Psychological assessment method based on large language model Agent

Abstract

The invention belongs to artificial intelligence, and relates to a psychological assessment method based on a large language model Agent, which comprises the following steps of: extracting user-level cue words from each round of dialogue with a user by utilizing a user psychological feature extraction module, connecting the user-level cue words with system-level cue words obtained when a large language model is loaded, performing entity recognition on the cue words spliced together through a large language model; matching the extracted entity nodes with symptom nodes in the psychological state knowledge graph by utilizing a user psychological state retrieval module, and retrieving all corresponding psychological state node sets; the Agent questioning module is used for initiating questioning according to the answer of the user; according to the psychological assessment report generation module, the large language model searches a self-scale with the highest similarity with the user state according to all features extracted by dialogue with the user, and sends a report with the highest similarity with the condition that the user completes the self-scale to the user; the use experience of the user and the effectiveness of the evaluation result are improved.

Description

Technical Field

[0001] The invention belongs to the fields of artificial intelligence, natural language processing and emotional computing, and in particular relates to a psychological assessment method based on a large language model Agent. Background Art

[0002] The earliest natural language processing (NLP) algorithms used to solve psychological assessment problems can be traced back to the 1990s, mainly relying on hand-crafted dictionaries and simple rules to identify and classify mental states in text. In 2002, machine learning methods were first used for emotion classification, including Naive Bayes, decision trees, and support vector machines (SVM), marking the entry of psychological assessment-related algorithms into the machine learning era. With the advent of deep learning, neural networks, especially recurrent neural networks (RNNs) and long short-term memory networks (LSTMs), have significantly improved the ability of psychological assessment by capturing contextual information in text.

[0003] In recent years, large language models (LLMs) based on the Transformer architecture, such as BERT and GPT, have greatly promoted the application of NLP in psychological assessment. BERT is pre-trained with a bidirectional Transformer and is able to understand the contextual relationship of each word in a sentence, which is particularly important in emotion recognition. By fine-tuning on the emotion annotation dataset, BERT has demonstrated efficient emotion classification capabilities. Models such as GPT-3 and GPT-4 have demonstrated strong capabilities in natural language generation and understanding tasks through generative pre-training, especially when dealing with complex emotional expressions. Although BERT and GPT have achieved remarkable results in psychological assessment, they still have some shortcomings, including insufficient contextual dependency, especially when dealing with long texts and multi-round conversations, which may lose contextual information; they are relatively fragile to adversarial samples and are easily affected by small perturbations in the input text, resulting in inaccurate psychological assessment results; and the model complexity and resource requirements are high, requiring a large amount of computing resources and data for training and fine-tuning, which poses a challenge to resource-limited environments in practical applications. Summary of the invention

[0004] In view of the problems existing in the prior art, the present invention proposes a psychological assessment method based on a large language model agent, including a user psychological feature extraction module, a user psychological state retrieval module, an agent questioning module, and a psychological assessment report generation module, wherein:

[0005] The user psychological feature extraction module is used to extract user-level prompt words from each round of dialogue with the user, connect the user-level prompt words with the system-level prompt words obtained when loading the large language model, perform entity recognition on the spliced ​​prompt words through the large language model, and summarize the entities obtained in each round of dialogue as user psychological features;

[0006] The user mental state retrieval module is used to match the extracted entity nodes with the symptom nodes in the mental state knowledge graph and retrieve the corresponding set of all mental state nodes;

[0007] Agent questioning module, used to initiate questions based on the user's answers;

[0008] In the psychological assessment report generation module, the large language model searches for the self-assessment scale that is most similar to the user's status based on all the features extracted from the conversation with the user, and sends the report that is most similar to the user's completed self-assessment scale to the user.

[0009] Furthermore, in the mental state knowledge graph, the information in the mental illness diagnosis manual is structured into nodes and edges in the knowledge graph, where the nodes in the knowledge graph correspond to mental states and the edges correspond to symptoms.

[0010] Furthermore, when the Agent follow-up module asks follow-up questions, it extracts keywords related to mental state symptoms in the conversation. When new keywords appear in the user's answer, it calls the secondary question vocabulary corresponding to the keywords and selects an unasked question from the vocabulary to ask the user; if the user does not appear in the conversation within the specified number of times, the questioning ends.

[0011] Furthermore, the acquisition of system-level prompt words includes: when the Agent questioning module conducts questioning, a corresponding psychological state set is obtained according to the symptom nodes obtained from the user's question and answer, and other symptom nodes corresponding to the corresponding psychological state set are used as system-level prompt words.

