A content recommendation method, a content recommendation device, and a computer storage medium

Abstract

This application proposes a content recommendation method, a content recommendation device, and a computer storage medium. The content recommendation method includes: determining a search scenario type based on a user's search request; matching the request semantic vector of the user's search request with the entity semantic vector of a knowledge entity database corresponding to the search scenario type to determine a semantic matching degree; determining a search intent type based on the entity keywords of the user's search request and the semantic matching degree; obtaining a knowledge quality score for the knowledge content from a knowledge provider; determining a ranking of the knowledge content based on the search intent type and the knowledge quality score; and displaying the content recommendation results for the user's search request according to the knowledge content ranking. Through the above content recommendation method, accurate classification of search scenario types and search intent types is achieved, enabling precise, dynamic, and differentiated knowledge recommendation.

Description

Technical Field

[0001] This application relates to the field of information retrieval technology, and in particular to a content recommendation method, a content recommendation device, and a computer storage medium. Background Technology

[0002] Traditional search engines employ a "index-match-rank" process, with ranking mechanisms driven primarily by keyword relevance and bidding amounts. However, in the matching and ranking process, traditional search engines lack differentiated adaptations for different scenarios and intent types, resulting in a static ranking mechanism with poor dynamic adaptability, unable to make real-time adjustments, and struggling to cope with the ever-changing market environment. Summary of the Invention

[0003] To address the aforementioned technical problems, this application proposes a content recommendation method, a content recommendation device, and a computer storage medium.

[0004] To address the aforementioned technical problems, this application proposes a content recommendation method, which includes: Determine the search scenario type based on the user's search request; The semantic vector of the user's search request is matched with the semantic vector of the entity in the knowledge entity database corresponding to the search scenario type to determine the semantic matching degree. The search intent type is determined based on the entity keywords of the user's search request and the semantic matching degree; The knowledge quality score of the knowledge content obtained from knowledge providers; The ranking of knowledge content is determined based on the search intent type and the knowledge quality score; The results of the content recommendations for the user's search request are displayed according to the ranking of the knowledge content.

[0005] The content recommendation method further includes: Obtain user profiles; Determine the user fit score between the user profile and the knowledge content; The step of determining the knowledge content ranking based on the search intent type and the knowledge quality score includes: Determine the weight combination based on the search intent type; The comprehensive score of the knowledge content is obtained by combining the user adaptation score and the knowledge quality score according to the weight combination. The knowledge content of all knowledge providers is sorted from highest to lowest based on the comprehensive score, resulting in the knowledge content ranking.

[0006] The content recommendation method further includes: Obtain the bid amount from the knowledge provider; The knowledge provider's base bidding score is determined based on the bid amount; The step of determining the knowledge content ranking based on the search intent type and the knowledge quality score includes: The comprehensive score of the knowledge content is obtained by combining the bidding base score, the user adaptation score, and the knowledge quality score according to the weighted combination. The knowledge content of all knowledge providers is sorted from highest to lowest based on the comprehensive score, resulting in the knowledge content ranking.

[0007] Among them, the knowledge content of all knowledge providers is sorted in descending order of the comprehensive score, and the knowledge content with the same comprehensive score is sorted in descending order of the knowledge quality score.

[0008] The content recommendation method, after determining the weight combination based on the search intent type, further includes: Determine the user's historical interaction history with content recommendation results; The revenue parameters of the weight combination are determined based on the interaction records; The weight combination is dynamically updated based on the said return parameters.

[0009] The revenue parameter is calculated and determined based on the user click-through rate in the interaction record, the return on investment of the knowledge provider, and / or user satisfaction.

[0010] The step of determining the revenue parameters of the weight combination based on the interaction records includes: The revenue weighting is determined by integrating the user click-through rate in the interaction records, the return on investment of the knowledge provider, and / or user satisfaction to obtain the revenue parameter; The revenue weights are dynamically optimized based on knowledge quality indicators, user experience indicators, and / or business bidding indicators.

[0011] The step of determining the search scenario type based on the user's search request includes: Determine the semantic parsing result of the user's search request; Extract scene features from all scenes; The semantic parsing results are matched with the scene features of all scenarios, and the scenario with the highest matching degree is determined as the search scenario type.

[0012] The step of determining the search intent type based on the entity keywords of the user's search request and the semantic matching degree includes: When the semantic matching degree is within a preset range and the entity keywords include service matching keywords, the search intent type is determined to be demand-oriented consultation type; When the semantic matching degree is higher than the upper limit of the preset range, and the entity keywords do not include commercial keywords, the search intent type is determined to be pure knowledge acquisition. When the semantic matching degree is lower than the lower limit of the preset range, and the entity keywords include product / decision keywords, the search intent type is determined to be product / service decision type.

[0013] Prior to obtaining the knowledge quality score of the knowledge content from the knowledge provider, the content recommendation method further includes: Assess the qualifications and compliance of all knowledge providers; Exclude knowledge providers that do not meet qualifications and compliance requirements; The dimensions for determining the compliance of qualifications include the entity's qualifications, domain authorization, and / or compliance records.

[0014] The knowledge quality score of the knowledge content obtained from the knowledge provider includes: Obtain the knowledge accuracy score of the knowledge content; Obtain the knowledge authority score of the knowledge content; Obtain the knowledge compliance score of the knowledge content; The knowledge quality score is obtained by fusing the knowledge accuracy score, the knowledge authority score, and the knowledge compliance score according to the scenario differentiation weight determined by the search scenario type.

