User portrait construction method based on knowledge graph and psychology knowledge pushing method

Abstract

The application provides a user portrait construction method and a psychology knowledge pushing method based on a knowledge graph. In the complex knowledge graph, not only data generated by user active behaviors are used to construct an explicit user portrait of the user, but also behaviors of other users associated with the user are used to construct an implicit user portrait of the user, so that a more comprehensive and more accurate user psychological portrait is formed, and then useful personalized psychological knowledge is accurately pushed to the user, so as to help the user discover and solve possible psychological problems as early as possible and avoid developing into a psychological disease.

Description

Technical Field

[0001] This invention belongs to the field of big data and artificial intelligence technology, and specifically relates to a method for constructing user profiles based on knowledge graphs and a method for pushing psychological knowledge. Background Technology

[0002] With socio-economic development, people are increasingly valuing both physical and mental health. However, the growing pressures of study, work, and life are negatively impacting the mental well-being of many. Mental illnesses often develop gradually; by the time people realize they have a mental illness and need treatment, it has often progressed significantly, causing irreversible damage to their physical and mental health and making a complete cure difficult. Therefore, proactively guiding people to recognize and address mental health issues before they develop can prevent minor psychological problems from escalating into serious illnesses. This is not only highly beneficial to individual health but also crucial for maintaining social stability and harmony.

[0003] With the development of big data and artificial intelligence technologies, mobile applications and computer programs often collect user information and push content that users are interested in based on artificial intelligence algorithms when people use their mobile phones or computers daily. For example, mobile applications such as Douyin, Kuaishou, and Toutiao push video or news content based on user preferences; mobile applications such as JD.com, Taobao, and Pinduoduo also push product links or advertisements based on user preferences or needs. This personalized content push is generally based on user profiles. By analyzing massive amounts of user data, the data is abstracted into tags, and these tags are then used to concretize the user image, ultimately forming a user profile. However, for some people with psychological problems, obtaining their user profiles is often very difficult because they are "too lazy" to actively express their feelings and seek solutions. For example, potential patients with depression often feel that nothing is meaningful or enjoyable before developing depression. They are "too lazy" to socialize in real life, thus lacking chat data with others as the basis for their user profiles; they are also "too lazy" to use mobile applications or computer programs on the internet, and even if they do, they never actively express opinions or communicate with others, thus lacking this data as the basis for their user profiles. It is evident that current technology lacks methods to construct comprehensive and accurate user profiles for individuals who already have psychological issues but have not yet developed into mental illnesses. This group is precisely the one most in need of attention and support. Moreover, given the large number of people in this group, early intervention to prevent the development of mental illnesses is often the most effective approach. Summary of the Invention

[0004] In view of the aforementioned shortcomings of existing technologies, the purpose of this invention is to provide a user profile construction method and a psychological knowledge delivery method based on knowledge graphs. This method not only utilizes data generated by user-initiated behavior to construct explicit user profiles within a vast knowledge graph, but also constructs implicit user profiles based on the behavior of other users associated with the user. This results in a more comprehensive and accurate user psychological profile, enabling the precise delivery of useful, personalized psychological knowledge to the user. This helps users identify and alleviate existing psychological problems early, preventing them from developing into mental illnesses. It is important to note that this invention aims to identify potential psychological problems and predict the potential risk of them developing into mental illnesses by constructing user psychological profiles; it is not intended to diagnose or treat mental illnesses. For diagnosis or treatment of mental illnesses, it is still necessary to seek help from professionals or specialized institutions.

[0005] To achieve the above objectives, on the one hand, the present invention provides a method for constructing user psychological profiles based on knowledge graphs, comprising the following steps:

[0006] (1) Collect user information from mobile applications and / or computer programs to form user description text; extract user emotional attributes and attribute values ​​from the description text to form description text with emotional information.

[0007] Furthermore, the sources of user information include, but are not limited to: text from user registration information (including real-name information) in mobile applications and / or computer programs, such as name, nickname, ID number, age, mobile phone number, major, job, grade, email address, personal signature, personal tags, friend list, mobile phone address book, and other basic information; content generated by users while using mobile applications and / or computer programs, such as chat logs, comments, updates, diaries, blogs, and other text content; and text that has been connected to the user, such as text content read by the user and video content viewed by the user. It should be noted that the collection of the above user information requires authorization from the user or the official entity. This invention opposes any practice of collecting the above user information through illegal or unauthorized means.

[0008] Furthermore, in the aforementioned sources of user information, non-textual information will first be converted into text information, while preserving emotion-related parameters such as speech rate, volume, and pitch as much as possible. Then, based on the aforementioned text information (including emotion-related parameters), the user's emotional attributes and attribute values ​​will be extracted to form descriptive text with emotional information. For example, in chat logs, users may send voice messages that are audio information, or selfie videos that are video information (containing audio information). While converting audio or video information into text information, the user's emotion-related parameters will be identified, such as speech rate (e.g., words per minute, percentage above / below average speech rate), volume (e.g., decibel value, percentage above / below average volume), and pitch (e.g., high-level, mid-level, low-level, rising, falling pitch), etc., which can reflect emotions. Converting audio or video information into text information using speech recognition technology is a very mature existing technology with high recognition efficiency and accuracy. Extracting emotional attributes from text information is also an existing technology, which can be achieved by most existing general artificial intelligence models, such as OpenAI's ChatGPT, Microsoft's Bing AI, Anthropic's Claude, xAI's Grok, Meta's Llama, ByteDance's Doubao, Baidu's Wenxin Yiyan, Alibaba's Tongyi Qianwen, Huawei's Pangu Big Model, Tencent's Hunyuan Big Model, iFlytek's Xinghuo Cognitive Big Model, and Kimi from Lunar Dark Side.