[0012] Furthermore, when recommending a self-examination form to a user, all mental state information of the user is obtained, the probability that the mental state information belongs to a mental symptom is calculated, and the self-examination form corresponding to the symptom with the probability greater than a set threshold is pushed to the user.

[0013] Furthermore, the probability that the user's mental state information belongs to a mental symptom is expressed as:

[0014]

[0015] Among them, U i represents the probability that the user's mental state information belongs to mental state i; u jrepresents the jth mental state of the user, N is the number of mental states of the user; w i,j is the correlation degree between the user's j-th mental state and symptom i.

[0016] Furthermore, the degree of association between the j-th mental state of the user and symptom i is the proportion of patients diagnosed with symptom i in the knowledge base who have mental state j.

[0017] Compared with the prior art, the agent-based psychological assessment method proposed in the present invention has the following advantages and positive effects:

[0018] 1. Enhanced context retention capability: Through multiple large language models (LLMs) loaded with specific prompt words and connected through thought chains, the present invention can maintain the consistency and coherence of context in multiple rounds of dialogue, avoiding the problem of context information loss that may occur in traditional models when processing long texts and complex dialogues, thereby improving the accuracy of psychological assessment.

[0019] 2. Improve the robustness of adversarial samples: By introducing knowledge graphs and vector knowledge bases, the present invention has higher robustness in identifying and processing adversarial samples, can effectively deal with small disturbances in the input text, and ensure the stability and professionalism of psychological assessment results.

[0020] 3. Efficient resource utilization: Compared with traditional large language models that require a lot of computing resources and data for training and fine-tuning, the present invention can be efficiently implemented in resource-limited environments through modular design and optimization strategies, reducing the demand for GPUs and training data and expanding its application scenarios.

[0021] 4. Personalized and accurate psychological assessment: Through the collaborative work of the user psychological reaction feature extraction module, the psychological state retrieval module, the agent questioning module, the scale self-assessment module and the psychological assessment report generation module, the present invention can provide personalized and accurate mental health assessment, generate detailed psychological assessment reports and targeted life suggestions, and improve the user experience and the effectiveness of the assessment results.

[0022] 5. Automation and intelligence: The present invention realizes a high degree of automation and intelligence in the psychological assessment process, reduces dependence on manual intervention, can be promoted and applied on a large scale, improves the accessibility and popularity of mental health services, and provides an effective technical means for the early detection and intervention of mental health problems. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] Figure 1 The present invention is a schematic diagram of the use of a large language model Agent-based psychological assessment method. DETAILED DESCRIPTION

[0024] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0025] The present invention proposes a psychological assessment method based on a large language model agent, comprising a user psychological feature extraction module, a user psychological state retrieval module, an agent questioning module, and a psychological assessment report generation module, wherein:

[0026] The user psychological feature extraction module is used to extract user-level prompt words from each round of dialogue with the user, connect the user-level prompt words with the system-level prompt words obtained when loading the large language model, perform entity recognition on the spliced ​​prompt words through the large language model, and summarize the entities obtained in each round of dialogue as user psychological features;

[0027] The user mental state retrieval module is used to match the extracted entity nodes with the symptom nodes in the mental state knowledge graph and retrieve the corresponding set of all mental state nodes;

[0028] Agent questioning module, used to initiate questions based on the user's answers;

[0029] In the psychological assessment report generation module, the large language model searches for the self-assessment scale that is most similar to the user's status based on all the features extracted from the conversation with the user, and sends the report that is most similar to the user's completed self-assessment scale to the user.

[0030] In this embodiment, a psychological assessment method based on a large language model agent mainly includes the following four parts:

[0031] 1. Multiple Large Language Models (LLMs) loaded with specific prompt words: These LLMs use pre-trained language models and are able to generate expected natural language responses in psychological assessment scenarios by loading specific prompt words.

[0032] 2. Knowledge graph based on professional mental illness diagnostic manuals: The knowledge graph structures the information in professional mental illness diagnostic manuals into nodes and edges. Nodes represent mental states and symptoms, and edges represent the relationships between them, helping the system to reason and retrieve.

[0033] 3. Knowledge base built based on professional psychological self-assessment scales: The knowledge base contains various standardized psychological self-assessment scales, which are used in actual psychological assessment scenarios to evaluate the user's psychological state through a standardized scoring system.

[0034] 4. Vector knowledge base based on professional mental illness treatment manuals: The vector knowledge base vectorizes the contents of the treatment manuals and is used to quickly retrieve relevant content when generating life advice.

