Inplanatable psychological anomaly prediction method of combined large language model

Abstract

The invention provides an interpretable psychological anomaly prediction method of a combined large language model, and belongs to the field of large models. Comprising the following steps: S1, realizing a dialogue webpage of an open problem by utilizing an API (Application Program Interface) of GPT-3. 5; s2, completing an interpretable feature extraction model by using an API (Application Program Interface) of GPT-4 to obtain an interpretable psychological feature set; s3, extracting a psychological condition feature data set by using an API (Application Program Interface) of GPT-4; and S4, constructing a psychological condition prediction model, and training the psychological feature data set by using a classifier to obtain the psychological condition prediction model. According to the invention, through organic combination of GPT-3. 5 and GPT-4, psychological data collection of dialogue data and picture data is realized by using high-concurrency dialogue of GPT-3. 5; the extraction of interpretable features is realized by utilizing offline analysis of GPT-4, the mining of the interpretable features of the large language model is fully utilized, and the prediction of the psychological condition of the user is realized through the trained psychological condition prediction model.

Description

Technical Field

[0001] The present invention belongs to the technical field of large models, and in particular relates to an interpretable psychological anomaly prediction method combining a large language model. Background Art

[0002] In order to quickly identify the user's psychological abnormalities and achieve high-speed initial screening, the existing technology often uses self-tests or peer assessment scales, or medical examinations to detect psychological abnormalities. With the development of AI, there are many conversational robots on the market that perceive psychological abnormalities.

[0003] Patent CN117609486A "An Intelligent Dialogue System in the Psychology Field" uses a large language model to model the user dialogue context and recommend psychology-related books and interesting activities to users. It does not extract images and interpretability features.

[0004] Patent CN117828054A "A dialogue method and device based on artificial intelligence psychological knowledge graph" The dialogue method based on artificial intelligence psychological knowledge graph includes obtaining current input information; obtaining prompt words matching the current input information according to the psychological knowledge graph; inputting the prompt words and the current input information into a preset large model to generate reply information. The promoter is set to use the large language model to realize the perception of the user's emotional state, but it does not achieve the interpretability feature, nor does it achieve the organic combination of multiple (gpt3.5 and gpt4.0) to support high concurrency.

[0005] Patent CN117744801A "Dialogue method, device and equipment based on artificial intelligence psychology big model" can identify the psychological scene of the user's input, and then generate prompt words corresponding to the reply strategy based on the hit psychological scene, so that the generative big model can generate more professional replies that fit psychology based on the prompt words and current input information. The extraction of interpretable features is not achieved.

[0006] Although many existing technologies use conversational robots to perceive psychological abnormalities, current conversational robots do not fully utilize the interpretable feature mining of large language models to achieve interpretable feature mining of text conversations and house-tree-person relationships. Summary of the invention

[0007] The purpose of the present invention is to provide an interpretable psychological anomaly prediction method combining a large language model, using GPT-3.5 as the main dialogue model and GPT-4 as the main model for analyzing and extracting interpretable features for psychological anomaly prediction, so as to solve the technical problem in the prior art that the dialogue robot does not fully utilize the interpretable feature mining of the large language model to perceive psychological anomalies.

[0008] To solve the above technical problems, the interpretable psychological abnormality prediction method of the combined large language model of the present invention comprises the following steps:

[0009] Step S1: Use the GPT-3.5 API to implement a dialogue webpage for open questions;

[0010] Step S2: Use the API of GPT-4 to complete the interpretable feature extraction model and obtain an interpretable psychological feature set;

[0011] Step S3: Use the GPT-4 API to complete the extraction of the psychological condition feature data set;

[0012] Step S4: construct a psychological condition prediction model, and use a classifier to train the psychological feature data set to obtain the psychological condition prediction model.

[0013] Furthermore, step S1 includes the following steps:

[0014] Step S11: Use GPT-3.5 to design a dialogue Promoter 1. The designed dialogue Promoter 1 can complete 5-10 rounds of dialogue, have open questions, and reflect certain psychological characteristics of users;

[0015] Step S12: Design a webpage system whose front end can input dialogues, save multiple rounds of dialogues, and upload a picture of a house, tree, and person;

[0016] Step S13: Collecting conversation data and picture data of people with mental abnormalities and people with normal mental health.

[0017] Further, step S2 includes the following steps:

[0018] Step S21: Use GPT-4 to extract the features of the conversation data and obtain the psychological features of the conversation; use GPT-4 to extract the features of the picture data and obtain the psychological features of the picture;

[0019] Step S22: Count the extracted dialogue psychological features and picture psychological features to obtain N psychological features with higher frequencies as an interpretable psychological feature set.

[0020] Further, step S3 includes the following steps:

[0021] Step S31: constructing a dialogue promoter 3 based on the interpretable psychological feature set, extracting all psychological features one by one, and forming a specific quantitative value of [0,1];

[0022] Step S32: extract interpretable psychological features from the collected conversation data and picture data of all psychologically abnormal and normal persons using the conversation promoter 3 to obtain a psychological condition feature data set.