[0015] The content recommendation method further includes, after obtaining the knowledge quality score of the knowledge content from the knowledge provider, the content quality score of the knowledge content from the knowledge provider. Knowledge content with a quality score below a preset threshold is excluded. To address the aforementioned technical problem, this application also proposes a content recommendation device, comprising a memory and a processor coupled to the memory; wherein the memory stores program data, and the processor executes the program data to implement the content recommendation method described above.

[0016] To address the aforementioned technical problems, this application also proposes a computer storage medium for storing program data, which, when executed by a computer, is used to implement the aforementioned content recommendation method.

[0017] Compared with existing technologies, the beneficial effects of this application are: the content recommendation device obtains the device operation request of the simulated device; selects a real device according to the device operation request; forwards the device operation request to the real device so that the real device performs the operation according to the device operation request and returns the execution data; and sends the execution data to the simulated device. Through the above content recommendation method, which does not rely on the simulated data of simulation software, but uses the data of the real device to create the simulated device, it can realistically reflect the operating status of the device, support real-time data interaction with the real device, and improve the effectiveness of testing and the efficiency of system debugging. Through the above content recommendation method, the search scenario type is determined according to the user's search request, thereby matching the knowledge entity library corresponding to the search scenario type. Limiting the matched knowledge entities to the search scenario type can effectively improve matching efficiency and accuracy; the search intent type is determined based on the knowledge entity library and the semantic matching results, eliminating the fixed combination of search scenario type and search intent type, and realizing the dynamic nature of the search intent type; the influence of knowledge quality score on knowledge content ranking is realized according to the search intent type, reflecting the differentiated processing of results by the search intent type. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. Wherein: Figure 1 This is a flowchart illustrating an embodiment of the content recommendation method provided in this application; Figure 2 This is a schematic diagram of the overall process of the search knowledge recommendation bidding ranking method that integrates AI deep semantic understanding, multi-dimensional knowledge quality assessment and dynamic bidding strategy provided in this application; Figure 3 This is a flowchart illustrating another embodiment of the content recommendation method provided in this application; Figure 4 This is a schematic diagram of an embodiment of the content recommendation system architecture provided in this application; Figure 5 This is a schematic diagram of the structure of an embodiment of the content recommendation device provided in this application; Figure 6 This is a schematic diagram of the structure of an embodiment of the computer storage medium provided in this application. Detailed Implementation

[0019] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of the embodiments. Based on the embodiments of this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of this application.

[0020] The terms “first,” “second,” “third,” “fourth,” etc. (if present) in the specification, claims, and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a particular order or sequence. It should be understood that such data can be interchanged where appropriate so that embodiments of the application described herein can be implemented, for example, in orders other than those illustrated or described herein. Furthermore, the terms “comprising” and “having,” and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0021] This application provides a search knowledge recommendation bidding ranking method and system that integrates AI (Artificial Intelligence) deep semantic understanding, multi-dimensional knowledge quality assessment, and dynamic bidding strategies. AI search, with its deep semantic understanding and knowledge integration and generation capabilities, is gradually replacing traditional keyword matching search logic and becoming the mainstream form in the field of information retrieval. This application can be widely applied to the knowledge service field of general AI search platforms, providing technical support for the commercial recommendation and ranking of knowledge content, and achieving precise matching of "user intent - knowledge entity - commercial value".

[0022] Please refer to the details. Figure 1 and Figure 2 , Figure 1 This is a flowchart illustrating an embodiment of the content recommendation method provided in this application. Figure 2 This is a schematic diagram of the overall process of the search knowledge recommendation bidding ranking method provided in this application, which integrates AI deep semantic understanding, multi-dimensional knowledge quality assessment and dynamic bidding strategy.

[0023] The content recommendation method of this application is applied to a content recommendation device, wherein the content recommendation device can be a server, a terminal device, or a system in which the server and the terminal device cooperate with each other. Accordingly, the various parts of the content recommendation device, such as various units, sub-units, modules, and sub-modules, can all be set in the server, all in the terminal device, or separately in the server and the terminal device.

[0024] Furthermore, the aforementioned server can be either hardware or software. When the server is hardware, it can be implemented as a distributed server cluster consisting of multiple servers, or as a single server. When the server is software, it can be implemented as multiple software programs or software modules, such as software or software modules used to provide distributed server functionality, or as a single software program or software module; no specific limitations are made here.

[0025] like Figure 2 As shown, the core idea of ​​the search knowledge recommendation bidding ranking method provided in this application is as follows: Based on AI deep semantic understanding, through precise intent classification and multi-dimensional dynamic weight allocation, a complete closed loop of "input-processing-output-feedback" is constructed to achieve the accuracy, dynamism, and differentiation of knowledge recommendation bidding ranking, balancing user experience, knowledge authority, and commercial value. This method progresses through five core steps, forming a full-link technical system of "user request parsing - bidding subject and knowledge preprocessing - dynamic weight calculation - differentiated ranking display - real-time feedback optimization." The core logic of each step is abstractly expressed through mathematical formulas, ensuring the rigor and feasibility of the solution.

[0026] like Figure 1 As shown, the specific steps are as follows: Step S11: Determine the search scenario type based on the user's search request.

[0027] In the embodiments of this application, such as Figure 2 As shown, after a user inputs a multimodal search request, the content recommendation device first performs basic format validation and cleaning on the user's search request. If it is an invalid request, it directly returns a prompt message; if it is a valid request, it generates a temporary user profile based on the user's historical data, providing a basis for subsequent intent analysis.

[0028] The content recommendation method in this application achieves in-depth mining and accurate classification of user needs through four levels of analysis: "profile construction, scene recognition, entity extraction, and intent determination," providing a core basis for subsequent differentiation strategies.