[0035] The above four main parts are integrated into a psychological assessment method based on a large language model Agent of the present invention to form various functional modules, namely, a user psychological feature extraction module, a user psychological state retrieval module, an Agent questioning module, and a psychological assessment report generation module, wherein:

[0036] The user psychological feature extraction module is mainly used to obtain the feature node word list and prompt words, and construct the psychological reaction feature set based on the acquired data, where:

[0037] 1. Feature node vocabulary, which consists of symptom nodes of all mental states in the knowledge graph, covering a wide range of psychological reaction characteristics; among them:

[0038] Prompt words include system-level prompt words and user-level prompt words. The system part is loaded as the system-level prompt words of LLM during inference to ensure that the model can recognize and extract psychological reaction features in the conversation; the user part is dynamically spliced ​​after receiving the user's answer, combined into a user-level query and passed into LLM for entity recognition;

[0039] The psychological reaction feature set stores the psychological reaction features extracted from the user in each round of dialogue. It serves as one of the short-term memories of the agent and supports the processing of subsequent modules.

[0040] 2. User mental state retrieval module, including knowledge graph, retrieval mechanism and transfer module, is used to match the extracted entity nodes with the symptom nodes in the mental state knowledge graph and retrieve the corresponding set of all mental state nodes, where:

[0041] The knowledge graph, namely the mental state knowledge graph, is composed of mental state nodes and symptom nodes. The mental state nodes represent different mental states, and the symptom nodes represent different psychological reactions. There is a many-to-many relationship between the two.

[0042] Retrieval mechanism: After each round of dialogue, the agent matches the symptom nodes of the knowledge graph based on the features in the psychological reaction feature set as entity nodes, and retrieves the mental state node set that contains all the symptoms in the feature set;

[0043] The transfer module selects the symptom nodes with the greatest discrimination that are not included in the psychological response feature set and passes them to the follow-up module to ensure the accuracy and pertinence of subsequent conversations.

[0044] 3. The Agent questioning module includes characteristic question prompt words generated according to the large language model, and generates questions in combination with prompt words; wherein:

[0045] Feature question prompts, including system-level prompts and user-level prompts; the system part is loaded as the system-level prompts of LLM during inference to ensure that the model can generate targeted questions based on the query. The user part is dynamically spliced ​​into the template by the symptom nodes obtained by the mental state retrieval module, and combined into a user-level query and passed into LLM;

[0046] Follow-up question generation: Generate targeted follow-up questions based on specific symptom nodes, guide users to provide more detailed psychological reaction information, and further improve the psychological reaction feature set.

[0047] 4. Scale self-assessment module, including the psychological self-assessment scale knowledge base, scale selection mechanism, scoring and evaluation, used to recommend the best self-assessment scale to the user based on the user's psychological response feature set; among which:

[0048] Psychological self-assessment scale knowledge base: contains a variety of standardized psychological self-assessment scales, which have been professionally certified and are scientific and practical;

[0049] Scale selection mechanism: After obtaining the most likely mental state of the user through multiple rounds of dialogue, the agent selects a targeted self-assessment scale from the knowledge base to ensure that the scale is highly relevant to the user's current mental state;

[0050] Scoring and evaluation: After the user completes the self-assessment scale, the Agent evaluates it according to a standardized scoring system to determine the severity of the user's mental state.

[0051] As an optional implementation scheme, when recommending a self-examination form to a user, all the user's mental state information is obtained, the probability that the mental state information belongs to a mental symptom is calculated, and the self-examination form corresponding to the symptom with a probability greater than a set threshold is pushed to the user. The probability that the user's mental state information belongs to a mental symptom is expressed as:

[0052]

[0053] Among them, U i represents the probability that the user's mental state information belongs to mental state i; u j represents the jth mental state of the user, N is the number of mental states of the user; w i,j is the correlation degree between the j-th mental state of the user and symptom i, and the correlation degree between the j-th mental state of the user and symptom i is the proportion of patients with mental state j in the knowledge base who are diagnosed with symptom i.

[0054] 5. Psychological assessment report generation module, including vector knowledge base, life advice prompts and assessment report generation; among which:

[0055] Vector knowledge base: The text of professional mental illness treatment manuals is segmented according to specific rules and stored in vector form through embedding, which supports fast retrieval and content generation;

[0056] Life suggestion prompts: including system-level prompts and user-level prompts. The system part is loaded as the system-level prompts of LLM during inference to ensure that the model can generate targeted suggestions based on the query. The Agent searches the vector knowledge base to obtain the reference content of the treatment method through the retrieval-augmented generation (RAG) method based on the scale score results and psychological reaction characteristics. The user part accepts the Agent scale score results, psychological reaction characteristics and treatment method reference content as parameters, and combines them into the user-level query and passes it to LLM.

