A method, system, and doll device for information recommendation based on sentiment semantics

Abstract

This invention discloses an information recommendation method, system, and doll device based on emotional semantics. The method determines the core interaction intent of a first user by inputting a complete dialogue context into an intent recognition model; real-time input information is then input into a text vector representation model to generate a query feature vector; when the interaction intent is determined to be a couple's wish-related intent, the identity information of a second user associated with the first user as a couple is obtained; user profile data of the second user is retrieved from a vector database, the user profile data containing the user's wish list feature vector set; similarity matching calculations are performed between the query feature vector and each wish feature vector in the wish list feature vector set to obtain a suitable wish feature vector, which is then converted into recommended wish information, and finally, corresponding response content is generated and fed back to the first user. This reduces user operation costs and enhances the naturalness of emotional semantic interaction.

Description

Technical Field

[0001] This invention relates to the field of intelligent human-computer interaction and intelligent information recommendation technology, and in particular to an information recommendation method, system and doll device based on emotional semantics. Background Technology

[0002] With the continuous development of artificial intelligence technology in the field of emotional interaction, emotional semantic interaction products have been widely used in scenarios such as maintaining intimate relationships. By understanding users' interaction intentions and matching relevant information, they provide personalized services to users, helping to improve the efficiency and experience of emotional communication in intimate relationships. In existing technologies, most intimate relationship information recommendation products adopt a single user-demand-driven implementation logic. For example, some smart assistant products allow users to manually upload wish lists and set reminders. When a user actively queries wish recommendations, the system only returns the user's previously preset information list based on their own wishes. However, existing technologies still cannot automatically identify users' implicit intentions regarding the connection between other users' wishes, such as those of their partners, during natural interactions. They can only make recommendations based on single user information, making it difficult to match the emotional needs and interaction scenarios between couples, resulting in a poor interactive experience. Summary of the Invention

[0003] This invention addresses the shortcomings of existing technologies by providing an information recommendation method based on sentiment semantics, comprising the following steps: Step S1: Receive real-time input information from the first user, and fuse the real-time input information with the historical dialogue content of the current session to construct a complete dialogue context; input the complete dialogue context into the trained intent recognition model to determine the core interaction intent of the first user; and input the real-time input information into the text vector representation model for semantic encoding to generate a query feature vector.

[0004] Step S2: When the interaction intent is determined to be a couple's wish association intent, the identity information of the second user who is associated with the first user as a couple is obtained. The second user has a couple relationship tag associated with the first user. Based on the identity information of the second user, the user profile data of the second user is searched from the vector database. The user profile data contains the user's wish list feature vector set. The wish list feature vector set is configured to be limited to query matching only with the user to which it belongs and users with a couple relationship tag associated with the user to which it belongs.

[0005] Step S3: Perform similarity matching calculations between the query feature vector and each wish feature vector in the wish list feature vector set to obtain the appropriate wish feature vector and convert it into recommended wish information. Based on the recommended wish information, generate the corresponding reply content to the real-time input information and feed it back to the first user.

[0006] Preferably, the wish list feature vector set is configured to be generated after semantic encoding based on the wish information pre-input by the second user through a text vector representation model.

[0007] Preferably, the wish list feature vector set is further configured to generate new wish feature vectors by semantically encoding the interaction content containing wishes collected by the second user during the dialogue interaction with the terminal through a text vector representation model, and then supplementing the user's original wish list feature vector set.

[0008] Preferably, the user profile data further includes the user's interest vector, preference vector, and behavioral feature vector; The interest vector is generated based on the user's pre-input information through a text vector representation model after semantic encoding, and is configured to be used for query matching by the current user and users with corresponding couple relationship tags, while rejecting queries from other users; The preference vector is generated based on the personal preference information extracted by the user during the dialogue and interaction with the terminal, and is configured to be available only for the user to query and match and to reject queries from other users. The behavioral feature vector is generated based on the historical dynamic behavioral features extracted by the user during the dialogue and interaction with the terminal, and is configured to be available only for the user to query and match and to reject queries from other users.