[0029] Specifically, the content recommendation device generates temporary user profile vectors based on users' historical behavior data (including but not limited to: search history, click preferences), device information (including but not limited to: terminal type, operating system), and geographic characteristics (including but not limited to: IP address, frequently used regions), quantifying the intensity of users' preferences in various fields.

[0030] Let user profile vector ,in, On behalf of the user in The intensity of preference for class labels The label categories provided in this application include, but are not limited to: "medical needs", "academic research", "consumer decision", and "cross-border procurement".

[0031] In one specific embodiment, if 30% of a user's historical searches are related to healthcare, 50% are related to consumer decisions, and 20% are related to academic research, then... .

[0032] It should be noted that the user profile vectors used in this application It can be continuously updated based on user behavior, making it a dynamically changing piece of information.

[0033] Furthermore, the content recommendation device of this application identifies the user's search scenario by semantic parsing of the user's search request and matching the scenario features of each search scenario, and assigns scenario weights according to the importance of the scenario (professionalism, sensitivity), with higher weights for more critical scenarios.

[0034] The search scenario set provided in this application is: Sc = {healthcare, academic research, cross-border shopping, financial consulting, general knowledge, ...}. It should be noted that the specific content of the search scenario set Sc = {sc1, sc2, sc3, ...} can be set according to actual circumstances and is not specifically limited here.

[0035] In one specific embodiment, ,in 1 represents a normal scenario, and 5 represents a highly sensitive / professional scenario. The specific S is related to the number of scenarios.

[0036] It should be noted that the above weights are not fixed values ​​and can be continuously updated and optimized based on user behavior during subsequent user interactions.

[0037] This application can utilize large-scale models to intelligently evaluate weights. The weighting instruction is: You are a professional user intent scene classifier. Your task is to identify the scene type based on the user's search request and assign the corresponding scene weight S. Specifically, scenario weights are assigned based on their importance (such as the sensitivity of question-and-answer knowledge to professional knowledge, commercial value, etc.), with more critical scenarios receiving higher weights. An open-source model was used to evaluate and assign weights of S=5 to healthcare / financial consulting, S=3 to cross-border shopping / corporate procurement, and S=1 to general knowledge. This section presents a weight configuration method based on a large model.

[0038] The scenarios mentioned above are all relatively high-level scenarios. For example, the medical and health / financial consulting scenario can be further refined into the medical and health scenario, with a weight of S=4.5, and the financial consulting scenario, with a weight of S=4.8, etc.

[0039] Step S12: Match the request semantic vector of the user's search request with the entity semantic vector of the knowledge entity base corresponding to the search scenario type to determine the semantic matching degree.

[0040] In the embodiments of this application, such as Figure 2 As shown, this application adopts a BiLSTM+vertical domain knowledge graph fusion model to extract the core knowledge entities and core intents in the search request, and calculates the matching degree between the intent and the knowledge entity database through cosine similarity.

[0041] The aforementioned knowledge entity database can be limited to those corresponding to the search scenario type determined in step S11. For example, if the search scenario type determined in step S11 is a medical and health scenario, then the knowledge entity database used for matching is the medical and health knowledge entity database, and other knowledge entity databases are not used for matching for the time being.

[0042] Specifically, the semantic vector of a user's search request and entity semantic vectors in the knowledge entity base The formula for calculating semantic matching degree is:

[0043] in, The semantic matching score is used to determine the semantic matching degree. and They are vectors and The value in the i-th dimension, where m is the total number of dimensions of the vector.

[0044] in, The closer the semantic matching degree is to 1, the higher the matching degree between the intent and the knowledge entity.

[0045] Step S13: Determine the search intent type based on the entity keywords of the user's search request and the semantic matching degree.

[0046] In this embodiment of the application, since the search scenario type and search intent type are not a fixed one-to-one correspondence, it is necessary to further determine the search intent type based on the semantic matching degree calculated in step S12 and the entity keywords in the user's search request. By introducing the entity information of the user's search request and the semantic matching information, the search intent type can be made more precise and differentiated.

[0047] Specifically, this application can classify user intent into three categories using a lightweight intent classification model (such as the FastText optimized model), providing a basis for differentiated ranking. The input of the lightweight intent classification model is the semantic matching degree and search scenario type determined in the above steps, and the output is the search intent type.

[0048] In one specific embodiment, the intent type set is: Y = {pure knowledge acquisition type (Y1), demand-oriented consultation type (Y2), product / service decision type (Y3), ...}.

[0049] The determination strategy / method is as follows: When the semantic matching degree is within a preset range and the entity keywords include service matching keywords, the search intent type is determined to be demand-oriented consultation type.

[0050] For example, It also includes service matching keywords (such as "which one", "where", "how to choose"), for example: "Which hospital should I go to if my eyes are uncomfortable?" or "What service provider should a company look for for cross-border customs clearance?"

[0051] When the semantic matching degree is higher than the upper limit of the preset range, and the entity keywords do not include commercial keywords, the search intent type is determined to be pure knowledge acquisition.

[0052] For example, Furthermore, there are no clear commercial keywords (such as "what it is", "principle", "definition"), for example: "What is universal gravitation?" or "Interpretation of cross-border e-commerce policies".

[0053] When the semantic matching degree is lower than the lower limit of the preset range, and the entity keywords include product / decision keywords, the search intent type is determined to be product / service decision type.

[0054] For example, It also contains product / decision keywords (such as "cost-effectiveness", "recommendation", "purchase"), for example: "high-cost-effective photovoltaic module brands" and "recommendation of compliant suppliers for cross-border procurement".

