Network information content governance method for mental health and mental health construction based on large model

Abstract

The invention discloses a network information content governance method for psychological health and mental health construction based on a large model, and the method comprises the steps: constructing a domain knowledge system for psychological and mental health construction, carrying out the subdivision of the domain knowledge system into a plurality of sub-domains, and injecting domain knowledge on the basis of a general large model, thereby constructing a network information content governance domain large model; automatically generating and updating a domain knowledge base based on the domain large model, and dividing the domain knowledge base into a positive viewpoint knowledge base and a negative viewpoint knowledge base; and comparing the network information content to be detected with the viewpoints of the viewpoint knowledge base, and dividing the network information content into prohibition, warning, permission, high-priority push, secondary-priority push and general push according to a detection result. The real-time performance, coverage, accuracy and deepness of network information content recognition can be improved, the method is suitable for real-time monitoring and ecological management of numerous and complicated multi-mode network information content in the all-media era, and technical support is provided for promoting social psychology and mental health construction.

Description

Technical Field

[0001] The present invention belongs to the technical field of information security, and specifically relates to an artificial intelligence governance technology for network information content oriented to mental health and mental hygiene construction. Background Art

[0002] In today's era of diversified media, new media, social media, self-media, and all-media are interdependent and complementary in function, jointly constituting a diversified media ecosystem: new media emphasizes digital interaction, social media focuses on social interaction, self-media is for independent creation, and all-media is for resource integration. Ideal network information content and its dissemination can be that self-media can provide positive in-depth professional content to meet users' needs for obtaining positive information content; social media provides healthy interactive communication functions to meet users' needs for positive social and entertainment; new media further enriches the ways and channels of information dissemination through its diverse forms and means. While the ever-emerging network information content and the ever-changing forms of expression enrich people's lives, new requirements and challenges are posed to the governance of network information content.

[0003] Many network information contents are gradually leading people's values, cognitions, thoughts, psychology, wills, emotions, and behaviors astray with wrong ideas, and there is an urgent need for accurate identification and intervention to achieve large-scale real-time in-depth governance.

[0004] Regarding the existing technologies for intervening in network information content in the aspect of mental health and mental hygiene construction, on the one hand, it focuses on pushing positive content: including content pushing technologies based on artificial intelligence psychology large models [Huang Li CN202311791910.1], reading-nourishing and heart-healthy chatbots based on large language models [Wang Xiuhong CN202311562873.7], etc. On the other hand, it is for the governance of bad content: currently, there are mainly dynamic identification of bad content [Zhu Dingju CN202410773074.2], taking the output of bad types as a part of the reserved variable set of input data, and knowing the probability that the prediction result belongs to the preset bad type by the probability of the bad type set in the input variable; identification of borderline pornographic content [Zhu Dingju CN202410411858.0], training a multi-sample pornographic identification model to identify pornographic content, and then testing the model with pornographic content, non-pornographic content, and borderline pornographic content to obtain the corresponding output probability range to judge whether the content to be detected is borderline pornographic.

[0005] In summary, the existing technologies have the following deficiencies.

[0006] (1) Large models have demonstrated excellent performance in natural language processing tasks, being able to understand and generate complex text content based on context. They are promoting training and scenario applications in vertical fields such as psychology and reading nourishment. In terms of positive content push for network information users, they can promote users' mental health and mental hygiene construction. However, their application in the governance of negative content for network information publishers needs to be expanded and deepened, and they are not refined enough in the positive push of complex network information content.

[0007] (2) In the existing technology for network negative content governance, mainly through multi-sample training to construct an identification model, and then test the model, and determine whether it is bad content by outputting probabilities. This type of method cannot handle complex context relationships, let alone deeply understand the core concepts and methods of the field. The quality of the model itself is generally not high, and the coverage, accuracy, and depth of its content identification are all insufficient, and there is a need to deeply understand negative content and conduct fine-grained classification and accurate identification detection. Summary of the Invention

[0008] The purpose of the present invention is to provide a network information content governance method based on large models for mental health and mental hygiene construction, so as to construct a domain large model for mental and mental hygiene construction, realize in-depth understanding and fine-grained classification of the information theme, information view, and publisher's view of multi-modal network information content, so as to improve the coverage, accuracy, and depth of the identification of negative and positive content, effectively deepen the ecological governance of network information content, better purify the network information environment, protect the mental health and mental hygiene of network information users from being polluted, and promote its positive construction.

[0009] To solve the above technical problems and achieve the above purpose, the specific technical solutions adopted by the present invention are as follows.