[0009] Meanwhile, existing technologies also offer many methods for identifying user emotions through audio, such as Chinese patent applications CN109935240A and CN118571265A. Using these technologies, audio or video information (including audio) does not need to be converted into text information first; instead, the user's emotional attributes and values ​​can be directly extracted from the audio information. Alternatively, existing technologies also offer many methods for directly identifying user emotions through video, such as Chinese patent applications CN108491764A, CN116453024A, and CN112492397A. Using these technologies, video information does not need to be converted into audio or text information first; instead, the user's emotional attributes and values ​​can be directly extracted from the video information. Adding these emotional attributes and values ​​to the user's descriptive text creates descriptive text with emotional information.

[0010] Furthermore, the aforementioned emotional attributes include happiness, anger, sadness, surprise, fear, disgust, jealousy, shyness, guilt, anxiety, and frustration. The attribute values ​​represent the degree of these emotions and can be assigned different values ​​according to the degree, such as 0 to 10, with larger values ​​indicating a deeper degree.

[0011] For example, among the emotions of "sadness," "grief" has a higher emotional intensity than "sadness." Therefore, the attribute value of "grief" is higher than that of "sadness." For instance, the attribute of the word "sadness" is grief, with an attribute value of 5; while the attribute of the word "sadness" is also grief, but its attribute value is 7.

[0012] Alternatively, the value of the emotional attribute can be changed by using emphasizing or weakening words. For example, in Chinese, adverbs such as "very," "super," "extremely," and "…died" are emphasizing words, which strengthen the attribute value; while words like "a little," "somewhat," and "slightly" are weakening words, which weaken the attribute value. For example, the attribute of the word "sad" is grief, with a value of 5; the attribute of the word "sad to death" is adjusted to 6 because of the emphasizing word "…died"; and the attribute of the word "a little sad" is adjusted to 4 because of the weakening word "a little."

[0013] Alternatively, parameters that reflect emotions, such as speech rate, volume, and pitch, can also alter the assigned value of emotional attributes. For example, the attribute of the word "happy" is happiness, with a value of 5. If a user increases their volume and pitch when saying "happy," the attribute value can be adjusted to 6 or 7 based on the degree of increase in volume and pitch.

[0014] (2) Extract information from the descriptive text containing emotional information to form a triple in the form of “entity-relationship-entity” or “entity-attribute-attribute value”. The information extraction includes entity recognition, relation recognition and attribute recognition. The extracted information is stored in a relational database.

[0015] Furthermore, the aforementioned entity recognition, relationship recognition, and attribute recognition can already be achieved using mature artificial intelligence algorithms in the existing technology. For example, Chinese patent application CN115982379A employs a method combining an improved convolutional neural network and an improved conditional random field model. The descriptive text is input into the improved convolutional neural network to obtain a feature map, which is then input into the improved conditional random field model for sequence labeling, used to identify entities of interest in the descriptive text.

[0016] Furthermore, the aforementioned relational database is selected from one or more relational databases such as MySQL, Oracle Database, Microsoft SQL Server, PostgreSQL, IBM DB2, MariaDB, SQLite, and Informix, with MySQL being the preferred choice.

[0017] Furthermore, since the descriptive text containing emotional information already includes emotional attributes and attribute values, the "entity-attribute-attribute value" triple representing the user's emotion can be directly extracted, i.e., the "entity-emotional attribute-attribute value" triple, such as "Zhang San-frustrated-5".

[0018] Furthermore, for the triple "entity-attribute-attribute value", the attribute value also includes time information, that is, the generation time of the text information extracted from the triple. For example, "Li Si-Depressed-5 (202412011505)" indicates that the text information was generated at 15:05 on December 1, 2024. Since psychological states and mental illnesses are dynamic, recording the time information corresponding to psychological states is crucial for identifying potential mental illnesses. For example, if a user exhibits potential depressive symptoms for a certain period, their user profile might be labeled "potentially depressed"; however, if these symptoms disappear after a while, the "potentially depressed" label in their user profile might be removed or its probability reduced. Another example is bipolar disorder, a mental disorder characterized by both manic and depressive episodes. During a manic episode, patients may exhibit symptoms such as elevated mood, high energy, and increased activity; while during a depressive episode, they may experience symptoms such as depressed mood, loss of interest, reduced activity, and self-blame and guilt. Therefore, potential patients with bipolar disorder may experience these diametrically opposed symptoms a considerable period before the onset of the disorder, which is closely related to time.

[0019] Furthermore, based on the user's "entity-emotion attribute-attribute value" triple containing time information, an emotion map of that user over a period of time (e.g., 1 minute, 1 hour, 1 day, 1 week, 1 month, 1 year, etc.) can be drawn. In the emotion map, different emotion attributes are marked with different colors; the higher the attribute value of a certain emotion attribute, the darker the corresponding color. The emotion map can intuitively and visually reflect a user's emotions over a period of time (images of the emotion map) or emotional changes (videos formed by multiple emotion map images). Existing artificial intelligence image recognition models can already effectively identify and compare these images and videos, such as deep convolutional neural networks with Residual Network (ResNet) architectures, such as the ResNet-50 network structure.

[0020] (3) Perform knowledge fusion on the triples, resolve the coreference of different triples about the same entity from multiple sources to map them to the correct entity, and disambiguate the same-name triples representing different entities to resolve the ambiguity generated by the same-name triples.

[0021] Furthermore, the aforementioned knowledge fusion can already be achieved using mature artificial intelligence algorithms in existing technologies. For example, unsupervised clustering methods based on encyclopedic knowledge can be used to resolve ambiguities arising from triples with the same name.

[0022] (4) The merged triples are processed to form a structured and networked knowledge system; the processed triples are stored in a graph database to form a primary knowledge graph.