[0055] The mathematical expression for the lightweight intent classification model is as follows:

[0056] in, Ultimate Intent Type Based on experience, intent categories are currently divided into three types: pure knowledge acquisition, demand-oriented consultation, and product / service decision-making.

[0057] It should be noted that this application can classify all cases other than the three judgment strategies mentioned above as pure knowledge acquisition. Alternatively, in other implementations, users can specify and set relevant types, judgment thresholds, or judgment strategies, etc.

[0058] Step S14: Obtain the knowledge quality score of the knowledge content from the knowledge provider.

[0059] In this application embodiment, the knowledge content submitted by the knowledge provider can be quantitatively scored from three dimensions: accuracy, authority, and compliance, to generate an initial knowledge quality score.

[0060] Scoring Dimensions Strategies / Methods: Knowledge accuracy score (∈[0,100]) is determined based on the fact-checking model; 100 points are awarded for no errors.

[0061] Knowledge authority score (∈[0,100]), associated provider qualification level, top-tier hospitals / industry leaders get 100 points.

[0062] Knowledge compliance score (∈[0,100]), based on the judgment of the security audit platform, 100 points are awarded for no sensitive / illegal content.

[0063] This application can obtain a knowledge quality score for knowledge content by fusing the scores of the above three dimensions with fixed weights.

[0064] Furthermore, this application can also configure dynamic weights differently based on the type of search scenario. For example: Healthcare / Educational Scenarios: .

[0065] Consumer / Cross-border E-commerce Scenarios: .

[0066] Common scenarios: .

[0067] The requirements for the above dynamic weights are: Therefore, the knowledge quality score of the knowledge content in this application can be calculated as follows:

[0068] Furthermore, before calculating the knowledge quality score for each knowledge provider's content, the provider's qualifications can be verified. For example, this application could establish a three-tiered qualification review system to comprehensively verify the knowledge provider's main qualifications, domain authorization, and compliance records, allowing only compliant entities to participate in the ranking.

[0069] The review dimensions include, but are not limited to: entity qualifications (business license, medical institution practice license, etc.), field authorization (medical knowledge requires certification from a top-tier hospital, educational content requires school operating qualifications, etc.), and compliance records (no record of illegal promotion in the past year).

[0070] The qualification compliance determination strategy is as follows: when all three levels of review are passed, C=1; when any level of review fails, C=2.

[0071] Its mathematical expression is:

[0072] in, As the main qualification, For domain licensing, For compliance records.

[0073] Finally, this application outputs a compliance identifier C (1 = compliant, 0 = non-compliant) for each knowledge provider. The knowledge provider can then proceed to the next stage of the process. Knowledge providers were excluded.

[0074] Furthermore, this application also provides filtering rules for the knowledge content of knowledge providers, by setting a score threshold to exclude knowledge content that is below the score threshold, i.e. The knowledge content will then proceed to the next stage of the process.

[0075] Then, this application will transform the knowledge content that has passed the quality screening into a standardized structured format, associate core knowledge entities with intent tags, and improve the efficiency of subsequent matching.

[0076] Structured format: ,in To associate knowledge entities, To adapt to intent types, This is for structured knowledge content, such as text, charts, and knowledge cards.

[0077] Association determination strategy / method: based on semantic matching degree Associate core knowledge entities with intent types.

[0078] Mathematical expression: .

[0079] in, For candidate knowledge set, For a full knowledge entity database, This is a full library of intent types.

[0080] Step S15: Determine the ranking of knowledge content based on the search intent type and the knowledge quality score.

[0081] In this embodiment, the application constructs a dynamic weight calculation model based on intent type and scenario characteristics, generates a comprehensive ranking score based on the knowledge quality score calculated in step S14, and realizes differentiated weight allocation.

[0082] Furthermore, this application can also construct a dynamic weight calculation model based on intent type and scenario characteristics, and generate a comprehensive ranking score by combining the scores of three dimensions: bidding amount, knowledge quality, and user adaptation, thereby achieving differentiated weight allocation.

[0083] It should also be noted that users can select any two of the three dimensions mentioned above to calculate the overall ranking score, which will not be elaborated on here.

[0084] This application requires standardization of the three dimensions of bidding base score, knowledge quality score, and user adaptation score to eliminate differences in units and ensure that the scores of each dimension can be directly weighted and calculated.

[0085] Raw score standardization strategy method: B: Base bid score (∈[0,1000]), which is the score after standardizing the bid amount of the knowledge provider (the higher the bid, the higher the score). The bid amount can be the amount of advertising money that the knowledge provider gives to the platform.

[0086] Q: Knowledge quality score (∈[0,100]), which is the initial knowledge quality score calculated in step S14.

[0087] User fit score (∈[0,1]), calculated based on the matching degree between temporary user profile and knowledge content. .

[0088] The standardized formulas for the scores of the above three dimensions are as follows:

[0089] in, , , These are the minimum and maximum scores for this dimension, respectively, after standardization. ∈[0,1].

[0090] Finally, we obtain the standardized scores for the three dimensions mentioned above. .

[0091] This application can integrate the scores of the above dimensions with fixed weights to obtain a comprehensive ranking score for each knowledge content, and rank the knowledge content of all knowledge providers according to the comprehensive ranking score.

[0092] Specifically, this application is ranked according to its overall score. Sort the data in descending order to generate the final ranking sequence. For scores in the same category, rank the data based on knowledge quality. Secondary sorting. In other implementations, secondary sorting can also be performed based on scores from other dimensions, which will not be elaborated here.