[0023] Furthermore, when predicting psychological abnormalities for new users, we first use the dialogue webpage of open questions implemented by GPT-3.5 to obtain the user's dialogue data and image data, and then use the dialogue promoter3 built by GPT-4 to extract psychological features from the dialogue data and image data to obtain the user's psychological features. The user's psychological features are input into the trained psychological prediction model to obtain the psychological prediction results.

[0024] Compared with the prior art, the present invention has the following beneficial technical effects: the present invention realizes the psychological data collection of conversation data and image data by organically combining GPT-3.5 and GPT-4, and utilizes the high concurrent conversation of GPT-3.5; realizes the extraction of interpretable features by utilizing the offline analysis of GPT-4, and makes full use of the interpretable feature mining of the large language model. The psychological condition of the user is predicted by the trained psychological condition prediction model. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required for use in the description of the embodiments of the present invention will be briefly introduced below. Obviously, the accompanying drawings in the following description are only some embodiments of the present invention. For ordinary technicians in this field, other accompanying drawings can be obtained based on these accompanying drawings without paying creative labor.

[0026] Figure 1 This is a diagram of the method for predicting abnormal psychological characteristics of the present invention. DETAILED DESCRIPTION

[0027] 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.

[0028] The most well-known large language model at present is OpenAI's ChatGPT series, which can complete image understanding and conversation tasks. However, the more effective GPT-4 has a longer interaction waiting time and cannot support high concurrency. The corresponding weaker GPT-3.5 can respond quickly and support high concurrency.

[0029] In view of the technical problem in the prior art that conversational robots do not fully utilize the explainable feature mining of large language models to perceive psychological abnormalities, the present invention combines the high concurrency of GPT-3.5 and the high performance of GPT-4. The present invention uses GPT-3.5 as the main model for conversation, and GPT-4 as the main model for analyzing and extracting explainable features for psychological abnormality prediction.

[0030] The interpretable psychological anomaly prediction method based on a combined large language model proposed in the present invention comprises the following steps:

[0031] Step S1: Use the GPT-3.5 API to implement a dialogue webpage for open questions.

[0032] Specifically, step S11: use GPT-3.5 to design dialogue Promoter 1, and the designed dialogue Promoter 1 can achieve the ability to complete 5-10 rounds of dialogue, have open-ended questions, and reflect certain psychological characteristics of users.

[0033] Before the conversation started, the dialogue Promoter 1 was designed as follows: You are now a psychological counselor. Please ask me 5 open-ended questions to understand my mental health status. Please ask me one question at a time. At the end, please say to me, "Please use the drawing function on your computer to draw a house, a person, and a tree, and upload the photo." Now please start asking me mental health questions.

[0034] Add "Please continue to ask me mental health questions." or "Please continue to ask me the next mental health question" or "Any more questions?" after the user's answer.

[0035] An example of a dialogue with Promoter 1 is as follows (the framed part is the promoter):

[0036]

[0037] Step S12: Design a web page system whose front end can input dialogues, save multiple rounds of dialogues, and upload a picture of a house, tree, and person.

[0038] Step S13: Collecting conversation data and picture data of people with mental abnormalities and people with normal mental health.

[0039] Step S2: Use the GPT-4 API to complete the interpretable feature extraction model and obtain an interpretable psychological feature set.

[0040] Specifically, step S21: use GPT-4 to complete the free feature extraction of dialogue data to obtain dialogue psychological features; use GPT-4 to complete the free feature extraction of image data to obtain image psychological features.

[0041] After collecting enough conversation data, we designed a dialogue promoter 2 to ask GPT-4 about the content of the conversation and obtain the psychological characteristics of the conversation. The designed dialogue promoter 2 is: What psychological characteristics of the respondent are reflected in the following conversation? Please answer in the following format: Extract psychological dimensions: qualitative conclusions, such as sleep quality: average.

[0042] An example of a conversation promoter 2 is as follows (the solid line is the promoter, and the dotted line is the conversation data collected by GPT-3.5):

[0043]

[0044] The content before the “:” in the GPT-4 reply is used as the extracted conversation psychological features.

[0045] After collecting enough image data, we designed a dialogue promoter2 to ask GPT-4 about the image data to obtain the psychological characteristics of the image. The designed dialogue promoter2 is: What psychological characteristics of the painter are reflected in the following picture? Please answer in the following format: Extract psychological dimensions: qualitative conclusions, such as self: low.

[0046] Step S22: Count the extracted conversation psychological features and picture psychological features to obtain N psychological features with higher frequencies as an interpretable psychological feature set, such as anxiety, depression, sleep, interpersonal relationships, etc.

[0047] Step S3: Use GPT-4’s API to complete the extraction of the psychological condition feature data set.

