A social guidance method and system

Abstract

The application discloses a social guidance method and system, belonging to the technical field of game socialization, which utilizes a pre-trained large language model to perform intent recognition processing on multi-modal game data of a target object in a game, to obtain a target social intent type of the target object and a confidence of the target social intent type; based on the target social intent type and the confidence, a target guidance strategy level for the target object is determined from a plurality of predefined guidance strategy levels; and according to the target guidance strategy level and the target social intent type, target guidance content for the target social intent type of the target object is generated. The application can effectively improve the accuracy and reliability of object social intent recognition, avoid invalid guidance or misguidance, reduce interference with the object's game experience, and at the same time make the guidance strategy more targeted and adaptive, improve the object's acceptance of the guidance content and the guidance conversion efficiency.

Description

Technical Field

[0001] This application relates to the field of game social technology, specifically to a social guidance method and system. Background Technology

[0002] Currently, most online games use traditional, rule-based methods for social guidance. Social intent recognition relies solely on simple keyword searches and fixed behavioral trigger conditions, failing to deeply interpret the user's colloquial expressions and implicit social needs within the game. This results in inaccurate intent recognition, high rates of false positives and false negatives, and an inability to accurately capture the user's true social needs.

[0003] Based on this, existing social guidance solutions for online games typically categorize social needs into two types: those with social intentions and those without. Furthermore, the guidance strategies for those with social intentions are often simplistic, either frequently displaying aggressive guidance interfaces that disrupt normal gameplay and severely damage the user's experience, or employing weak guidance methods that are ineffective. Clearly, the social guidance strategies in these technologies fail to accurately identify the user's social intent, resulting in poor guidance effectiveness. Summary of the Invention

[0004] This invention provides a social guidance method and system, aiming to solve the problem that game social guidance strategies in related technologies have poor guidance effects due to the inability to accurately identify the social intentions of the target audience.

[0005] Firstly, a social guidance method is provided, including the following steps: A pre-trained large language model is used to perform intent recognition processing on multimodal game data of target objects in the game, so as to obtain the target social intent type of the target object and the confidence level of the target social intent type; Based on the target social intent type and the confidence level, a target guidance strategy level for the target object is determined from multiple predefined guidance strategy levels; Based on the target guidance strategy hierarchy and the target social intent type, target guidance content is generated for the target object's target social intent type.

[0006] Secondly, a social guidance system is also provided, including: The social intent recognition module is used to perform intent recognition processing on the multimodal game data of the target object in the game using a pre-trained large language model, so as to obtain the target social intent type of the target object and the confidence level of the target social intent type; The guidance strategy determination module is used to determine the target guidance strategy level for the target object from multiple predefined guidance strategy levels based on the target social intent type and the confidence level; The guidance module is used to generate target guidance content for the target social intent type of the target object based on the target guidance strategy hierarchy and the target social intent type.

[0007] Thirdly, an electronic device is also provided, including a memory and a processor, wherein a computer program is stored on the memory, and when executed by the processor, the computer program implements the steps of any one of the methods described above.

[0008] Fourthly, a computer-readable storage medium is also provided, on which a computer program is stored, the computer program being loaded by a processor to perform the steps of any of the methods described above.

[0009] Beneficial effects: This application uses a pre-trained large language model combined with knowledge of the game social domain to accurately identify the target social intent type of an object and obtain its confidence level. Based on the intent type and confidence level, it determines the appropriate target guidance strategy level and finally generates target guidance content that fits the object's needs, personal characteristics and real-time game scenario. This can effectively improve the accuracy and reliability of the object's social intent identification, avoid ineffective or misleading guidance, reduce interference with the object's game experience, and make the guidance strategy more targeted and adaptable, thereby improving the object's acceptance of the guidance content and the guidance conversion efficiency. Attached Figure Description

[0010] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0011] Figure 1 This is a flowchart of a social guidance method provided by an exemplary embodiment of this disclosure; Figure 2 This is a flowchart of another social guidance method provided by an exemplary embodiment of this disclosure; Figure 3 This is a flowchart of another social guidance method provided by an exemplary embodiment of this disclosure; Figure 4 This is a flowchart of another social guidance method provided by an exemplary embodiment of this disclosure; Figure 5 This is a flowchart of another social guidance method provided by an exemplary embodiment of this disclosure; Figure 6 This is a flowchart of another social guidance method provided by an exemplary embodiment of this disclosure; Figure 7 This is a flowchart of another social guidance method provided by an exemplary embodiment of this disclosure; Figure 8 This is an exemplary flowchart of a social guidance method provided by an exemplary embodiment of the present disclosure; Figure 9 This is a schematic diagram of the functional modules of the social guidance system provided in an exemplary embodiment of this disclosure. Detailed Implementation

[0012] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0013] In the description of this application, it should be understood that the terms "center," "longitudinal," "lateral," "length," "width," "thickness," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," and "outer," etc., indicating orientation or positional relationships based on the orientation or positional relationships shown in the accompanying drawings, are used only for the convenience of describing this application and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of this application. Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, features defined with "first" and "second" may explicitly or implicitly include one or more of the stated features. In the description of this application, "a plurality of" means two or more, unless otherwise explicitly specified.

[0014] "A and / or B" includes the following three combinations: A only, B only, and a combination of A and B.

[0015] The use of "applies to" or "configured to" in this application implies open and inclusive language, which does not exclude the applicability to or configuration to devices performing additional tasks or steps. Additionally, the use of "based on" implies openness and inclusivity, because processes, steps, calculations, or other actions "based on" one or more of the stated conditions or values ​​may in practice be based on additional conditions or values ​​beyond those stated.

[0016] In this application, the term "exemplary" is used to mean "used as an example, illustration, or description." Any embodiment described as "exemplary" in this application is not necessarily to be construed as being more preferred or advantageous than other embodiments. The following description is provided to enable any person skilled in the art to make and use this application. Details are set forth in the following description for purposes of explanation. It should be understood that those skilled in the art will recognize that this application can be made without using these specific details. In other instances, well-known structures and processes are not described in detail to avoid obscuring the description of this application with unnecessary detail. Therefore, this application is not intended to be limited to the embodiments shown, but is consistent with the broadest scope of the principles and features disclosed in this application.

[0017] The social guidance method and system disclosed in this application aim to solve the technical problems in existing game social guidance, such as low accuracy of intent recognition, lack of targeted guidance strategy matching, and poor adaptability of guidance content to the needs of the target audience. It achieves accurate recognition of the target audience's social intent by combining a large language model with knowledge of the game social domain, completes intelligent matching of guidance strategy levels based on confidence calibration and intent priority, and generates personalized guidance content by combining the target audience profile and game scene.

[0018] On one hand, this embodiment provides a social guidance method applicable to hardware carriers with data processing and model running capabilities, such as game servers, game intelligent analysis platforms, or game cloud servers. The hardware carrier is equipped with a pre-trained large language model, which is a large language model pre-trained on a general natural language corpus. Preferably, it is a locally deployed open-source large language model or a private version of a commercial large language model, ensuring the security of game data and the real-time nature of intent recognition. It also supports fine-tuning or enhancement of prompts by incorporating domain knowledge. The method of this embodiment is applicable to various online games with social gameplay, including role-playing games, casual competitive games, and social mobile games, enabling real-time social guidance for single or multiple target objects within the game.

[0019] Please see Figure 1 This embodiment provides a social guidance method, which specifically includes the following steps: Step S101: Use a pre-trained large language model to perform intent recognition processing on the multimodal game data of the target object in the game, and obtain the target social intent type and the confidence level of the target social intent type.