[0023] Furthermore, the knowledge processing includes ontology construction, knowledge reasoning, and quality assessment. Specifically, ontology construction is based on a shared understanding of a domain, extracting a general vocabulary; knowledge reasoning acquires new knowledge or conclusions through various methods; and quality assessment quantifies the credibility of knowledge, ensuring the quality of the knowledge base by discarding knowledge with low confidence.

[0024] Furthermore, the graph database mentioned above is selected from one or more graph databases such as Neo4j, JanusGraph, OrientDB, ArangoDB, Titan, Virtuoso, Stardog, TigerGraph, AllegroGraph, Amazon Neptune, HugeGraph, and GeaBase, with Neo4j being the preferred database.

[0025] (5) Based on the primary knowledge graph, based on a certain person entity (i.e., user), extract the triples related to the psychological state generated by its active behavior to form an explicit knowledge graph about the person entity; extract the triples related to the psychological state of the person entity from the triples of other person entities associated with it to form an implicit knowledge graph about the person entity.

[0026] Furthermore, the aforementioned triplets related to psychological states generated by proactive behavior refer to triplets of keywords related to psychological states that are actively expressed, described, or evaluated by the individual entity itself.

[0027] Furthermore, the triples related to the mental state of the aforementioned other entities refer to the triples of keywords related to the mental state of the entity expressed, described, or evaluated by other people (who have a relationship with the entity that is close or distant, direct or indirect).

[0028] Furthermore, the explicit knowledge graph described above is like the trunk and branches of a tree, above ground, representing a person's subjective expression, description, or evaluation of their own psychological state. The implicit knowledge graph, on the other hand, is like the roots of this tree, below ground, representing the objective expression, description, or evaluation of the psychological state of the person (those with whom the person has a relationship, whether close or distant, direct or indirect). Because everyone's self-perception differs in reality, a person's subjective self-evaluation is often one-sided, even erroneous, and frequently differs from the evaluations made by others. Moreover, due to the complexity and multifaceted nature of human nature, a person often exhibits drastically different states in front of different groups of people; therefore, the evaluations of a person by others of varying degrees of closeness or distance are often different, even contradictory. However, existing psychological knowledge graph technologies do not distinguish between subjective and objective evaluations of psychological states, therefore, the user profiles constructed based on them are often incomplete and inaccurate.

[0029] (6) Based on the explicit knowledge graph, explicit user features are mined, and an explicit user profile of the person entity is constructed according to the explicit user features; based on the implicit knowledge graph, implicit user features are mined, and an implicit user profile of the person entity is constructed according to the implicit user features; the explicit user profile and the implicit user profile together constitute a user psychological profile of the person entity, and the user psychological profile includes user psychological tags.

[0030] Furthermore, regarding how to mine user characteristics and build user profiles through knowledge graphs, there are already mature artificial intelligence algorithms in existing technologies, such as Chinese patent application CN115982379A.

[0031] Furthermore, each user's psychological label includes a corresponding probability value, representing the probability that the user may potentially suffer from the corresponding mental illness. Please note that this invention aims to identify potential psychological problems by constructing a complete user psychological profile, and to predict and prevent the potential risk of these problems developing into mental illnesses; it is not intended to diagnose mental illnesses. For diagnosis or treatment of mental illnesses, it is still necessary to seek help from professionals or specialized institutions.

[0032] Furthermore, the aforementioned user psychological labels (including probability values) can be obtained by an artificial intelligence image recognition model to identify and compare the user's emotion map. For example, this invention also provides a method for identifying and comparing a user's emotion map using a deep convolutional neural network with a deep residual network architecture to obtain user psychological labels, comprising the following steps:

[0033] A. Users diagnosed with various mental illnesses (diagnosed by professionals or professional institutions) are used as sample users. Emotional maps (including time information) of the sample users are obtained before and after the onset of the illness (e.g., 2 years before the onset of the illness and 1 year after the diagnosis) as the sample dataset.

[0034] B. Establish the backbone network of the artificial intelligence data analysis module, and pre-train the backbone network of the artificial intelligence data analysis module using the sample dataset;

[0035] C. Transfer the weights of the backbone network of the pre-trained artificial intelligence data analysis module to the analysis network of the artificial intelligence data analysis module, and iteratively train the analysis network on the sample dataset;

[0036] D. Obtain the emotion map of the user to be analyzed over a period of time. Use the emotion map of the user to be analyzed (including time information) as the input of the analysis network obtained by iterative training in step D to obtain the psychological label of the user to be analyzed, that is, the probability that the user to be analyzed may suffer from various mental illnesses. In essence, it is the similarity between the psychological map of the user to be analyzed and the psychological map of the sample users.

[0037] On the other hand, the present invention also provides a method for pushing psychological knowledge based on the above-mentioned user psychological profiles (including explicit user profiles and implicit user profiles), including the following steps:

[0038] S1: Extract explicit and implicit psychological tags related to psychological state from the explicit and implicit user profiles of the same user, and compare whether there is a contradiction between the explicit and implicit psychological tags: if there is no contradiction, push psychological knowledge related to the explicit psychological tag with the highest probability and the implicit psychological tag with the highest probability; if there is a contradiction, proceed to S2.

[0039] Furthermore, the existence of contradictions between the aforementioned psychological labels refers to whether the mental illnesses corresponding to these labels have completely opposite and incompatible symptoms. If so, it means that the two mental illnesses are generally contradictory and incompatible, and the user may have schizophrenia. For example, patients with histrionic personality disorder often exhibit excessive emotionality and drama in social situations; while patients with schizoid personality disorder exhibit extreme coldness and aloofness in social situations. The symptoms of the two are opposite and incompatible. If a user's explicit psychological label is "hitrionic personality disorder," but their implicit psychological label is "schizoid personality disorder," then the user may have schizophrenia.