[0009] Preferably, the query feature vector is compared with the behavioral feature vector, preference vector, and interest vector of the first user, and the interest vector and couple's wish vector of the second user for similarity matching calculation. Based on the matching results, the most similar wish feature vectors of a preset number are obtained and converted into recommended wish information. Based on the recommended wish information, the corresponding reply content is generated and fed back to the first user.

[0010] Preferably, step S2 further includes: When the couple's wish association intent includes the requirement for related recommendations, the second user's wish list feature vector set is retrieved based on the couple relationship tag. The current time node, the realization conditions of each item in the wish list, and the first user's historical behavior data are combined to determine whether the recommendation triggering conditions are met. If they are met, the corresponding reply content is generated based on the recommended wish and the adapted recommended content and fed back to the first user.

[0011] Preferably, the current time node includes statutory holidays, couples' anniversaries, the second user's birthday, and the second user-defined time node in the wish list. The real-time time is obtained through the time calibration module and accurately matched with the time node.

[0012] This invention also discloses an information recommendation system based on sentiment semantics, comprising: The information receiving module is configured to receive real-time input information from the first user, merge the real-time input information with the historical dialogue content of the current session to construct a complete dialogue context, pass the complete dialogue context into a trained intent recognition model to determine the core interaction intent of the first user, and input the real-time input information into a text vector representation model for semantic encoding to generate a query feature vector.

[0013] The intent determination module is configured to, when the interaction intent is determined to be a couple's wish-related intent, obtain the identity information of a second user who is a couple associated with the first user, the second user having a couple relationship tag associated with the first user; based on the identity information of the second user, search for the user profile data of the second user from the vector database, the user profile data containing the user's wish list feature vector set; the wish list feature vector set is configured to be limited to query matching only with the user to which it belongs and users with a couple relationship tag associated with the user to which it belongs.

[0014] The matching feedback module is configured to perform similarity matching calculations between the query feature vector and each wish feature vector in the wish list feature vector set, obtain the appropriate wish feature vector and convert it into recommended wish information, and generate corresponding reply content based on the recommended wish information to feed back to the first user.

[0015] The present invention also discloses a doll device, including an audio acquisition device, a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the information recommendation method based on emotional semantics as described above.

[0016] The present invention also discloses a computer-readable storage medium that, when an executable computer program in the storage medium is executed by a processor, enables the implementation of the information recommendation method based on sentiment semantics as described in the preceding claim.

[0017] This invention discloses an information recommendation method, system, and doll device based on emotional semantics. It receives real-time input from a first user, integrates this input with historical dialogue content of the current session to construct a complete dialogue context, and inputs this complete dialogue context into a trained intent recognition model to determine the first user's core interaction intent. The real-time input is then input into a text vector representation model for semantic encoding to generate a query feature vector. When the interaction intent is determined to be a couple's wish-related intent, the system obtains the identity information of a second user associated with the first user as a couple, the second user having a couple relationship tag associated with the first user. Based on the second user's identity information, the system searches a vector database for the second user's user profile data, which includes the user's wish list feature vector set. This wish list feature vector set is configured to be limited to matching with the user and users with a couple relationship tag associated with the user. The system performs similarity matching calculations between the query feature vector and each wish feature vector in the wish list feature vector set, obtains suitable wish feature vectors, and converts them into recommended wish information. Based on the recommended wish information, it generates a response content corresponding to the real-time input and feeds it back to the first user. This technology solves the problems of existing technologies that require explicit user actions to trigger recommendations and lack the ability to match the wishes of both partners. It enables automatic identification of the wishes of couples through natural interaction, reducing user operation costs, improving the naturalness of interaction, and enhancing the emotional and practical value of information recommendation services.

[0018] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0019] The accompanying drawings, which are included to provide a further understanding of the invention and form part of this application, illustrate exemplary embodiments of the invention and, together with their description, serve to explain the invention and do not constitute an undue limitation thereof. In the drawings: Figure 1 This is a schematic diagram illustrating the specific process of an information recommendation method based on sentiment semantics disclosed in an embodiment of the present invention.

