question requires information on three fields—time range, frequency, and intensity—and the response only contains frequency and intensity information but lacks time range information, then the coverage confidence is low.

[0088] Penalized confidence is a metric used to measure the reliability of a user's natural language response. It is evaluated from the dimensions of vague expression and semantic inconsistency across rounds. Vague expression refers to the use of uncertain terms such as "maybe," "more or less," "okay," or "I can't say for sure" in the response. Semantic inconsistency across rounds refers to a significant discrepancy between the current response and the user's expression in previous rounds, such as saying "I've been sleeping very well lately" in the previous round but saying "I have insomnia almost every day" in the current round. The evaluation results from all dimensions are combined to obtain the penalized confidence.

[0089] For any candidate option level currently being processed, the system first determines the coverage confidence of the user's natural language answer to each preset scoring field based on the set of preset scoring fields required to be covered by the question item. Simultaneously, it comprehensively evaluates the answer from the dimensions of cross-round semantic contradictions and expression ambiguity to obtain a penalty confidence score. Then, using the first and second matching degrees corresponding to that candidate option level as a positive evaluation basis, and combining the coverage confidence score and penalty confidence score, it performs a comprehensive calculation to determine the candidate score for that candidate option level. The system executes the above process for each candidate option level separately.

[0090] In one feasible implementation, for user u regarding the question item For each answer r, the system calculates the matching relationship between the answer and the question's semantics, option level, and user response anchor (i.e., personalized expression anchor), and generates a candidate scoring function. To support multiple rating fields participating in the rating, let's set... For the title The set of rating fields involved, For field m in the question item The weights in, and The formula for calculating the candidate scoring function can be expressed as follows: .

[0091] in, This represents the overall vector representation of the user's answer r. Indicates the terms The semantic vector formed together with its kth standard option level, This represents a semantic vector extracted from the answer r that is related to the rating field m. sim represents the smoothed individualized response anchor vector of user u on rating field m, options k, etc., where sim represents the similarity function and α, β, γ, and δ are weight parameters.

[0092] This indicates that the answer r corresponds to the question item. The overall coverage (i.e., coverage confidence) of the necessary fields can be expressed as:

[0093] in, This represents the confidence level of the response r in covering field m, with a value ranging from 0 to 1. =1 indicates that the information in field m is explicitly present. =0 indicates that information in field m is missing. Coverage confidence. Penalized confidence scores are used to correct candidate scores based on information completeness and expression reliability. They are independent of each other.

[0094] The penalty term representing the semantic contradiction or uncertain expression across rounds in the response r can be represented as:

[0095] in, This indicates a penalty for semantic contradictions across rounds. Indicating uncertainty and punishment and The weight of the penalty item.

[0096] Thus, this embodiment of the application improves the credibility of candidate scores by introducing coverage confidence and penalty confidence, thereby enhancing the credibility of the scores from the perspectives of information completeness and expression reliability. Coverage confidence evaluates the proportion of explicitly stated scoring fields in a user's answer to all required scoring fields for the question. Higher coverage confidence reflects the completeness of the answer's information. If an answer lacks one or more required scoring fields, the system lowers the candidate score accordingly to suppress unreliable output due to insufficient information. Penalty confidence evaluates the reliability of the user's answer in terms of expression, applying negative evaluation from two dimensions: first, whether the answer contains evasive or uncertain expressions such as "I don't know," "I can't say," or "maybe"; second, whether the user's current answer is inconsistent or contradictory to information expressed in previous dialogue rounds. The penalty confidence, derived from these two dimensions, applies a corresponding penalty to answers with lower expression quality. Working in conjunction with coverage confidence, it corrects candidate scores from the perspectives of information sufficiency and expression reliability.

[0097] Through the above mechanism, the system can avoid giving seemingly clear but actually unfounded rating results when the user's answer is insufficient or has low reliability.

[0098] In this embodiment, step S305 may include: Step B30: Normalize each candidate score to obtain the candidate probability of each candidate option level.

[0099] Normalization refers to the process of converting the candidate ratings of each candidate option level into probability values, so that the sum of the probabilities of all candidate option levels is 1, thus facilitating relative comparisons between levels. Candidate probability refers to the probability value corresponding to each candidate option level after normalization, representing the likelihood that a user's answer belongs to that level.