[0040] S2: Push psychological knowledge about schizophrenia to the user and other users associated with the user, and compare the probabilities of psychological tags in the user's explicit user profile and implicit user profile. If the probability of the most probable explicit psychological tag in the explicit user profile is much greater than the probability of the most probable implicit psychological tag in the implicit user profile, then push psychological knowledge related to the most probable explicit psychological tag to the user. If the probability of the most probable implicit psychological tag in the implicit user profile is much greater than the probability of the most probable explicit psychological tag in the explicit user profile, then push psychological knowledge related to the most probable implicit psychological tag to the user. If the probability of the most probable explicit psychological tag in the explicit user profile is equal to or the same as the probability of the most probable implicit psychological tag in the implicit user profile, then push both the most probable explicit psychological tag and the most probable implicit psychological tag to the user.

[0041] Furthermore, "significantly greater than" means that the probability (higher) of a psychological label in one user profile is twice or more than the probability (lower) of a psychological label in another user profile. "Equal probabilities" means that the probability (higher) of a psychological label in one user profile is one to two times (excluding twice) the probability (lower) of a psychological label in another user profile.

[0042] S3: If the number and / or probability of the user's explicit and implicit psychological tags are both 0, then push psychological knowledge about mental illnesses such as depression and bipolar disorder (especially during depressive episodes) to the user and other users associated with the user.

[0043] The user profile construction method and psychology knowledge push method based on knowledge graph of the present invention have at least the following beneficial technical effects:

[0044] (1) This invention innovatively collects user information from mobile applications and / or computer programs to form user descriptive text, extracts user emotional attributes and attribute values ​​from it to form descriptive text with emotional information, and then draws user emotional maps; and uses a deep convolutional neural network with residual network architecture to identify and compare user emotional maps, thereby obtaining user psychological labels with probability values, that is, the probability that the user may have various mental illnesses, and realizes quantitative prediction of the user's potential disease possibility.

[0045] (2) Based on the complex primary knowledge graph, not only is the explicit user profile of the user constructed using the data generated by the user's active behavior, but also the implicit user profile of the user is constructed based on the behavior of other users associated with the user, thereby forming a more comprehensive and accurate user psychological profile, and thus enabling the precise delivery of useful and personalized psychological knowledge to the user to help prevent and solve the user's psychological problems and avoid them from developing into mental illnesses.

[0046] (3) Extract explicit and implicit psychological tags from the explicit user profile and implicit user profile of the same user respectively, and determine whether there is a contradiction between them and which psychological tag is dominant, so as to accurately push the most useful personalized psychological knowledge to the user.

[0047] (4) For “invisible” users whose number and / or probability of explicit and implicit psychological tags are both 0, a corresponding psychological knowledge push program is also provided to help them to the greatest extent. Attached Figure Description

[0048] Figure 1 This is an example of chat history generated by a user while using a mobile application, as described in a preferred embodiment of the present invention.

[0049] Figure 2 This is a preferred embodiment of the present invention, showing a user's mood map over a 7-hour period.

[0050] Figure 3 This is a preferred embodiment of the present invention, showing a user's mood map over 7 days.

[0051] Figure 4 This is an example of a primary knowledge graph formed in a preferred embodiment of the present invention;

[0052] Figure 5 This is a schematic diagram of the explicit knowledge graph and the implicit knowledge graph formed in a preferred embodiment of the present invention;

[0053] Figure 6 This is a flowchart illustrating a preferred embodiment of the user psychological profile construction method based on knowledge graphs according to the present invention.

[0054] Figure 7 This is a flowchart illustrating a preferred embodiment of the psychological knowledge push method based on user profiles according to the present invention. Detailed Implementation

[0055] The embodiments of the present invention will be described in detail below. The embodiments described below are implemented based on the technical solution of the present invention, and detailed implementation methods and specific operation processes are given. However, the protection scope of the present invention is not limited to the embodiments described below.

[0056] like Figure 4 As shown, in a preferred embodiment, the present invention provides a method for constructing user psychological profiles based on knowledge graphs, including the following steps:

[0057] (1) Collect user information from mobile applications and / or computer programs to form user description text; extract user emotional attributes and attribute values ​​from the description text to form description text with emotional information.

[0058] The sources of user information include, but are not limited to: text in user registration information (including real-name information) for mobile applications and / or computer programs, such as name, nickname, ID number, mobile phone number, major, job, age, grade, email address, personal signature, personal tags, friend list, mobile phone address book and other basic information; content generated by users while using mobile applications and / or computer programs, such as chat history, comments, status updates, diaries, blogs and other text content; and text that has been connected to the user, such as text content that the user has read and video content that the user has viewed.

[0059] In the aforementioned user information sources, non-text information will first be converted into text information, while preserving emotion-related parameters such as speech rate, volume, and pitch as much as possible. Then, based on the aforementioned text information (including emotion-related parameters), the user's emotional attributes and attribute values ​​will be extracted to form descriptive text with emotional information. For example, in chat logs, users may send voice messages that are audio information, or selfie videos that are video information (including audio information). While converting audio or video information into text information, the user's emotions will be identified or parameters that reflect emotions will be preserved, such as speech rate (e.g., words per minute, percentage above / below average speech rate), volume (e.g., decibel value, percentage above / below average volume), and pitch (e.g., high-level, mid-level, low-level, rising, falling pitch). Converting audio or video information into text information using speech recognition technology is a very mature existing technology with high recognition efficiency and accuracy. Extracting emotional attributes from text information is also an existing technology, which can be achieved by most existing general artificial intelligence models, such as OpenAI's ChatGPT, Microsoft's Bing AI, Anthropic's Claude, xAI's Grok, Meta's Llama, ByteDance's Doubao, Baidu's Wenxin Yiyan, Alibaba's Tongyi Qianwen, Huawei's Pangu Big Model, Tencent's Hunyuan Big Model, iFlytek's Xinghuo Cognitive Big Model, and Kimi from Lunar Dark Side.

