Method, device and electronic equipment for generating user interest tags

Abstract

The present disclosure provides a user interest label generation method, relates to the technical field of computers, and particularly relates to the technical fields of large models, user interest labels, supervised models, agents, and the like. The specific implementation scheme is as follows: based on user data of a target user, at least one candidate interest label is generated through a large language model; the candidate interest label is input into a cognitive bias simulator to obtain a bias report for representing an evaluation result of the candidate interest label on a preset psychological cognitive bias dimension; historical contradictory case data related to the target user is retrieved from a preset counterfactual experience replay pool; the historical contradictory case data is case data in which system evaluation is inconsistent with user feedback in a historical generation process; at least based on the bias report and the historical contradictory case data, the large language model is driven to perform adversarial mutation on the candidate interest label to generate at least one mutated target interest label.

Description

Technical Field

[0001] This disclosure relates to the field of computer technology, and in particular to the fields of large models, user interest tags, supervised models, and intelligent agents. Background Technology

[0002] In the field of personalized content recommendation, user interest tags are the core foundation for achieving accurate distribution and improving user experience. Traditional user interest generation solutions mostly rely on large language models to generate interest tags in a single instance, or only conduct objective evaluation and positive optimization based on user explicit clicks, views, and other behavioral data and content popularity. Such solutions can only learn users' surface behavioral preferences and cannot reach users' deep psychological motivations. The generated interest tags are prone to homogenization and cannot meet the recommendation system's requirements for the accuracy and diversity of interest tags. Summary of the Invention

[0003] This disclosure provides a method, apparatus, and electronic device for generating user interest tags to solve at least one of the above-mentioned technical problems.

[0004] According to a first aspect of this disclosure, a method for generating user interest tags is provided, wherein the method includes: Based on the user data of the target users, at least one candidate interest tag is generated through a large language model; The candidate interest tags are input into the cognitive bias simulator to obtain a bias report that characterizes the evaluation results of the candidate interest tags on a preset psychological cognitive bias dimension. Retrieve historical conflict case data related to the target user from a pre-set counterfactual experience replay pool; wherein, the historical conflict case data are case data where the system evaluation and user feedback are inconsistent during the historical generation process; Based at least on the deviation report and the contradictory case data, the large language model is driven to perform adversarial mutation on the candidate interest tags to generate at least one mutated target interest tag.

[0005] According to a second aspect of this disclosure, an apparatus for generating user interest tags is provided, wherein the apparatus includes: The generation module is used to generate at least one candidate interest tag based on the user data of the target user through a large language model. The bias module is used to input the candidate interest tags into the cognitive bias simulator and obtain a bias report that characterizes the evaluation results of the candidate interest tags on a preset psychological cognitive bias dimension. The retrieval module is used to retrieve historical conflict case data related to the target user from a preset counterfactual experience replay pool; wherein, the historical conflict case data are case data in which the system evaluation and user feedback are inconsistent during the historical generation process; The mutation module is used to drive the large language model to perform adversarial mutation on the candidate interest tags based at least on the deviation report and the contradictory case data, so as to generate at least one mutated target interest tag.

[0006] According to a third aspect of this disclosure, an electronic device is provided, comprising: At least one processor; and A memory communicatively connected to the at least one processor; wherein, The memory stores instructions that can be executed by the at least one processor, which, when executed by the at least one processor, enables the at least one processor to perform the above-described method.

[0007] According to a fourth aspect of this disclosure, a non-transitory computer-readable storage medium is provided storing computer instructions, wherein the computer instructions are used to cause the computer to perform the method described above.

[0008] According to a fifth aspect of this disclosure, a computer program product is provided, comprising a computer program that, when executed by a processor, implements the method described above.

[0009] It should be understood that the description in this section is not intended to identify key or essential features of the embodiments of this disclosure, nor is it intended to limit the scope of this disclosure. Other features of this disclosure will become readily apparent from the following description. Attached Figure Description

[0010] The accompanying drawings are provided to better understand this solution and do not constitute a limitation of this disclosure. Wherein: Figure 1 This is a flowchart illustrating a method for generating user interest tags according to the first embodiment of this disclosure; Figure 2 This is an exemplary process diagram of S101; Figure 3 This is an exemplary process diagram of S102; Figure 4 This is an exemplary flowchart of S103; Figure 5 This is an exemplary flowchart of S104; Figure 6 This is a schematic diagram of the structure of a user interest tag generation device provided in the second embodiment of this disclosure; Figure 7 This is a block diagram of an electronic device used to implement the methods of the embodiments of this disclosure. Detailed Implementation

[0011] The exemplary embodiments of this disclosure are described below with reference to the accompanying drawings, including various details of the embodiments to aid understanding, and should be considered merely exemplary. Therefore, those skilled in the art will recognize that various changes and modifications can be made to the embodiments described herein without departing from the scope and spirit of this disclosure. Similarly, for clarity and brevity, descriptions of well-known functions and structures are omitted in the following description.

[0012] Where there is no conflict, the various embodiments of this disclosure and the features thereof in the embodiments may be combined with each other.

[0013] As used herein, the term “and / or” includes any and all combinations of one or more related enumerated entries.

[0014] The terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit this disclosure. As used herein, the singular forms “a” and “the” are also intended to include the plural forms, unless the context clearly indicates otherwise.

[0015] Unless otherwise specified, all terms used herein (including technical and scientific terms) have the same meaning as commonly understood by one of ordinary skill in the art. It will also be understood that terms such as those defined in commonly used dictionaries should be interpreted as having a meaning consistent with their meaning in the context of the relevant art and this disclosure, and will not be interpreted as having an idealized or overly formal meaning, unless expressly so defined herein.

[0016] The method for generating user interest tags according to this disclosure can be executed by an electronic device such as a terminal device or a server. The terminal device can be an in-vehicle device, user equipment (UE), mobile device, user terminal, terminal, cellular phone, cordless phone, personal digital assistant (PDA), handheld device, computing device, wearable device, etc. The method can be implemented by a processor calling computer-readable program instructions stored in memory. Alternatively, the method for generating user interest tags provided in this disclosure can be executed by a server.

[0017] In the first disclosed embodiment, see Figure 1 , Figure 1 The diagram shows a flowchart of a method for generating user interest tags according to a first embodiment of this disclosure. The method includes: S101. Based on the user data of the target user, generate at least one candidate interest tag through a large language model.

[0018] In this step, the target user refers to the user whose exclusive interest tag needs to be generated; user data refers to relevant data used to characterize the target user's features and interests, including user profiles, user historical interest data, and currently popular content data, etc., which are not limited here; large language model refers to a large language model with semantic understanding, content generation, and logical reasoning capabilities; candidate interest tags refer to interest labels initially generated based on user data that have not yet undergone bias evaluation and optimization; preset refers to the parameters, rules, or data sets that are pre-set according to business needs and model training rules before the method is executed.

[0019] In practice, user data of the target users is first collected and organized. After the user data is integrated into the dynamic prompt information, it is input into the large language model. The large language model generates at least one candidate interest tag based on user profile, historical interest behavior and current popular content characteristics, thus completing the initial generation of interest tags and providing basic data for subsequent deviation evaluation and variation optimization.

[0020] S102. Input the candidate interest labels into the cognitive bias simulator to obtain a bias report that characterizes the evaluation results of the candidate interest labels on the preset psychological cognitive bias dimension.

[0021] In this step, the cognitive bias simulator is a model specifically designed to simulate the subjective cognitive distortion of interest tags by users under different psychological cognitive states. It evaluates the input candidate interest tags from multiple preset psychological cognitive bias dimensions, and it is not a traditional scoring evaluation model. The preset psychological cognitive bias dimensions refer to the dimensions that are pre-set to measure the user's psychological cognitive bias. The bias report is a file that records the evaluation results of the candidate interest tags on each psychological cognitive bias dimension, and is used to reflect the user's psychological bias risk in the tags.

[0022] In practice, the candidate interest tags generated in S101 are input into a cognitive bias simulator. The simulator evaluates the candidate interest tags based on preset psychological cognitive bias dimensions. Each dimension can have a predicted score, which represents the potential strength of a candidate interest tag in triggering user psychological bias in that dimension. The cognitive bias simulator generates corresponding counterfactual evaluation descriptions based on the predicted scores of each dimension, such as a piece of natural language text explaining that "although this tag matches the user's historical interests, the user has recently been exposed to too much similar content, resulting in aesthetic fatigue bias, and the actual click-through rate may decrease." The predicted scores and the counterfactual evaluation descriptions are used together as the evaluation result to form the bias report. This bias report can intuitively reflect the potential problems of candidate interest tags at the user's psychological level, providing a bias basis for subsequent adversarial mutation.