[0020] Figure 2 This is a schematic diagram of the structure of an information recommendation system based on sentiment semantics disclosed in an embodiment of the present invention. Detailed Implementation

[0021] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the described embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0022] Unless otherwise defined, the technical or scientific terms used herein shall have the ordinary meaning as understood by one of ordinary skill in the art to which this invention pertains. The terms “first,” “second,” and similar terms used in the specification and claims of this patent application do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Similarly, the terms “an” or “a” and similar terms do not indicate a limitation of quantity, but rather indicate the presence of at least one.

[0023] In this embodiment, as shown in the appendix Figure 1 As shown, an information recommendation method based on sentiment semantics is disclosed, including the following steps.

[0024] Step S1: Receive real-time input information from the first user, and fuse the real-time input information with the historical dialogue content of the current session to construct a complete dialogue context; input the complete dialogue context into the trained intent recognition model to determine the core interaction intent of the first user; and input the real-time input information into the text vector representation model for semantic encoding to generate a query feature vector.

[0025] The system obtains real-time input information from the first user through a voice interaction interface. In the conversation instance created after the user activates the corresponding interactive function, the real-time input information is concatenated with the historical dialogue content of the current conversation. The historical dialogue content includes interaction summaries from past conversations, various requests mentioned by the user, corresponding responses, and interaction scenario-related information. The system obtains the core information elements involved in the current conversation, including key requests, information expression focus, and scenario-related tendencies in the real-time input information, as well as key content from past interactions that has semantic, scenario, or preference-related relationships with the current input information. The system constructs a complete dialogue context corresponding to the current conversation. The complete dialogue context includes the information currently input by the user, key related content formed during past interactions, historical interaction fragments that have potential relationships with the current conversation scenario, and integrates preference-related clues related to the user and relevant social objects.

[0026] The generated complete dialogue context is fed into the trained intent recognition model to identify and determine the core interaction intent of the first user, extract the textual semantic content of the real-time input information, and input this textual semantic content into the text vector representation model. The real-time input information is then semantically encoded and transformed to generate a query feature vector that matches the real-time input information. The text vector representation model decomposes the dialogue context into basic semantic units through word segmentation and encodes and transforms each semantic unit to obtain a high-dimensional semantic feature vector.

[0027] In one embodiment, the step of decomposing the dialogue context into basic semantic units through word segmentation and encoding each semantic unit to obtain a high-dimensional semantic feature vector specifically includes the following steps: Step S101: Integrate the real-time input information, historical dialogue interaction summary, user preference association clues and social object related association information contained in the complete dialogue context, and perform standardized preprocessing on the integrated text data. The standardized preprocessing includes removing redundant interference information and unifying text format and encoding specifications.

[0028] Step S102: The standardized preprocessed dialogue context text is decomposed into a structured form using a word segmentation algorithm. The continuous text is split into basic semantic units consisting of words and phrases according to semantic logic. Units with semantic representation value are selected and retained, while function words and stop words that have no actual semantic contribution are removed.

[0029] Step S103: Input the selected basic semantic units into a preset text vector representation model. The text vector representation model is used to numerically transform and extract features from the semantic connotation of each basic semantic unit. Combine the semantic weights and relationships of each basic semantic unit in the dialogue context to generate a high-dimensional semantic feature vector that can represent the semantic features of the complete dialogue context.

[0030] The intent recognition model incorporates a hierarchical intent classification system, which includes intent categories and corresponding subcategories such as query / consultation, operation execution, demand feedback, and information confirmation. The query / consultation category includes subcategories such as scenario-related needs (e.g., searching for restaurants, entertainment, travel plans); socially relevant needs (e.g., inquiries about couples' wishes, friend interaction compatibility, and social object preferences); and general information needs (e.g., news and general knowledge queries, which are non-scenario-based and non-social information retrieval requests). The operation execution category includes subcategories such as message forwarding (e.g., explicit message forwarding, implicit message forwarding candidates, etc., cross-user information transmission operations); social interaction triggering (e.g., meeting social objects, collaborating on activities, fulfilling the wishes of associated objects); and configuration adjustment (e.g., modifying account permission settings, updating preference information, editing wish lists, etc., system-related operations). The feedback category includes a subcategory of recommended content feedback, such as accepting recommendations, rejecting recommendations, and modifying recommended content, reflecting feedback requests for scene services or social reminders; a subcategory of operation result feedback, such as confirmation of forwarding results and feedback on interaction effects, reflecting the status after an operation is executed; and a subcategory of function optimization feedback, such as feedback on the accuracy of intent recognition and the suitability of recommendations. The information confirmation category includes a subcategory of social relationship confirmation, such as verifying the validity of a romantic relationship and confirming the status of friend associations; a subcategory of time node confirmation, such as confirming holiday suitability, confirming anniversary associations, and confirming the matching of wish completion times; a subcategory of permission scope confirmation, such as verifying data access permissions and confirming the scope of information sharing; and a subcategory of content accuracy confirmation, such as verifying the content of forwarded messages and validating wish list information. The intent recognition model performs correlation analysis on the complete dialogue context of the input, matches intent category features from a feature dimension, and determines the degree of fit with each intent category center through distance calculation. It also excludes abnormal situations such as semantic conflicts and meaningless input, locating the core interaction request of the first user in the dialogue scenario and outputting an interaction intent result that perfectly matches the classification system.