[0010] A network information content governance method based on large models for mental health and mental hygiene construction, characterized by including the following steps:

[0011] Step 1, construct a domain large model for network information content governance for mental health and mental hygiene construction;

[0012] Step 2, generate and update a domain view knowledge base based on the domain large model; including a negative view knowledge base and a positive view knowledge base;

[0013] Step 3, extract the view of the network information publisher to be detected based on the domain large model to obtain the publisher's view;

[0014] Step 4, compare the publisher's view with the views in the negative view knowledge base to obtain the corresponding similarity S1;

[0015] If S1 > 50%, mark the network information published by this publisher as "forbidden" and do not allow it to be uploaded to the network; and associate it with the network information content published by other social media or self-media accounts of this publisher through the ID, and conduct key tracking, monitoring and governance on it;

[0016] If 25% ≤ S1 ≤ 50%, mark the network information published by this publisher as "warning", temporarily allow it to be uploaded to the network and conduct further tracking and monitoring; and associate it with the network information content published by other social media or self-media accounts of this publisher through the ID, and conduct key tracking, monitoring and governance on it;

[0017] If S1 < 25%, mark the network information published by this publisher as "permitted" and allow it to be uploaded to the network;

[0018] Step Five: When the similarity between the publisher's view and the negative view knowledge base is < 25%, further compare the publisher's view with the positive view knowledge base to obtain the corresponding similarity S2;

[0019] If S2 > 85%, mark the network information published by this publisher as "high-priority push", allow it to be uploaded to the network and perform high-priority push;

[0020] If 70% ≤ S2 ≤ 85%, mark the network information published by this publisher as "sub-priority push", allow it to be uploaded to the network and perform sub-priority push;

[0021] If 50% < S2 < 70%, mark the network information published by this publisher as "general push", allow it to be uploaded to the network and perform general push.

[0022] The construction of the domain large model includes the following processes

[0023] Process One, data collection and preprocessing: Collect publicly available data in the fields that have an impact on mental health and mental hygiene construction, including text, audio, voice video, and convert both voice video and audio content into text; clean, denoise and manually annotate the text; classify the network information content that can promote the positive construction of mental health and mental hygiene as positive content, and vice versa as negative content; construct a domain knowledge system as external brain knowledge; establish the judgment rules for positive and negative views, and conduct fine-grained association between "publisher - information theme - information view - publisher view", and then store them as different files according to different sub-domains respectively, so as to construct a domain knowledge system; the publisher refers to an individual or organization that publishes specific information content to a specific channel or platform during the information dissemination process;

[0024] Process 2, Domain Knowledge Injection: Select a general large model as the base model; according to different sub - domain types, automatically convert the constructed domain knowledge into multi - turn conversations that match different sub - domain types to form domain instruction supervision data, and after review and correction, load it into the base model;

[0025] Process 3, Model Vertical Domain Training: Use domain instruction supervision data to perform secondary instruction supervision fine - tuning on the base model, enabling the base model to learn domain knowledge, thereby constructing a domain large model to deeply understand the core concepts and methods of the domain; Use the multiple - choice evaluation method to evaluate the model: If the evaluation effect is ideal, terminate the training and save the model file, otherwise re - select hyperparameters and retrain;

[0026] Process 4, Model Online Application: Deploy the domain large model online to provide server - side call services; For the network information content that has been published or is about to be published and needs to be detected, based on the domain large model, understand the network information content and automatically generate complex text content, accurately extract the information theme, information view, and publisher's view of the network information content, and compare them with the views in the domain view knowledge base to identify whether the publisher's view is a negative view or a positive view, achieve intelligent interaction, and store historical conversation records; Monitor and manage the network information content that has been published or will be published by the information publisher with a negative view.

[0027] The domain public data is: Network information content publicly available on various social media or self - media based on new media technology; The data types include text, audio, and video; The network information content is network information content that may affect mental health and mental hygiene construction; The sub - domain is the division of information content according to the information theme, including: psychology, education, emotion, love, marriage.

[0028] The judgment rules for positive and negative views are as follows:

[0029] Network information content is divided into negative information and positive information according to its information theme; Negative information is a theme that is not conducive to mental health and mental hygiene construction; Positive information is a theme that can promote mental health and mental hygiene construction, including positive mental construction, family education, marital relationship, gender relationship, interpersonal relationship, and parent - child relationship.