[0100] For example, after obtaining the scores of each candidate, the probability that the answer r belongs to the level of the k-th option is obtained by Softmax normalization, which is expressed as:

[0101] in, This indicates that the answer r was selected as a question item. The probability of the k-th option's level, where K represents the highest option level for that question. The system then further calculates the probability based on each... The distribution of candidate options is used to calculate the separation degree and determine whether candidate scores should be included in the formal evaluation results.

[0102] Step B40: Determine the first candidate probability with the largest probability and the second candidate probability with the largest probability among all the candidate probabilities, and determine the difference between the first candidate probability and the second candidate probability.

[0103] The first candidate probability is the maximum among all candidate probabilities, representing the most likely option level. The second candidate probability is the second largest among all candidate probabilities, representing the second most likely option level.

[0104] For example, the difference between the first candidate probability and the second candidate probability can be expressed as: ,in, This represents the separation degree of candidate options, which is the difference between the probability of the largest candidate and the probability of the second largest candidate. This indicates the level of the candidate option with the highest probability. This indicates the ranking of the candidate option with the second highest probability.

[0105] Step B50: If the difference is greater than or equal to the second preset threshold, the candidate option level corresponding to the first candidate probability is taken as the target option level of the question item. Step B60: If the difference is less than the second preset threshold, at least one counterfactual answer is generated based on the natural language answer corresponding to the question item to form a counterfactual set, and the target option level of the question item is determined from each of the candidate option levels based on the counterfactual set.

[0106] The second preset threshold is a preset threshold value used to determine whether the difference between the first candidate probability and the second candidate probability is significant enough. When the difference is greater than or equal to the threshold, the level corresponding to the first candidate probability is considered to have a significant advantage and can be directly determined as the unique best matching level. When the difference is less than the threshold, it indicates that the probabilities between the two levels are close, and the system has difficulty in directly determining the optimal level. A counterfactual verification mechanism needs to be introduced to assist in decision-making.

[0107] The system first normalizes each candidate score to obtain the candidate probability corresponding to each candidate option level. Then, it determines the maximum value among the candidate probabilities as the first candidate probability and the second largest value as the second candidate probability, and calculates the difference between the two. Next, it compares the difference with a second preset threshold. When the difference is greater than or equal to the second preset threshold, the candidate option level corresponding to the first candidate probability is directly determined as the target option level of the question. When the difference is less than the second preset threshold, at least one counterfactual answer is generated based on the natural language answer corresponding to the question, forming a counterfactual set. Based on the counterfactual set, the target option level of the question is re-determined from each candidate option level.

[0108] Thus, this embodiment of the application obtains the probability distribution of each candidate option level by normalizing the candidate scores, and makes a judgment based on the difference between the highest probability and the second highest probability. When the difference is significant, the target option level is directly determined, achieving efficient and reliable scoring decision-making. When the probability difference is large enough, it indicates that the system has a high degree of confidence in a certain level, and directly determining that level as the target option level ensures decision-making efficiency. When the probability difference is insufficient, it indicates that the system has difficulty distinguishing between two or more levels. In this case, a counterfactual set mechanism is introduced for auxiliary verification. This hierarchical decision-making mechanism ensures the accuracy of scoring while only initiating the computationally expensive counterfactual verification process when necessary, achieving a balance between accuracy and processing efficiency.

[0109] In this embodiment, step B60 may include: Step B601: Determine the counterfactual candidate probability for each counterfactual answer in the counterfactual set.

[0110] Counterfactual responses refer to one or more alternative response versions generated by the system based on the user's original natural language response. These versions do not change the core facts but alter the expression. Changes in expression include paraphrasing, tone adjustment, background information adjustment, or order of expression. A counterfactual set is a collection of one or more of these counterfactual response versions. The counterfactual candidate probability refers to the probability distribution of each counterfactual response in the counterfactual set belonging to each candidate option level, calculated using the same calibration process as the original response (extracting semantic vectors, determining matching degree, determining candidate scores, and normalization).

[0111] For each counterfactual answer in the counterfactual set, the system extracts the semantic vector of the counterfactual answer and determines the probability distribution of the counterfactual answer at each candidate option level according to the same candidate score calculation and normalization process as the original answer, which is used as the counterfactual candidate probability of the counterfactual answer.

[0112] Step B602: For each counterfactual response, determine the third candidate probability, which is the largest among all counterfactual candidate probabilities corresponding to the counterfactual response, and obtain the counterfactual option level corresponding to the third candidate probability.

[0113] The third candidate probability refers to the highest probability value among the probability distributions of candidate option levels corresponding to a counterfactual response. It represents the likelihood of the counterfactual response being the most likely option level after calibration. The counterfactual option level refers to the candidate option level corresponding to the third candidate probability, that is, the temporary target level determined after calibration of the counterfactual response. This level is not necessarily the same as the target option level of the original response.