[0020] Specifically, the system collects real-time game behavior data of all objects within the game. When a target object triggers any social-related action in a social context, its multimodal game data is extracted. The social context can include public channels, private channels, guild channels, etc. Multimodal game data represents a collection of various types and sources of data reflecting the target object's game state, behavioral characteristics, and social tendencies. For example, it may include text chat data and voice chat-to-text data collected in real-time from the game client; it may also include structured behavioral event data extracted from game server logs, such as "entering a dungeon" and "gifting items"; and it may include attribute data representing the object's real-time state, such as level, geographical location, and equipment information. This embodiment of the invention does not limit the specific source and format of the multimodal game data.

[0021] For example, social-related operations triggered by the target object include, but are not limited to, viewing other objects' profiles, creating or disbanding a team, adding friends, or staying in the game's social scene for a long time; the object generates help-seeking behavior when performing game operations (such as failing a dungeon challenge, getting stuck on a game task, or suffering a series of defeats in PVP competition); the object opens the game's social function panel (such as the marriage tree, the mentor-apprentice system, or the guild recruitment panel).

[0022] Then, a pre-trained large language model is used to perform intent recognition processing on the multimodal game data of the target object within the game. First, the multimodal game data generated by the target object in the game is normalized and features are extracted. Then, the processed multimodal game data is input into the pre-trained large language model. The model performs comprehensive analysis based on learned semantic understanding, behavioral patterns, and social intent features to automatically identify the most matching social intent type of the object, i.e., the target social intent type, and outputs the confidence score corresponding to the target social intent type. The pre-trained large language model refers to a general-purpose large model fine-tuned with game domain knowledge. It has built-in game social vocabulary, language rules, and scene-corresponding appeal libraries, which can identify in-game slang, colloquial expressions, and implicit appeals, thus avoiding the limitations of traditional keyword matching. The confidence score ranges from 0 to 1, with a higher value indicating a higher degree of confidence in the target social intent type.

[0023] Step S102: Based on the target social intent type and confidence level, determine the target guidance strategy level for the target object from multiple predefined guidance strategy levels.

[0024] Specifically, based on preset social intent type level and confidence interval division rules, the target social intent type is classified into the corresponding guidance strategy level, and the confidence level is matched to the corresponding confidence interval. Then, the social intent type level and confidence interval are comprehensively judged, and the guidance strategy level that best matches the current object state is selected from multiple preset guidance strategy levels. Finally, the target guidance strategy level for the target object is determined.

[0025] Step S103: Generate target guidance content for the target audience based on the target guidance strategy hierarchy and the target social intent type.

[0026] Specifically, based on the determined target guidance strategy level and the target social intent type of the target audience, corresponding guidance content is generated to ensure that the form and intensity of the guidance content are adapted to the target guidance strategy level, thereby ensuring that the guidance content matches the social needs of the target audience.

[0027] Furthermore, based on the target guidance content, the guidance actions corresponding to the target guidance strategy level are executed.

[0028] For example, for targets seeking dungeon assistance, they are categorized into the mild guidance strategy level, with the target guidance content being online teammate matching prompts for the dungeon, and the guidance action being the display of online teammate matching prompts for the dungeon seeking via a sidebar; for targets seeking in-game intimacy, they are categorized into the strong guidance strategy level, with the target guidance content being guidance content related to the marriage scenario, and the guidance action being the display of marriage scenario-related guidance content via a pop-up visual display; for targets engaging in random interactions, they are categorized into the silent guidance strategy level, with no target guidance content, and only the random interaction requests of the target target are recorded in the background without any guidance display on the front-end page.

[0029] This embodiment uses a pre-trained large language model combined with knowledge of the game's social domain to accurately identify the target social intent type of an object and obtain its confidence level. Based on this intent type and confidence level, it determines the appropriate target guidance strategy level, and finally generates target guidance content that fits the object's needs, personal characteristics, and real-time game scenario. This application can effectively improve the accuracy and reliability of object social intent identification, avoid ineffective or misleading guidance, reduce interference with the object's game experience, and make the guidance strategy more targeted and adaptable. It also improves the object's acceptance of the guidance content and the guidance conversion efficiency, taking into account both system operating efficiency and object experience, further optimizing the overall effect of in-game social guidance, and meeting the actual needs of game social scenarios.

[0030] In some embodiments, in step S102, a target guidance strategy level for the target object is determined from multiple predefined guidance strategy levels based on the target social intent and confidence level. (See also...) Figure 2 ,include: Step S201: Select the target pre-trained confidence calibrator corresponding to the target social intent from multiple pre-trained confidence calibrators. The target social intent type includes the target macro intent type and the target sub-intent type of the target object under the macro intent type.

[0031] Specifically, after determining the target social intent and confidence level of the target object, multiple pre-trained confidence calibrators are retrieved. These pre-trained confidence calibrators are all associated with different types of social intent, and each model is specifically trained for a particular macro-intent type and its subordinate sub-intent types to adapt to the confidence calibration requirements of different social intents. Then, based on the specific type of the currently determined target social intent, i.e., the target macro-intent type and the corresponding target sub-intent type, precise matching and filtering are performed among the multiple pre-trained confidence calibrators. Finally, the pre-trained confidence calibrator that completely corresponds to the target social intent and can accurately adapt to its intent features is selected and determined as the target pre-trained confidence calibrator.

[0032] Furthermore, to more accurately identify the true social intentions of the target audience, this embodiment categorizes social intention types into macro-intention types and sub-intention types under macro-intention types. For example, macro-intention types include, but are not limited to, seeking close social relationships, seeking growth assistance, and finding playmate communities. Specifically, the "seeking close social relationships" type includes sub-intention types such as, but are not limited to, game couples, sworn brothers, and emotional support, aiming to establish emotional connections or fixed companionship within the game, thereby increasing the emotional stickiness of the target audience and reducing churn; the "seeking growth assistance" type includes sub-intention types such as, but are not limited to, being guided through dungeons or quests, answering gameplay questions, providing equipment matching or development advice, and player-versus-player (PVP) skill guidance, aiming to obtain game knowledge, skills, or support for clearing levels, thereby helping the target audience overcome difficulties and improve their experience; the "finding playmate communities" type includes sub-intention types such as, but are not limited to, finding fixed teammates, guild recruitment, casual chat partners, and temporary team formation for activities, aiming to find partners for game activities or casual conversation, thereby enhancing social activity and forming a community.

[0033] For cases where the target social intent type includes the target macro intent type and the target sub-intent type of the target object under the macro intent type, the model is first matched and filtered from multiple pre-built and trained confidence calibrators according to the macro intent type and target sub-intent type corresponding to the target social intent. The pre-trained confidence calibrator that matches the target sub-intent type, social scenario and confidence calibration logic of the current target social intent is selected and determined as the target pre-trained confidence calibrator to ensure that the subsequent calibration process is consistent with the business logic of the current social intent.

[0034] For example, the steps of pre-building and training multiple confidence calibrators include: An independent confidence calibrator is created for each type of social intent, with confidence scores limited to a range of [0,1]. Optionally, the confidence calibrator employs an isotonic regression algorithm, which does not require a pre-defined data distribution and is adapted to the non-linear characteristics of game social data. A large amount of labeled real-world datasets, including predicted confidence, real labels, and predicted intent types, are collected as training data for the confidence calibrator.