[0030] Both the publisher's view and the information view are divided into negative views and positive views: When holding a negative attitude towards negative information or a positive attitude towards positive information, it is determined as a positive view; When holding a positive attitude towards negative information or a negative attitude towards positive information, it is determined as a negative view;

[0031] Judgment of positive and negative content: Comprehensive judgment is made according to the information theme, information view, and publisher's view;

[0032] When the publisher has his own views on the information theme and information opinion: If the publisher's view is a positive view, the online information content published by the publisher is marked as positive content; if the publisher's view is a negative view, the online information content published by the publisher is marked as negative content.

[0033] When the publisher does not have his own views on the information theme and information opinion: If the information opinion is a positive view, the online information content is marked as positive content; if the information opinion is a negative view, the online information content is marked as negative content.

[0034] The general large model is any one of the large language models that have been open-sourced at home and abroad and the large language models that have not been open-sourced temporarily.

[0035] The social media is a medium that spreads through user-generated content based on user relationships and interactions; the user group of social media is mainly young people and socially active people, and its influence is mainly reflected in social interaction and information dissemination; the online information content of social media includes the information published on social media by brand social media accounts, government social media accounts, media social media accounts, legally registered non-governmental organizations (NGOs) social media accounts, and non-profit organization social media accounts.

[0036] The self-media mainly relies on personal computers or mobile phones to create and share content on personal platforms; KOL: Key Opinion Leader, a group that masters rich and professional product and brand information in a specific field, has great influence and sales promotion ability, and is trusted and loved by specific audiences; KOC: Key Opinion Consumer, a group that influences the purchase behavior of relatives, friends and fans and is closer to the audience; the self-media has great freedom and flexibility in content creation and dissemination; the online information content of the self-media includes the information published by KOL, KOC or other individuals on personal blogs, personal websites, personal spaces, and social media.

[0037] The new media is a medium that spreads through various Internet applications; it includes social media and self-media, as well as other media derived, innovated, and developed with the development of digital technology: websites, blogs, podcasts, Vlogs, emails, live broadcasts, mobile apps, mini-programs, online games; the user group of the new media is extensive, and its influence is mainly reflected in information dissemination and public opinion guidance; the online information content of the new media includes the information published in the form of new media.

[0038] The present invention has beneficial effects.

[0039] (1) Provide a large model in the field of network information content governance to achieve the automatic generation and update of the domain view knowledge base. The present invention constructs a high-quality knowledge system in the domain manually and injects domain knowledge into a general large model for vertical domain training, thereby constructing a large model in the field of network information content governance for mental health and mental hygiene construction. The domain knowledge base is divided into a positive correlation point knowledge base and a negative view knowledge base, and the automatic generation and update of the view knowledge base in the field of mental health and mental hygiene construction are further realized by using the domain large model. It provides effective AI technical support for the governance of network information content.

[0040] (2) Deepen and expand the intensity of network information content governance. By constructing a large model in the field of mental health and mental hygiene construction, the present invention can deeply understand the core concepts and methods in the domain, process complex context relationships, and perform fine-grained classification and accurate identification of the information theme, information view, publisher's view, etc. of network information content, and compare them with the constructed domain view knowledge base. The recognition results are further divided into: prohibited, warning, allowed, high-priority push, sub-priority push, general push, etc., thus greatly improving the timeliness, coverage, accuracy and depth of the recognition and governance of domain network information content, and refining the governance level.

[0041] (3) Refine the complex network information content and improve the accuracy, coverage and depth of governance. The present invention extracts the information theme of network information content and divides the field of mental health and mental hygiene construction into multiple sub-domains, including: psychology, education, emotion, love, marriage, thus refining the complex network information content, dividing it into sub-domains, including positive view knowledge bases and negative view knowledge bases for each sub-domain, and storing them separately. This technical solution improves the sub-domain coverage of network information content governance in different aspects of mental and mental hygiene construction, as well as the recall rate and precision rate of in-depth recognition of content in each sub-domain.

[0042] (4) Promote social mental health and mental hygiene construction in the all-media era. The technical solution of the present invention can detect and govern multi-modal information such as text, video, and audio. It can not only detect information to be published, but also perform real-time monitoring and governance on multi-modal network information content such as social media and self-media. It can not only precisely govern the information content on the network that is harmful to mental and mental hygiene construction, greatly maintaining the security of network information content, but also recommend network information content with positive construction functions for information users. It can realize all-round real-time monitoring and governance of network information content in the all-media era and promote social mental health and mental hygiene construction. Brief Description of the Drawings

[0043] Figure 1 It is the overall idea diagram of the technical solution of the present invention;

[0044] Figure 2 Schematic diagram for fine-grained recognition and associated storage of publisher - information topic - information view - publisher view, etc.