[0031] Step S2: When the interaction intent is determined to be a couple's wish association intent, the identity information of the second user who is associated with the first user as a couple is obtained. The second user has a couple relationship tag associated with the first user. Based on the identity information of the second user, the user profile data of the second user is searched from the vector database. The user profile data contains the user's wish list feature vector set. The wish list feature vector set is configured to be limited to query matching only with the user to which it belongs and users with a couple relationship tag associated with the user to which it belongs.

[0032] When the core interaction intent is determined to be a couple's wish association intent, the identity information of the second user who has established a couple relationship with the first user is retrieved. The second user's information dimension is pre-configured with a couple relationship tag that is uniquely bound to the first user. Based on the second user's identity information, the vector database is searched and matched with its corresponding user profile data. The user profile data contains the user's wish list feature vector set. The exclusive query and matching permissions for the wish list feature vector set are configured and controlled, and the scope of opening this permission is defined. The query and matching permissions are only open to the user to which the wish list feature vector set belongs, and the associated users who have established a couple relationship with the user to which the set belongs and are bound with the corresponding couple relationship tag in the information dimension.

[0033] Specifically, when the interaction intent is determined and the result is a couple's wish-related intent, the first user account configuration information is retrieved. The account configuration information includes structured data actively submitted and confirmed by the user during the registration phase or system initialization process. Specifically, it includes a couple relationship tag configured by the user, which serves as a credential identifying the couple's relationship and has unique relevance and identifiability; it also includes a long-term request record pre-entered by the user and organized according to the request type, priority, and other dimensions, i.e., a wish list. The wish list contains various matters that the user clearly expects to achieve and related conditions; at the same time, it includes the information authorization scope set by the user according to their own information security needs, i.e., data access permission configuration, which clarifies the accessible subjects, access scenarios, and usage restrictions of various account information.

[0034] Using the aforementioned couple relationship tag as the search keyword, a standardized data interface is used to access a vector database storing long-term stable user information. This vector database employs a structured storage architecture to classify, archive, and map various types of long-term information submitted by users. Through preset tag matching rules, all user relationship tags stored in the vector database are traversed, compared, and their relevance verified. A second user is selected who has a unique and valid association with the first user's couple relationship tag. The wish list information, pre-entered and structured by the second user during registration or initialization, is extracted. This wish list information includes details of the second user's various long-term requests, associated scenarios, priority indicators, and related supplementary explanations.

[0035] The tag matching rules include unique binding verification rules, association status validity judgment rules, and tag format consistency matching rules. Specifically, the unique binding verification rule stipulates that the couple relationship tag is a two-way unique binding identifier, with each couple-related user corresponding to a unique couple relationship tag. The couple relationship tag forms a one-to-one mapping with the user's unique identity identifier, prohibiting multiple users from binding the same tag or one user from binding multiple tags. The association status validity judgment rule verifies the association status identifier corresponding to the couple relationship tag, retaining only valid status tags that have been bound and have not triggered the unbinding process, and removing invalid status tags such as pending binding, unbound, or bound invalid. The tag format consistency matching rule verifies the consistency of character encoding, identifier composition, and format specifications between the search keywords and the couple relationship tags stored in the database, uniformly using a preset standardized format for matching and comparison.