[0035] The steps for training the confidence calibrator by category include: first, according to the intent type, traversing the social intents corresponding to each real data in the labeled real dataset, and filtering the predicted confidence and real labels of all samples under each type of social intent; then, counting the number of samples for each type of social intent, and for social intent types that meet the preset number of samples, performing fitting training of the confidence calibrator for that social intent type using the corresponding sample predicted confidence and real labels, so that the confidence calibrator learns the mapping relationship between the original confidence and real labels under that social intent type, thereby obtaining the pre-trained confidence calibrator corresponding to that social intent type, to ensure the effectiveness of the training results.

[0036] Furthermore, an incremental learning approach is adopted to iteratively update the confidence calibrator online. For each type of social intent, new samples are continuously collected and added to the training data, and the model parameters are updated regularly. For example, the fitting training step is re-executed every 100 new samples to continuously iterate and update the pre-trained confidence calibrator, thereby further improving the calibration accuracy, adapting to scenarios such as game version updates and changes in target language, and ensuring long-term effectiveness.

[0037] Step S202: Use the target pre-trained confidence calibrator to calibrate the confidence level and obtain the calibrated confidence level.

[0038] Specifically, in order to eliminate the identification bias of the pre-trained large language model in the social intent of the target object, this embodiment uses a confidence calibrator to calibrate the confidence of the output of the large model. The pre-trained confidence calibrator is trained based on a large amount of labeled social data of game objects. It adopts the isotonic regression algorithm to train independent confidence calibrators for different macro intents, so as to avoid calibration bias of different intent types and ensure that the calibration results are accurate and reliable.

[0039] The steps for calibrating confidence using a target pre-trained confidence calibrator include: First, the system checks if a pre-trained confidence calibrator corresponding to the target social intent type exists in the confidence calibrator library. If no calibrator exists for this intent type, calibration is not required, and the original confidence score is directly output as the calibrated confidence score. If a pre-trained confidence calibrator exists for the target social intent type, the original confidence score value is input into this pre-trained confidence calibrator as a two-dimensional array. The model then corrects the bias of the confidence score based on learned business features, confidence score distribution patterns, and social scenarios, outputting a confidence score value calibrated using the isotonic regression algorithm as the initial calibration value. The confidence calibrator, trained using the isotonic regression algorithm, is essentially a monotonically increasing piecewise linear function used to map the original confidence score output by the model to a calibrated confidence score that is closer to the true probability. The calibration process is as follows: the original confidence score is input into the calibrator, which uses its internally stored piecewise function parameters to find the interval containing the original confidence score and calculates the calibrated confidence score through linear interpolation. For example, for a calibrator targeting the intention of "seeking intimate relationships", if the original confidence level is 0.75, the calibrated confidence level output after calibrator mapping is 0.82; if the original confidence level is 0.60, the calibrated confidence level output after calibration is 0.55.

[0040] Then, the initial calibration values ​​are validated to ensure that the final output calibrated confidence score is limited to the range of 0 to 1. For example, if the calibration value is less than 0, it is forcibly corrected to 0; if the calibration value is greater than 1, it is forcibly corrected to 1; if it is within the range of 0 to 1, the initial calibration value is retained; finally, the calibrated confidence score after boundary constraints is returned. This calibrated confidence score is a quantified value that is closer to the true probability, representing the matching probability between the target social intent type and the object's true social needs, eliminating the bias of overconfidence or underconfidence inherent in the large language model itself.

[0041] Step S203: Based on the intent priority and post-calibration confidence corresponding to the target social intent, determine the target guidance strategy level of the target object from multiple predefined guidance strategy levels.

[0042] Specifically, predefined intent priority classification rules are retrieved. In this embodiment, intent priorities are only defined for specific needs, and macro-intents do not have independent priorities. It is easy to understand that macro-intents are only used to reflect the overall priority tendency and do not directly participate in the hierarchical determination, avoiding judgment bias caused by uniformly classifying macro-intents. Optionally, sub-intent types are divided into three priority levels. High-priority intents include types with strong immediacy and urgent needs, such as game couples, sworn brothers, dungeon help requests, finding fixed teammates, and finding intimate relationships in the game; medium-priority intents include types with non-immediate and essential needs, such as emotional support, guidance on combat skills, equipment matching and development consultation, guild recruitment, and temporary team formation for activities; low-priority intents include types with ambiguous needs, such as finding casual chat partners and random interactions.

[0043] Based on the business type of the target social intent, the urgency of the target's needs, and the system's processing priority, the intent priority corresponding to the target social intent is determined. Then, combined with the magnitude and distribution range of the calibrated confidence score, the calibrated confidence score is compared with at least one preset confidence score threshold. In conjunction with the intent priority corresponding to the intent type, the guidance strategy level that best matches the current intent priority, calibrated confidence score, and target state is selected from multiple predefined guidance strategy levels and determined as the target guidance strategy level for the current target target.

[0044] For example, the predefined boot strategy hierarchy is divided into three levels: the first boot strategy level, the second boot strategy level, and the third boot strategy level, corresponding to strong boot, mild boot, and silent boot. Each level sets clear calibration confidence thresholds and priority conditions. The specific division and implementation rules are as follows: The first guidance level is the highest intensity guidance, triggered when the target social intent belongs to a high-priority sub-intent type and the calibrated confidence score is higher than the first confidence threshold (e.g., 0.8). This level is for individuals with clear and urgent needs and extremely high credibility of the recognition results. The corresponding guidance action is to trigger a proactive guidance interface on the front end, such as generating a front-end pop-up window. The pop-up window is located in a non-full-screen position on the game interface, without obscuring the core operation area. It includes clear guidance text, a one-click jump button, and may include basic virtual reward incentives. Clicking the pop-up window will directly jump to the corresponding social function interface, such as the dungeon matching interface or the marriage matching interface, eliminating the need for manual object search and improving guidance conversion efficiency.

[0045] The second guidance level is of medium intensity. The triggering condition is that the target social intent belongs to a medium-priority sub-intent type, and the calibrated confidence level is lower than the first confidence threshold but higher than the second confidence threshold. In other words, there is no strict restriction on intent priority; both high and medium priority intents can be triggered, while low-priority intents will not trigger this level. This level is for objects with relatively clear needs and moderate recognition result credibility. The corresponding guidance action is to trigger front-end prompts, such as generating a floating bar on the side of the game interface or a red dot for social function icons. The floating bar disappears automatically after 5 seconds by default, and the red dot reminder can be manually clicked to view. There are no forced pop-ups, and it does not interfere with the object's normal game operation. The guidance text is concise, only providing core information. Clicking it redirects to the corresponding function interface, balancing guidance effectiveness and user experience.

[0046] The third guidance level is low-intensity guidance, triggered when the post-calibration confidence score falls below the second confidence threshold (e.g., 0.5). This level is triggered regardless of the intent priority. This level is for objects with low recognition confidence and vague intent. There is no front-end visual guidance content; the corresponding guidance action is recorded in the background. For example, only the object's target social intent type and calibration value are recorded in the background and incorporated into the object's social profile for indirect guidance such as contextualized recommendations and friend recommendations. No active intervention is performed to avoid ineffective guidance due to misjudgment, thus maximizing the protection of the object's gaming experience.

[0047] This embodiment achieves confidence calibration by introducing a pre-trained confidence calibrator to eliminate biases in the original data. It also achieves qualitative screening through intent priority, enabling the determination of guidance strategy levels to accurately match the social needs of objects with different recognition credibility and different urgency of demands. This ensures that high-priority and high-value demands are strongly guided, while low-priority and ambiguous demands are not overly interfered with, achieving a balance between the accuracy of guidance and the user experience. This effectively solves the shortcomings of guidance strategies that rely solely on a single numerical value or simple rules in related technical solutions, which are inaccurate.