[0114] For each counterfactual response, the system determines the maximum value from the counterfactual candidate probabilities corresponding to that response as the third candidate probability, and determines the candidate option level corresponding to the third candidate probability as the counterfactual option level of that response.

[0115] Step B603: The option level that appears most frequently among the various counterfactual option levels is taken as the target option level of the question item.

[0116] Frequency of occurrence refers to the number of times or proportion of each counterfactual option level appear in all counterfactual responses. The option level with the highest frequency of occurrence is the option level that is commonly referred to by the largest number of counterfactual responses. This level is used as the target option level finally determined by the system, representing the most stable matching result after verification from multiple perspectives.

[0117] The system counts the frequency of occurrence of each counterfactual option level in the counterfactual set and determines the option level with the highest frequency; then, the option level with the highest frequency is taken as the final target option level for that question.

[0118] In one feasible implementation, the system includes a counterfactual consistency verification agent, which generates multiple counterfactual response versions that do not change the core facts based on the user's current response r, forming a counterfactual set: ,in, This is the original answer. to These are answer versions that have undergone semantic rewriting, weakened tone, obscured background information, adjusted expression order, or removed non-key information.

[0119] To prevent counterfactual responses from altering the user's original intent, the system first performs a core factual consistency check on each version of the counterfactual response: ),in, This represents the core fact vector extracted from the original response r. The core facts include information such as the time range, frequency, intensity, duration, and scope of influence related to the scale score. This represents the original answer r and the counterfactual answer. The core factual consistency score between them. When If the counterfactual response version is less than the core fact consistency threshold, the system discards the counterfactual response version and does not use it for counterfactual consistency calculation.

[0120] For each one that passes the core fact consistency check The system calculates the probability distribution of candidate scores. And obtain the corresponding candidate option levels: ,in, This represents the ranking of the j-th counterfactual answer version among the candidate options for the i-th item. This indicates that the level with the highest probability is selected from all candidate option levels k, which is the counterfactual option level mentioned above.

[0121] Furthermore, the system calculates a counterfactual consistency index:

[0122] in, This indicates the number of counterfactual response versions that passed the core fact consistency check. This indicates the proportion of consistency between the original answer and each valid counterfactual answer version in the candidate option level, and 1[·] represents the indicator function. The higher the score, the less susceptible the candidate's rating is to slight changes in wording, tone, or background information. =0, then the system does not calculate. The system will then directly transfer the question to the calibration clarification or expert review process. The system can also calculate the scoring volatility.

[0123] in, This represents the variance of the candidate rank set. If... Below the consistency threshold or Above the fluctuation threshold If the score is not found, it means that the current candidate score is unstable and cannot be directly written into the official evaluation results.

[0124] The system is based on the separation degree of candidate options Counterfactual consistency index Rating volatility Necessary field coverage and risk status Together, they decide the next action to take on the i-th question item. .in, This represents the current risk state score output by the security boundary agent, ranging from 0 to 1, with higher values ​​indicating higher risk. This indicates the risk threshold.

[0125] The system determines G_i according to the following priority: First, when the candidate score is unstable and At that time, the system will The assessment will be designated as "expert review or risk triage," and further follow-up questions during the regular assessment will be prohibited. Instability in candidate scores refers to... or Second, when candidate scores are unstable and At that time, the system will The item is designated as "deferred scoring," and its score is marked as unstable in the assessment record. Third, when a candidate score is stable, but... or ,and At that time, the system will The question is identified as "calibration clarification," and a minimum clarification question is generated based on the missing field or candidate level difference. Fourth, when... , , , and At that time, the system will This is designated as the "official score," and allows the current candidate option's level to be written into the official assessment result. Among these, , , and These are thresholds for candidate separation, consistency, volatility, and field coverage, respectively. Through this gating function, the system only writes the score into the formal evaluation result if the candidate score is stable, the field coverage is sufficient, and the risk status allows; otherwise, it generates a clarification question, a deferred scoring flag, or an expert review flag.

[0126] When the stability gating function outputs "calibration clarification," the system generates a minimal clarification question based on the reasons for the scoring instability. For example, if the time range field is missing, only the time range is clarified; if the frequency field is missing, only the frequency is clarified; if the candidate options lack separation, clarification is provided around the key differences between adjacent candidate levels; if the current expression conflicts with the user's existing response anchor, the strength of the current expression relative to the existing anchor is inquired. This clarification question is not intended to induce the user to choose a specific answer, but rather to supplement the minimum necessary information for scoring.