[0060] The aforementioned emotional attributes include happiness, anger, sadness, surprise, fear, disgust, jealousy, shyness, guilt, anxiety, and frustration. The attribute value represents the degree of these emotions, and different values ​​can be assigned according to the degree, such as 0 to 10, with a larger value indicating a deeper degree.

[0061] For example, among the emotions of "sadness," "grief" has a higher emotional intensity than "sadness." Therefore, the attribute value of "grief" is higher than that of "sadness." For instance, the attribute of the word "sadness" is grief, with an attribute value of 5; while the attribute of the word "sadness" is also grief, but its attribute value is 7.

[0062] Alternatively, the value of the emotional attribute can be changed by using emphasizing or weakening words. For example, in Chinese, adverbs such as "very," "super," "extremely," and "…died" are emphasizing words, which strengthen the attribute value; while words like "a little," "somewhat," and "slightly" are weakening words, which weaken the attribute value. For example, the attribute of the word "sad" is grief, with a value of 5; the attribute of the word "sad to death" is adjusted to 6 because of the emphasizing word "…died"; and the attribute of the word "a little sad" is adjusted to 4 because of the weakening word "a little."

[0063] Alternatively, parameters that reflect emotions, such as speech rate, volume, and pitch, can also alter the assigned value of emotional attributes. For example, the attribute of the word "happy" is happiness, with a value of 5. If a user increases their volume and pitch when saying "happy," the attribute value can be adjusted to 6 or 7 based on the degree of increase in volume and pitch.

[0064] In a preferred embodiment, such as Figure 1 As shown, user Li Si generated a chat log while using a mobile application, which included both text and voice messages. The voice messages were first converted into text (or saved with emotion-related parameters) through speech recognition. Then, the emotional attributes within the text messages were identified and assigned values. For example, in Li Si's message "Mom, I'm so annoyed lately," "annoyed" could be assigned the emotional attributes of anxiety, frustration, and anger, with values ​​of 5, 3, and 1 respectively. Therefore, the text message could be represented as "Mom, I'm so annoyed lately (anxiety 5, frustration 3, anger 1)." Emotions are often a direct reflection of a user's psychological state and can be used to help determine if a user has psychological problems. Therefore, a large amount of emotion-related user data can help assess the likelihood of a user potentially suffering from various mental illnesses.

[0065] (2) Extract information from the descriptive text containing emotional information to form a triple in the form of “entity-relationship-entity” or “entity-attribute-attribute value”. The information extraction includes entity recognition, relation recognition and attribute recognition. The extracted information is stored in a relational database.

[0066] The aforementioned entity recognition, relationship recognition, and attribute recognition can already be achieved using mature artificial intelligence algorithms in existing technologies. For example, Chinese patent application CN115982379A employs a method combining an improved convolutional neural network and an improved conditional random field model. The descriptive text is input into the improved convolutional neural network to obtain a feature map, which is then input into the improved conditional random field model for sequence labeling, used to identify entities of interest in the descriptive text.

[0067] The relational databases mentioned above are selected from one or more of the following relational databases: MySQL, Oracle Database, Microsoft SQL Server, PostgreSQL, IBM DB2, MariaDB, SQLite, Informix, etc., with MySQL being the preferred choice.

[0068] For example, regarding the above Figure 1 Information can be extracted from chat logs to form multiple "entity-relationship-entity" triples such as "Zhang San-mother-son-Li Si", "Li Si-lover-Wang Wu", and "Zhang San-neighbor-Zhao Liu"; it can also form multiple "entity-attribute-attribute value" triples such as "Zhang San-surprise-1", "Li Si-anxiety-5", "Li Si-frustration-3", and "Li Si-anger-1".

[0069] For the "entity-attribute-attribute value" triple, the attribute value also includes time information. For example, the attribute value of the triple "Zhang San-surprise-1" above also includes time information, that is, the generation time of the text information extracted from the triple, which is the sending time of the sentence "Si Zai, why haven't you video chatted with your parents lately?" If the sending time is 18:06 on September 28, 2024, then the triple "Zhang San-surprise-1" above, after including the time information, can be saved as "Zhang San-surprise-1(202409281806)".

[0070] Because psychological states and mental illnesses are dynamic, recording the time information corresponding to psychological states is crucial for identifying potential mental illnesses. For example, if a user exhibits potential depressive symptoms for a certain period, their user profile might be labeled "potentially depressed." However, if these symptoms disappear after some time, the "potentially depressed" label might be removed from their user profile or its probability reduced. Another example is bipolar disorder, a mental disorder characterized by both manic and depressive episodes. During a manic episode, patients exhibit elevated mood, increased energy, and increased activity; while during a depressive episode, they experience depressed mood, decreased interest, reduced activity, and feelings of guilt and self-blame. Therefore, potential bipolar disorder patients may experience these contrasting symptoms for a considerable period before developing the disorder, demonstrating a close correlation with time.

[0071] Based on the user's "entity-emotion attribute-attribute value" triple containing time information, an emotion map of that user over a period of time (e.g., 1 minute, 1 hour, 1 day, 1 week, 1 month, 1 year, etc.) can be drawn. In the emotion map, different emotion attributes are marked with different colors; the higher the attribute value, the darker the corresponding color. Emotion maps can intuitively and visually reflect a user's emotions over a period of time (images of the emotion map) or emotional changes (videos formed by multiple emotion map images). Existing artificial intelligence image recognition models can already identify and compare these images and videos quite well, such as deep convolutional neural networks with deep residual network architectures, such as the ResNet-50 network structure. For example, Figure 2 and Figure 3 These are the sentiment maps for a user over 7 hours (12:00-18:00 on December 1, 2024) and 7 days (December 1-7, 2024), respectively. A larger-scale sentiment map can be obtained by summing multiple smaller-scale sentiment maps.