[0093] Mathematical expression: in The recommended result is the one with the highest overall score, and k is the number of recommended results (default k=10).

[0094] This application can also integrate the scores of the above dimensions based on the differentiated weights of search intent types to obtain a comprehensive ranking score for each knowledge content, and rank the knowledge content of all knowledge providers based on the comprehensive ranking score.

[0095] Specifically, this application is based on the final intent type. We implement a differentiated weight allocation strategy to ensure that business recommendations are accurately matched with user needs.

[0096] Pure knowledge acquisition category ( ): Weighting strategy: .

[0097] Core logic: Block commercial recommendations and sort only by knowledge quality and user suitability.

[0098] Demand-oriented consulting ( ): Weighting strategy: Healthcare / Education scenarios Consumption scenarios .

[0099] Core logic: Prioritize the authority of knowledge while taking into account commercial adaptability.

[0100] Product / Service Decision-Making ( ): Weighting strategy: Consumer / Cross-border e-commerce scenarios Enterprise procurement scenarios .

[0101] Core logic: Balance commercial value with user fit and improve conversion efficiency.

[0102] Mathematical expression: .

[0103] This application uses a weighted summation formula to calculate the comprehensive ranking score of each knowledge content based on the standardized score and the final weight coefficients determined above.

[0104] Overall score formula:

[0105] Score range: The higher the value, the higher the ranking.

[0106] Furthermore, the final weight coefficients determined above can be regarded as initial values ​​of weight coefficients associated with the search intent type. This application can also adjust the weight coefficients of the three dimensions in real time through the reinforcement learning PPO (Proximal Policy Optimization) algorithm. The weights change dynamically with time t, user feedback, and scenario type.

[0107] Please refer to the details. Figure 3 , Figure 3 This is a flowchart illustrating another embodiment of the content recommendation method provided in this application.

[0108] like Figure 1 As shown, the specific steps are as follows: Step S21: Determine the user's historical content recommendation results interaction records.

[0109] In this application embodiment, the application needs to determine the user's interaction records with the historical content recommendation results for each search intent type, including but not limited to user click-through rate, user satisfaction, etc. In addition, the application can also determine the return on investment of the knowledge provider as a powerful reference data for the dynamic weight allocation of search intent types.

[0110] Step S22: Determine the revenue parameters of the weight combination based on the interaction records.

[0111] In this embodiment of the application, the revenue function for calculating the revenue parameters is:

[0112] in, User click-through rate at time t : Return on investment for the bidding party at time t User satisfaction at time t (based on feedback rating).

[0113] The weight combination in the above return function must satisfy: The specific numerical values ​​can be dynamically configured by the platform or the user.

[0114] Step S23: Dynamically update the weight combination according to the revenue parameters.

[0115] In this embodiment of the application, the scheme for dynamically updating the weights of various search intent types is as follows:

[0116]

[0117]

[0118] in, For learning rate, This is the gradient of the profit function.

[0119] Step S16: Display the recommended content results of the user's search request according to the ranking of the knowledge content.

[0120] In this application embodiment, in addition to adopting conventional multimodal display methods, this application can also implement differentiated ranking strategies and display rules based on comprehensive ranking scores and intent types to ensure the compliance, transparency and adaptability of recommendation results.

[0121] Specifically, this application can execute differentiated display logic for different search intent types, balancing user experience and business value.

[0122] Pure knowledge acquisition class ): Content displayed: Only authoritative knowledge content will be displayed, and all commercial recommendation labels and commercial information will be blocked.

[0123] Additional information: Indicate the source of the knowledge (such as "academic institution", "authoritative publication", "industry standard") and the knowledge quality score Q.

[0124] Mathematical expression:

[0125] in, To demonstrate the results, For knowledge content, As a source of knowledge, As a score for knowledge quality, This is a function for displaying pure knowledge.

[0126] Demand-oriented consulting ( ): Display content: The top 3 bidding results are marked with the "Bidding Recommendation" label, and subsequent displays are non-bidding authoritative knowledge content.

[0127] Additional information: Display the provider's compliance qualifications (such as "Top-Tier Hospital Certification", "Cross-border E-commerce Qualification", "AEO Certification"). The priority of displaying these qualifications is relative to... Positive correlation.

[0128] Mathematical expression:

[0129] in, For functions that display paid results, For the search results, For the compliance qualifications of the search results, Functions for displaying standard / authoritative results. The knowledge source for the search results.

[0130] Product / Service Decision-Making ( ): Display content: The top 5 bidding results are marked with the "Bidding Recommendation" label, and subsequent non-bidding content with high user relevance is displayed.

[0131] Additional information: Showcases the core selling points of the product / service, user reviews, and compliance qualifications, categorized by... Displayed in sorted order.

[0132] Mathematical expression:

[0133] in, For product display functions, As the core selling point, For user reviews, For compliance qualifications, To adapt to the display function, For user adaptation.

[0134] This application implements differentiated ranking strategies and display rules based on comprehensive ranking scores and intent types to ensure the compliance, transparency, and suitability of the recommendation results.

[0135] Furthermore, this application can also optimize the display layout and interaction logic for different terminals such as PC, mobile, and mini-programs to ensure a consistent user experience.

[0136] For example, this application can adjust the content layout according to the terminal screen size and interaction method, including but not limited to: displaying concise knowledge cards first on mobile devices and displaying complete text and image content on PC devices.

[0137] Furthermore, such as Figure 2 As shown, this application can also construct a real-time feedback loop based on streaming computing and reinforcement learning, dynamically updating the ranking score according to user behavior and competitor strategies, ensuring continuous optimization of the ranking effect.