[0127] Thus, this embodiment generates multiple counterfactual answers that do not change the core facts, and after calibrating each counterfactual answer, it counts the frequency of the corresponding counterfactual option level, using the level with the highest frequency as the target option level, thereby achieving cross-validation of the original scoring results. When the probability of the original answer is close among two or more candidate levels, the counterfactual verification mechanism examines the stability of each level under different expression conditions through multi-angle expression transformations. If a certain level is selected in most counterfactual answers, the reliability of that level is further confirmed; if the level distribution of each counterfactual answer is relatively scattered, it indicates that the scoring of the item itself is unstable, and the system can mark the item as a low-confidence score in the evaluation record. Through the above mechanism, the system can reduce the randomness of scoring caused by expression factors such as wording or tone, and improve the stability and reliability of the target option level determination.

[0128] Furthermore, the system also includes a deep dialogue and guidance agent. After completing the scale assessment interaction and calibration scoring, the system activates the deep dialogue and guidance agent. This agent extracts keywords from the target questions with abnormal scores on the psychological assessment scale and the user's historical dialogue during the aforementioned interaction phase (the phase of interaction using the first and second agents). A knowledge retrieval agent then searches a preset professional psychology knowledge base based on these keywords to obtain search results. The deep dialogue and guidance agent also acquires the user's language features and contextual information from historical dialogues. Combining the search results, language features, and contextual information, it generates new conversational questions for deep guidance interaction with the user. The content of the deep guidance interaction is constrained by preset safety boundaries and professional knowledge retrieval results. The safety boundary agent identifies high-risk cues in the user's expression and outputs a risk status score based on the risk status. When the risk status score reaches a gating threshold, the system prohibits further ordinary assessment follow-up questions and outputs a suspended scoring mark, safety boundary language, manual review, or risk diversion action. The candidate output review module reviews the outputs before the system sends a response to check for issues such as diagnostic overreach, false reassurance, over-certainty, leading questions, privacy breaches, or inconsistencies with professional evidence.

[0129] The system also includes a knowledge retrieval agent, which plays a crucial role in ensuring the professionalism, security, and reliability of all interactive content. This agent aims to address the technical issues of the "illusion" inherent in large language models and the limitations of knowledge. During the aforementioned interaction phase, when other agents need to generate professional responses, the knowledge retrieval agent is activated, retrieving relevant knowledge and injecting it into the context to guide the system in generating professional, accurate, and ethically compliant responses. For the assessment phase, the knowledge retrieval agent can also provide evidentiary constraints on the meaning, applicability boundaries, taboo conditions, and report expression of scale items, but it does not directly replace the response anchor calibration and counterfactual consistency verification processes.

[0130] The system also includes a data analysis agent, which is responsible for summarizing and transforming data from all preceding interaction stages and generating a verifiable, calibrated assessment report. Its processing includes not only the user's original dialogue, scale mapping results, and risk indicators, but also individualized response anchors, candidate score probability distributions, candidate option separation, counterfactual consistency indicators, score volatility, stability gating results, clarification records, reasons for deferred scoring, and expert review markers. Specifically, the data analysis agent integrates the trust initiation and calibration samples from stage one, the natural language assessment and stability gating results from stage two, and the in-depth dialogue and guidance data from stage three to automatically generate an instant psychological assessment report. This report not only provides scale dimension scores and risk warnings but also indicates whether each formal score has passed individualized response anchor calibration, counterfactual consistency verification, and stability gating. For items that fail stability gating, the report displays a deferred scoring or expert review marker to prevent low-stability scores from being disguised as definitive conclusions. As the foundation for report generation, the intelligent agent automatically processes and analyzes dialogue data across all stages, extracts key psychological features, and combines scale scoring results, response anchor tables, and stability verification results to generate structured personal assessment records. These records are used for subsequent review, trend observation, and auxiliary analysis in authorized scenarios. Furthermore, in multiple assessment scenarios, the system can also display indicators such as changes in user response anchors, changes in the stability of scores for different items, the proportion of items withheld from scoring, and the proportion of expert review and corrections. These indicators help authorized professionals determine whether user expression habits have changed significantly and whether the system's scoring remains consistently stable.