[0036] In one embodiment, the process of traversing, comparing, and verifying the relevance of all user relationship tags stored in the vector database using preset tag matching rules specifically includes the following steps: Step S201: Standardize and preprocess the search keywords and database tag data. Use the first user's couple relationship tag as the search keyword, and complete the standardization process of character encoding and sequence arrangement according to the preset format specifications. At the same time, retrieve the full amount of user relationship tag data from the user relationship tag storage partition of the vector database through the standardized data interface, and perform structured classification indexing on the retrieved tag data according to the user's unique identity.

[0037] Step S202: Based on the preset tag matching rules, perform format consistency and precise character sequence matching on the standardized search keywords and the full set of relation tags in the vector database, filter out the tag candidate set that meets the character matching requirements, perform unique binding verification and association status validity determination on all tags in the tag candidate set, and obtain the tag matching result that conforms to the tag matching rules.

[0038] Step S203: Perform a uniqueness check on the obtained tag matching results to confirm that there is only one valid couple relationship tag that matches the search keyword. Based on the preset association mapping relationship of the couple relationship tag in the vector database, extract the unique identity of the second user bound to the couple relationship tag and the corresponding basic user association information, and filter the second user who has a unique binding with the first user and whose association status is valid.

[0039] A query request is sent to the vector database based on the identity information of the second user who is associated with the first user in a romantic relationship. The query request is used to obtain the second user's wish list data.

[0040] After identifying the second user who is bound to the first user and whose association is valid, the unique identifier information corresponding to the second user is extracted. This identifier information is index data used in the vector database to distinguish user entities. Based on this identifier information, a standardized query request is constructed. This query request carries the second user's unique identifier information and the target retrieval field of the wish list data. The query request is sent to the vector database through a preset database access interface. The query is performed in the structured storage partition of the vector database associated with the second user's identifier to obtain the wish list data that the second user has pre-entered and categorized.

[0041] In one embodiment, the step of sending the query request to the vector database through a preset database access interface, querying the structured storage partition associated with the second user's identity in the vector database, and obtaining the wish list data pre-entered and categorized by the second user specifically includes the following steps: Step S301: Construct a standardized query request based on the second user's unique identification information and complete the pre-permission verification. The pre-permission verification includes encapsulating the second user's unique identification information, the target retrieval field set of the wish list data, and the first user's permission verification identifier in the query request according to the access protocol specifications preset by the vector database. At the same time, it retrieves the data access permission configuration in the first user's account configuration information and the information sharing authorization rules of the second user.

[0042] The information sharing authorization rule is a permission configuration set by the second user for long-term structured information in their own account, and for associated users who are bound to them by a romantic relationship tag. It is used to regulate the scope of data access, access scenarios, and the validity of access permissions.

[0043] Step S302: Send a query request and locate the target structured storage partition through a preset database access interface. Transmit the query request with completed permission verification to the vector database. After receiving the query request, the vector database matches the second user's unique identification information encapsulated in the query request and locates the target storage partition. The target storage partition is a structured storage partition bound to the second user's unique identification information.

[0044] Step S303: Perform data retrieval and structured extraction within the target storage partition. In the located second user structured storage partition, retrieve the wish list data based on the target retrieval field in the query request, match the wish list structured storage entries associated with the second user's identity information, and obtain the complete wish list data.

[0045] In another embodiment, when the core interaction intent is determined to be a friend social intent, the friend user group of the first user is obtained, and the friend users in the friend user group have friend relationship tags associated with the first user; user profile data corresponding to the first user and each friend user are searched from the vector database, and the user profile data includes the user's interest vector and behavioral feature vector.

[0046] The user profile data is a high-dimensional semantic feature dataset corresponding to the user, stored in a vector database. The high-dimensional semantic feature dataset includes interest vectors and behavioral feature vectors.

[0047] The interest vector is generated based on the user's pre-input information through a text vector representation model and is configured to be used for query matching by the current user and users with preset tags, while rejecting queries from other users. The preset tags include friend relationship tags.

[0048] The behavioral feature vector is generated based on the historical dynamic behavioral features extracted by the user during the dialogue and interaction with the terminal, and is configured to be available only for the user to query and match and to reject queries from other users.