[0057] Evaluation report generation: The agent performs the above reasoning on all symptoms in the psychological reaction feature set, and combines the scale scoring results to generate a comprehensive psychological evaluation report, including the user's mental health status analysis and specific life suggestions.

[0058] As an optional implementation, the five modules are connected in a thought chain, and the user's mental health status is evaluated based on multiple rounds of conversations. The specific process is as follows:

[0059] 101. Dialogue: Enter the user psychological reaction feature extraction module to identify and store the user's psychological and response features;

[0060] 102. Mental state retrieval: Based on the extracted features, the knowledge graph is retrieved to determine the possible mental state;

[0061] 103. Precise follow-up questions: Generate targeted follow-up questions through the follow-up question module to further refine the user's psychological reaction characteristics;

[0062] 104. Scale self-assessment: According to the mental state determined by multiple rounds of dialogue, a suitable self-assessment scale is selected, and the user is rated after completing the self-assessment;

[0063] 105. Generate report: Combine all the information and generate detailed psychological assessment report and life suggestions through the assessment report generation module.

[0064] Specifically, this embodiment provides a specific implementation process of steps 101 to 105, such as Figure 1 , including the following steps:

[0065] 200. Start the Agent, which randomly asks the user initial questions from a preset list of questions. These questions are designed to guide the user into a state of psychological assessment of the subject.

[0066] 201. User answer: The user answers the questions raised by the Agent based on his / her recent mental state. The answer may involve psychological reactions in terms of emotions, behaviors, thoughts, etc.

[0067] 202. Psychological reaction feature extraction stage:

[0068] 202-A: Feature extraction prompt word loading: After receiving the user's answer, the Agent enters the user psychological reaction feature extraction module. The feature extraction prompt words of this module include system part and user part: the system part is loaded as the system-level prompt word of LLM during reasoning to ensure that the model can recognize and extract psychological reaction features in the conversation; the user part is dynamically spliced ​​after receiving the user's answer to form a user-level query and passed to LLM for entity recognition;

[0069] 202-B: Psychological reaction feature recognition: LLM uses feature extraction prompts to infer and identify and extract psychological reaction features from the user's answers. These features may include emotional reactions such as irritability, loss of appetite, insomnia, and behavioral or thinking characteristics related to specific mental states.

[0070] 202-C Feature storage: The identified psychological reaction features are stored in the psychological reaction feature set as the short-term memory of the Agent, providing basic data for the processing of subsequent modules.

[0071] 203. Mental state retrieval stage:

[0072] 203-A: Knowledge graph matching: read all psychological reaction features from short-term memory, match the corresponding symptom nodes in the knowledge graph according to these features, and the many-to-many relationship between mental state nodes and symptom nodes contained in the knowledge graph enables the agent to retrieve possible mental state node sets based on the features;

[0073] 203-B: Symptom node selection: After finding the mental state node that contains all the psychological reaction features, the agent selects new symptom nodes with the highest discrimination that are not included in the psychological reaction feature set. These new symptom nodes will be used for the next step of precise questioning.

[0074] 204. Precision Questioning Stage:

[0075] 204-A: Loading follow-up question prompts: After the Agent obtains the new symptom node, it enters the user psychological state retrieval module. The system part of the characteristic follow-up question prompt is loaded as the system-level prompt of the LLM during reasoning to ensure that the model can generate targeted follow-up questions based on the query. The user part is combined into a user-level query and passed into the LLM by dynamically splicing the new symptom node into the template;

[0076] 205-B: Generate follow-up questions: Generate targeted follow-up questions based on new symptom nodes to guide users to provide more detailed psychological response information. The content of the follow-up questions may involve the user's specific feelings, behaviors or thinking patterns in a specific situation;

[0077] 205-C: Feedback collection and processing: The user's answer to the follow-up question is processed again by the psychological reaction feature extraction module, and the new features are added to the psychological reaction feature set.

[0078] 206. Iteration phase:

[0079] 206-A: Multiple loops: The process of steps 202 to 205 is repeatedly executed. The system gradually improves the psychological reaction feature set by continuously extracting and analyzing the user's psychological reaction features. Each iteration generates more accurate questions based on new symptom nodes to obtain more comprehensive information about the user's psychological state.

[0080] 206-B: Mental state confirmation: Finally, through multiple rounds of dialogue and iterative retrieval, the agent finds the mental state node that best matches the user's psychological response characteristics.

[0081] 207. Scale self-assessment stage:

[0082] 207-A: Rating scale selection: After the Agent obtains the most likely mental state node of the user, it enters the scale self-assessment module. The system selects targeted self-assessment scales from the knowledge base. These scales have been professionally certified and are scientific and practical.