[0072] (3) Perform knowledge fusion on the triples, resolve the coreference of different triples about the same entity from multiple sources to map them to the correct entity, and disambiguate the same-name triples representing different entities to resolve the ambiguity generated by the same-name triples.

[0073] The aforementioned knowledge fusion can already be achieved using mature artificial intelligence algorithms in existing technologies. For example, unsupervised clustering methods based on encyclopedic knowledge can be used to resolve ambiguities arising from triples with the same name.

[0074] For example, coreference resolution is performed on the triples extracted from the chat record, which can identify that "Sz Tsai" and "Li Si" co-refer to the same entity, so they can be mapped to the same entity Li Si; "Wang Wu" and "Xiao Wang" co-refer to the same entity, so they can be mapped to the same entity Wang Wu.

[0075] (4) Perform knowledge processing on the fused triples to form a structured and networked knowledge system; use a graph database to store the triples after knowledge processing to form a primary knowledge graph.

[0076] The knowledge processing includes ontology construction, knowledge reasoning, and quality assessment. Among them, the ontology construction extracts a general term based on the common understanding of a field; the knowledge reasoning obtains new knowledge or conclusions through various methods; the quality assessment quantifies the credibility of the knowledge and ensures the quality of the knowledge base by discarding the knowledge with lower confidence.

[0077] The above graph database is selected from one or several of the graph databases such as Neo4j, JanusGraph, OrientDB, ArangoDB, Titan, Virtuoso, Stardog, TigerGraph, AllegroGraph, Amazon Neptune, HugeGraph, GeaBase, etc., preferably the Neo4j database.

[0078] Figure 4 is part of the primary knowledge graph formed by Figure 1 the chat record. Among them, the "insomnia", "difficulty waking up", "poor appetite", "exercise inertia", etc. shown by Li Si belong to health status or physical symptoms. These physical symptoms may be related to physical or mental diseases and can be used to assist in judging whether the user may have psychological problems or a potential tendency to develop mental diseases. What Li Si said, "unlucky", is a self-evaluation of his own state, which can also reflect the user's psychological state. Therefore, it can also be used to assist in judging whether the user has psychological problems or a potential tendency to develop mental diseases. The "depression tendency" shown by Zhao Liu is an evaluation of others made in the conversation between Zhang San and Li Si, and can also be used to assist in judging whether Zhao Liu has psychological problems or a potential tendency to develop mental diseases.

[0079] (5) Based on the primary knowledge graph, for a certain person entity (i.e., the user), extract the triples related to the mental state generated by his active behavior to form an explicit knowledge graph about the person entity; extract the triples related to the mental state of the person entity from the triples of other person entities associated with it to form an implicit knowledge graph about the person entity.

[0080] The aforementioned triplets related to psychological states generated by proactive behavior refer to triplets of keywords related to psychological states that are actively expressed, described, or evaluated by the individual entity itself. For example, the sentence "Mom, I'm so annoyed lately" is Li Si's own proactive emotional expression, where "annoyed" is a keyword related to psychological state. Another example is "Why am I so unlucky lately?", which is a self-evaluation of one's state and can also reflect the user's psychological state, where "unlucky" is a keyword related to psychological state.

[0081] In the aforementioned triplets of other related entities, the triplets related to the psychological state of that entity refer to triplets containing keywords related to the psychological state expressed, described, or evaluated by others (who have a relationship with that entity that is close or distant, direct or indirect). For example, the statement "Sizi, Mom knows you're under a lot of pressure" is said directly by Zhang San to Li Si; "under a lot of pressure" is Zhang San's evaluation of Li Si, which is a keyword related to the psychological state. Another example is the statement "Almost depressed" which is Zhang San's evaluation of Zhao Liu in their conversation; "depressed" is Zhang San's evaluation of Zhao Liu from another person's perspective, which is also a keyword related to the psychological state.

[0082] by Figure 5 Using a tree analogy, the explicit knowledge graph 1 is like the trunk and branches of a tree, above ground, representing a person's subjective expression, description, or evaluation of their own psychological state. The implicit knowledge graph 2, below ground, represents the roots of the tree, expressing the objective expression, description, or evaluation of the psychological state of the person (those with whom the person has a relationship, whether close or distant, direct or indirect). Because everyone's self-perception differs in reality, a person's subjective self-evaluation is often one-sided, even erroneous, and frequently differs from the evaluations made by others. Furthermore, due to the complexity and multifaceted nature of human nature, a person often displays drastically different states in front of different groups of people; therefore, the evaluations of a person by others of varying degrees of closeness or distance are often different, even contradictory. However, existing psychological knowledge graphs do not distinguish between subjective and objective evaluations of psychological states, therefore, the user profiles constructed based on them are often incomplete and inaccurate.

[0083] (6) Based on the explicit knowledge graph, explicit user features are mined, and an explicit user profile of the person entity is constructed based on the explicit user features; based on the implicit knowledge graph, implicit user features are mined, and an implicit user profile of the person entity is constructed based on the implicit user features, wherein the user psychological profile includes user psychological tags.

[0084] Regarding how to mine user characteristics and build user profiles through knowledge graphs, there are already mature artificial intelligence algorithms in existing technologies, such as Chinese patent application CN115982379A.