[0049] The user profile data also includes the user's preference vector, which is generated based on personal preference information extracted during the user's interaction with the terminal and is configured to be available only for the user to query and match and to reject queries from other users.

[0050] The query feature vector is matched with the first user's behavior feature vector and the interest vectors of each friend user. Based on the matching results, friend users with high social relevance to the current friend's social intention are selected as target social objects. Friend inquiry information is generated based on the real-time interaction information. After receiving confirmation feedback from the first user on the target social object and the friend inquiry information, the friend inquiry information is sent to the device where the target social object is located.

[0051] The query feature vector is matched with the first user's behavior feature vector, interest vector, and preference vector, and then with the interest vectors of each friend user to calculate similarity. Based on the matching results, a preset number of friend users with the highest similarity are selected as candidate social objects. Candidate friend inquiry information is generated based on the real-time interaction information. The candidate social objects and candidate friend inquiry information are output to the first user for confirmation. After receiving the confirmation feedback, the confirmed friend inquiry information is sent to the device where the target social object is located.

[0052] The real-time interactive information includes user query information and activity invitation intention. Based on the activity invitation intention, a friend inquiry message is generated. After receiving the target social object confirmed by the first user based on each candidate social object, the friend inquiry message is sent to the device where the target social object is located.

[0053] The behavioral feature vector is configured to extract dialogue information and interaction information related to friends from the historical interaction content between the user and the text big model based on preset rules, and then transform it into a dynamic behavioral feature vector through the text vector representation model.

[0054] The real-time interactive information also includes scenario requirement information; obtain structured service information matching the scenario requirements, the structured service information including candidate activity venues, activity times or candidate venue details; After receiving confirmation feedback from the first user regarding the target social object and structured service information, a friend request message is generated based on the user's willingness to participate in the activity and the structured service information, and then sent to the device where the target social object is located.

[0055] The query feature vector is compared with the first user's behavior feature vector, interest vector, and preference vector, and then with the interest vectors of each friend user for similarity calculation. A friend compatibility scoring model is constructed by combining the first user's historical shared activity records and interaction frequency data with each friend, and the compatibility score of each friend user is output. It is then determined whether the match meets the preset friend matching recommendation rules, and a set number of friend users with the highest compatibility scores are selected as candidate social objects from those who meet the preset rules. The information of each candidate social object is fed back to the first user, and the target social object is determined based on the first user's selection.

[0056] In this embodiment, the wish list feature vector set is configured to be generated after semantic encoding based on the wish information pre-input by the second user through a text vector representation model.

[0057] Based on the wish information pre-entered by the second user, the wish information is input into a text vector representation model, and semantic encoding conversion processing is performed on the wish information. Based on the result of the encoding conversion processing, corresponding feature vectors are generated and integrated to form a feature vector set.

[0058] In this embodiment, the wish list feature vector set is further configured to generate new wish feature vectors by semantically encoding the interaction content containing wishes collected by the second user during the dialogue interaction with the terminal through a text vector representation model, and then supplementing the user's original wish list feature vector set.

[0059] The wish list feature vector set is also configured to collect interactive content containing wish-related expressions generated during the continuous natural dialogue interaction between the second user and the terminal, extract the text semantic information carried in the interactive content, input the obtained text semantic information into the text vector representation model, perform semantic encoding conversion processing on the text semantic information of the interactive content, generate a new wish feature vector based on the result of the encoding conversion processing, and integrate the new wish feature vector into the original wish list feature vector set of the second user.

[0060] In this embodiment, the user profile data also includes the user's interest vector, preference vector, and behavioral feature vector; The interest vector is generated based on the user's pre-input information through a text vector representation model and is configured to be used for query matching by the current user and users with corresponding romantic relationship tags, while rejecting queries from other users.

[0061] Specifically, the user's pre-entered information is input into a text vector representation model and semantically encoded to generate an interest vector. Query matching permissions are configured for the interest vector, allowing the user and users bound to the corresponding couple relationship tag to perform query matching, while restricting query matching operations for all other users.