[0023] Among them, the preset psychological cognitive bias dimensions include at least one of the following: novelty bias dimension, controversy bias dimension, emotional tendency bias dimension, conformity pressure bias dimension, aesthetic fatigue bias dimension, and curiosity bias dimension.

[0024] Furthermore, the predefined psychological cognitive bias dimensions refer to a set of predefined and measurable indicators used to quantify users' subjective psychological reactions. The cognitive bias simulator evaluates candidate interest tags from multiple perspectives based on these dimensions to capture the irrational cognitive distortions that users may exhibit in real interactions.

[0025] Specifically, the pre-defined dimensions of psychological cognitive bias include at least one of the following: Novelty Bias Dimension: This dimension measures the novelty of candidate interest tags relative to the target user's historical content exposure. Tags with low novelty may cause user fatigue, while tags with high novelty may be ignored due to a lack of cognitive basis. The Cognitive Bias Simulator evaluates this dimension's score by calculating the semantic differences between candidate tags and the user's historical tags.

[0026] Controversy Bias Dimension: This dimension assesses the degree of controversy surrounding the topics covered by candidate interest tags. Highly controversial tags may stimulate users' desire to discuss, but they may also cause users to actively avoid them to avoid conflict. The cognitive bias simulator predicts the bias score for this dimension based on the degree of disagreement among opinions regarding the tags in the public corpus and users' historical sensitivity to controversial content.

[0027] Emotional Tendency Bias Dimension: This dimension is used to determine the emotional polarity (e.g., positive, negative, or neutral) evoked by candidate interest tags. Tags with different emotional tendencies have significantly different impacts on users' willingness to click. The cognitive bias simulator extracts the emotional features of tags through a sentiment analysis model and combines this with users' historical emotional preferences to calculate the score for this dimension.

[0028] Conformity pressure bias dimension: This dimension assesses the social impact of candidate interest tags due to their popularity within a group. Even if a tag essentially matches a user's interests, if it's too niche, users may hesitate to click for fear of deviating from the group; conversely, overly popular tags may receive abnormally high click rates due to the "herd mentality." The cognitive bias simulator combines statistics on tag popularity across the entire internet or social circles with users' historical reaction patterns to popular / niche content to calculate the bias score for this dimension.

[0029] Aesthetic fatigue bias dimension: This dimension measures the degree of repetition between candidate interest tags and content that users have recently and frequently encountered. When users see tags with similar styles, topics, or expressions multiple times in a row, they experience psychological fatigue, leading to a decrease in actual click intent. The cognitive bias simulator evaluates this dimension's score by calculating the semantic similarity and frequency of occurrence between candidate tags and users' recent historical tags.

[0030] Curiosity Bias Dimension: This dimension assesses the ability of candidate interest tags to stimulate users' desire to explore. Tags containing rare, mysterious, or counterintuitive elements are more likely to arouse users' curiosity, thus pushing them beyond their usual interest boundaries. The cognitive bias simulator predicts the score for this dimension based on the density of rare words in the tags, the strength of cross-domain associations, and users' historical response preferences to novel content.

[0031] The cognitive bias simulator calculates a predicted score for each candidate interest tag across one or more of the aforementioned dimensions. The scores for each dimension, along with corresponding counterfactual assessment descriptions, are combined to form a bias report for use in subsequent adversarial mutation steps. Examples of counterfactual assessment descriptions include: the tag is too novel, exceeding the user's cognitive safety zone, and may be ignored; or, the tag presents strong herd pressure, causing users to be curious but hesitant to click.

[0032] S103. Retrieve historical conflict case data related to the target user from the preset counterfactual experience replay pool; wherein, the historical conflict case data are case data in which the system evaluation and user feedback are inconsistent during the historical generation process.

[0033] In this step, a counterfactual experience replay pool is pre-constructed. This pool is a storage space for case data generated during the historical generation process where the system evaluation results and actual user feedback are inconsistent. System evaluation refers to the quantitative scores given to individual interest tags by a cognitive bias simulator or other evaluation modules in the past; user feedback refers to the actual behavioral data exhibited by users after using the interest tag for online recommendations, such as clicks, viewing time, and interactions. When there is a significant difference between the system evaluation score and user feedback, the case is marked as a historical contradictory case, and the data is stored in the counterfactual experience replay pool. For example, a case might have a high system evaluation score but poor actual user feedback (a case of missed opportunity fear), or a low system evaluation score but the user later accidentally expressed interest (an case of unexpected positive outcome). In step S103, based on the target user's user identifier, historical contradictory cases related to that user are queried from the counterfactual experience replay pool. If multiple cases are found, they can be sorted according to a preset sorting rule, and a preset number of cases (e.g., the first 3 or 5) are taken as the retrieval results.

[0034] In practice, based on the user identifier of the target user, the corresponding historical contradictory case data is accurately retrieved from the preset counterfactual experience replay pool. A specified number of valid cases can be selected according to the preset case sorting rules. These cases record the situation where the model evaluation in the historical generation contradicts the user's actual needs, providing counterfactual learning basis for subsequent adversarial mutations.

[0035] S104. Based at least on deviation reports and historical conflict case data, drive the large language model to perform adversarial mutations on candidate interest tags to generate at least one mutated target interest tag.

[0036] In this step, adversarial mutation refers to the tag optimization method that corrects psychological cognitive biases of tags based on deviation reports, avoids historical errors by combining contradictory case data, and explores users' potential interests; target interest tags refer to the final interest tags that, after adversarial mutation, eliminate cognitive biases and fit the users' real needs.

[0037] In practice, deviation reports are transformed into deviation prompts, and contradictory case data are transformed into case prompts. After the adversarial prompt words are integrated and constructed, they are input into a large language model to drive the model to perform adversarial mutation on candidate interest tags, generate at least one mutated target interest tag, and complete the final optimization and generation of interest tags.

[0038] The method disclosed herein generates candidate interest tags based on user data of target users, obtains a bias report representing the assessment results of psychological cognitive biases using a cognitive bias simulator, and retrieves historical contradictory cases related to users from a counterfactual experience replay pool. Then, based at least on the bias report and historical contradictory case data, it drives a large language model to perform adversarial mutation, generating mutated target interest tags. This method enables the generated interest tags to effectively resist psychological biases such as conformity and aesthetic fatigue, while also drawing on lessons learned from failures or unexpected surprises in historical contradictory cases. This overcomes the limitations of traditional methods that rely solely on successful experiences and superficial popularity, enhancing the novelty and diversity of interest discovery and alleviating the information cocoon problem.

[0039] See in some examples Figure 2 , Figure 2 An exemplary flow diagram of S101 is shown. S101 includes: S1011. Obtain user data of the target user, including at least one of user profile, historical interest data, and currently popular content data.

[0040] In this step, the target user refers to the user whose exclusive interest tags need to be generated; user data refers to comprehensive data used to characterize the target user's identity features, behavioral preferences, and content environment; user profile refers to feature tags pre-constructed based on the target user's basic information, specifically including information that can represent the user's basic attributes such as gender, age, city, and life status; historical interest data refers to the target user's historical behavior data within the platform collected according to a preset statistical period, specifically the user's topic interest data for the past week; currently popular content data refers to content data with high preference among the target user's group across the entire network, collected according to a preset scope; preset refers to the data collection period, collection scope, and filtering rules pre-set according to business needs and data processing specifications before performing data acquisition operations.

[0041] In practice, according to the preset data acquisition rules, one or more of the target user's user profile, historical interest data, and current popular content data are collected to comprehensively obtain the basic information that supports the generation of interest tags, ensuring that the tags generated subsequently match the user's real characteristics and current content trends.

[0042] S1012. Construct dynamic prompt words based on user data. Dynamic prompt words include at least one of the following: task description, user profile, historical interest data, and currently popular content data.