[0045] Figure 3 Schematic diagram for constructing the knowledge system in the field of the present invention. Detailed implementation manners

[0046] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0047] Embodiment 1: Network information content governance method based on large model

[0048] The network information content governance method based on large model of the present invention is as Figure 1 shown, and specifically includes the following processes.

[0049] Process 1.1: Construct a knowledge system for the field

[0050] Construct a knowledge system for the field of network information content governance for mental health and mental hygiene construction, and divide it into sub - fields such as psychology, education, emotion, love, marriage, etc.

[0051] Process 1.2: Construct a large model for the field

[0052] Construct a large model for the field of network information content governance for mental health and mental hygiene construction, which specifically includes the following processes.

[0053] Domain knowledge injection: Based on the constructed domain knowledge system, form domain instruction supervision data and conduct manual review and correction; select ChatGLM as the base model and inject the domain instruction supervision data into the base model;

[0054] Model vertical domain training: Use the domain instruction supervision data to perform secondary instruction supervision fine - tuning on the base model, so that the base model learns domain knowledge, thereby constructing a large model for the field of network information content governance to deeply understand the core concepts and methods of the field and adapt to specific tasks in the field of network information content governance for mental health and mental hygiene construction;

[0055] Model comprehensive evaluation: Evaluate the large model for the field after continuous pre - training and optimization with a large amount of domain data by using the multiple - choice evaluation method: If the evaluation effect is ideal, terminate the training and save the model file; otherwise, re - select hyperparameters and re - train to obtain the large model for the field. At the same time, evaluate the domain professional knowledge and application scenarios to test the applicability of the model in actual business.

[0056] Process 1.3: Form a knowledge base of domain views

[0057] Based on the constructed domain large model, generate and update the domain opinion knowledge base, including a negative opinion knowledge base and a positive opinion knowledge base; and store them separately according to different sub-domains such as psychology, education, emotion, love, marriage, etc.

[0058] Process 1.4 detects the published content

[0059] Online deploy the domain large model to provide server call services; for the network information content that has been published or is about to be published to be detected, based on the constructed domain large model, understand the network information content and automatically generate complex text content, and accurately extract the information theme, information opinion, and publisher's opinion of the network information content.

[0060] Compare the publisher's opinion with the opinions in the negative opinion knowledge base to obtain the corresponding similarity S1.

[0061] If S1 ≥ 50%, then mark this network information as "forbidden" and do not allow it to be uploaded to the network.

[0062] If 25% < S1 < 50%, then mark this network information as "warning", temporarily allow it to be uploaded to the network, and further track and detect it.

[0063] If S1 ≤ 25%, then mark this network information as "permitted" and allow it to be uploaded to the network; at this time, further compare the publisher's opinion with the positive opinion knowledge base to obtain S2; if S2 ≥ 85%, then mark this network information as "high-priority push", allow it to be uploaded to the network, and push it with high priority; if 70% < S2 < 85%, then mark this network information as "secondary-priority push", allow it to be uploaded to the network, and push it with secondary priority; if 50% ≤ S2 ≤ 70%, then mark this network information as "general push", allow it to be uploaded to the network, and push it generally.

[0064] Example 2 Construction of the domain knowledge system for "governance of network information content for mental health and mental hygiene construction" The construction of the domain knowledge system of the present invention is as Figure 2 and Figure 3 shown, and specifically includes the following processes.

[0065] Process 2.1 Collect domain data

[0066] Collect new media data that will affect mental health and mental hygiene construction, such as Figure 2 the information published on social media and self-media as shown, including information in different modalities such as text, video, and audio, clean, denoise, and label it, and divide it into sub-domains according to psychology, education, emotion, love, marriage, etc.

[0067] Process 2.2 Identify positive and negative content in the collected domain data

[0068] Precisely extract and establish the fine-grained association relationships among the publisher - information theme - information view - publisher's view based on the network information content, as Figure 3 shown.

[0069] Taking the governance of network information content related to euthanasia as an example, identify whether the network information is positive or negative content, so as to determine whether the governance method for this network information content is prohibition, warning or permission; if it is permission, is it high-priority push, low-priority push, general push, or just permission to publish without push.

[0070] For example, publisher B publishes a network information content on a certain social media. The information content is about the whole process of publisher A's plan to euthanize herself after getting cancer and shows that she enjoys this process. Publisher B expresses a negative attitude towards this event in the content he publishes.