[0083] 207-B: Users complete a self-assessment scale and answer specific questions about their mental state. The scale questions cover different dimensions of mental health, including emotions, cognition, behavior, etc.

[0084] 207-C: Scoring assessment. After the user completes the self-assessment, the Agent will evaluate according to the standardized scoring system to determine the severity of the user's mental state. The scoring results will be used to generate subsequent psychological assessment reports.

[0085] 208. Evaluation report generation stage:

[0086] 208-A: Vector knowledge base retrieval: After obtaining the user's scale score results and psychological reaction feature set, the agent enters the psychological assessment report generation module. The system part of the life suggestion prompt word is loaded as the system-level prompt word of the LLM to ensure that the model has the ability to generate targeted suggestions. Through the retrieval enhanced generation (RAG) method, the agent retrieves the treatment method reference content in the vector knowledge base;

[0087] 208-B: Generate life suggestions: The user part accepts the Agent scale score results, psychological reaction characteristics and treatment method reference content as parameters, combines them into user-level queries and passes them into LLM. The model generates life suggestions for different symptoms and mental states based on these parameters;

[0088] 208-C: Comprehensive report analysis: The agent performs the above reasoning on all symptoms in the psychological reaction feature set, and combines the scale scoring results to generate a comprehensive psychological assessment report. The report includes an analysis of the user's mental health status, a specific description of the symptoms, an assessment of the severity, and life suggestions for each symptom.

[0089] 209. In the report delivery phase, the final psychological assessment report is delivered to the user through the Agent interface, so that the user can intuitively understand his or her own psychological state. The suggestions and assessment results in the report provide users with specific action guidance to help them manage and improve their mental health.

[0090] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A psychological assessment method based on a large language model agent, characterized in that: It includes user psychological feature extraction module, user psychological state retrieval module, agent questioning module, scale self-assessment module, and psychological assessment report generation module, among which: The user psychological feature extraction module is used to extract user-level prompt words from each round of dialogue with the user, connect the user-level prompt words with the system-level prompt words obtained when loading the large language model, perform entity recognition on the spliced ​​prompt words through the large language model, and summarize the entities obtained in each round of dialogue as user psychological features; The user mental state retrieval module is used to match the extracted entity nodes with the symptom nodes in the mental state knowledge graph and retrieve the corresponding set of all mental state nodes; Agent questioning module, used to initiate questions based on the user's answers; In the self-assessment module, the large language model searches for the self-assessment scale with the highest similarity to the user's status based on all the features extracted from the conversation with the user; The psychological assessment report generation module will send the report with the highest similarity to the user's completed self-scale to the user.

2. A psychological assessment method based on a large language model Agent according to claim 1, characterized in that: In the mental state knowledge graph, the information in the mental illness diagnostic manual is structured into nodes and edges in the knowledge graph, where the nodes in the knowledge graph correspond to mental states and the edges correspond to symptoms.

3. A psychological assessment method based on a large language model Agent according to claim 1, characterized in that: When the Agent follow-up question module asks follow-up questions, it extracts keywords related to mental state symptoms in the conversation. When new keywords appear in the user's answer, it calls the secondary questioning vocabulary corresponding to the keywords and selects an unasked question from the vocabulary to ask the user; if the user does not appear in the conversation within the specified number of times, the questioning ends.

4. A psychological assessment method based on a large language model Agent according to claim 3, characterized in that: The acquisition of system-level prompt words includes: when the Agent questioning module conducts questioning, the corresponding psychological state set is acquired according to the symptom nodes obtained from the user's question and answer, and other symptom nodes corresponding to the corresponding psychological state set are used as system-level prompt words.

5. A psychological assessment method based on a large language model Agent according to claim 1, characterized in that: When recommending a self-examination form to a user, all mental state information of the user is obtained, the probability that the mental state information belongs to a mental symptom is calculated, and the self-examination form corresponding to the symptom with a probability greater than a set threshold is pushed to the user.

6. A psychological assessment method based on a large language model Agent according to claim 1, characterized in that: The probability that the user's mental state information belongs to a mental symptom is expressed as: Among them, U i represents the probability that the user's mental state information belongs to mental state i; u j represents the jth mental state of the user, N is the number of mental states of the user; w i,j is the correlation degree between the user's j-th mental state and symptom i.

7. A psychological assessment method based on a large language model Agent according to claim 6, characterized in that: The degree of association between the user's j-th mental state and symptom i is the proportion of patients diagnosed with symptom i in the knowledge base who have mental state j.

Images