[0043] In this step, dynamic prompts refer to prompts that are flexibly adjusted based on real-time user data of the target user. They are not fixed and are used to guide the large language model to perform precise generation tasks. The task description refers to the pre-set text instructions used to clarify the core task of the large language model in generating user interest tags. The preset refers to the prompt splicing template and content combination rules that are pre-set according to the model interaction rules before constructing the prompts.

[0044] In practice, based on the user data obtained by S1011, a template is built according to the preset prompt words. The task description, user profile, historical interest data, and current popular content data are integrated to generate exclusive dynamic prompt words that are suitable for the current target user. The dynamic prompt words can be adjusted in real time according to changes in user data, avoiding the problem of rigid generation results caused by fixed prompt words.

[0045] S1013. Input the dynamic prompt words into the large language model to generate candidate interest tags.

[0046] In this step, candidate interest tags refer to initial interest identifiers generated based on dynamic prompts but not yet subjected to cognitive bias assessment and adversarial mutation optimization. Specifically, the dynamic prompts constructed in S1012 are input into the large language model. The large language model, based on the task instructions, user characteristics, historical interest preferences, and currently popular content information in the dynamic prompts, performs an interest tag generation operation, outputting at least one candidate interest tag, thus completing the initial generation of interest tags.

[0047] The above steps, by acquiring user data step by step, constructing personalized dynamic prompts, and driving a large language model to generate candidate interest tags, can fully combine user basic attributes, historical behavior, and real-time trends to generate initial tags. By using dynamic prompts instead of traditional fixed prompts, the personalization and adaptability of candidate interest tag generation are greatly improved, accurately covering users' explicit interest features. This provides reliable basic data for subsequent cognitive bias simulation, counterfactual case retrieval, and adversarial mutation optimization, ensuring that the initial stage of the entire interest tag generation process is efficient and accurate.

[0048] See in some examples Figure 3 , Figure 3 An exemplary flow diagram of S102 is shown. S102 includes: S1021. Input the candidate interest labels into the cognitive bias simulator. The cognitive bias simulator calculates the predicted scores of the candidate interest labels on each dimension from the preset set of psychological cognitive bias dimensions.

[0049] In this step, the preset set of psychological cognitive bias dimensions is a set of multiple dimensions of user psychological cognitive bias. This set includes at least one of the following: novelty bias dimension, controversy bias dimension, emotional tendency bias dimension, conformity pressure bias dimension, aesthetic fatigue bias dimension, and curiosity bias dimension. The prediction score refers to the quantitative value used to characterize the degree of bias calculated by the cognitive bias simulator for each psychological cognitive bias dimension for candidate interest tags.

[0050] In practice, the generated candidate interest tags are input into the cognitive bias simulator. The simulator calls the preset set of psychological cognitive bias dimensions and performs quantitative calculations on each candidate interest tag on each of the psychological cognitive bias dimensions included in the set. This yields the predicted score for the candidate interest tag on each dimension, thus completing the quantitative assessment of the psychological cognitive bias of the candidate interest tag.

[0051] For example, the predicted score can be obtained in the following way: The cognitive bias simulator extracts semantic features from the input candidate interest tags, transforming the tag text into vector features that the model can recognize. These vector features can fully represent the semantic connotation, attribute type, and associated content of the tags. For multiple preset dimensions of psychological cognitive bias, quantitative calculations are performed through independent prediction branches. Combined with the patterns of manually labeled samples learned in the pre-training stage, the original prediction scores of each dimension are output.

[0052] According to the preset scoring mapping rules, the original predicted scores are standardized to a score range of 0 to 2, where 0 points correspond to a poor level with excessive deviation, 1 point corresponds to a passing level with moderate deviation, and 2 points correspond to an excellent level with extremely low deviation. The standardized score is the final predicted score of the candidate interest tag on that dimension.

[0053] During the model training phase, the cognitive bias simulator learns the aforementioned weight parameters through supervised fine-tuning. Training data can consist of manually labeled sample interest tags and their labeling scores across various dimensions. During training, a mean squared error loss function or a cross-entropy loss function can be used to minimize the difference between the model's predicted score and the manually labeled score. After multiple rounds of iterative training, once the model converges, it can be used to predict new candidate interest tags.

[0054] Taking the cognitive bias dimension as an example of novelty bias, the cognitive bias simulator calculates the predicted score for the candidate interest tag "outdoor camping" on this dimension. This score is based on the semantic comparison between the tag text and the user's historical interest tag set (e.g., "tent selection" and "trekking pole recommendation"). If a large number of similar tags have appeared in the past, the novelty score output by the model is low, such as 0.25; if no similar tags have appeared in the past, the score is high, such as 0.85. Finally, the scores of each dimension are integrated into vector form or key-value pair form as the core quantitative information in the bias report.

[0055] S1022. Generate the corresponding counterfactual assessment description based on the predicted score.

[0056] In this step, the counterfactual assessment description refers to the description content generated in natural language based on the predicted scores of each dimension. It is used to characterize the reverse feedback, potential risks, or biased performance that may occur under the influence of user psychological cognitive biases of candidate interest tags. It is not a conventional evaluation of the quality of the tags, but a counterfactual judgment given by simulating the user's true psychological state.

[0057] In practice, based on the predicted scores of each psychological cognitive bias dimension obtained from S1021, and combined with the preset description generation rules, a counterfactual assessment description that precisely matches the predicted score is generated. For example, for a high predicted score in the aesthetic fatigue bias dimension, a description is generated that "although this tag is in line with popular trends, users have recently viewed too much similar content, resulting in aesthetic fatigue bias, and the actual click-through rate may decrease." For a high predicted score in the conformity pressure bias dimension, a description is generated that "this tag is biased towards niche areas, and users may experience conformity pressure bias, and dare not click easily but have curiosity and interest in their hearts." This intuitively presents the potential problems of candidate interest tags at the user's subjective psychological level.

[0058] S1023. Use the predicted score and counterfactual assessment description as the assessment results to generate a bias report.

[0059] In this step, the evaluation result refers to the complete deviation evaluation content, which consists of the predicted scores of candidate interest tags in each psychological cognitive bias dimension and the corresponding counterfactual evaluation descriptions; the deviation report refers to the document formed after integrating the evaluation results, which is used to comprehensively reflect the psychological cognitive bias of candidate interest tags.

[0060] In practice, the predicted scores of each dimension obtained in S1021 are integrated with the counterfactual assessment description generated in S1022 and used as the deviation assessment result of the candidate interest tags. Based on the assessment result, a corresponding deviation report is generated. The deviation report fully records the quantitative data and qualitative analysis of the psychological cognitive deviation of the candidate interest tags and clearly presents the various deviation risks of the tags.

[0061] For example, the cognitive bias simulator is pre-trained in the following manner: Obtain manually labeled training datasets, which include sample interest labels and their labeling scores on multiple preset dimensions of psychological cognitive bias; The initial model was trained using a supervised fine-tuning method to obtain a cognitive bias simulator.

[0062] First, a manually labeled training dataset is obtained. The training data includes sample interest labels and their labeled scores on multiple preset psychological cognitive bias dimensions. The labeled scores are obtained by manually standardizing the degree of deviation of sample interest labels on each psychological cognitive bias dimension according to preset scoring standards. Then, a supervised fine-tuning method is used to train the initial model. The manually labeled training dataset is input into the initial model, and the model parameters are optimized through supervised fine-tuning. This enables the model to accurately simulate user psychological cognitive biases, calculate bias dimension prediction scores, and generate counterfactual evaluation descriptions. Finally, a cognitive bias simulator that has been trained is obtained.

[0063] The above steps, through multi-dimensional psychological cognitive bias quantification and scoring, generating counterfactual assessment descriptions and integrating them into a bias report, combined with a cognitive bias simulator with manually labeled data and supervised fine-tuning training, break through the limitations of traditional assessment models that only score from objective dimensions. It can accurately simulate various psychological cognitive biases of users, comprehensively presenting the risk of label bias in a combination of quantitative and qualitative methods, providing accurate bias guidance for subsequent adversarial mutations, effectively solving the problems of traditional interest label generation ignoring users' deep psychological motivations and having a single assessment dimension, and significantly improving the comprehensiveness, pertinence and relevance of interest label assessment to users' true psychology.

[0064] See in some examples Figure 4 , Figure 4 An exemplary flow diagram of S103 is shown. S103 includes: S1031. Based on the user identifier of the target user, query the historical conflict case data corresponding to the target user in the counterfactual experience replay pool.