[0071] Mark this network information content: The information theme (about euthanasia and suicide) is negative information; the information view (in this embodiment, it is the view of publisher A): Publisher A has a positive attitude towards the negative content, so the information view is a negative view, and the publisher A's view is also a negative view, thus determining that the content published by publisher A is negative content. Publisher B has a negative attitude towards the negative information and its negative view published by publisher A, that is, it is determined that: the publisher B's view is a positive view, and the content published by publisher B is positive content.

[0072] When comparing with the negative view knowledge base later, the content published by publisher A is detected to have S1 = 80%, which reaches more than 50%, so the relevant content published by publisher A is determined to be "prohibited"; after comparing the content published by publisher B with the negative knowledge base, it is detected that S1 = 10%, which satisfies S1 < 25%, and then it is further compared with the positive view knowledge base, and it is detected that S2 = 90%, which is greater than 80%, so the relevant content published by publisher B is determined to be "high-priority push", as Figure 1 shown. In this embodiment, the content published by publisher A is determined to be negative content; the content published by publisher B is determined to be positive content.

[0073] Finally, it is identified that: Publisher A - Information Theme (Euthanasia, i.e., negative information) - Information Viewpoint (Affirmative attitude towards negative information, i.e., negative viewpoint) - Publisher A's Viewpoint (Same as information viewpoint in this case, i.e., negative viewpoint) - Negative Content (S1 = 80% ≥ 50%) - Governance Method (Prohibition). Publisher B - Information Theme (Euthanasia, i.e., negative information) - Information Viewpoint (Affirmative attitude towards negative information, i.e., negative viewpoint) - Publisher B's Viewpoint (Negative attitude towards negative information, i.e., positive viewpoint) - Positive Content (S1 = 10% < 25% and S2 = 90% ≥ 85%) - Governance Method (Allowed and with high - priority push).

[0074] Example 3: Governance of Online Information Content Related to Marriage Relationships

[0075] There is a live - streamer A who specifically broadcasts content related to marriage relationships on social media platforms and self - media platforms at night. It seems like marriage counseling, but actually teaches how to attract the other party; when one spouse has an affair, how the other party can quickly transfer property to their own hands and get divorced as soon as possible, etc. The number of people in the live - stream room reaches more than ten thousand every night. The live - streamer interacts warmly with the fans, sells corresponding courses at the same time, and leaves a replay video of the live - stream.

[0076] In this example, it is identified that the information theme (divorce) is negative information; the publisher and live - streamer A have an affirmative attitude towards the negative information, that is, both the information viewpoint and the publisher's viewpoint are negative viewpoints; thus, it is determined as negative content. Comparing with the negative knowledge base, the similarity S1 reaches 97%, and S1 is greater than 50%. This part of the online information content is marked as "prohibited"; finally, it is obtained that: Publisher (Live - streamer A) - Information Theme (Negative information) - Information Viewpoint (Negative viewpoint) - Publisher's Viewpoint (Negative viewpoint) - Negative Content (S1 = 97% > 50%) - Governance Method (Prohibition). Through the ID of live - streamer A, all related social media platforms and self - media platforms are associated, and other content published by him / her is key - tracked and monitored.

[0077] Another anchor B comments on the video content of anchor A. Anchor B advocates that both parties should be loyal in a marriage relationship, and no one should destroy other people's families; when there is a conflict in the relationship between husband and wife, they should persuade them to reconcile rather than to separate. The couple should tolerate each other in the marriage relationship, grow together, and recognize the trauma of divorce on children. It is identified that anchor B has a negative attitude towards negative information (divorce, etc.) and a positive attitude towards positive information (mutual loyalty, tolerance, and growth in marriage), so the network information content released by anchor B is determined to be positive content. The comparison result with the negative view knowledge base is S1=12%, and the further comparison result with the positive view knowledge base is S2=93%, so the network information released by anchor B is marked as "allowed" and "high priority push". Finally, it is concluded that: publisher (anchor B)-information topic (negative information)-information view (negative view)-publisher (anchor B) view (positive view)-positive content (S1=12%, S2=93%)-governance method (allowed, high priority push).

[0078] Example 4 Network information content management related to family education

[0079] Parent-child relationship problems in family education begin to gradually manifest themselves during adolescence as children grow older. If not dealt with promptly and correctly, they will continue to the next stage of life and continue to affect their college years, work, family life, and even old age.

[0080] For example, publisher A publicly posted the same message on multiple social media accounts. The message was about a female writer's suicide, and detailed the writer's literary works and the films and TV series adapted from her works, as well as the profound influence of the works on a generation of people at that time. Publisher A expressed his deep remembrance of the deceased.