[0048] In some embodiments, please refer to Figure 3 The method also includes: Step S301: Based on the target social intent type and post-calibration confidence, a reinforcement learning selection algorithm is used to screen several combinations of guidance strategies corresponding to the target guidance strategy level to determine the optimal guidance strategy combination. The guidance strategy combination is any combination of guidance copy, incentive reward, and trigger timing.

[0049] Specifically, based on the determined target social intent type (including macro-intent type and target sub-intent type) and the calibrated confidence level, a pre-trained reinforcement learning selection algorithm is invoked. The algorithm uses the appeal features corresponding to the target social intent type and the recognition credibility reflected by the calibrated confidence level as input features to perform multi-dimensional evaluation and screening of all guidance strategy combinations under the target guidance strategy level in the hierarchical strategy adapter. Each guidance strategy combination includes different combinations of guidance copywriting style, incentive reward type or amount, and trigger timing window.

[0050] The reinforcement learning selection algorithm combines the reward values ​​of different strategy combinations in historical guidance data, such as conversion effect, object acceptance, and experience feedback, to iteratively calculate and score the suitability of each strategy combination with the current target object. It prioritizes the selection of strategy combinations that not only meet the appeal matching degree of the target social intent type, but also match the recognition reliability corresponding to the calibrated confidence level, while taking into account the object experience and guidance conversion efficiency. Finally, the combination with the highest comprehensive score is determined as the optimal guidance strategy combination.

[0051] For example, consider the Multi-Armed Bandi (MAB) algorithm, a classic reinforcement learning framework where each arm represents a combination of guiding policies, or selectable policies, used to dynamically select the optimal combination of guiding policies among multiple arms, achieving a balance between exploring new policies and utilizing known policies.

[0052] Based on this, this embodiment selects a multi-armed trial algorithm for the reinforcement learning selection algorithm. Therefore, in step S301, based on the target social intent type and the post-calibration confidence, the reinforcement learning selection algorithm is used to screen several combinations of guidance strategies corresponding to the target guidance strategy level to determine the optimal guidance strategy combination, including: Step M101: Model several combinations of guidance strategies corresponding to the target guidance strategy level as strategy arms in the multi-armed probing algorithm.

[0053] Specifically, each guidance strategy combination under the target guidance strategy level is composed of any combination of different options in three dimensions: guidance copy, incentive reward, and trigger timing. Each independent guidance strategy combination is modeled as a strategy arm in a multi-armed probing algorithm, establishing a one-to-one correspondence between guidance strategy combinations and strategy arms, so that the characteristic performance of each strategy arm is consistent with the copy, reward, and trigger timing configuration of the corresponding guidance strategy combination.

[0054] Step M102: Initialize the number of attempts and the number of successful gains for each strategy arm, and define the comprehensive gain for each strategy arm. The comprehensive gain includes immediate gain and long-term gain. Immediate gain refers to the target object's immediate interactive behavior in response to guidance, while long-term gain refers to the target object's successful establishment of social relationships after accepting guidance.

[0055] Specifically, for each strategy arm, the number of attempts and the number of successful gains are initialized as statistical parameters, and the comprehensive gains of each strategy arm are defined to include two parts: immediate gains and long-term gains. Immediate gains refer to the immediate interactive behavior of the target audience after receiving the guidance content, such as clicking on the guidance or viewing guidance details; long-term gains refer to the behavioral outcome of the target audience successfully establishing a social relationship after accepting the guidance.

[0056] Step M103: Calculate the overall benefit of each strategy arm based on the target object's feature information, the target social intent type, and the post-calibration confidence.

[0057] Specifically, the feature information of the target object, the appeal features corresponding to the target social intent type, and the credibility of intent recognition reflected by the calibrated confidence level are extracted. The target object feature information includes object level (e.g., VIP level), occupation, gaming habits, etc., the appeal features correspond to the specific needs and tendencies of the target social intent, and the recognition credibility reflects the reliability of the intent recognition result. The extracted multi-dimensional information is matched and analyzed with the guidance strategy combinations corresponding to each strategy arm. Combined with the existing trial records of each strategy arm, the actual performance of immediate and long-term benefits is comprehensively considered to calculate the comprehensive benefit value corresponding to each strategy arm. The higher the comprehensive benefit value, the better the adaptability and effect of the corresponding guidance strategy combination. In one embodiment of this disclosure, the comprehensive benefit R is calculated using the following formula: R = α * R_immediate + β * R_long_term. Wherein, R_immediate is the immediate benefit, which is 1 if the guidance is clicked, otherwise 0; R_long_term is the long-term benefit, which is 1 if a social relationship is successfully established within a preset time after guidance, otherwise 0; α and β are weighting coefficients, for example, α = 0.4, β = 0.6. Target object characteristics (such as VIP level) can be used to adjust weights. For example, a higher α value can be set for a high VIP level target object to emphasize immediate feedback compared to a low VIP level target object. The target social intent type and the calibrated confidence level can be used to estimate the expected value of R_long_term. For example, for strategy arm A, targeting a high VIP level target object with the intent "game couple" and a calibrated confidence level of 0.9, its historical immediate return is 0.5 and its long-term return is 0.3, then its estimated overall return R = 0.4*0.5 + 0.6*0.3 = 0.38. For strategy arm B targeting the same target object, its historical immediate return is 0.3 and its long-term return is 0.2, then its estimated overall return R = 0.4*0.3 + 0.6*0.2 = 0.24.

[0058] Step M104: Based on the overall returns of each strategy arm, select several candidate guidance strategy combinations.

[0059] Specifically, the overall return value of each strategy arm is calculated, and all strategy arms are ranked according to their overall return value. Based on preset screening rules, several strategy arms with the highest overall return values ​​are selected, and the corresponding guidance strategy combinations are determined as candidate guidance strategy combinations. It is worth noting that the screening process balances return advantage and strategy diversity, retaining both currently performing strategy combinations and some potentially effective ones, providing sufficient candidate samples for subsequent iterative optimization.

[0060] Step M105: Based on the feedback data of the target objects corresponding to each candidate guidance strategy combination, iteratively update the number of attempts and the number of successful returns of the candidate strategy arms corresponding to each candidate guidance strategy combination, and repeat the step of calculating the comprehensive return of the candidate strategy arms to determine the optimal guidance strategy combination.

[0061] Specifically, each candidate guidance strategy combination is deployed to the target audience, and feedback data from the target audience on each candidate strategy combination is collected in real time. This feedback data includes whether immediate interaction occurs, whether guidance is accepted, and whether social relationships are successfully established, which is used to evaluate the actual effect of each candidate strategy combination. Then, based on the collected feedback data, the number of attempts and the number of successful gains for the corresponding strategy arm of each candidate strategy combination are iteratively updated. If a candidate strategy combination generates immediate or long-term gains after execution, the number of successful gains for the corresponding strategy arm is updated synchronously, and the number of attempts is accumulated. Then, the comprehensive gain calculation step in step M103 is repeated to recalculate the comprehensive gain value of each candidate strategy arm. After multiple rounds of iterative updates and gain evaluations, the candidate guidance strategy combination with the highest comprehensive gain value, which is suitable for the characteristics of the target audience, matches the target's social intent, and identifies the needs corresponding to the credibility, while taking into account both guidance effect and audience experience, is finally selected as the optimal guidance strategy combination.