[0065] In this step, the user identifier refers to the unique identification information of the target user within the system, used to accurately match the user's exclusive historical data. In practice, the target user's unique user identifier is used as the search keyword to perform a matching query operation in a preset counterfactual experience replay pool. This accurately retrieves all historical conflict case data corresponding to the target user, ensuring that the retrieved cases are highly correlated with the target user's historical behavior and interests, providing exclusive counterfactual reference for subsequent adversarial mutations.

[0066] S1032. Sort the retrieved historical conflict case data according to the preset case sorting rules, and select a preset number of historical conflict case data from the sorting results as the search results.

[0067] The preset case sorting rules include: sorting cases from most recent to oldest based on their generation time, and / or sorting them from highest to lowest based on their similarity to the current scene.

[0068] In this step, the preset case sorting rules refer to the pre-defined sorting criteria used to prioritize historical conflict case data. Specifically, this includes sorting cases from most recent to oldest based on their generation time, and / or sorting cases from highest to lowest based on their similarity to the current interest tag generation scenario. The preset quantity refers to the number of case data entries to be filtered and extracted based on the model's input performance. The specific value can be set according to the adversarial mutation requirements, for example, a quantity of 5. The retrieval results refer to the effective historical conflict case data that, after sorting and filtering, is ultimately used to drive the large language model to perform adversarial mutations.

[0069] In practice, all historical conflict case data obtained from query S1031 are first sorted according to the preset case sorting rules. The time sorting rule or the scene similarity sorting rule can be used alone, or the two rules can be combined for comprehensive sorting. Then, according to the preset number, the case data with the highest priority are selected from the sorted results and determined as the final result of this search. This ensures that the selected cases are timely and suitable for the scene, and improves the accuracy of subsequent mutation optimization.

[0070] The historical conflict case data specifically includes the first historical conflict case data and the second historical conflict case data.

[0071] The first type of historical contradictory case data refers to case data where the system evaluation score is higher than a preset first threshold, but the user feedback score is lower than a preset second threshold. The system evaluation score is a quantitative evaluation score calculated by the model for historical interest tags according to preset standards. The preset first threshold is a pre-set score threshold used to determine whether the system evaluation result is excellent. The user feedback score is a quantitative score of feedback on historical interest tags calculated based on the user's actual online behavior. The preset second threshold is a pre-set score threshold used to determine whether the user feedback result is poor. This type of case represents contradictory data that the model judges to be excellent but the actual acceptance by users is low.

[0072] The second type of historical contradictory case data refers to case data where the system evaluation score is lower than the preset third threshold, but the user feedback score is higher than the preset fourth threshold. The preset third threshold is a pre-set score threshold used to determine that the system evaluation result is poor, and the preset fourth threshold is a pre-set score threshold used to determine that the user feedback result is excellent. This type of case represents contradictory data that the model judges as poor but the users actually accept.

[0073] The first to third preset thresholds can be set as needed, and there are no further restrictions.

[0074] The above steps involve precise querying using user identifiers and sorting and filtering historical contradictory case data according to preset rules. This clearly defines two core contradictory case types and fully extracts counterfactual cases from the target user's history where system evaluations and user feedback contradict each other. This breaks through the limitation of traditional experience pools that only store successful cases. It can uncover users' true potential interests from high-quality contradictory cases of failure and unexpected events, providing a real and effective counterfactual learning basis for subsequent adversarial mutations. This effectively enhances the counterfactual exploration capability of interest tag generation, avoids tag homogenization and information cocoon problems, and improves the degree to which interest tags match users' real needs.

[0075] See in some examples Figure 5 , Figure 5 An exemplary flow diagram of S104 is shown. S104 includes: S1041. Convert the predicted scores of each psychological cognitive bias dimension in the bias report into bias prompts described in natural language.

[0076] In this step, the deviation prompt is a message that converts the quantified predicted score into natural language to guide the model in correcting biases. Specifically, according to preset score conversion rules, the predicted scores for each psychological cognitive bias dimension in the deviation report are converted into natural language deviation prompts one by one. These prompts clearly indicate the specific biases present in the candidate interest tags, such as: "This tag exhibits aesthetic fatigue bias; the user has already viewed too much similar content," or "This tag exhibits conformity pressure bias; users are prone to resistance due to niche attributes," providing clear textual guidance for the model to correct tag biases.

[0077] S1042. Convert historical conflict case data into case prompts described in natural language.

[0078] In this step, the historical contradiction case data may include the first historical contradiction case data that received a high score in the system evaluation but a low score in user feedback, and the second historical contradiction case data that received a low score in the system evaluation but a high score in user feedback; the case prompts are prompts that convert the historical contradiction case data into natural language form to guide the model to avoid historical errors or to explore potential interests.

[0079] In practice, according to the preset case conversion template, the filtered historical contradictory case data is converted into case prompts in natural language. The case prompts can clearly present the core information of the historical contradictory cases, such as: "Historical case: The system evaluates it as a high-quality label, but the actual user feedback is poor. It is necessary to avoid generating similar labels," or "Historical case: The system evaluates it as a common label, but the actual user feedback is good. It can be used as a reference to generate similar potential interest labels." This allows the model to intuitively learn from historical counterfactual experiences and provides a reference for adversarial mutations.

[0080] S1043. Integrate deviation prompts and case prompts to construct adversarial prompt words.

[0081] In this step, adversarial cue words refer to exclusive cue words that integrate the need for bias correction with counterfactual experience to drive the model to perform adversarial mutations; integration refers to combining bias cue words with case cue words according to the preset cue word splicing logic to form a complete and logically coherent cue text.

[0082] In practice, based on the preset prompt word construction rules, the deviation prompts obtained in S1041 and the case prompts obtained in S1042 are organically integrated to construct adversarial prompt words that combine deviation correction guidance and counterfactual experience reference. These prompt words can guide the model to no longer simply follow popular trends and historical success experiences, but to generate labels in the direction of resisting cognitive bias and exploring potential interests, thus achieving the core purpose of adversarial evolution.

[0083] S1044. Input the adversarial cue words and candidate interest labels into the large language model to guide the large language model to generate the mutated target interest labels.

[0084] In this step, adversarial mutation refers to the generation method of adjusting and optimizing candidate interest tags based on adversarial prompts to resist psychological cognitive biases and conform to the user's real potential interests; the mutated target interest tag refers to the final interest tag that has been optimized by adversarial mutation to eliminate bias problems and meet the user's real needs.

[0085] In practice, the constructed adversarial prompts and the candidate interest tags generated in the previous steps are input into the large language model. Guided by the adversarial prompts, the large language model performs adversarial mutation operations on the candidate interest tags, breaking through the limitations of traditional generation logic, generating at least one mutated target interest tag, and completing the final optimized generation of interest tags.

[0086] This example fully implements a label optimization process based on cognitive biases and counterfactual experience by transforming biased data and contradictory cases into natural language prompts, integrating and constructing adversarial prompt words, and driving the model to perform adversarial mutations. It abandons the traditional model that relies solely on positive experience to generate labels, effectively guiding the model to correct psychological cognitive biases and avoid historical contradictions. The generated target interest labels are more in line with users' deep psychological needs, significantly improving the accuracy, diversity, and innovation of interest labels. It fundamentally alleviates the problems of label homogenization and information cocoons, and strengthens the interest exploration capabilities of the recommendation system.

[0087] In some examples, after S101 and before S102, this method also includes: Step 1: Calculate the similarity between candidate interest tags and current popular content data; The process involves calculating the similarity between candidate interest tags and currently popular content data. In this step, candidate interest tags are initial interest identifiers generated from target user data using a large language model; currently popular content data is pre-collected content with high preference across the entire network for the target user's demographic; similarity is obtained through multimodal calculation and is a quantitative value representing the degree of matching between candidate interest tags and currently popular content data; a preset similarity threshold is a pre-set critical value used to determine whether a candidate interest tag meets the popularity matching requirements, and can be set as needed, for example, to 0.2. In practice, multimodal calculation is used to calculate the similarity between each candidate interest tag and the currently popular content data one by one. This calculation method balances online execution efficiency and cost consumption, meeting the actual deployment requirements of this application.

[0088] Step 2: Remove candidate interest tags with similarity lower than the preset similarity threshold, and retain candidate interest tags with similarity higher than or equal to the preset similarity threshold.