[0062] The preference vector is generated based on the personal preference information extracted by the user during the dialogue and interaction with the terminal, and is configured to be available only for the user to query and match and to reject queries from other users. A preference vector is generated based on personal preference-related information extracted from the user's dialogue interaction with the terminal. Query matching permissions are configured for the preference vector, allowing query matching operations only to the current user and restricting query matching operations for all other users.

[0063] The behavioral feature vector is generated based on the historical dynamic behavioral features extracted by the user during the dialogue and interaction with the terminal, and is configured to be available only for the user to query and match and to reject queries from other users.

[0064] A feature vector is generated based on historical dynamic behavioral features extracted from the dialogue interaction between the user and the terminal. Query matching permissions are configured for the behavioral feature vector, allowing query matching operations only to the current user and restricting query matching operations for all other users.

[0065] In this embodiment, step S2 further includes: When the couple's wish association intent includes the requirement for related recommendations, the second user's wish list feature vector set is retrieved based on the couple relationship tag. The current time node, the realization conditions of each item in the wish list, and the first user's historical behavior data are combined to determine whether the recommendation triggering conditions are met. If they are met, the corresponding reply content is generated based on the recommended wish and the adapted recommended content and fed back to the first user.

[0066] Specifically, when the couple's wish association intent includes a recommendation request, the second user's wish list feature vector set is retrieved from the vector database based on the couple's relationship tag. The current time node information, the realization condition information corresponding to each wish in the wish list, and the first user's historical behavior data are integrated and extracted. After comprehensive analysis of various types of information, it is determined whether the recommendation trigger condition has been met. If the recommendation trigger condition has been met, the recommended wish is selected from the second user's wish list feature vector set. The recommended wish is matched with the corresponding suitable recommendation content. Based on the recommended wish and the suitable recommendation content, a response content that semantically matches the real-time input information is generated, and the generated response content is pushed back to the first user.

[0067] In this embodiment, the current time node includes statutory holidays, couples' anniversaries, the second user's birthday, and the second user-defined time node in the wish list. The real-time time is obtained through the time calibration module and accurately matched with the time node.

[0068] The current time nodes include national statutory holidays, couple anniversaries set after the first user and the second user bind their relationship, birthday information stored by the second user during the information entry stage, and various time nodes set by the second user for each wish during the wish list construction process. Real-time time information is collected through the time calibration module. After the collected real-time time information is processed in a standardized format, it is accurately matched and verified with the basic information of various preset time nodes according to the dimensions, including consistency verification of time nodes and determination of matching results.

[0069] The built-in time calibration module synchronizes with a standard time source to obtain real-time time. Key time elements such as date, month, year, and special time markers are extracted from the real-time time. The extracted time elements are compared with preset legal holiday time databases, couple anniversary data, second user birthday records, and custom time node information in wish lists. The real-time time is accurately matched with various target time nodes, providing accurate time support for determining subsequent recommendation trigger conditions.

[0070] If the triggering condition is met, the preference vectors of the first user and the second user are fused together, the similarity of their common preferences is calculated, and the structured service information matching the second user's wish type is obtained through the scene service and recommendation module. Combining the structured service information, the common preference similarity, and the dialogue context, recommendation feedback information containing wish-related prompts and scene-based recommendation content is generated.

[0071] Specifically, when the triggering condition is met, the preference vector of the first user and the constructed preference vector of the second user are retrieved. Feature fusion processing is performed on the two types of vectors, integrating the feature representations of the first and second users in dimensions such as long-term interests, dynamic behavioral tendencies, and historical activity preferences. Vector association analysis is used to calculate the similarity of their shared preferences. Based on the type attributes of each wish in the second user's wish list, such as dining experiences, leisure and entertainment, travel planning, and holiday commemorations, a data request is initiated to the scene service and recommendation module. The scene service and recommendation module retrieves and obtains corresponding structured service information according to wish type matching rules. This structured service information includes standardized data such as the service provider, scene adaptation conditions, service content, and associated execution elements. Further integrating the structured service information, the similarity result of shared preferences, and the constructed complete dialogue context of the current session, these three types of information are input into the dynamic prompt word construction module. This constructs structured context prompt information containing wish association logic, shared preference basis, and scene service details. The content generation module then transforms this information into natural language recommendation feedback information, which includes wish association prompts and scene-specific recommendation content.