[0062] This example addresses multiple guidance strategy combinations that utilize arbitrary combinations of guiding text, incentive rewards, and trigger timing. A multi-armed trial optimizer employs a multi-armed trial algorithm, taking the immediate interaction of the target audience with the guidance as the immediate benefit and the successful establishment of a social relationship after the target audience accepts the guidance as the long-term benefit. The parameters of each guidance strategy combination are dynamically updated based on the benefit results. This comprehensively considers the suitability of each guidance strategy combination with the target audience, performing a comprehensive scoring and selection based on three dimensions: guiding text, incentive rewards, and trigger timing. Strategy combinations that do not match the current social scenario are eliminated, ultimately selecting the optimal guidance strategy combination that highly matches the target social intent type and calibrated confidence level.

[0063] Step S302: Determine the target guidance strategy level based on the optimal guidance strategy combination.

[0064] Specifically, after obtaining the optimal combination of guidance strategies, the system retrieves the pre-stored association mapping rules between guidance strategy combinations and guidance strategy levels. These rules pre-define the guidance strategy level to which different guidance strategy combinations belong. For example, a strong guidance level corresponds to a strategy combination with high incentives and immediate triggering, a mild guidance level corresponds to a strategy combination with lightweight copywriting and non-forced triggering, and a silent guidance level corresponds to a strategy combination with no front-end display and no back-end recording.

[0065] By matching the incentive intensity, triggering method, display format and mapping rules of the optimal guidance strategy combination, the predefined guidance strategy level to which the optimal combination belongs is located, and this level is determined as the target guidance strategy level for the current target object. The target guidance strategy level can accurately match the target object's needs and current state.

[0066] In some embodiments, step S103 generates target guidance content tailored to the target social intent type of the target object based on the target guidance strategy hierarchy and the target social intent type. (See also...) Figure 4 ,include: Step S401: Based on the target sub-intent type, obtain the basic guidance template from the pre-built guidance template library, wherein the pre-built guidance template library stores templates associated with different sub-intent categories and different guidance strategy levels.

[0067] Specifically, multiple guidance templates are pre-built according to the different relationships between different sub-intent categories and different guidance strategy levels. Each guidance template corresponds to a different sub-intent type and guidance strategy level. The multiple guidance templates are classified and stored to form a pre-built guidance template library.

[0068] Then, the guide content generator selects and retrieves the associated basic guide template from the pre-built guide template library based on the currently determined target sub-intent type, ensuring that the core direction of the template is consistent with the specific social needs of the target.

[0069] Step S402: Based on the object profile of the target object, adapt and adjust at least some of the content in the basic guide template to obtain the adapted guide content.

[0070] Specifically, the object profile information of the target object is obtained, wherein the object profile information may include at least one of the following: object level, permission level, online duration, and social history.

[0071] Based on the target audience's profile information, at least some content in the basic onboarding template is personalized and adjusted, including: adjusting at least one of the following based on one or more of the target audience's profile: the wording of the onboarding content, the type of virtual reward, or the display icon, to obtain the adapted onboarding content.

[0072] For example, for novice users, the tutorial text is simplified and basic operation prompts are added; for advanced users, the professionalism of the text is optimized and the type and amount of incentive rewards are adjusted; for users who prefer concise prompts, redundant content in the template is streamlined, making the adjusted tutorial content more in line with the target users' gaming habits, preferences, and behavioral characteristics, forming personalized tutorial content that is highly matched with the target users. One embodiment of this application predefines an adaptation adjustment rule library. For example, if the user's permission level is lower than 30, the text keyword "expert" in the template is replaced with "helpful user," and the virtual reward is replaced with "newbie gift pack"; if the user's VIP level is greater than or equal to 5, the prefix "esteemed VIP user," is added at the beginning of the text, and the reward amount is increased by 20%; if the user's social history shows that they have rejected overly strong guidance, a mild variant text is used and the pop-up prompt effect is canceled. The adjustment process includes: traversing various dimensions of the user profile, querying matching rules in the adjustment rule library, and applying them sequentially to the basic tutorial template. For example, the basic template text is "Looking for an expert to help you pass the game?" For example, for a target with a permission level of 25, a VIP level of 3, and no special social history, the "permission level below 30" rule is applied, replacing "expert" with "helpful person," resulting in the adapted copy "Looking for a helpful person to help you pass the level?", while the reward is changed to "newbie gift pack." For another target with a permission level of 80, a VIP level of 6, and who has previously refused forced guidance, the "VIP level greater than or equal to 5" and "previously refused forced guidance" rules are applied, resulting in the adapted copy "Honorable VIP, would you like to see other people also looking for teammates?", increasing the reward amount by 20%, and using a sidebar prompt.

[0073] Step S403: Based on the real-time game scene information of the target object, enhance the adapted guide content with a contextualized approach to generate the target guide content.

[0074] Specifically, the system obtains real-time game scene information about the target object, including at least one of the following: the map type currently in which the target object is located and information about other objects whose distance from the target object meets preset conditions. The information about other objects whose distance from the target object meets preset conditions includes at least one of the following: ongoing game tasks, current team status, and real-time in-game activities.

[0075] Next, it is determined whether the map type currently in which the target object is located and the target sub-intention type satisfy a predefined association relationship. This predefined association relationship refers to the pre-defined correspondence rules between the map type, the target sub-intention type, and the guidance scene. If the association relationship is satisfied, the scene enhancement rules corresponding to the map type currently in which the target object is located are retrieved. The text in the adapted guidance content is then optimized and enhanced according to these rules. This can be achieved by adding scene-related phrases or adding real-time scene dynamic information, or a combination of both. This ensures that the guidance content naturally integrates into the target object's current game scene, avoiding a disconnect between the guidance content and the scene. Ultimately, this generates target guidance content that is both tailored to the target object's personal characteristics and adapted to the real-time game scene. The scene enhancement rules corresponding to the map type currently in which the target object is located are pre-defined guidance content optimization rules that match the map type, based on the attributes of different game map types and the corresponding social needs. These rules are used to enhance the adapted guidance content in a contextualized way, making it more suitable for the current map scene and better meeting the needs of the target object.

[0076] Based on this real-time game scenario information, the presentation and display of the guidance content are adjusted to ensure that the content blends naturally into the current game scenario. For example, when a user fails a dungeon challenge, the guidance content is supplemented with relevant help prompts based on the dungeon scenario; when a user is in the team lobby scenario, the social matching prompts in the guidance content are enhanced to avoid the guidance content becoming disconnected from the scenario. Ultimately, this generates targeted guidance content that is both relevant to the user's needs and adapted to their personal characteristics and the real-time scenario.

[0077] In some embodiments, before performing step S102, step S104 is further included, specifically: The system checks whether the target object is within the boot cooldown period. If the system detects that the target object is not within the boot cooldown period, step S102 is executed. The boot cooldown period is the cooling duration triggered by the last booting of the target object based on the specified boot strategy level.

[0078] Specifically, to avoid repeatedly performing high-intensity guidance operations on the same target object within a short period, and to prevent excessive disruption to the object due to frequent guidance, ensuring the object's normal gaming experience remains undisturbed; simultaneously, to reduce the computational overhead of the system repeatedly performing intent recognition, confidence calibration, and guidance strategy determination processes, thereby improving the overall system efficiency and stability, this embodiment sets differentiated cooldown periods for different guidance strategy levels. The guidance cooldown period is the duration triggered by the last guidance of the target object based on the specified guidance strategy level, with higher guidance strategy levels having longer cooldown periods. By setting cooldown periods, the frequency of guidance triggers can be reasonably controlled while ensuring guidance effectiveness, making the guidance behavior more in line with the object's acceptance habits, thereby improving the object's acceptance and conversion rate of the guidance content.