[0089] In this process, candidate interest tags with similarity scores below a preset similarity threshold are removed, while those with similarity scores at or above the threshold are retained. Removal refers to excluding candidate interest tags that do not adequately match popular content from subsequent processing, while retention refers to keeping candidate interest tags that meet the popularity requirements for later cognitive bias assessment and adversarial mutation. Specifically, the similarity calculation result of each candidate interest tag is compared one by one with the preset similarity threshold. Tags with insufficient fit are filtered out, and only candidate interest tags that meet the matching criteria for current popular content are retained. This ensures that candidate tags have basic popularity adaptability while reducing the amount of data processing required for subsequent bias simulation and optimization, thus improving the overall efficiency of the process.

[0090] In some examples, prior to S101, this method also includes: Step 1: Retrieve historical success case data related to the target user from the pre-set success case experience pool; Specifically, this step involves retrieving historical successful case data related to the target user from a pre-built and continuously updated database. This database is specifically designed to store high-quality interest tag cases that received excellent system evaluation scores and positive online user feedback during the historical generation process. The historical successful case data consists of interest tag cases that are highly relevant to the target user, have been generated in the past, and have achieved satisfactory practical application results—in other words, the target user's past successful generation experience. In practice, based on the target user's unique user identifier, a precise matching search is performed within the pre-built successful case experience pool to obtain the user's exclusive historical high-quality generated case data.

[0091] Step 2: Construct positive prompt words based on historical success case data, and input the positive prompt words and user data into the large language model to generate candidate interest tags.

[0092] Furthermore, positive prompts are constructed based on historical successful case data, and these prompts, along with user data, are input into a large language model to generate candidate interest tags. In this step, positive prompts are information constructed from historical successful case data to guide the large language model in learning high-quality generation experiences; they are an important supplement to dynamic prompts. User data includes at least one of user profiles, historical interest data, and current popular content data. Specifically, the retrieved historical successful case data is converted into natural language prompt content to construct positive prompts adapted to the current user. These positive prompts are then integrated with the target user's user data and input into the large language model. This allows the model to combine the user's historical successful generation experience to more accurately generate initial candidate interest tags, improving the quality and relevance of the generated initial tags.

[0093] The above steps employ a dual optimization approach: adding positive guidance based on historical success cases before generating candidate interest tags and filtering based on popularity and similarity after generation. On one hand, this leverages the experience of successful cases to strengthen the model's ability to learn from users' high-quality interests, improving the accuracy of initial candidate tags. On the other hand, similarity filtering ensures the tags match current popular content while reducing processing pressure in subsequent steps, balancing online performance and cost control. This dual optimization mechanism further refines the generation and filtering process of interest tags, ensuring that the target interest tags obtained through cognitive bias simulation and adversarial mutation not only align with users' historical success experiences and adapt to real-time content trends but also break through surface-level interest limitations to uncover deeper needs, comprehensively improving the stability, accuracy, and overall efficiency of interest tag generation.

[0094] In some examples, after S104, this method also includes: Step 1: Input the mutated target interest label as a new candidate interest label into the cognitive bias simulator for iteration to generate new target interest labels. Stop the iteration when the generated new target interest labels meet the preset equilibrium conditions.

[0095] In this step, the mutated target interest label refers to the interest label obtained after adversarial mutation optimization in step S104, which has initially corrected the psychological cognitive bias; the new candidate interest label refers to the target interest label obtained in the previous iteration as the initial label to be optimized in this iteration, which is used to re-enter the bias assessment and mutation process; iteration refers to repeatedly executing the complete optimization process of bias assessment, counterfactual case retrieval, and adversarial mutation, and continuously improving the quality of interest labels through multiple cycles; the new target interest label refers to the interest label generated after each iteration optimization, which has a lower degree of bias and is more in line with user needs; the preset equilibrium condition refers to the iteration stopping condition pre-set before the method is executed, combined with the adversarial evolution logic, model performance, and user interest matching requirements, representing that the target interest label has achieved a balance between cognitive bias simulation and adversarial correction.

[0096] Here are some examples of ways to define equilibrium conditions: Example 1: Equilibrium condition based on bias score convergence Set a convergence threshold, for example, 0.05. In the t-th and t-1-th iterations, calculate the comprehensive bias score output by the cognitive bias simulator, for example, the weighted average of the predicted scores for each dimension. If the absolute value of the difference between the comprehensive scores of the two iterations is less than the convergence threshold, the label quality is considered to have stabilized and the equilibrium condition is met. For example, if the comprehensive score in the first iteration is 0.72 and in the second iteration is 0.74, the difference is 0.02 and less than 0.05, then the iteration stops.

[0097] Example 2: Equilibrium conditions based on the semantic stability of tags Set a stable similarity threshold, for example, 0.95. Calculate the semantic similarity between the target interest tags generated in round t and round (t-1). If the similarity is greater than or equal to the stable similarity threshold, it is considered that subsequent mutations can no longer bring substantial improvement, and the iteration stops.

[0098] Example 3: Equilibrium condition based on maximum number of iterations To avoid infinite loops consuming computational resources, a maximum number of iterations is preset, for example, 7. When the actual number of iterations reaches this limit, the iteration is forcibly stopped regardless of whether the above conditions are met.

[0099] In practical applications, the above conditions can be combined: for example, stop immediately when the maximum number of iterations is reached; otherwise, prioritize determining whether the deviation score converges or the label semantics are stable, and stop iterating when either condition is met.

[0100] In practice, the mutated target interest labels obtained in the previous round are used as new candidate interest labels to be optimized again. They are then input into the cognitive bias simulator to perform bias evaluation. S102-S104 are repeated to continuously generate better new target interest labels until the newly generated target interest labels reach the preset equilibrium conditions, and the simulation and anti-simulation balance is completed, at which point the iteration process stops.

[0101] In some examples, for each iteration in step one, this method also includes: Sub-step 1: From the multiple target interest tags generated in the current round, select a preset number of target interest tags according to the preset evaluation indicators, and use them as seeds for the next round of adversarial mutation; Here, the multiple target interest tags generated in the current round refer to several optimized interest tags generated through adversarial mutation in a single iteration; the preset evaluation index refers to the pre-set comprehensive evaluation criteria used to select high-quality tags, which are set in combination with dimensions such as tag deviation degree, user demand fit, and popularity adaptability; the preset quantity refers to the number of high-quality tags selected in each round, which is set in advance according to the efficiency of the evolutionary algorithm, model input constraints, and online performance requirements. The specific value can be set as needed, for example, 2, and is not limited here; the seed refers to the high-quality target interest tags obtained after screening, which serve as the basic optimization material for the next round of adversarial mutation; adversarial mutation refers to the operation of correcting and optimizing tags based on deviation reports and historical contradictory case data to drive the large language model.

[0102] In each iteration's seed selection step, the target interest tags generated in the current round need to be sorted according to preset evaluation metrics to select the optimal tags as seeds for the next round of mutation. Several evaluation metrics are listed below as examples: Example 1: Overall score of the cognitive bias simulator The predicted scores for each psychological cognitive bias dimension output by the cognitive bias simulator are weighted and summed to obtain a comprehensive score. The weights can be preset according to business objectives. For example, if the system wants to prioritize resisting "aesthetic fatigue bias," its weight can be set to 0.5, and the sum of the weights of other dimensions can be set to 0.5. The higher the comprehensive score, the better the label performs in resisting each psychological bias. In the calculation, the score of each dimension is multiplied by the weight corresponding to that dimension, and then all products are added together to obtain the comprehensive score.

[0103] Example 2: Relevance score to user's historical interests The mean cosine similarity between the target interest tag and the user's historical interest tag set is calculated using a semantic encoding model. This metric ensures that the selected seeds do not deviate from the user's basic interest range. For example, if the user's historical interest tags are "new energy vehicles" and "charging pile reviews," the relevance score of the tag "electric vehicle range test" may be as high as 0.8, while "traditional fuel vehicle maintenance" may only be 0.2. Tags with higher relevance scores are prioritized during seed selection.

[0104] In practical applications, the above multiple indicators can be combined into a comprehensive score, and then sorted from high to low according to the comprehensive score, and the first preset number of tags can be selected as seeds.

[0105] In practice, for the multiple target interest tags generated in a single iteration, a scoring and ranking are completed based on preset evaluation indicators. A preset number of high-quality tags with the highest ranking are selected as seeds for the next round of adversarial mutation, ensuring that the iteration is continuously optimized based on high-quality tags.