[0138] Specifically, this application uses a streaming computing engine (Flink 1.15) to collect user behavior data and competitor bidding strategy change data in real time, providing input for feedback adjustment.

[0139] User behavior data: Click feedback FC (click = 1, no click = 0), dwell time feedback FS (normalized to [0, 1], the longer the dwell time, the higher the score), conversion feedback FT (complete consultation / purchase = 1, otherwise = 0).

[0140] Competitive strategy data: Competitive bidding change Where t is the current time and t-1 is the previous time.

[0141] Data collection frequency: User behavior data is collected in real time, while competitor data is collected every 3 seconds.

[0142] Mathematical expression: .

[0143] Then, this application integrates multi-dimensional user behavior data to generate user feedback coefficients, quantifying user satisfaction with the recommendation results.

[0144] Specifically, the feedback weight configuration strategy is: δ1+δ2+δ3=1, with the default values ​​being δ1=0.3, δ2=0.4, and δ3=0.3. It should be noted that the above weight configuration can be adjusted according to the scenario, and no specific limitations are specified here.

[0145] Feedback coefficient formula: , where F(t)∈[0,1], and a higher value represents a higher level of user satisfaction.

[0146] This application combines user feedback coefficients with changes in competitor bidding to update the overall ranking score every 3 seconds, enabling real-time dynamic adjustment of the ranking.

[0147] Coefficient configuration strategy: positive incentive coefficient λ∈[0.05, 0.2] (rewarding high satisfaction results), competitor influence coefficient μ∈[0.03, 0.15] (balancing the impact of competitor bidding).

[0148] Dynamic update formula: Total(t+1) = Total(t) + λF(t) μΔB(t).

[0149] Constraint: Total(t+1)∈[0,1] (scores below 0 are assigned 0, scores above 1 are assigned 1).

[0150] Finally, this application can determine the updated overall ranking score Total(t+1) and the new ranking sequence R(t+1).

[0151] In another specific embodiment, this application can also generate a three-dimensional evaluation report of "knowledge quality-user experience-business ROI" daily to optimize the reward function and weight configuration strategy of the reinforcement learning model.

[0152] The three-dimensional evaluation index scheme is as follows: Knowledge quality indicator: Average knowledge quality score

[0153] User experience metrics: Average user feedback coefficient T represents the duration of the day, in seconds.

[0154] Business ROI Metric: Average Bidding , where n is the number of bidders.

[0155] Optimization of the payoff function: Adjustment based on three-dimensional evaluation indicators , , .

[0156] Finally, this application can output a three-dimensional evaluation report and an optimized profit function according to a cycle. Updated weight configuration strategies, etc.

[0157] For the content recommendation method in the above embodiments, this application adopts a layered architecture design of application layer-service layer-storage layer. The responsibilities of each layer module are clearly defined and they work together to ensure the accuracy, dynamism and differentiation of the entire knowledge recommendation bidding ranking process.

[0158] Please refer to the details. Figure 4 , Figure 4 This is a schematic diagram of an embodiment of the content recommendation system architecture provided in this application.

[0159] like Figure 4 As shown, the content recommendation system architecture provided in this application is specifically implemented through the collaborative operation of the application layer, service layer, and storage layer. The specific functions of each layer module are described below: The application layer mainly includes the F101 user interaction module, the F102 bidding party (i.e. knowledge provider) operation module, the F103 intent parsing and scene recognition module, and the F104 ranking display and strategy configuration module.

[0160] The F101 user interaction module is responsible for receiving multimodal user requests, displaying ranking results, and collecting user behavior data. It supports multi-platform adaptation, including PC, mobile, and mini-programs, and optimizes content layout and interaction logic according to terminal characteristics; it also collects user clicks, dwell time, conversions, and other behavioral data in real time to provide input for the feedback loop.

[0161] The F102 bidding platform operations module focuses on uploading bidding platform qualifications, managing bids, submitting promotional content, and viewing data reports. It provides functions for checking qualification review progress and monitoring and reminding bidders of competitor strategies; it also supports bidders in adjusting their bids and content strategies based on evaluation reports to improve promotional effectiveness.

[0162] The F103 intent parsing and scene recognition module is responsible for accurately determining user search intent and identifying search scenes. It integrates a BiLSTM model with a vertical domain knowledge graph to extract core knowledge entities and calculate semantic matching degree; it also completes scene weight assignment and intent type classification according to preset rules, providing a core basis for subsequent weight calculation.

[0163] The F104 ranking display and strategy configuration module is responsible for generating comprehensive ranking sequences, executing differentiated display rules, and configuring system parameters. It outputs corresponding display content based on user intent types and marks commercial recommendations with compliance tags. It also supports platform operators in configuring core indicators such as scenario weights, security review thresholds, and reinforcement learning model parameters.

[0164] The service layer mainly includes the F201 qualification and knowledge review engine, the F202 dynamic weight calculation engine, the F203 real-time feedback and strategy optimization engine, and the F204 knowledge structuring processing engine.

[0165] The F201 Qualification and Knowledge Audit Engine serves as a core component of compliance screening. Relying on the storage layer qualification standard library and knowledge quality rule library, it conducts three-level verification of the bidder's qualifications, domain authorization, and compliance records. At the same time, it performs multi-dimensional quality scoring on knowledge content, outputting compliance identifiers and quality scores to filter out non-compliant entities and content.

[0166] The F202 dynamic weight calculation engine is responsible for standardizing the scores of each item and dynamically adjusting the weight coefficients. It eliminates the differences in the dimensions of the scores of each dimension through the extreme value method; it optimizes the weight coefficients in real time based on the reinforcement learning PPO algorithm, combined with user feedback and business ROI; and it executes differentiated weight strategies according to intent type and scenario characteristics to output a comprehensive ranking score.