[0131] The system also includes a privacy and access control module, independent of the agent framework. This module manages all data, employing multi-layered authentication, encrypted data storage, and fine-grained hierarchical access mechanisms to ensure that users' original dialogue data, response anchor data, counterfactual verification data, and analysis reports are only visible to specific roles with explicit authorization. This module utilizes encrypted data storage, encrypted transmission, anonymization, access auditing, authorization revocation, minimum necessary sharing, and training data isolation mechanisms. Samples used for controlled optimization preferentially employ anonymized or de-identified data; unauthorized original dialogues, response anchor tables, and counterfactual response versions must not be included in the model training or policy optimization process.

[0132] For example, such as Figure 3The diagram shows the system architecture. The mental health assessment system includes a multi-agent collaboration module and a privacy and access control module. The multi-agent collaboration module includes a trust initiation and context-aware agent (i.e., the first agent), a dynamic assessment agent (i.e., the second agent), a deep dialogue and guidance agent (i.e., the third agent), a knowledge retrieval agent (i.e., the fourth agent), and a data analysis agent (i.e., the fifth agent). The system comprises the following components: First agent acquires basic user information to generate prompts, creates a safe environment for open-ended dialogue, and outputs the first assessment result (main concerns, emotional state); Second agent converts the original questions from the psychological scale into conversational questions, extracts answers from the user's natural language responses and scores them, outputting the second assessment result; Third agent determines the target question based on the second assessment result, extracts keywords from the first assessment result, calls the fourth agent to retrieve the professional knowledge base, and simultaneously obtains the language features and contextual information of historical dialogues to generate new conversational questions for in-depth guidance, outputting the third assessment result; Fifth agent integrates the first, second, and third assessment results to generate a structured personal psychological profile (emotional polarity, type of concern, risk indicators) and personalized intervention suggestions, outputting the final result in a visual report; During multiple assessments, the fifth agent also generates a mental health trajectory map through time-series comparison to display the changing trends of each indicator; The privacy and access control module independently manages the encrypted storage and hierarchical access control of all data.

[0133] This application also provides a mental health assessment device; please refer to... Figure 4 The mental health assessment device includes: Anchor point construction module 10 is used to collect language expression samples of users to be evaluated and to build a personalized expression anchor point table for users to be evaluated based on the language expression samples. The personalized expression anchor point table is used to indicate the language expression anchor points of users to be evaluated at each option level of a preset scoring field. The interaction module 20 is used to interact with the user to be tested based on a preset psychological assessment form to obtain the user's natural language response. The calibration module 30 is used to calibrate the natural language response according to the personalized expression anchor table, obtain the target option level of the user to be evaluated for each item in the psychological assessment table, and generate assessment results based on each target option level.

[0134] Optionally, the anchor point construction module 10 is further configured to: Engage in trust-aware interaction with the user to be evaluated to obtain language expression samples from the user to be evaluated; The language expression samples are used to characterize the personal expression scale of the user to be evaluated to preset rating fields, which include frequency, intensity, persistence and scope of influence fields.

[0135] Optionally, the anchor point construction module 10 is further configured to: The language expression samples are semantically encoded to obtain the corresponding expression vectors; According to the option levels of each of the preset rating fields, the expression vectors are grouped and aggregated to obtain multiple vector groups; Based on each of the vector groupings, the personalized expression anchor points of the user to be evaluated at each option level of each of the preset rating fields are determined, so as to construct the personalized expression anchor point table based on each of the personalized expression anchor points.

[0136] Optionally, the anchor point construction module 10 is further configured to: For any first option level of each of the preset rating fields, determine the number of vectors in the first vector group corresponding to the first option level; If the number of vectors is lower than a first preset threshold, determine the second option level that is adjacent to the first option level among the option levels of the preset rating field; Based on the second vector grouping corresponding to the second option level and the first vector grouping, the personalized expression anchor point of the first option level is determined.

[0137] Optionally, the interaction module 20 is further configured to: Transform the original questions in the pre-set psychological assessment scale into conversational questions; Each of the conversational questions is output sequentially to the user to be evaluated, and the user's response in natural language is obtained as the natural language answer.

[0138] Optionally, the calibration module 30 is further configured to: For any item in the psychological assessment form, extract the first semantic vector of the natural language answer corresponding to the item, and the second semantic vector of the level of each candidate option of the item; Extract the third semantic vector related to the preset scoring field from the natural language answer corresponding to the question item, and obtain the anchor vector corresponding to each candidate option level of the question item in the personalized expression anchor table; Determine the first matching degree between the first semantic vector and the second semantic vector, and determine the second matching degree between the third semantic vector and the anchor vector; Candidate scores for each candidate option level are determined based on the first matching degree and the second matching degree; Based on each candidate score, the target option level for each question item is determined.