[0106] Sub-step two: Use the seed as a new candidate interest label and perform adversarial mutation in the next iteration.

[0107] Here, new candidate interest tags refer to the seed tags selected in this round as the initial tags to be optimized in the next iteration; adversarial mutation in the next iteration refers to repeatedly executing the complete process of bias report generation, counterfactual case retrieval, adversarial prompt word construction, and tag mutation. In specific implementation, the selected seed tags are directly used as new candidate interest tags for the next iteration, re-entering the adversarial mutation process to continuously optimize and correct the tags, thereby achieving the cyclical evolution of interest tags.

[0108] This disclosure fully implements a multi-level cyclical verification method of generation, evaluation, and evolution by adding an iterative optimization process after the target interest tags are generated and performing high-quality seed selection in each iteration. The iterative process, through continuous deviation simulation and adversarial correction, allows the target interest tags to gradually reach an equilibrium state between simulation and anti-simulation, minimizing the impact of user psychological cognitive biases. The seed selection step in each iteration retains high-quality tags as the basis for evolution and matches the optimization logic of the evolutionary algorithm, ensuring both the quality of tag optimization and improving iteration efficiency. Through multiple iterations, the interest exploration and iterative optimization effects of the recommendation system are strengthened.

[0109] In some examples, after S104, this method also includes: Step 1: Obtain online content recall results based on target interest tags and collect real-time user feedback data.

[0110] In this step, the online content retrieval result refers to the set of relevant content, such as matched and popular content, retrieved for the target user after applying the target interest tags to the online recommendation strategy. Real-time feedback data refers to the actual behavioral feedback data generated by the target user in response to the online content retrieval result, including clicks, dwell time, interactions, and skips; this is the core basis for reflecting the actual application effect of the tags. In practice, the target interest tags are input into the strategy retrieval stage of the online recommendation system to generate content retrieval results for the target user, while simultaneously collecting real-time user behavioral feedback on the retrieved content to obtain the most authentic user interest verification data.

[0111] Step 2: Update the historical contradictory case data in the counterfactual experience replay pool based on real-time feedback data.

[0112] In this step, the system evaluation results of the target interest tags in this online application are compared with the real-time feedback data. If there is an inconsistency, the tag and its corresponding data are added to the counterfactual experience replay pool as new historical contradiction case data, thereby completing the dynamic update of the replay pool data and continuously enriching the counterfactual learning materials.

[0113] Step 3: Iteratively optimize the cognitive bias simulator and large language model using the updated counterfactual experience replay pool.

[0114] Iterative optimization refers to the process of adjusting and updating the model's parameters and inference logic based on newly added real-world contradictory case data to improve model performance. In practice, data from the updated counterfactual experience replay pool is used as supplementary training material to fine-tune the parameters of both the cognitive bias simulator and the large language model. This makes the cognitive bias simulator's ability to simulate biases more closely resemble real user psychology, and makes the large language model's generation and adversarial mutation capabilities more aligned with users' actual interests and needs, thus achieving continuous model evolution.

[0115] In some examples, following S104, this method also includes an evaluation and guidance step, specifically including: Step 1: The supervised judge model evaluates the target interest tags from multiple evaluation dimensions to obtain evaluation results; among them, the multiple evaluation dimensions include at least two of the following: comprehensiveness dimension, accuracy dimension, and expansion dimension.

[0116] In this step, a supervised evaluation model is used to assess the target interest tags from multiple evaluation dimensions to obtain the evaluation results. The supervised evaluation model refers to a dedicated evaluation model trained using manually labeled data through supervised fine-tuning. The multiple evaluation dimensions are pre-defined dimensions used to comprehensively measure the quality of interest tags, specifically including at least two of the following: comprehensiveness, accuracy, and extensibility. For example, the specific meanings and judgment examples for each dimension are as follows: The comprehensiveness dimension is used to assess whether the generated interest tags cover all important interest dimensions of the target user, reflecting the overall coverage of the user's interests. For example, if the target user's historical interests include outdoor camping, new energy vehicles, and home renovation, a comprehensiveness score of 2 points is given if the generated tags cover all three interests, and 0 points if they only cover one. The accuracy dimension is used to assess... The accuracy of a single interest tag is assessed based on whether it accurately reflects the user's actual behavior, avoiding tags that are too broad or off-target. For example, if a user has only recently browsed content related to new energy vehicle reviews, generating a tag like "new energy vehicle reviews" would score 2 points in accuracy; generating a vague tag like "cars" would score 1 point; and generating a tag like "gasoline cars" would score 0 points. The extensibility dimension is used to evaluate whether interest tags make reasonable inferences about potential interests based on the user's existing behavior, rather than being limited to superficial behavior. For example, if a user's historical behavior is only browsing outdoor camping pictures and text, generating tags like "camping equipment selection" or "niche camping spot guide" would score 2 points in extensibility; repeatedly generating the tag "outdoor camping" would score 1 point; and generating tags unrelated to camping would score 0 points.

[0117] The evaluation results are quantitative scores and qualitative evaluations of the output of the supervised judge model for each dimension. The scoring criteria include three levels: poor, passable, and excellent.

[0118] In practice, the target interest label is input into the trained supervised referee model, and the model completes the standardized evaluation according to the preset multi-evaluation dimensions, and outputs a complete evaluation result containing the scores and evaluations of each dimension.

[0119] Step 2: Use the evaluation results as examples for few-shot learning to guide the large language model to optimize its output in subsequent generation.

[0120] In this step, "few-shot learning examples" refers to using high-quality evaluation results as demonstration samples to provide prompts for the large language model to quickly learn optimization directions. "Preset" refers to the example transformation and input rules pre-defined according to the few-shot learning rules before model application. In practice, the evaluation results output by the supervised judge model are converted into few-shot learning examples in natural language form and integrated into the prompt word system of the large language model. This allows the large language model to continuously optimize the quality of its output content by referencing the standards of high-quality examples when performing candidate interest tag generation and adversarial mutation operations.

[0121] The above steps, by adding online feedback loops and model iteration optimization steps, as well as multi-dimensional evaluation and small-sample guidance steps, fully construct a "generation-evaluation-evolution-feedback-iteration" process. Real-time online feedback dynamically updates the counterfactual experience replay pool, allowing the cognitive bias simulator and the large language model to continuously learn from contradictory cases of real users, constantly improving the accuracy of bias simulation and interest generation. The evaluation results of the multi-dimensional supervised judge model serve as small-sample learning examples, providing standardized optimization guidance for the large language model, taking into account the comprehensiveness, accuracy, and scalability of interest tags. This mechanism enables the entire interest tag generation system to continuously self-evolve, effectively solving the problems of lagging feedback and singular optimization directions in traditional solutions, further breaking down the homogenization of interest tags and information cocoons, and comprehensively improving the interest exploration capabilities and content distribution effectiveness of the recommendation system.

[0122] The method disclosed herein combines large language model generation, cognitive bias simulator evaluation, counterfactual experience replay pool to recall contradictory cases, and adversarial mutation to achieve verifiable and autonomously evolving user interest tags. This method not only focuses on the similarity between generated content and popular content but also simulates deep-seated psychological cognitive biases in users, thereby generating interest tags that truly penetrate users' surface behavior. Simultaneously, by storing and utilizing contradictory cases with high system evaluation but poor feedback, and low system evaluation but good feedback, the system can learn from failures and unexpected events, continuously expanding the boundaries of interest exploration. Ultimately, this significantly improves the accuracy, diversity, and long-term user satisfaction of the recommendation system, and possesses strong robustness and interpretability.

[0123] In the disclosed second embodiment, see Figure 6 ,like Figure 1 The principle shown Figure 6 This illustration shows a user interest tag generation apparatus 60 according to a second embodiment of the present disclosure, wherein the apparatus includes: The generation module 601 is used to generate at least one candidate interest tag based on the user data of the target user through a large language model. The deviation module 602 is used to input candidate interest labels into the cognitive deviation simulator and obtain a deviation report that characterizes the evaluation results of the candidate interest labels on a preset psychological cognitive deviation dimension. The retrieval module 603 is used to retrieve historical conflict case data related to the target user from a preset counterfactual experience replay pool; wherein, the historical conflict case data are case data in which the system evaluation and user feedback are inconsistent during the historical generation process; The mutation module 604 is used to drive the large language model to perform adversarial mutations on candidate interest tags based at least on deviation reports and historical contradictory case data, generating at least one mutated target interest tag.