[0167] The F203 real-time feedback and strategy optimization engine is responsible for processing user behavior data, calculating feedback coefficients, and dynamically updating rankings. Based on the Flink streaming computing framework, it collects and analyzes user behavior data and competitor bidding changes in real time; generates user feedback coefficients to drive dynamic updates to the overall ranking score; and outputs a daily three-dimensional evaluation report on "knowledge quality, user experience, and business ROI" to optimize the reward function of the reinforcement learning model.

[0168] The F204 knowledge structuring engine is responsible for standardizing and transforming knowledge content that has passed the initial quality screening. It associates core knowledge entities with intent tags, transforming unstructured text into knowledge cards in a unified format; and establishes a mapping relationship between knowledge content and user intent and scenario, improving the efficiency of subsequent matching and weight calculation.

[0169] The storage layer F301 mainly includes a user and bidder database, a vertical domain knowledge graph library, a weight and strategy configuration library, a knowledge quality and security rule library, and a behavior and feedback data warehouse.

[0170] The user and bidder database stores historical user behavior data and temporary profile vectors; it also retains bidder qualification information, bidding records, promotional content, and review results. Utilizing MySQL 8.0, it supports high-concurrency read and write operations, with a data backup cycle of one hour, ensuring the security and availability of core user and bidder data.

[0171] This vertical domain knowledge graph library aggregates knowledge entities and relationships from multiple fields, including healthcare, academic research, and cross-border shopping. Built on the Neo4j database, it supports rapid querying of knowledge entities and dynamic updates of graph relationships, providing core knowledge support for intent parsing and scene recognition.

[0172] The weight and policy configuration library stores weight configuration rules for various scenarios, intent classification criteria, reinforcement learning model parameters, and differentiated weight policies. It employs a Redis+MySQL hybrid storage architecture, with Redis storing real-time weight data to ensure read / write response times within 5ms; and MySQL storing historical policy data to support policy iteration and review.

[0173] The knowledge quality and security rule base stores a knowledge quality scoring matrix, security audit rules, and a blacklist of prohibited content. It covers three scoring standards: accuracy, authority, and compliance, and supports dynamic rule updates; providing a unified basis for the qualification and knowledge audit engine to ensure content quality and compliance.

[0174] The behavior and feedback data warehouse stores user behavior data, competitor bidding changes, feedback coefficient calculation results, and daily evaluation reports. It employs a Kafka + MongoDB architecture, with Kafka handling real-time streaming data and MongoDB storing historical data for one year, providing data support for strategy optimization and model training.

[0175] This application's content recommendation method innovatively employs a dual-drive weight allocation strategy based on "intent-scenario," dynamically adjusting the weight ratios of commerce, quality, and user fit for different intent types. For example, for pure knowledge acquisition intents, commerce weight is completely shielded to ensure users receive authoritative information; for product decision-making intents, commercial value and user experience are balanced to improve conversion efficiency. Compared to traditional solutions, user satisfaction is increased by over 35%, and the rate of achieving authoritative knowledge content standards is increased by 50%.

[0176] This application's content recommendation method constructs a two-tiered screening system of "three-level qualification verification + three-dimensional knowledge scoring" to eliminate unqualified entities and substandard content from the source. Simultaneously, it dynamically configures quality scoring weights based on scenario characteristics; for example, authoritative scoring is prioritized in medical scenarios, while compliance scoring is prioritized in consumer scenarios. Compared to traditional solutions, the violation rate of bid-based content is reduced by 60%, and user trust in recommended content increases by 45%.

[0177] This application's content recommendation method is based on a streaming computing engine to build a real-time feedback loop, updating ranking scores every 3 seconds to dynamically adapt to user behavior and competitor changes. A daily 3D evaluation report is generated to reverse-optimize strategies, achieving continuous iteration from "user needs - ranking results - strategy optimization." Compared to traditional solutions, ranking response time is reduced from 24 hours to 3 seconds, and business ROI is improved by over 25%.

[0178] This application's content recommendation method features customized display schemes for three intent types, clearly marking commercial recommendations with identifiers, displaying compliance qualifications and knowledge quality scores to ensure users' right to know; it also achieves multi-platform adaptation, prioritizing concise knowledge cards on mobile devices and displaying complete text and image content on PCs. Compared to traditional solutions, user click-through rates are increased by 30%, and the consistency of multi-platform experience reaches 98%.

[0179] Those skilled in the art will understand that, in the above-described method of the specific implementation, the order in which each step is written does not imply a strict execution order and does not constitute any limitation on the implementation process. The specific execution order of each step should be determined by its function and possible internal logic.

[0180] To implement the above content recommendation method, this application also proposes a content recommendation device, which can be found in the following details. Figure 5 , Figure 5 This is a schematic diagram of an embodiment of the content recommendation device provided in this application.

[0181] The content recommendation device 400 in this embodiment includes a processor 41, a memory 42, an input / output device 43, and a bus 44.

[0182] The processor 41, memory 42, and input / output device 43 are respectively connected to the bus 44. The memory 42 stores program data, and the processor 41 is used to execute the program data to implement the content recommendation method described in the above embodiments.

[0183] In this embodiment, processor 41 can also be referred to as a CPU (Central Processing Unit). Processor 41 may be an integrated circuit chip with signal processing capabilities. Processor 41 can also be a general-purpose processor, a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components. The general-purpose processor can be a microprocessor, or processor 41 can be any conventional processor.