[0085] For example, an AI algorithm can assign explicit psychological tags such as "anxiety disorder," "dysthymia," and "depression" to Li Si's explicit user profile based on keywords, emotional attributes, and attribute values ​​related to psychological states such as "insomnia," "difficulty waking up," "poor appetite," "motor inertia," "depression," "bad luck," and "breakup tendency," etc. Similarly, an AI algorithm can assign implicit psychological tags such as "depression" to Zhao Liu's implicit user profile based on keywords related to psychological states such as "depressive tendency," etc. These psychological tags are the result of qualitative analysis of users based on the aforementioned keywords in the knowledge graph.

[0086] In addition to qualitative analysis, each user psychological tag in this invention also includes a corresponding probability value, i.e., quantitative analysis, representing the probability that the user may potentially suffer from the corresponding mental illness. The user psychological tag containing the probability value can be obtained by an artificial intelligence image recognition model identifying and comparing the user's emotion map. For example, using a deep convolutional neural network with a deep residual network architecture to identify and compare the user's emotion map to obtain the user psychological tag containing the probability value includes the following steps:

[0087] A. Users diagnosed with various mental illnesses (diagnosed by professionals or professional institutions) are used as sample users. Emotional maps (including time information) of the sample users are obtained before and after the onset of the illness (e.g., 2 years before the onset of the illness and 1 year after the diagnosis) as the sample dataset.

[0088] B. Establish the backbone network of the artificial intelligence data analysis module, and pre-train the backbone network of the artificial intelligence data analysis module using the sample dataset;

[0089] C. Transfer the weights of the backbone network of the pre-trained artificial intelligence data analysis module to the analysis network of the artificial intelligence data analysis module, and iteratively train the analysis network on the sample dataset;

[0090] D. Obtain the emotion map of the user to be analyzed over a period of time. Use the emotion map of the user to be analyzed (including time information) as input to the analysis network obtained through iterative training in step D to obtain the psychological label of the user to be analyzed, that is, the probability that the user to be analyzed may suffer from various mental illnesses. In essence, it is the similarity between the psychological map of the user to be analyzed and the psychological maps of the sample users. For example, using Li Si's emotion map over a period of time as input to the analysis network obtained through iterative training, Li Si's psychological labels are "anxiety disorder (31%)", "dysthymia disorder (17%)", and "depression (8%)", indicating that the potential probabilities of Li Si suffering from anxiety disorder, dysthymia disorder, and depression are 31%, 17%, and 8%, respectively.

[0091] like Figure 7 As shown, in a preferred embodiment, the present invention also provides a method for pushing psychological knowledge based on the above-mentioned user psychological profiles (including explicit user profiles and implicit user profiles), including the following steps:

[0092] S1: Extract explicit and implicit psychological tags related to psychological state from the explicit and implicit user profiles of the same user, and compare whether there is a contradiction between the explicit and implicit psychological tags: if there is no contradiction, push psychological knowledge related to the explicit psychological tag with the highest probability and the implicit psychological tag with the highest probability; if there is a contradiction, proceed to S2.

[0093] The existence of contradictions between the aforementioned psychological labels refers to whether the mental illnesses corresponding to these labels have completely opposite and incompatible symptoms. If so, it means the two mental illnesses are generally contradictory and incompatible, and the user may have schizophrenia. For example, patients with histrionic personality disorder often exhibit excessive emotionality and drama in social situations; while patients with schizoid personality disorder exhibit extreme coldness and aloofness in social situations. The symptoms are opposite and incompatible. If a user's explicit psychological label is "hitrionic personality disorder," but their implicit psychological label is "schizoid personality disorder," then that user may have a potential risk of developing schizophrenia.

[0094] S2: Push psychological knowledge about schizophrenia to the user and other users associated with the user, and compare the probabilities of psychological tags in the user's explicit user profile and implicit user profile. If the probability of the most probable explicit psychological tag in the explicit user profile is much greater than the probability of the most probable implicit psychological tag in the implicit user profile, then push psychological knowledge related to the most probable explicit psychological tag to the user. If the probability of the most probable implicit psychological tag in the implicit user profile is much greater than the probability of the most probable explicit psychological tag in the explicit user profile, then push psychological knowledge related to the most probable implicit psychological tag to the user. If the probability of the most probable explicit psychological tag in the explicit user profile is equal to or the same as the probability of the most probable implicit psychological tag in the implicit user profile, then push both the most probable explicit psychological tag and the most probable implicit psychological tag to the user.

[0095] The term "significantly greater than" means that the probability of a psychological label in one user profile (higher) is twice or more than the probability of a psychological label in another user profile (lower). The term "comparable probabilities" means that the probability of a psychological label in one user profile (higher) is one to two times (excluding twice) the probability of a psychological label in another user profile (lower).

[0096] For example, if the most probable psychological label in a user's explicit user profile is "anxiety disorder (30%)", while the most probable psychological label in their implicit user profile is "dysthymia disorder (10%)", then psychological knowledge related to "anxiety disorder" will be pushed to that user.

[0097] S3: If the number and / or probability of the most probable explicit psychological tag and implicit psychological tag of a certain user are both 0, then push psychological knowledge about mental illnesses such as depression and bipolar disorder (especially during depressive episodes) to that user and other users associated with that user.

[0098] The preferred embodiments of the present invention have been described in detail above. It should be understood that those skilled in the art can make numerous modifications and variations based on the concept of the present invention without creative effort. Therefore, all technical solutions that can be obtained by those skilled in the art based on the concept of the present invention through logical analysis, reasoning, or limited experimentation on the basis of existing technology should be within the scope of protection defined by the claims.