[0124] In some examples, the deviation module 602 is specifically used for: Input the candidate interest labels into the cognitive bias simulator, and calculate the predicted scores of the candidate interest labels on each dimension from the preset set of psychological cognitive bias dimensions. Generate a corresponding counterfactual assessment description based on the predicted score; The predicted score and counterfactual assessment description are used as evaluation results to generate a bias report.

[0125] In some examples, the cognitive bias simulator in bias module 602 is pre-trained in the following manner: Obtain manually labeled training datasets, which include sample interest labels and their labeling scores on multiple preset dimensions of psychological cognitive bias; The initial model was trained using a supervised fine-tuning method to obtain a cognitive bias simulator.

[0126] In some examples, the pre-defined psychological cognitive bias dimensions include at least one of the following: novelty bias dimension, controversy bias dimension, emotional tendency bias dimension, conformity pressure bias dimension, aesthetic fatigue bias dimension, and curiosity bias dimension.

[0127] In some examples, the retrieval module 603 is specifically used for: Based on the user identifier of the target user, query the historical conflict case data corresponding to the target user in the counterfactual experience replay pool; According to the preset case sorting rules, the retrieved historical conflict case data is sorted, and a preset number of historical conflict case data in the sorted results are selected as the search results.

[0128] In some examples, the historical conflict case data includes first historical conflict case data and second historical conflict case data; The first set of historical conflict case data consists of cases whose system evaluation scores are higher than the preset first threshold and whose user feedback scores are lower than the preset second threshold. The second set of historical conflict case data consists of cases where the system evaluation score is lower than the preset third threshold, but the user feedback score is higher than the preset fourth threshold.

[0129] In some examples, mutation module 604 is specifically used for: Convert the predicted scores of each psychological cognitive bias dimension in the bias report into bias prompts described in natural language; Convert historical conflict case data into case prompts described in natural language; By integrating deviation prompts and case prompts, adversarial prompt words are constructed; The adversarial cue words and candidate interest labels are input into the large language model, which then guides the model to generate mutated target interest labels.

[0130] In some examples, the generation module 601 is specifically used for: Acquire user data of the target users, including at least one of user profiles, historical interest data, and currently popular content data; Dynamic prompts are constructed based on user data. The dynamic prompts include at least one of the following: task description, user profile, historical interest data, and currently popular content data. The dynamic prompts are input into a large language model to generate candidate interest tags.

[0131] In some examples, the device also includes a similarity module for: Calculate the similarity between candidate interest tags and current popular content data; Candidate interest tags with similarity below a preset similarity threshold are removed, while candidate interest tags with similarity above or equal to the preset similarity threshold are retained.

[0132] In some examples, the device also includes a success experience module for: Retrieve historical success case data related to the target user from a pre-defined pool of success case experiences; Positive prompts are constructed based on historical success case data, and these positive prompts are input into a large language model along with user data to generate candidate interest tags.

[0133] In some examples, the device also includes an iteration module for: The mutated target interest label is used as a new candidate interest label and input into the cognitive bias simulator for iteration to generate new target interest labels. The iteration stops when the generated new target interest labels meet the preset equilibrium conditions.

[0134] In some examples, the device also includes a seed module for: From the multiple target interest tags generated in the current round, a preset number of target interest tags are selected according to preset evaluation indicators and used as seeds for the next round of adversarial mutation; Use the seed as a new candidate interest label and perform adversarial mutations in the next iteration.

[0135] In some examples, the device also includes an online feedback module for: Online content retrieval results are obtained based on target interest tags, and real-time user feedback data is collected; The historical contradictory case data in the counterfactual experience replay pool is updated based on real-time feedback data. The cognitive bias simulator and large language model are iteratively optimized using an updated counterfactual experience replay pool.

[0136] The acquisition, storage, and application of user personal information involved in the technical solution disclosed herein comply with the provisions of relevant laws and regulations and do not violate public order and good morals.

[0137] According to embodiments of this disclosure, this disclosure also provides an electronic device, a readable storage medium, and a computer program product.

[0138] Computer instructions stored on a non-transitory computer-readable storage medium are used to cause a computer to perform the above-described method.

[0139] Computer program products include computer programs that, when executed by a processor, implement the methods described above.

[0140] Figure 7 A schematic block diagram of an example electronic device 700 that can be used to implement embodiments of the present disclosure is shown. The electronic device is intended to represent various forms of digital computers, such as laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. The electronic device may also represent various forms of mobile devices, such as personal digital processors, cellular phones, smartphones, wearable devices, and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely illustrative and are not intended to limit the implementation of the present disclosure described and / or claimed herein.

[0141] like Figure 7 As shown, device 700 includes a computing unit 701, which can perform various appropriate actions and processes based on a computer program stored in read-only memory (ROM) 702 or a computer program loaded from a storage unit into random access memory (RAM) 703. RAM 703 may also store various programs and data required for the operation of device 700. The computing unit 701, ROM 702, and RAM 703 are interconnected via bus 704. Input / output (I / O) interface 705 is also connected to bus 704.

[0142] Multiple components in device 700 are connected to I / O interface 705, including: input unit 706, such as keyboard, mouse, etc.; output unit 707, such as various types of monitors, speakers, etc.; storage unit 708, such as disk, optical disk, etc.; and communication unit 709, such as network card, modem, wireless transceiver, etc. Communication unit 709 allows device 700 to exchange information / data with other devices through computer networks such as the Internet and / or various telecommunications networks.

[0143] The computing unit 701 can be a variety of general-purpose and / or special-purpose processing components with processing and computing capabilities. Some examples of the computing unit 701 include, but are not limited to, a central processing unit (CPU), a graphics processing unit (GPU), various special-purpose artificial intelligence (AI) computing chips, various computing units running machine learning model algorithms, a digital signal processor (DSP), and any suitable processor, controller, microcontroller, etc. The computing unit 701 performs the various methods and processes described above, such as the methods described above. For example, in some embodiments, the methods described above can be implemented as a computer software program tangibly contained in a machine-readable medium, such as storage unit 708. In some embodiments, part or all of the computer program can be loaded and / or installed on device 700 via ROM 702 and / or communication unit 709. When the computer program is loaded into RAM 703 and executed by the computing unit 701, one or more steps of the methods described above can be performed. Alternatively, in other embodiments, the computing unit 701 can be configured to perform the methods described above by any other suitable means (e.g., by means of firmware).

[0144] Various embodiments of the systems and techniques described above herein can be implemented in digital electronic circuit systems, integrated circuit systems, field-programmable gate arrays (FPGAs), application-specific integrated circuits (ASICs), application-specific standard products (ASSPs), systems-on-a-chip (SoCs), payload-programmable logic devices (CPLDs), computer hardware, firmware, software, and / or combinations thereof. These various embodiments may include implementations in one or more computer programs that can be executed and / or interpreted on a programmable system including at least one programmable processor, which may be a dedicated or general-purpose programmable processor, capable of receiving data and instructions from a storage system, at least one input device, and at least one output device, and transmitting data and instructions to the storage system, the at least one input device, and the at least one output device.

[0145] The program code used to implement the methods of this disclosure may be written in any combination of one or more programming languages. This program code may be provided to a processor or controller of a general-purpose computer, special-purpose computer, or other programmable data processing apparatus, such that when executed by the processor or controller, the program code causes the functions / operations specified in the flowcharts and / or block diagrams to be implemented. The program code may be executed entirely on a machine, partially on a machine, as a standalone software package partially on a machine and partially on a remote machine, or entirely on a remote machine or server.

[0146] In the context of this disclosure, a machine-readable medium can be a tangible medium that may contain or store a program for use by or in conjunction with an instruction execution system, apparatus, or device. A machine-readable medium can be a machine-readable signal medium or a machine-readable storage medium. A machine-readable medium can be, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any suitable combination of the foregoing. More specific examples of machine-readable storage media include electrical connections based on one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination of the foregoing.