[0184] This application also provides a computer storage medium; please refer to the following: Figure 6 , Figure 6 This is a schematic diagram of a computer storage medium according to an embodiment of the present application. The computer storage medium 600 stores a computer program 61, which, when executed by a processor, is used to implement the content recommendation method of the above embodiment.

[0185] When the embodiments of this application are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) or processor to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0186] The above description is merely an embodiment of this application and does not limit the patent scope of this application. Any equivalent structural or procedural transformations made using the content of this application's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of this application.

Claims

1. A content recommendation method, characterized in that, The content recommendation method includes: Determine the search scenario type based on the user's search request; The semantic vector of the user's search request is matched with the semantic vector of the entity in the knowledge entity database corresponding to the search scenario type to determine the semantic matching degree. The search intent type is determined based on the entity keywords of the user's search request and the semantic matching degree; The knowledge quality score of the knowledge content obtained from knowledge providers; The ranking of knowledge content is determined based on the search intent type and the knowledge quality score; The results of content recommendations based on the knowledge content are displayed according to the ranking of the user's search request.

2. The content recommendation method according to claim 1, characterized in that, The content recommendation method also includes: Obtain user profiles; Determine the user fit score between the user profile and the knowledge content; The step of determining the knowledge content ranking based on the search intent type and the knowledge quality score includes: Determine the weight combination based on the search intent type; The comprehensive score of the knowledge content is obtained by combining the user adaptation score and the knowledge quality score according to the weight combination. The knowledge content of all knowledge providers is sorted from highest to lowest based on the comprehensive score, resulting in the knowledge content ranking.

3. The content recommendation method according to claim 2, characterized in that, The content recommendation method also includes: Obtain the bid amount from the knowledge provider; The knowledge provider's base bidding score is determined based on the bid amount; The step of determining the knowledge content ranking based on the search intent type and the knowledge quality score includes: The comprehensive score of the knowledge content is obtained by combining the bidding base score, the user adaptation score, and the knowledge quality score according to the weight combination. The knowledge content of all knowledge providers is sorted from highest to lowest based on the comprehensive score, resulting in the knowledge content ranking.

4. The content recommendation method according to claim 3, characterized in that, The knowledge content of all knowledge providers is sorted in descending order of the comprehensive score, and the knowledge content with the same comprehensive score is sorted in descending order of the knowledge quality score.

5. The content recommendation method according to claim 2, characterized in that, After determining the weight combination based on the search intent type, the content recommendation method further includes: Determine the user's historical interaction history with content recommendation results; The revenue parameters of the weight combination are determined based on the interaction records; The weight combination is dynamically updated based on the said return parameters.

6. The content recommendation method according to claim 5, characterized in that, The revenue parameters are calculated and determined based on the user click-through rate in the interaction records, the knowledge provider's return on investment, and / or user satisfaction.

7. The content recommendation method according to claim 6, characterized in that, The step of determining the revenue parameters of the weight combination based on the interaction records includes: The revenue weighting is determined by integrating the user click-through rate in the interaction records, the return on investment of the knowledge provider, and / or user satisfaction to obtain the revenue parameter; The revenue weights are dynamically optimized based on knowledge quality indicators, user experience indicators, and / or business bidding indicators.

8. The content recommendation method according to claim 1, characterized in that, Determining the search scenario type based on the user's search request includes: Determine the semantic parsing result of the user's search request; Extract scene features from all scenes; The semantic parsing results are matched with the scene features of all scenarios, and the scenario with the highest matching degree is determined as the search scenario type.

9. The content recommendation method according to claim 1, characterized in that, Determining the search intent type based on the entity keywords of the user's search request and the semantic matching degree includes: When the semantic matching degree is within a preset range and the entity keywords include service matching keywords, the search intent type is determined to be demand-oriented consultation type; When the semantic matching degree is higher than the upper limit of the preset range, and the entity keywords do not include commercial keywords, the search intent type is determined to be pure knowledge acquisition. When the semantic matching degree is lower than the lower limit of the preset range, and the entity keywords include product / decision keywords, the search intent type is determined to be product / service decision type.

10. The content recommendation method according to claim 1, characterized in that, Before obtaining the knowledge quality score of the knowledge content from the knowledge provider, the content recommendation method further includes: Assess the qualifications and compliance of all knowledge providers; Exclude knowledge providers that do not meet qualifications and compliance requirements; The dimensions for determining the compliance of qualifications include the entity's qualifications, domain authorization, and / or compliance records.

11. The content recommendation method according to claim 1, characterized in that, The knowledge quality score of the knowledge content obtained from the knowledge provider includes: Obtain the knowledge accuracy score of the knowledge content; Obtain the knowledge authority score of the knowledge content; Obtain the knowledge compliance score of the knowledge content; The knowledge quality score is obtained by fusing the knowledge accuracy score, the knowledge authority score, and the knowledge compliance score according to the scenario differentiation weight determined by the search scenario type.

12. The content recommendation method according to claim 11, characterized in that, After obtaining the knowledge quality score of the knowledge content from the knowledge provider, the content recommendation method further includes: Exclude knowledge content whose knowledge quality score is lower than a preset score threshold.

13. A content recommendation device, characterized in that, The content recommendation device includes a memory and a processor coupled to the memory; The memory is used to store program data, and the processor is used to execute the program data to implement the content recommendation method as described in any one of claims 1 to 12.

14. A computer storage medium, characterized in that, The computer storage medium is used to store program data, which, when executed by the computer, is used to implement the content recommendation method as described in any one of claims 1 to 12.

Images