[0079] Before executing step S102, the boot history of the target object is queried and read to obtain the boot strategy level used in the last boot operation of the target object, the actual execution time of the last boot, and the preset cooldown period corresponding to the boot strategy level. Then, the time interval is calculated based on the current system time and the last boot execution time. The time interval is compared with the cooldown period of the corresponding boot strategy level to determine whether the current time is still within the range covered by the cooldown period, thereby completing the detection of whether the target object is in the boot cooldown period.

[0080] In some embodiments, step S103 involves processing the multimodal game data of the target object using a pre-trained large language model to obtain the target social intent type and the confidence level of the target social intent type. (See also...) Figure 5 ,include: Step S501: Based on pre-built game social domain knowledge, perform structured processing on multimodal game data to generate domain knowledge-enhanced prompt words.

[0081] Specifically, firstly, pre-constructed game social domain knowledge is retrieved, including in-game social vocabulary, dialogue rules, the correspondence between scenarios and requests, and the logic of object social behavior. Then, the collected multimodal game data of the target object is cleaned and organized, invalid information is removed, and data of different types and formats is uniformly converted into a standardized structured form. Finally, the structured multimodal game data is fused and spliced ​​with the retrieved game social domain knowledge, and domain constraint information is supplemented by the social scenarios corresponding to the data to form domain knowledge-enhanced prompt words that include actual object data and domain knowledge support. This ensures that the prompt words can not only fully reflect the object's game behavior and expression, but also help the large language model accurately understand the social requests in the game scenario with the help of domain knowledge.

[0082] Step S502: Use a pre-trained large language model to process domain knowledge-enhanced prompts to obtain the target social intent type.

[0083] Specifically, domain-knowledge-enhanced prompts are input into a pre-trained large language model. The model combines its learned general semantic understanding capabilities with the game-related social domain knowledge incorporated into the prompts to parse them. The pre-trained large language model extracts key information from the prompts, including the semantics of the chat content, the needs reflected in the behavioral events, and the current state characteristics of the target. It also combines domain knowledge to identify expressions and implicit demands within the game, avoiding misjudgments of intent due to deviations from the game context. Through comprehensive analysis, reasoning, and matching of this information, the model ultimately outputs a target social intent type consistent with the target target's actual social needs, and simultaneously outputs the confidence level corresponding to this target social intent type. The confidence level is used to characterize the model's certainty about the intent recognition result, thus achieving accurate identification of the target target's social intent.

[0084] In some embodiments, multimodal game data includes chat data, behavioral event data, and object state data. Chat text data includes, but is not limited to, the content of chat text sent by the target object, the sending time, the sending channel type, and the chat dialogue context. Game behavioral event data includes, but is not limited to, the target object's in-game action sequence, the time of the action, and the game scene in which the action occurred. Examples include entering or exiting a dungeon, the number of times a dungeon challenge is failed, creating or joining a team, gifting game items, the duration of time spent in scenes such as the main city square, the marriage tree, or the dungeon entrance, and the number of times information about other objects is viewed. Real-time object state data includes the target object's static game attributes and real-time dynamic state. Static attributes include, but are not limited to, the object's level, game class, VIP payment level, game registration time, and historical social behavior records. Real-time dynamic state includes, but is not limited to, the object's current game scene, team status, online time, current game quest progress, and dungeon challenge progress.

[0085] Step S502 involves structuring the multimodal game data based on pre-built game social domain knowledge to generate domain knowledge-enhanced prompts. Please refer to [link / reference]. Figure 6 ,include: Step S601: Based on chat data and / or behavioral event data, obtain preliminary social intent recognition results.

[0086] Specifically, chat data and behavioral event data of the target object are extracted. The chat data is subjected to preliminary semantic analysis to identify socially related keywords, tone tendencies, and core demands. The behavioral event data is subjected to feature analysis to determine whether the object's behavior points to specific social needs (such as frequently initiating team-up requests, asking about social-related operations, etc.). Based on the feature analysis results of chat data and behavioral event data, or by combining the feature association results of these two types of data, a rough preliminary social intent identification result is obtained to clarify the general direction of the object's social demands.

[0087] Step S602: Based on the preliminary social intent recognition results, retrieve relevant domain knowledge fragments from the pre-constructed game social domain knowledge graph.

[0088] Specifically, a game social domain knowledge graph (DKG) is constructed in advance using data such as game social scene entities, social intent categories, sub-intent types, object behavior characteristics, chat semantic keywords, social rule constraints, and the relationships between various game social scene entities. This ensures that the game social domain knowledge graph covers the classification information of different social scenes in the game, the hierarchical relationship between macro social intents and sub-intents, the object behavior event characteristics that can represent various social intents, the keywords and expressions in the corresponding chat text, the object state attributes under different social intents, and the association mapping relationship between various behaviors, texts, states and corresponding social intents. It also includes general rules and scene constraint information for social interaction in the game.

[0089] Based on the preliminary social intent recognition results, a pre-built game social domain knowledge graph is invoked. Using the preliminary social intent as the search keyword, a targeted search is performed in the knowledge graph to filter out domain knowledge fragments that are highly relevant to the preliminary intent and can help to accurately understand the object's needs. Knowledge content that is irrelevant to the current intent is removed to ensure that the retrieved domain knowledge fragments can accurately match the object's preliminary social needs.

[0090] Step S603: Assemble chat data, behavioral event data, object state data, and related domain knowledge fragments to generate domain knowledge-enhanced prompts.

[0091] Specifically, chat data, behavioral event data, and object status data are standardized to ensure consistent data formats, complete information, and a clear presentation of the object's content, behavioral characteristics, and current game state. For example, the target object's chat data is first semantically parsed to serve as a prompt describing the object's needs; secondly, behavioral event data is organized into behavioral sequence descriptions and appended to the prompt describing the object's needs to demonstrate the object's actual operational characteristics; finally, object status data is organized into status descriptions and appended to the behavioral sequence descriptions to supplement the object's current basic game information.

[0092] Then, the retrieved relevant domain knowledge fragments are placed at the end of the prompt words as knowledge constraint descriptions to supplement the knowledge and rule descriptions in the game's social scenario. Through the above hierarchical and orderly splicing and integration, the final result is a domain knowledge-enhanced prompt word that is complete in information, logically clear, contains both actual object data and domain knowledge support, and is encapsulated in a dynamic prompt word assembler to ensure that the subsequent pre-trained large language model can accurately understand the object's social intent in the game scenario.

[0093] In other embodiments, obtaining the target object's multi-turn dialogue history within a preset time window includes: using the game's dialogue data storage module to filter all dialogue records corresponding to the target object's unique identifier and whose timestamps fall within the preset time window, where the preset time window can be pre-configured according to the game's social scenario requirements, such as the most recent 10 minutes or the most recent 30 minutes; then, sorting the filtered dialogue records and integrating them according to the chronological order of the dialogue occurrences to form the target object's multi-turn dialogue history, where the dialogue history covers all relevant interactive content such as dialogues initiated by the target object, dialogues received, and group chats participated in; finally, performing preliminary analysis of the multi-turn dialogue history to extract key information such as key semantics and social appeal tendencies in the dialogue, and synchronously linking it to the target object's real-time game scenario information and social intent data to ensure that the guidance content is more in line with the target object's immediate social needs and dialogue context.

[0094] In this embodiment, the multimodal game data includes chat data, behavioral event data, and object state data, as well as multi-turn dialogue history. This embodiment assembles the chat data, behavioral event data, object state data, multi-turn dialogue history of the target object, and related domain knowledge fragments from the multimodal game data to generate domain knowledge-enhanced prompts.