[0081] Publisher B added his own comments and personal opinions on the incident. In addition to expressing deep regret, he believed that the female writer's suicide was her understanding of life, a kind of transcendence that was respectable and admirable, just as admirable as her works.

[0082] Publisher C sorted out the original family events of the female writer and found that when she was a child, her two younger brothers were lost. Her parents said to her at that time: How come it wasn’t you who was lost? Later, when she went to school, her grades were not as good as her sister. Her parents looked down on her and humiliated her in every way. They asked her why she didn’t die and told her that she must go to college in the future, otherwise it would be a disgrace to her father who was a professor! When she was in college, she fell in love with her teacher. When her parents knew about this relationship, they were even more severely humiliated. Publisher C analyzed her original family. The writer was bullied by her parents since she was a child, which caused great emotional damage. She did not receive real love from her parents. She was constantly hurt and had no right to control herself. Everything was decided by her parents. She attempted suicide several times during her growth but failed. Publisher C expressed his own opinion that the female writer fulfilled her wish to control her own life in her last years. The impact of family education on children's lives, her story is awakening and regrettable. It also points out that parents should give their children the right kind of love, learn to respect their children, be fair to each other, not be overbearing, and give their children the power of love so that they can have a sense of security and belonging. If the female writer had good parents who understood family education during her growth, she would not have been so sad in her old age and wanted to commit suicide to show that she could make her own decisions.

[0083] In this embodiment, the theme of the network information content is identified as suicide, and the information theme (first category information theme): negative information. Publisher A does not hold his own point of view; Publisher B holds a positive view on the negative information (believing that suicide is her understanding of life, a kind of transcendence, respectable and admirable, just like her works are admirable), and the view of publisher B is identified as a negative view. Publisher C objectively sorted out the reasons behind the whole incident and disclosed the family education information of the female writer. The information theme (second category information theme) is: in family education, parents are strong, disrespectful, unfair, incomprehensible, and unsupportive, and the second category information theme is identified as negative information; Publisher C holds a negative attitude towards both types of negative information, that is, the view of publisher C is a positive view.

[0084] Since publisher A only described an objective negative information and did not hold any opinion on the information topic, it was allowed to be published. The opinions of publishers B and C were compared with those of the negative opinion knowledge base, and publisher B's S1 = 50%, marked as "warning" and followed up for detection; publisher C's S1 = 20%, marked as "allowed", and further compared with the opinions of the positive opinion knowledge base, publisher C's S2 = 85%, marked as "high priority push".

[0085] Example 5: Governance of Internet Information Content Related to Psychological Construction

[0086] A college student at a well-known university (publisher A) publicly released his current inner emptiness on his B station, Xiaohongshu, WeChat video account, WeChat public account, and Weibo. Learning has never been a problem for him since he was a child. He has been among the best in elementary school, junior high school, high school, and now in college. He was also the top scorer in science in his province in the college entrance examination. He majored in physics in college, and felt that it was just like this in the top universities that others dream of. It was meaningless. He was too lazy to go to the classroom, and he could learn the courses by replaying. He wanted to play mobile games in the dormitory all day long, and he could pass five levels and kill six generals in the game. He didn't distinguish between day and night every day. He could only find a little temporary happiness in the game, but after the game, he felt that playing games was actually meaningless, but there was no point in doing anything else except games, including the basketball he loved before. My mother was so angry that she blocked me on WeChat, but I didn't think it was important. I feel that I have no motivation to do anything now. I have worked so hard all the way from childhood to adulthood and got into a good university, but it is just like this. I wonder if other students have the same situation? I also feel that something is wrong with me, and I don't know how to break through this inner state?

[0087] Publisher B reposted the content published by Publisher A on his own Xiaohongshu and commented on it with his own views. Publisher B believes that being empty-hearted is a normal state. Life is meaningless. Don't think it's strange. Let everything go with the flow and follow your heart. Don't do what you don't want to do. Do what you want to do. People should enjoy freedom. Live life to the fullest. Just live one day at a time. It's enough to have food and drink. Why think so much? Isn't it tiring?

[0088] Publisher C forwarded the content published by Publisher A and Publisher B on his own B station and commented on it. Publisher C pointed out that the current hollow state is common among young people, and they need to reconstruct their value system and recognize their social mission, so as to stimulate their inner driving force and find lasting satisfaction and happiness. The real sense of happiness and value comes from altruistic giving, not selfish demands. Blindly selfish demands will make the heart become more and more scarce, and thus feel more empty. The more greedy, the less satisfied. Young people need to reconstruct their value system and have an altruistic mindset, so that they can find the value of existence and the meaning of life, thereby generating a positive inner drive to do things, and their hearts are no longer empty.