Claims

1. A method for constructing user psychological profiles based on knowledge graphs, characterized in that, Includes the following steps: (1) Collect user information from mobile applications and / or computer programs to form user description text; extract the user's emotional attributes and attribute values ​​from the description text to form description text with emotional information; (2) Extract information from the descriptive text containing emotional information to form a triple in the form of "entity-relationship-entity" or "entity-attribute-attribute value". The information extraction includes entity recognition, relation recognition and attribute recognition. The extracted information is stored in a relational database. Based on the user's "entity-emotional attribute-attribute value" triple containing time information, draw an emotion map of the user over a period of time. In the emotion map, different emotional attributes are marked with different colors. The larger the attribute value of a certain emotional attribute, the darker the corresponding color. (3) Perform knowledge fusion on the triples, combining different triples about the same entity from multiple sources. The group performs coreference resolution to map to the correct entity, and disambiguates triples with the same name that represent different entities to resolve ambiguities caused by triples with the same name. (4) Process the knowledge of the merged triples to form a structured and networked knowledge system; A graph database is used to store the processed triples of knowledge, forming a primary knowledge graph. (5) Based on the primary knowledge graph, extract the triples related to the psychological state generated by the active behavior of a certain person entity to form an explicit knowledge graph about the person entity; extract the triples related to the psychological state of the person entity from the triples of other person entities associated with it to form an implicit knowledge graph about the person entity. (6) Based on the explicit knowledge graph, mine explicit user features and construct an explicit user profile of the person entity based on the explicit user features; Based on the implicit knowledge graph, implicit user characteristics are mined, and an implicit user profile of the person entity is constructed according to the implicit user characteristics. The explicit user profile and the implicit user profile together constitute a user psychological profile of the person entity. The user psychological profile includes user psychological tags, which are obtained by an artificial intelligence image recognition model to identify and compare the user's emotion map.

2. The user psychological profile construction method based on knowledge graph according to claim 1, characterized in that, The sources of user information in step (1) include: text in the user registration information of mobile applications and / or computer programs; content generated by the user during the use of mobile applications and / or computer programs; text that has been connected with the user; among the sources of user information, non-text information will first be converted into text information and retain parameters related to emotion; then, the user's emotional attributes and attribute values ​​will be extracted according to the descriptive text to form the descriptive text with emotional information.

3. The user psychological profile construction method based on knowledge graph according to claim 2, characterized in that, The user's emotional attributes and attribute values ​​are extracted from the descriptive text and implemented through a general artificial intelligence model; the general artificial intelligence model is selected from one or more of ChatGPT, Bing AI, Claude, Grok, Llama, Doubao, Wenxin Yiyan, Tongyi Qianwen, Pangu Big Model, Hunyuan Big Model, Xinghuo Cognitive Big Model, and Kimi.

4. The user psychological profile construction method based on knowledge graph according to claim 1, characterized in that, The relational database mentioned in step (2) is selected from one or more of the following relational databases: MySQL, Oracle Database, Microsoft SQL Server, PostgreSQL, IBM DB2, MariaDB, SQLite, and Informix.

5. The user psychological profile construction method based on knowledge graph according to claim 1, characterized in that, The attribute value of the triple "entity-attribute-attribute value" mentioned in step (2) also includes time information, that is, the generation time of the text information extracted from the triple.

6. The user psychological profile construction method based on knowledge graph according to claim 1, characterized in that, The graph database mentioned in step (4) is selected from one or more of the following graph databases: Neo4j, JanusGraph, OrientDB, ArangoDB, Titan, Virtuoso, Stardog, TigerGraph, AllegroGraph, Amazon Neptune, HugeGraph, and GeaBase.

7. The user psychological profile construction method based on knowledge graph according to claim 1, characterized in that, The triples related to psychological state generated by the active behavior mentioned in step (5) refer to the triples of keywords related to psychological state that are actively expressed, described or evaluated by the person entity itself; the triples of other persons associated with it that are related to psychological state refer to the triples of keywords related to psychological state that are expressed, described or evaluated by other persons who have a relationship with the person entity.

8. The user psychological profile construction method based on knowledge graph according to claim 1, characterized in that, The user psychological label mentioned in step (6) contains a corresponding probability value, which represents the probability that the user may potentially suffer from the corresponding mental illness.

9. A method for pushing psychological knowledge based on user psychological profiles according to any one of claims 1-8, characterized in that, Includes the following steps: S1: Extract explicit and implicit psychological tags related to psychological state from the explicit and implicit user profiles of the same user, and compare whether there is a contradiction between the explicit and implicit psychological tags: if there is no contradiction, push psychological knowledge related to the explicit psychological tag with the highest probability and the implicit psychological tag with the highest probability; if there is a contradiction, proceed to S2. S2: Push psychological knowledge about schizophrenia to the user and other users associated with the user, and compare the probabilities of psychological tags in the user's explicit user profile and implicit user profile. If the probability of the most probable explicit psychological tag in the explicit user profile is much greater than the probability of the most probable implicit psychological tag in the implicit user profile, then push psychological knowledge related to the most probable explicit psychological tag to the user. If the probability of the most probable implicit psychological tag in the implicit user profile is much greater than the probability of the most probable explicit psychological tag in the explicit user profile, then push psychological knowledge related to the most probable implicit psychological tag to the user. If the probability of the most probable explicit psychological tag in the explicit user profile is equal to or the same as the probability of the most probable implicit psychological tag in the implicit user profile, then push both the most probable explicit psychological tag and the most probable implicit psychological tag to the user. S3: If the number and / or probability of both the user's explicit and implicit psychological tags are 0, then push psychological knowledge about depression and bipolar disorder to the user and other users associated with that user; where, In step S2, "far greater than" means that, compared to the probability of the most probable explicit psychological tag in a user's explicit user profile, the probability of the larger psychological tag is twice or more than the probability of the smaller psychological tag; "equivalent probability" means that, compared to the probability of the most probable implicit psychological tag in a user's explicit user profile, the probability of the larger psychological tag is one to two times, but not more than two times, the probability of the smaller psychological tag.

Images