[0095] It is worth noting that the step of assembling chat data, behavioral event data, object state data, the target object's multi-turn dialogue history, and related domain knowledge fragments to generate domain knowledge-enhanced prompts differs from the step of assembling chat data, behavioral event data, object state data, and related domain knowledge fragments to generate domain knowledge-enhanced prompts only in that: the multi-turn dialogue history of the target object is semantically parsed in chronological order and concatenated with the object request description appended to the prompt, along with the chat data, behavioral event data, and object state data, to form the object request description information of the prompt; and the related domain knowledge fragments are placed at the end of the prompt as knowledge constraint descriptions, together generating the domain knowledge-enhanced prompts.

[0096] In some embodiments, the method further includes a step of optimizing domain knowledge-enhanced prompts; see [link to relevant documentation]. Figure 7 Specifically, it includes: Step 701: Based on the feedback data of the target object after execution guidance, calculate at least one evaluation metric for the version corresponding to the domain knowledge-enhanced prompt words.

[0097] Specifically, feedback data from the target audience after the guided session is collected. This feedback data includes at least one or more of the following: whether the target audience clicked on the guided interface, whether they entered the recommended functional system, and whether they successfully established a social relationship within a preset time window. Then, based on the collected feedback data, corresponding evaluation indicators are calculated. If the feedback data includes records of clicking on the guided interface and entering the recommended functional system, the guided acceptance rate is calculated, which is the ratio of the number of target audiences who clicked on the guided session and entered the functional system to the total number of target audiences who received the guided session. If the feedback data includes the establishment of social relationships within the preset time window, the social relationship establishment conversion rate is calculated, which is the ratio of the number of target audiences who successfully established social relationships to the total number of target audiences who received the guided session. If the feedback data includes multiple items, multiple corresponding evaluation indicators are calculated simultaneously.

[0098] Step 702: With the goal of maximizing the evaluation index, use an optimization algorithm to search for parameters of multiple adjustable components that constitute the domain knowledge enhancement prompt words, and obtain the optimized component configuration; based on the optimized component configuration, generate a new version of the domain knowledge enhancement prompt words.

[0099] Specifically, with the goal of maximizing evaluation metrics, an optimization algorithm is used to conduct a comprehensive parameter search on multiple adjustable components that constitute the domain knowledge-enhanced prompts. The effects of different component parameter combinations are tested one by one, and the component configuration that achieves the optimal evaluation metrics is selected. The adjustable components refer to the constituent parts of the domain knowledge-enhanced prompts that can be adjusted and optimized by modifying their parameters to optimize the prompt effect. Based on the optimized component configuration obtained by the selection, a new version of the domain knowledge-enhanced prompts is regenerated. This retains the original domain knowledge and better adapts to the needs of the target audience, thereby improving the guidance effect and the accuracy of social intent recognition, and thus better adapting to the dynamic changes of game social scenarios and the iteration of target audience needs.

[0100] Optionally, after completing the optimization and adjustment, assign a new version identifier to the domain knowledge enhancement prompts of the new version, record the content of the version update, the optimization direction and the corresponding analysis basis, and retain the prompts and related analysis data of the previous version to facilitate subsequent continuous comparison and optimization.

[0101] Please see Figure 8 , Figure 8 The diagram shown is an exemplary flowchart of a social guidance method, which specifically includes: First, the data input layer on the game server collects real-time chat streams, object behavior events, and object state snapshots to construct multimodal game data, which is then input into the knowledge enhancement and intelligent analysis layer on the game server. In the knowledge enhancement and intelligent analysis layer, a context manager maintains the target object's session context. Combined with game social domain knowledge provided by the domain knowledge graph, a dynamic prompt word assembler generates domain knowledge-enhanced prompt words. A large language model inference engine then identifies social intent and outputs confidence scores, which are then calibrated by a confidence calibrator. Next, in the adaptive guidance layer on the game server, a hierarchical strategy adapter matches the target guidance strategy level. A multi-armed trial optimizer selects the optimal combination of guidance strategies, and finally, a guidance content generator generates suitable guidance content and pushes it to the game client. Finally, in the iterative optimization layer on the game server, an effect tracker collects object feedback and guidance effect data. After prompt word heatmap analysis, an automated controlled experiment (A / B testing) framework iteratively optimizes the dynamic prompt words and guidance strategies, forming a complete closed-loop game social guidance process.

[0102] On the other hand, this embodiment provides a social guidance system, please refer to... Figure 9 ,include: The social intent recognition module 901 is used to perform intent recognition processing on the multimodal game data of the target object in the game using a pre-trained large language model, and to obtain the target social intent type and the confidence level of the target social intent type. The guidance strategy determination module 902 is used to determine the target guidance strategy level for the target object from multiple predefined guidance strategy levels based on the target social intent type and confidence level; The guidance module 903 is used to generate target guidance content for the target social intent type of the target audience based on the target guidance strategy hierarchy and the target social intent type.

[0103] This embodiment also provides an electronic device, including a memory and a processor. In a specific example, the memory stores a computer program, and when the computer program is executed by the processor, it implements the method of any of the above embodiments.

[0104] This embodiment also provides a computer-readable storage medium having a computer program stored thereon, the computer program being loaded by a processor to perform the steps of any of the methods in the above embodiments.

[0105] In the embodiments of this application, the storage medium may be a magnetic disk, an optical disk, a read-only memory (ROM), or a random access memory (RAM), etc.

[0106] It should be noted that, in the data processing stage, the technical solution of this application has strictly limited the scope of data collection to the minimum necessary to achieve the technical objectives, preventing the acquisition of irrelevant information. For any user information to be collected, the data subject will be clearly informed and their consent obtained. Furthermore, technologies such as encrypted storage and access control are employed to strengthen data security and ensure the security and compliance of the entire data processing process. The technical model and decision-making mechanism are based on objective technical parameters and do not introduce unnecessary parameters such as gender or age that may lead to discrimination, resolutely eliminating algorithmic discrimination and upholding public order and good morals. In addition, the specification fully describes the technical implementation methods, application scenarios, and compliance protection details. The claims are consistent with the content of the specification, key compliance designs are clear and verifiable, and the overall technical design is guided by the protection of public interests and adherence to social ethics, without any circumstances that harm public interests or violate public order and good morals.

[0107] In the above embodiments, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.

[0108] The above provides a detailed description of a social guidance method and system provided by the embodiments of this application. Specific examples have been used to illustrate the principles and implementation methods of this application. The description of the above embodiments is only for the purpose of helping to understand the method and core ideas of this application. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of this application. Therefore, the content of this specification should not be construed as a limitation of this application.

Claims

1. A social guidance method, characterized in that, Includes the following steps: A pre-trained large language model is used to perform intent recognition processing on the multimodal game data of the target object to obtain the target social intent type of the target object and the confidence level corresponding to the target social intent type; Based on the target social intent type and the confidence level, a target guidance strategy level for the target object is determined from multiple predefined guidance strategy levels; Based on the target guidance strategy hierarchy and the target social intent type, target guidance content is generated for the target object.

2. The social guidance method according to claim 1, characterized in that, The step of determining the target guidance strategy level for the target object from multiple predefined guidance strategy levels based on the target social intent and the confidence level includes: Select a target pre-trained confidence calibrator corresponding to the target social intent from multiple pre-trained confidence calibrators. The target social intent type includes a target macro intent type and the target sub-intent type of the target object under the macro intent type. The confidence level is calibrated using the target pre-trained confidence calibrator to obtain the calibrated confidence level; Based on the intent priority corresponding to the target social intent and the post-calibration confidence level, the target guidance strategy level of the target object is determined from multiple predefined guidance strategy levels.

3. The social guidance method according to claim 2, characterized in that, The multiple pre-trained confidence calibrators are trained independently for each macro-intent category based on labeled historical datasets.