[0089] In this embodiment, it is identified that the information theme (hollowness, nothingness, lack of happiness) of the network information content published by publisher A is a negative theme; the information viewpoint is a negative viewpoint; the viewpoint of publisher A (feeling confused about his own situation, feeling that it is wrong, and wanting to break through) is a positive viewpoint, and the network information content published by publisher A is positive content, marked as "allowed", and further compared with the positive viewpoint knowledge base, and S2=25% is obtained.

[0090] Publisher B has a positive attitude towards negative themes and information views, believing that life is inherently meaningless and people should indulge in pleasure. It is identified that Publisher B's view is a negative view. Further comparison with the views in the negative view knowledge base shows that S1 = 55% which is greater than 50%. The online information content published by Publisher B is marked as "Prohibited", and other online information content published by Publisher B's ID is monitored.

[0091] Publisher C has a negative attitude towards negative themes and information views, pointing out that people need to reconstruct their value system, shift from self-interest to altruism, so as to find their sense of mission, stimulate the internal drive to do things, and realize the meaning of life. It is identified that Publisher C's view is a positive view; further comparison with the positive view knowledge base shows that S2 = 85%. The online information content published by Publisher C is marked as "Allowed" and "High-priority push".

[0092] Example Six: Governance of Online Information Content

[0093] Currently, there are many short dramas on the Internet. Many people have distorted their views on love and marriage due to watching too many toxic dramas.

[0094] If Publisher A releases such short dramas and has a non-committal attitude towards the plot content, then it is determined that the relevant online information released by Publisher A has a negative information theme; the information view is a negative view; the publisher's view (the publisher releases negative information and has a default attitude of not commenting on the information content, and its public release means advocating) is a negative view; comparison with the negative view library shows that S1 = 45%, which is less than 50%; the online information content released by Publisher A is marked as "Warning".

[0095] Publisher B also releases such short dramas, but has comments on the plot content and has a negative attitude. Then it is determined that the relevant online information content released by Publisher B: the information theme is negative information; the information view is a negative view; Publisher B's view is a positive view; it is identified that the online information content released by Publisher B is positive content; comparison with the negative view library shows that S1 = 25%; at the same time, comparison with the positive view library shows that S2 = 50%. So the online information content released by Publisher B is marked as "Allowed".

Claims

1. A network information content governance method based on a big model for mental health and mental hygiene construction, characterized by It includes the following steps: Step 1, construct a large model in the field of network information content governance for mental health and mental hygiene construction; Step 2, generate and update the domain opinion knowledge base based on the domain large model; It includes a negative opinion knowledge base and a positive opinion knowledge base; Step 3, extract the opinion of the network information publisher to be detected based on the domain large model to obtain the publisher's opinion; Step 4, compare the publisher's opinion with the opinions in the negative opinion knowledge base to obtain the corresponding similarity S1; If S1>50%, then mark the network information published by this publisher as "forbidden" and do not allow it to be uploaded to the network; And Associate by ID with the network information content published by other social media or self-media accounts of this publisher, and conduct key tracking, monitoring and governance on it; If 25%≤S1≤50%, then mark the network information published by this publisher as "warning", temporarily allow it to be uploaded to the network, and further track and detect it; and associate by ID with the network information content published by other social media or self-media accounts of this publisher, and conduct key tracking, monitoring and governance on it; If S1<25%, then mark the network information published by this publisher as "permitted" and allow it to be uploaded to the network; Step 5: When the similarity between the publisher's opinion and the negative opinion knowledge base is <25%, further compare the publisher's opinion with the positive opinion knowledge base to obtain the corresponding similarity S2; If S2>85%, then mark the network information published by this publisher as "high-priority push", allow it to be uploaded to the network, and push it with high priority; If 70%≤S2≤85%, then mark the network information published by this publisher as "sub-priority push", allow it to be uploaded to the network, and push it with sub-priority; If 50%<S2<70%, then mark the network information published by this publisher as "general push", allow it to be uploaded to the network, and push it generally.