[0139] Optionally, the calibration module 30 is further configured to: For any candidate option level of the question item, determine the coverage confidence of the natural language answer corresponding to the question item to the preset scoring field, and determine the penalty confidence of the natural language answer corresponding to the question item, wherein the penalty confidence is evaluated from the dimensions of cross-round semantic contradiction and expression ambiguity; Based on the first matching degree, the second matching degree, the coverage confidence degree, and the penalty confidence degree corresponding to the candidate option level, the candidate score of the candidate option level is determined.

[0140] Optionally, the calibration module 30 is further configured to: Normalize each candidate score to obtain the candidate probability of each candidate option level; Determine the first candidate probability with the largest probability and the second candidate probability with the second largest probability among all the candidate probabilities, and determine the difference between the first candidate probability and the second candidate probability; If the difference is greater than or equal to the second preset threshold, the candidate option level corresponding to the first candidate probability is taken as the target option level of the question item; If the difference is less than a second preset threshold, at least one counterfactual answer is generated based on the natural language answer corresponding to the question item, forming a counterfactual set, and the target option level of the question item is determined from each of the candidate option levels based on the counterfactual set.

[0141] Optionally, the calibration module 30 is further configured to: Determine the counterfactual candidate probability for each counterfactual answer in the counterfactual set; For each counterfactual response, determine the third candidate probability, which is the largest among all counterfactual candidate probabilities corresponding to the counterfactual response, and obtain the counterfactual option level corresponding to the third candidate probability; The option level that appears most frequently among the various counterfactual option levels is taken as the target option level for the question item.

[0142] The mental health assessment device provided in this application, employing the mental health assessment method described in the above embodiments, can solve the technical problem of how to improve the accuracy of mental health assessment results. Compared with the prior art, the beneficial effects of the mental health assessment device provided in this application are the same as those of the mental health assessment method provided in the above embodiments, and other technical features in the mental health assessment device are the same as those disclosed in the methods of the above embodiments, and will not be repeated here.

[0143] This application provides a mental health assessment device, which includes: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, which are executed by the at least one processor to enable the at least one processor to perform the mental health assessment method in Embodiment 1 above.

[0144] The following is for reference. Figure 5 The diagram illustrates a structural schematic suitable for implementing the mental health assessment device in the embodiments of this application. The mental health assessment device in the embodiments of this application may include, but is not limited to, mobile terminals such as mobile phones, laptops, digital broadcast receivers, PDAs (Personal Digital Assistants), PMPs (Portable Media Players), in-vehicle terminals (e.g., in-vehicle navigation terminals), and fixed terminals such as digital TVs and desktop computers. Figure 5 The mental health assessment device shown is merely an example and should not impose any limitations on the functionality and scope of use of the embodiments of this application.

[0145] like Figure 5 As shown, the mental health assessment device may include a processing unit 1001 (e.g., a central processing unit, a graphics processing unit, etc.), which can perform various appropriate actions and processes according to a program stored in a read-only memory 1002 or a program loaded from a storage device 1003 into a random access memory 1004. The random access memory 1004 also stores various programs and data required for the operation of the mental health assessment device. The processing unit 1001, the read-only memory 1002, and the random access memory 1004 are interconnected via a bus 1005. An input / output interface 1006 is also connected to the bus. Typically, the following systems can be connected to the input / output interface 1006: input devices 1007 including, for example, a touchscreen, touchpad, keyboard, mouse, image sensor, microphone, accelerometer, gyroscope, etc.; output devices 1008 including, for example, a liquid crystal display (LCD), speaker, vibrator, etc.; storage devices 1003 including, for example, magnetic tape, hard disk, etc.; and communication devices 1009. Communication device 1009 allows the mental health assessment device to communicate wirelessly or wiredly with other devices to exchange data. Although the figure shows mental health assessment devices with various systems, it should be understood that implementing or having all of the systems shown is not required. More or fewer systems may be implemented alternatively.

[0146] Specifically, according to the embodiments disclosed in this application, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments disclosed in this application include a computer program product comprising a computer program carried on a computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via a communication device, or installed from storage device 1003, or installed from read-only memory 1002. When the computer program is executed by processing device 1001, it performs the functions defined in the methods of the embodiments disclosed in this application.