[0048] Step S31: Construct a dialogue promoter 3 based on the interpretable psychological feature set to extract all psychological features one by one and form a specific quantitative value of [0,1]. The designed dialogue promoter 3 is: In the following dialogue, does the respondent have a problem with a certain psychological feature? Please reply with a specific value from 0 to 1. The larger the value, the more problematic it is. Please only reply with a specific value. A certain psychological feature is a psychological feature in the interpretable psychological feature set.

[0049] An example of a conversation promoter 3 is as follows (the solid line is the promoter, and the dotted line is the conversation data collected by GPT-3.5):

[0050]

[0051]

[0052] Step S32: extract interpretable psychological features from the collected conversation data and picture data of all psychologically abnormal and normal persons using the conversation promoter 3 to obtain a psychological condition feature data set.

[0053] Specifically, the structure of the psychological condition feature dataset is as follows: [anxiety, depression, sleep, self, ...] The corresponding vector is [0.2, 0.3, 0.4, 0.2, ...].

[0054] Step S4: construct a psychological condition prediction model, and use a classifier to train the psychological feature data set to obtain the psychological condition prediction model.

[0055] Specifically, according to the data of the conversation users, the users are divided into different psychological condition outcome indicators (such as two categories of good and bad, or four categories of anxiety, depression, anxiety with depression, and normal), and then the psychological condition feature data set and the psychological condition outcome indicators are trained using any classification model (such as SVM / logistic regression / MLP / RF, etc.) to obtain a trained psychological condition prediction model.

[0056] When predicting psychological abnormalities for new users, we first use the open question dialogue webpage implemented by GPT-3.5 to obtain the user's dialogue data and image data, and then use the dialogue promoter3 built by GPT-4 to extract psychological features from the dialogue data and image data to obtain the user's psychological features. The user's psychological features are input into the trained psychological prediction model to obtain the psychological prediction results.

[0057] The present invention organically combines GPT-3.5 and GPT-4, and utilizes the high concurrent conversation of GPT-3.5 to realize the psychological data collection of conversation data and image data; and utilizes the offline analysis of GPT-4 to realize the extraction of interpretable features. , Make full use of the interpretable feature mining of the large language model. The psychological condition of the user is predicted through the trained psychological condition prediction model.

[0058] It is to be understood that the present invention is described by some embodiments, and it is known to those skilled in the art that various changes or equivalent substitutions may be made to these features and embodiments without departing from the spirit and scope of the present invention. In addition, under the teachings of the present invention, these features and embodiments may be modified to adapt to specific circumstances and materials without departing from the spirit and scope of the present invention. Therefore, the present invention is not limited by the specific embodiments disclosed herein, and all embodiments falling within the scope of the claims of this application are within the scope of protection of the present invention.

Claims

1. An interpretable psychological anomaly prediction method combining a large language model, characterized in that: The steps include: Step S1: Use the GPT-3.5 API to implement a dialogue webpage for open questions; Step S2: Use the API of GPT-4 to complete the interpretable feature extraction model and obtain an interpretable psychological feature set; Step S3: Use the GPT-4 API to complete the extraction of the psychological condition feature data set; Step S4: construct a psychological condition prediction model, and use a classifier to train the psychological feature data set to obtain the psychological condition prediction model.

2. The interpretable psychological anomaly prediction method based on a combined large language model according to claim 1, characterized in that: Step S1 includes the following steps: Step S11: Use GPT-3.5 to design a dialogue Promoter 1. The designed dialogue Promoter 1 can complete 5-10 rounds of dialogue, have open questions, and reflect certain psychological characteristics of users; Step S12: Design a webpage system whose front end can input dialogues, save multiple rounds of dialogues, and upload a picture of a house, tree, and person; Step S13: Collecting conversation data and picture data of people with mental abnormalities and people with normal mental health.

3. The interpretable psychological anomaly prediction method based on a combined large language model according to claim 2, characterized in that: Step S2 includes the following steps: Step S21: Use GPT-4 to extract the features of the conversation data and obtain the psychological features of the conversation; use GPT-4 to extract the features of the picture data and obtain the psychological features of the picture; Step S22: Count the extracted dialogue psychological features and picture psychological features to obtain N psychological features with high frequency as an interpretable psychological feature set.

4. The interpretable psychological anomaly prediction method based on a combined large language model according to claim 3, characterized in that: Step S3 includes the following steps: Step S31: constructing a dialogue promoter 3 based on the interpretable psychological feature set, extracting all psychological features one by one, and forming a specific quantitative value of [0,1]; Step S32: extract interpretable psychological features from all collected conversation data and picture data of psychologically abnormal and normal persons using conversation promoter 3 to obtain a psychological condition feature data set.

5. The interpretable psychological anomaly prediction method based on a combined large language model according to claim 4, characterized in that: When predicting psychological abnormalities for new users, we first use the open question dialogue webpage implemented by GPT-3.5 to obtain the user's dialogue data and image data, and then use the dialogue promoter3 built by GPT-4 to extract psychological features from the dialogue data and image data to obtain the user's psychological features. The user's psychological features are input into the trained psychological prediction model to obtain the psychological prediction results.

Images