2. According to claim 1, a network information content governance method based on a large model for mental health and mental hygiene construction is characterized in that The construction of the domain large model includes the following process Process 1, data collection and preprocessing: collect publicly available data in the field that has an impact on mental health and mental hygiene construction, including text, audio, voice video, and convert the voice video and audio content into text; clean, denoise and manually annotate the text; classify the network information content that can promote the positive construction of mental health and mental hygiene as positive content, and vice versa as negative content; construct a domain knowledge system as external brain knowledge; establish judgment rules for positive and negative opinions, and conduct fine-grained association between "publisher - information topic - information opinion - publisher's opinion", and then store them as different files according to different sub-domains respectively, so as to construct a domain knowledge system; the publisher refers to an individual or organization that publishes specific information content to a specific channel or platform during the information dissemination process; Process 2, domain knowledge injection: select a general large model as the base model; according to different sub-domain types, automatically convert the constructed domain knowledge into multi-round dialogues that match different sub-domain types to form domain instruction supervision data, and load it into the base model after review and correction; Process 3: Model vertical domain training: Use domain instruction supervision data to perform secondary instruction supervision fine-tuning on the base model, so that the base model can learn domain knowledge, thereby building a large domain model to deeply understand the core concepts and methods of the domain; use multiple choice evaluation to evaluate the model: if the evaluation effect is ideal, terminate the training and save the model file, otherwise reselect the hyperparameters and retrain; Process 4: Online application of the model: deploy the domain big model online and provide the calling service of the server; for the network information content that has been published or is about to be published to be detected, based on the domain big model, understand the network information content and automatically generate complex text content, accurately extract the information theme, information viewpoint and publisher viewpoint of the network information content, and compare them with the viewpoints of the domain viewpoint knowledge base to identify whether the publisher's viewpoint is negative or positive, realize intelligent interaction, and store historical conversation records; monitor and manage the network information content that has been published or will be published by the identified information publisher with negative views.

3. According to claim 2, a network information content governance method based on a large model for mental health and mental hygiene construction is characterized in that The public data in the field are: network information content disclosed by various social media or self-media based on new media technology; data includes text, audio, video and other multi-modal data; Online information content that may affect mental health and mental hygiene; The sub-fields are divisions of information content according to information themes, including: psychology, education, emotion, love, and marriage.

4. According to claim 2, a network information content governance method based on a large model for mental health and mental hygiene construction is characterized in that The judgment rules for the positive and negative views are: The content of online information is divided into negative information and positive information according to its information theme; negative information refers to themes that are not conducive to mental health and mental hygiene construction; positive information refers to themes that can promote mental health and mental hygiene construction, including positive psychological construction, family education, marital relations, gender relations, interpersonal relations, and parent-child relations; The publisher's opinions and information opinions are divided into negative opinions and positive opinions: when a negative attitude is held towards negative information, or a positive attitude is held towards positive information, it is judged as a positive opinion; when a positive attitude is held towards negative information, or a negative attitude is held towards positive information, it is judged as a negative opinion; The judgment of the positive content and the negative content is as follows: a comprehensive judgment is made based on the information subject, the information viewpoint and the publisher's viewpoint; when the publisher holds his own view on the information subject and the information viewpoint: if the publisher's viewpoint is a positive viewpoint, the network information content published by the publisher is identified as positive content; if the publisher's viewpoint is a negative viewpoint, the network information content published by the publisher is identified as negative content; when the publisher does not hold his own view on the information content and the information viewpoint: if the information viewpoint is a positive viewpoint, the network information content is identified as positive content; if the information viewpoint is a negative viewpoint, the network information content is identified as negative content.

5. According to claim 2, a network information content governance method based on a large model for mental health and mental hygiene construction is characterized in that The general large model is any one of the large language models that have been open sourced at home and abroad and the large language models that have not yet been open sourced.

6. According to claim 3, a network information content governance method based on a large model for mental health and mental hygiene construction is characterized by: The social media is a medium that disseminates content through user-generated content based on user relationships and interactions; social media network information content includes information posted on social media by brand social media accounts, government social media accounts, media social media accounts, legally registered non-governmental organizations (NGOs) social media accounts, and non-profit organization social media accounts; The self-media mentioned above mainly rely on personal computers or mobile phones to create and share content on personal platforms; KOL: Key opinion leaders, a group that has rich and professional product and brand information in a specific field, has great influence and sales ability, and is trusted and loved by a specific audience; KOC: Key opinion consumers, a group that influences the purchasing behavior of friends, relatives and fans and is closer to the audience; self-media network information content includes information published by KOL, KOC or other individuals on personal blogs, personal websites, personal spaces, and social media; The new media mentioned above refers to the media that is disseminated through various Internet applications, including social media and self-media, as well as other media that are derived, innovated, and developed with digital technology: websites, blogs, podcasts, Vlogs, emails, live broadcasts, mobile apps, mini-programs, and online games; new media network information content includes information released in the form of new media.

Images