[0147] The mental health assessment device provided in this application, employing the mental health assessment method described in the above embodiments, can solve the technical problem of how to improve the accuracy of mental health assessment results. Compared with the prior art, the beneficial effects of the mental health assessment device provided in this application are the same as those of the mental health assessment method provided in the above embodiments, and other technical features of this mental health assessment device are the same as those disclosed in the previous embodiment method, and will not be repeated here.

[0148] It should be understood that the various parts disclosed in this application can be implemented using hardware, software, firmware, or a combination thereof. In the description of the above embodiments, specific features, structures, materials, or characteristics can be combined in any suitable manner in one or more embodiments or examples.

[0149] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0150] This application provides a computer-readable storage medium having computer-readable program instructions (i.e., a computer program) stored thereon, the computer-readable program instructions being used to execute the mental health assessment method in the above embodiments.

[0151] The computer-readable storage medium provided in this application may be, for example, a USB flash drive, but is not limited to, electrical, magnetic, optical, electromagnetic, infrared, or semiconductor systems or devices, or any combination thereof. More specific examples of computer-readable storage media may include, but are not limited to: electrical connections having one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof. In this embodiment, the computer-readable storage medium may be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system or device. The program code contained on the computer-readable storage medium may be transmitted using any suitable medium, including but not limited to: wires, optical cables, RF (Radio Frequency), etc., or any suitable combination thereof.

[0152] The aforementioned computer-readable storage medium may be included in the mental health assessment device; or it may exist independently and not be assembled into the mental health assessment device.

[0153] The aforementioned computer-readable storage medium carries one or more programs that, when executed by a mental health assessment device, cause the mental health assessment device to: collect language expression samples from a user to be assessed; establish a personalized expression anchor table for the user to be assessed based on the language expression samples, wherein the personalized expression anchor table is used to indicate the language expression anchor points of the user to be assessed at each option level in a preset rating field; perform scale assessment interaction with the user to be assessed based on a preset mental health assessment table to obtain the user's natural language responses; calibrate the natural language responses according to the personalized expression anchor table to obtain the user's target option level for each item in the mental health assessment table, and generate assessment results based on each target option level.

[0154] Computer program code for performing the operations of this application can be written in one or more programming languages ​​or a combination thereof, including object-oriented programming languages ​​such as Java, Smalltalk, and C++, and conventional procedural programming languages ​​such as the "C" language or similar programming languages. The program code can be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving remote computers, the remote computer can be connected to the user's computer via any type of network—including a Local Area Network (LAN) or a Wide Area Network (WAN)—or can be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0155] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of this application. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0156] The modules described in the embodiments of this application can be implemented in software or hardware. The names of the modules do not necessarily limit the functionality of the unit itself.

[0157] The readable storage medium provided in this application is a computer-readable storage medium that stores computer-readable program instructions (i.e., computer programs) for executing the above-described mental health assessment method, thereby solving the technical problem of how to improve the accuracy of mental health assessment results. Compared with the prior art, the beneficial effects of the computer-readable storage medium provided in this application are the same as the beneficial effects of the mental health assessment method provided in the above embodiments, and will not be repeated here.

[0158] This application provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the mental health assessment method described above.

[0159] The computer program product provided in this application can improve the accuracy of mental health assessment results. Compared with the prior art, the beneficial effects of the computer program product provided in this application are the same as the beneficial effects of the mental health assessment method provided in the above embodiments, and will not be repeated here.

[0160] The above are merely preferred embodiments of this application and do not limit the patent scope of this application. Any equivalent structural or procedural transformations made using the content of this application's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent scope of this application.

Claims

1. A method for assessing mental health, characterized in that, The mental health assessment methods include: Collect language expression samples from users to be evaluated, and establish a personalized expression anchor point table for users to be evaluated based on the language expression samples. The personalized expression anchor point table is used to indicate the language expression anchor points of users to be evaluated at each option level of a preset scoring field. Based on a preset psychological assessment form, the scale assessment interaction is conducted with the user to be assessed to obtain the user's natural language response. The natural language responses are calibrated based on the personalized expression anchor table to obtain the target option level of the user to be evaluated for each item in the psychological assessment table, and assessment results are generated based on each target option level.

2. The mental health assessment method as described in claim 1, characterized in that, The steps for collecting language expression samples from users to be evaluated include: Engage in trust-aware interaction with the user to be evaluated to obtain language expression samples from the user to be evaluated; The language expression samples are used to characterize the personal expression scale of the user to be evaluated to preset rating fields, which include frequency, intensity, persistence and scope of influence fields.