[0147] To provide interaction with a user, the systems and techniques described herein can be implemented on a computer having: a display device for displaying information to the user (e.g., a CRT (cathode ray tube) or LCD (liquid crystal display) monitor); and a keyboard and pointing device (e.g., a mouse or trackball) through which the user provides input to the computer. Other types of devices can also be used to provide interaction with the user; for example, feedback provided to the user can be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user can be received in any form (including sound input, voice input, or tactile input).

[0148] The systems and technologies described herein can be implemented in computing systems that include backend components (e.g., as a data server), or computing systems that include middleware components (e.g., an application server), or computing systems that include frontend components (e.g., a user computer with a graphical user interface or web browser through which a user can interact with implementations of the systems and technologies described herein), or any combination of such backend, middleware, or frontend components. The components of the system can be interconnected via digital data communication of any form or medium (e.g., a communication network). Examples of communication networks include local area networks (LANs), wide area networks (WANs), and the Internet.

[0149] Computer systems can include clients and servers. Clients and servers are generally located far apart and typically interact via communication networks. Client-server relationships are created by computer programs running on the respective computers and having a client-server relationship with each other. Servers can be cloud servers, servers in distributed systems, or servers incorporating blockchain technology.

[0150] It should be understood that the various forms of processes shown above can be used to rearrange, add, or delete steps. For example, the steps described in this disclosure can be executed in parallel, sequentially, or in different orders, as long as the desired result of the technical solution disclosed in this disclosure can be achieved, and this is not limited herein.

[0151] The specific embodiments described above do not constitute a limitation on the scope of protection of this disclosure. Those skilled in the art should understand that various modifications, combinations, sub-combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this disclosure should be included within the scope of protection of this disclosure.

Claims

1. A method for generating user interest tags, wherein, The method includes: Based on the user data of the target users, at least one candidate interest tag is generated through a large language model; The candidate interest tags are input into the cognitive bias simulator to obtain a bias report that characterizes the evaluation results of the candidate interest tags on a preset psychological cognitive bias dimension. Retrieve historical conflict case data related to the target user from a pre-set counterfactual experience replay pool; wherein, the historical conflict case data are case data where the system evaluation and user feedback are inconsistent during the historical generation process; Based at least on the deviation report and the historical contradictory case data, the large language model is driven to perform adversarial mutation on the candidate interest tags to generate at least one mutated target interest tag.

2. The method according to claim 1, wherein, The step of inputting the candidate interest tags into a cognitive bias simulator to obtain a bias report characterizing the evaluation results of the candidate interest tags on a preset psychological cognitive bias dimension includes: The candidate interest tags are input into the cognitive bias simulator, and the cognitive bias simulator calculates the predicted scores of the candidate interest tags on each dimension from a preset set of psychological cognitive bias dimensions. Generate a corresponding counterfactual assessment description based on the predicted score; The predicted score and the counterfactual assessment description are used as the assessment results to generate the bias report.

3. The method according to claim 1 or 2, wherein, The cognitive bias simulator was pre-trained using the following method: Obtain manually labeled training dataset, which includes sample interest labels and their labeling scores on multiple preset psychological cognitive bias dimensions; The initial model was trained using a supervised fine-tuning method to obtain the cognitive bias simulator.

4. The method according to claim 2 or 3, wherein, The preset psychological cognitive bias dimensions include at least one of the following: novelty bias dimension, controversy bias dimension, emotional tendency bias dimension, conformity pressure bias dimension, aesthetic fatigue bias dimension, and curiosity bias dimension.

5. The method according to any one of claims 1-4, wherein, The step of retrieving historical contradictory case data related to the target user from a preset counterfactual experience replay pool includes: Based on the user identifier of the target user, query the historical contradictory case data corresponding to the target user in the counterfactual experience replay pool; According to the preset case sorting rules, the retrieved historical conflict case data is sorted, and a preset number of historical conflict case data in the sorting results are selected as the search results.

6. The method according to any one of claims 1-5, wherein, The historical conflict case data includes first historical conflict case data and second historical conflict case data; The first set of historical conflict case data consists of case data whose system evaluation score is higher than a preset first threshold and whose user feedback score is lower than a preset second threshold. The second set of historical conflict case data consists of cases where the system evaluation score is lower than the preset third threshold, but the user feedback score is higher than the preset fourth threshold.

7. The method according to any one of claims 1-6, wherein, The process of driving the large language model to perform adversarial mutations on the candidate interest tags, based at least on the deviation report and the historical contradictory case data, to generate at least one mutated target interest tag, includes: The predicted scores for each psychological cognitive bias dimension in the bias report are converted into bias prompts described in natural language. Convert the historical conflict case data into case prompts described in natural language; By merging the deviation prompts and the case prompts, adversarial prompt words are constructed; The adversarial cue words and the candidate interest tags are input into the large language model, which then guides the large language model to generate the mutated target interest tags.

8. The method according to any one of claims 1-7, wherein, The user data based on the target user is used to generate at least one candidate interest tag through a large language model, including: Obtain the user data of the target user, wherein the user data includes at least one of user profile, historical interest data, and currently popular content data; Dynamic prompt words are constructed based on the user data, and the dynamic prompt words include at least one of the following: task description, user profile, historical interest data, and currently popular content data; The dynamic prompt words are input into the large language model to generate the candidate interest tags.

9. The method according to any one of claims 1-8, wherein, After generating at least one candidate interest tag using a large language model based on user data of the target user, and before inputting the candidate interest tag into a cognitive bias simulator to obtain a bias report characterizing the evaluation results of the candidate interest tag on a preset psychological cognitive bias dimension, the method further includes: Calculate the similarity between the candidate interest tags and the currently popular content data; Candidate interest tags with similarity below a preset similarity threshold are removed, while candidate interest tags with similarity above or equal to the preset similarity threshold are retained.

10. The method according to any one of claims 1-9, wherein, Before generating at least one candidate interest tag using a large language model based on the target user's user data, the method further includes: Retrieve historical success case data related to the target user from a pre-defined pool of success case experiences; Positive prompt words are constructed based on the historical success case data, and these positive prompt words are input into the large language model along with the user data to generate the candidate interest tags.

11. The method according to any one of claims 1-10, wherein, After the method, based at least on the deviation report and the historical contradictory case data, drives the large language model to perform adversarial mutation on the candidate interest tags to generate at least one mutated target interest tag, the method further includes: The mutated target interest label is used as a new candidate interest label and input into the cognitive bias simulator for iteration to generate a new target interest label. The iteration stops when the generated new target interest label meets a preset equilibrium condition.

12. The method according to claim 11, wherein, For each iteration round, the method further includes: From the multiple target interest tags generated in the current round, a preset number of target interest tags are selected according to preset evaluation indicators and used as seeds for the next round of adversarial mutation; The seed is used as the new candidate interest label to perform the adversarial mutation in the next iteration.

13. The method according to any one of claims 1-12, wherein, After the method, based at least on the deviation report and the historical contradictory case data, drives the large language model to perform adversarial mutation on the candidate interest tags to generate at least one mutated target interest tag, the method further includes: Based on the target interest tags, online content retrieval results are obtained, and real-time user feedback data is collected; The historical contradictory case data in the counterfactual experience replay pool is updated based on the real-time feedback data. The cognitive bias simulator and the large language model are iteratively optimized using an updated counterfactual experience replay pool.

14. A device for generating user interest tags, wherein, The device includes: The generation module is used to generate at least one candidate interest tag based on the user data of the target user through a large language model. The bias module is used to input the candidate interest tags into the cognitive bias simulator and obtain a bias report that characterizes the evaluation results of the candidate interest tags on a preset psychological cognitive bias dimension. The retrieval module is used to retrieve historical conflict case data related to the target user from a preset counterfactual experience replay pool; wherein, the historical conflict case data are case data in which the system evaluation and user feedback are inconsistent during the historical generation process; The mutation module is used to drive the large language model to perform adversarial mutation on the candidate interest tags based at least on the deviation report and the historical contradictory case data, so as to generate at least one mutated target interest tag.

15. An electronic device comprising: At least one processor; as well as A memory communicatively connected to the at least one processor; wherein, The memory stores instructions that can be executed by the at least one processor to enable the at least one processor to perform the method of any one of claims 1-13.

16. A non-transitory computer-readable storage medium storing computer instructions, wherein, The computer instructions are used to cause the computer to perform the method according to any one of claims 1-13.

17. A computer program product comprising a computer program that, when executed by a processor, implements the method according to any one of claims 1-13.

Images