4. The social guidance method according to claim 2, characterized in that, The step of determining the target guidance strategy for the target object from multiple predefined guidance strategy levels based on the intent priority corresponding to the target social intent and the calibrated confidence level includes: The calibrated confidence level is compared with at least one preset confidence threshold, and the target guidance strategy level is determined from the plurality of predefined guidance strategy levels in combination with the intent priority corresponding to the intent type.

5. The social guidance method according to claim 4, characterized in that, The method further includes: A reinforcement learning selection algorithm is used to screen several combinations of guidance strategies corresponding to the target guidance strategy level to determine the optimal combination of guidance strategies. The combination of guidance strategies can be any combination of guidance text, incentive rewards, and trigger timing. Based on the optimal combination of guidance strategies, the target guidance strategy level is determined.

6. The social guidance method according to claim 5, characterized in that, The reinforcement learning selection algorithm is a multi-armed trial algorithm. The step of using the reinforcement learning selection algorithm to screen several combinations of guidance strategies corresponding to the target guidance strategy level to determine the optimal guidance strategy combination includes: The several combinations of guidance strategies corresponding to the target guidance strategy level are respectively modeled as strategy arms in the multi-arm probing algorithm; Initialize the number of attempts and the number of successful rewards for each of the strategy arms, and define the comprehensive reward for each of the strategy arms, wherein the comprehensive reward includes immediate reward and long-term reward; Based on the feature information of the target object, the type of the target social intent, and the post-calibration confidence, the comprehensive benefit of each strategy arm is calculated; Based on the overall benefits of each strategy arm, several candidate guidance strategy combinations are selected; Based on the feedback data of the target object corresponding to each candidate guidance strategy combination, the number of attempts and the number of successful gains of the candidate strategy arm corresponding to each candidate guidance strategy combination are iteratively updated, and the step of calculating the comprehensive gain of the candidate strategy arm is repeatedly executed to determine the optimal guidance strategy combination.

7. The social guidance method according to claim 2, characterized in that, Based on the target guidance strategy hierarchy and the target social intent type, generate target guidance content for the target object's target social intent type, including: Based on the target sub-intent type, a basic guidance template is obtained from a pre-built guidance template library, wherein the pre-built guidance template library stores templates associated with different sub-intent categories and different guidance strategy levels; Based on the object profile of the target object, at least some of the content in the basic guide template is adapted and adjusted to obtain the adapted guide content. Based on the real-time game scene information of the target object, the adapted guidance content is enhanced with contextualization to generate target guidance content.

8. The social guidance method according to claim 7, characterized in that, The object profile of the target object includes at least one of the following: object level, permission level, online duration, and social history; Based on the object profile of the target object, at least some content in the basic guide template is adapted and adjusted, including: Based on one or more of the object profiles, at least one of the following adjustments is made to the text description of the guidance content, the type of virtual reward, or the display icon.

9. The social guidance method according to claim 7, characterized in that, The real-time game scene information of the target object includes at least one of the following: the map type where the target object is currently located and information about other objects whose distance from the target object meets preset conditions; Based on the real-time game scene information of the target object, the adapted guidance content is enhanced with contextualization, including: When the map type where the target object is currently located satisfies a predefined association relationship with the target sub-intention type, scene-related phrases and / or dynamic information are added to the text description in the adapted guidance content according to the scene enhancement rules corresponding to the map type.

10. The social guidance method according to claim 1, characterized in that, The step of determining the target guidance strategy level for the target object from multiple predefined guidance strategy levels based on the target social intent type and the confidence level includes: In response to detecting that the target object is not within the guidance cooldown period, the step of determining the target guidance strategy level for the target object from multiple predefined guidance strategy levels based on the target social intent type and the confidence level is executed; The bootstrapping cooldown period is the cooldown duration triggered by the last bootstrapping of the target object based on the specified bootstrapping strategy level.

11. The social guidance method according to claim 10, characterized in that, Different boot strategy levels have different cooldown times.

12. The social guidance method according to any one of claims 1-11, characterized in that, The process of using a pre-trained large language model to perform intent recognition processing on multimodal game data of target objects within the game, to obtain the target social intent type of the target object and the confidence level of the target social intent type, includes: Based on pre-constructed game social domain knowledge, the multimodal game data is structured to generate domain knowledge-enhanced prompt words; The domain knowledge-enhanced prompts are processed using the pre-trained large language model to obtain the target social intent type.

13. The social guidance method according to claim 12, characterized in that, The multimodal game data includes chat data, behavioral event data, and object state data. Based on pre-built game social domain knowledge, the multimodal game data undergoes structured processing to generate domain knowledge-enhanced prompts, including: Based on the chat data and / or the behavioral event data, a preliminary social intent recognition result is obtained; Based on the preliminary social intent recognition results, relevant domain knowledge fragments are retrieved from the pre-constructed game social domain knowledge graph; The chat data, behavioral event data, object state data, and associated domain knowledge fragments are assembled to generate domain knowledge-enhanced prompts.

14. The social guidance method according to claim 13, characterized in that, The method further includes: Obtain the multi-turn dialogue history of the target object within a preset time window; The chat data, behavioral event data, object state data, multi-turn dialogue history of the target object, and related domain knowledge fragments are assembled to generate domain knowledge-enhanced prompt words.

15. The social guidance method according to claim 1, characterized in that, The method further includes: Based on the target guidance content, execute the guidance action corresponding to the target guidance strategy level.

16. The social guidance method according to claim 2, characterized in that, The boot strategy hierarchy includes at least one of a first boot hierarchy, a second boot hierarchy, and a third boot hierarchy, wherein... The triggering condition for the first guidance level is that the post-calibration confidence level is higher than the first confidence level threshold, and the target social intent type belongs to a high-priority type; The triggering condition for the second guidance level is that the post-calibration confidence level is between the second confidence threshold and the first confidence threshold, and the first confidence threshold is greater than the second confidence threshold; The trigger condition for the third guidance level is that the second confidence level is lower than the second confidence level threshold.

17. The social guidance method according to claim 16, characterized in that, The guiding action corresponding to the first guiding level is to trigger the front-end active guiding interface; The guidance action corresponding to the second guidance level is to trigger a front-end prompt; The guidance actions corresponding to the third guidance level are recorded in the background.

18. The social guidance method according to any one of claims 15-17, characterized in that, The method further includes: Based on the feedback data of the target object after the execution guidance, calculate at least one evaluation index for the version corresponding to the domain knowledge enhanced prompt words; With the goal of maximizing the evaluation index, an optimization algorithm is used to search for parameters of multiple adjustable components that constitute the domain knowledge enhancement prompts, resulting in an optimized component configuration; based on the optimized component configuration, a new version of the domain knowledge enhancement prompts is generated.

19. The social guidance method according to claim 18, characterized in that, The feedback data includes at least one of the following: whether the target object clicks on the guidance interface, whether it enters the recommended function system, and whether it successfully establishes a social relationship within a preset time window; the evaluation indicators include at least one of the guidance acceptance rate and the social relationship establishment conversion rate.

20. A social guidance system, characterized in that, include: The social intent recognition module is used to perform intent recognition processing on the multimodal game data of the target object in the game using a pre-trained large language model, so as to obtain the target social intent type of the target object and the confidence level of the target social intent type; The guidance strategy determination module is used to determine the target guidance strategy level for the target object from multiple predefined guidance strategy levels based on the target social intent type and the confidence level; The guidance module is used to generate target guidance content for the target social intent type of the target object based on the target guidance strategy hierarchy and the target social intent type.

Images