3. The mental health assessment method as described in claim 1, characterized in that, The step of establishing a personalized expression anchor point table for the user to be evaluated based on the language expression sample includes: The language expression samples are semantically encoded to obtain the corresponding expression vectors; According to the option levels of each of the preset rating fields, the expression vectors are grouped and aggregated to obtain multiple vector groups; Based on each of the vector groupings, the personalized expression anchor points of the user to be evaluated at each option level of each of the preset rating fields are determined, so as to construct the personalized expression anchor point table based on each of the personalized expression anchor points.

4. The mental health assessment method as described in claim 3, characterized in that, The step of determining the personalized expression anchor point of the user to be evaluated at each option level of each preset rating field based on each of the vector groupings includes: For any first option level of each of the preset rating fields, determine the number of vectors in the first vector group corresponding to the first option level; If the number of vectors is lower than a first preset threshold, determine the second option level that is adjacent to the first option level among the option levels of the preset rating field; Based on the second vector grouping corresponding to the second option level and the first vector grouping, the personalized expression anchor point of the first option level is determined.

5. The mental health assessment method as described in claim 1, characterized in that, The step of interacting with the user to be assessed based on a preset psychological assessment form to obtain the user's natural language response includes: Transform the original questions in the pre-set psychological assessment scale into conversational questions; Each of the conversational questions is output sequentially to the user to be evaluated, and the user's response in natural language is obtained as the natural language answer.

6. The mental health assessment method as described in claim 1, characterized in that, The step of calibrating the natural language response based on the personalized expression anchor table to obtain the target option level of each item in the psychological assessment form for the user to be assessed includes: For any item in the psychological assessment form, extract the first semantic vector of the natural language answer corresponding to the item, and the second semantic vector of the level of each candidate option of the item; Extract the third semantic vector related to the preset scoring field from the natural language answer corresponding to the question item, and obtain the anchor vector corresponding to each candidate option level of the question item in the personalized expression anchor table; Determine the first matching degree between the first semantic vector and the second semantic vector, and determine the second matching degree between the third semantic vector and the anchor vector; Candidate scores for each candidate option level are determined based on the first matching degree and the second matching degree; Based on each candidate score, the target option level for each question item is determined.

7. The mental health assessment method as described in claim 6, characterized in that, The step of determining the candidate score for each candidate option level based on the first matching degree and the second matching degree includes: For any candidate option level of the question item, determine the coverage confidence of the natural language answer corresponding to the question item to the preset scoring field, and determine the penalty confidence of the natural language answer corresponding to the question item, wherein the penalty confidence is evaluated from the dimensions of cross-round semantic contradiction and expression ambiguity; Based on the first matching degree, the second matching degree, the coverage confidence degree, and the penalty confidence degree corresponding to the candidate option level, the candidate score of the candidate option level is determined.

8. The mental health assessment method as described in claim 6, characterized in that, The step of determining the target option level of the question item based on each of the candidate scores includes: Normalize each candidate score to obtain the candidate probability of each candidate option level; Determine the first candidate probability with the largest probability and the second candidate probability with the second largest probability among all the candidate probabilities, and determine the difference between the first candidate probability and the second candidate probability; If the difference is greater than or equal to the second preset threshold, the candidate option level corresponding to the first candidate probability is taken as the target option level of the question item; If the difference is less than a second preset threshold, at least one counterfactual answer is generated based on the natural language answer corresponding to the question item, forming a counterfactual set, and the target option level of the question item is determined from each of the candidate option levels based on the counterfactual set.

9. The mental health assessment method as described in claim 8, characterized in that, The step of determining the target option level of the question item from each of the candidate option levels based on the counterfactual set includes: Determine the counterfactual candidate probability for each counterfactual answer in the counterfactual set; For each counterfactual response, determine the third candidate probability, which is the largest among all counterfactual candidate probabilities corresponding to the counterfactual response, and obtain the counterfactual option level corresponding to the third candidate probability; The option level that appears most frequently among the various counterfactual option levels is taken as the target option level for the question item.

10. A mental health assessment device, characterized in that, The mental health assessment device includes: a memory, a processor, and a computer program stored in the memory and executable on the processor, the computer program being configured to implement the steps of the mental health assessment method as described in any one of claims 1 to 9.

11. A storage medium, characterized in that, The storage medium is a computer-readable storage medium, on which a computer program is stored, and when the computer program is executed by a processor, it implements the steps of the mental health assessment method as described in any one of claims 1 to 9.

Images