[0072] Step S3: Perform similarity matching calculations between the query feature vector and each wish feature vector in the wish list feature vector set to obtain the appropriate wish feature vector and convert it into recommended wish information. Based on the recommended wish information, generate the corresponding reply content to the real-time input information and feed it back to the first user.

[0073] The generated query feature vector is matched with each wish feature vector in the second user's wish list feature vector set for similarity calculation. Adaptive wish feature vectors with matching degree meeting the preset threshold are selected from the wish list feature vector set. The adapted wish feature vectors are subjected to semantic reverse parsing processing to convert the vector information into recommended wish information in natural language form. The semantic association between the recommended wish information and the real-time input information of the first user is integrated to generate response content that matches the real-time input information. The generated response content is then fed back to the first user through the interactive interface.

[0074] Specifically, the similarity matching calculation is performed between the query feature vector and each wish feature vector in the wish list feature vector set, which includes the following steps: Step S401: Retrieve the generated first user query feature vector, and simultaneously retrieve the full set of wish list feature vectors from the second user's identity identifier. Perform data format compliance verification and vector dimension consistency calibration on the retrieved two types of feature vectors. Classify, collect and structure the qualified vector data.

[0075] Step S402: The query feature vector is used as the reference vector, and it is precisely matched with each wish feature vector in the aggregated wish list feature vector set. Numerical similarity calculation is performed on each pair of matched feature vectors.

[0076] Step S403: For each wish feature vector and query feature vector, the similarity calculation result is independently encoded and accurately recorded in combination with the unique identifier of the corresponding wish feature vector. The similarity matching calculation results of all wish feature vectors and query feature vectors are sorted, classified and orderly integrated and summarized according to the similarity value. The summarized result data is persistently stored according to the data storage specification, and the result data is classified, archived and indexed.

[0077] In this embodiment, the query feature vector is matched with the first user's behavior feature vector, preference vector, and interest vector, as well as the second user's interest vector and couple's wish vector for similarity calculation. Based on the matching results, the most similar wish feature vectors are selected and converted into recommended wish information. Based on the recommended wish information, the corresponding reply content is generated and fed back to the first user.

[0078] The generated query feature vector is retrieved and its similarity is calculated by multi-dimensional matching with the first user's behavior feature vector, preference vector, and interest vector. At the same time, the query feature vector is also calculated by similarity matching with the second user's interest vector and couple's wish vector. The combined matching results are sorted and filtered, and a preset number of wish feature vectors with the highest similarity are selected. The selected wish feature vectors are semantically parsed and converted into recommended wish information in natural language form. Based on the recommended wish information, a response content that semantically matches the first user's real-time input information is generated and output to the first user.

[0079] In one embodiment, it is determined whether the couple's wish recommendation trigger condition is met. The trigger condition is configured to be at least one of the following: the current time node matches the preset completion time in the second user's wish list; the semantic feature vector of the first user's real-time input information and the semantic feature vector of the second user's wish have a similarity of up to a preset threshold; or there are activity records related to the second user's wish in the first user's historical behavior data.

[0080] If the triggering condition is met, the similarity matching calculation is performed between the query feature vector and each wish feature vector in the wish list feature vector set to obtain the appropriate wish feature vector and convert it into recommended wish information. Based on the recommended wish information, the corresponding reply content is generated and fed back to the first user.

[0081] Specifically, the triggering conditions for recommending couples' wishes are comprehensively determined. These triggering conditions include three judgment dimensions, and meeting any one of them is sufficient for the recommendation to take effect. First, the completion times of each preset wish in the second user's wish list are retrieved from the vector database, including specific dates, time periods, or time node-related constraints. These are compared with the currently retrieved time node to determine whether the current time node is within the preset completion time range or meets the time association requirements. Second, semantic feature vectors of the first user's real-time input information and the second user's wishes are constructed using text vector generation. The semantic feature vector of the real-time input information originates from the semantic encoding of the current dialogue context, while the semantic feature vector of the wishes is generated based on the structured description encoding of each request in the wish list. The similarity between the two types of vectors is calculated, and it is determined whether the similarity reaches a preset threshold standard. Third, the historical behavior data of the first user stored in the user preference management module is retrieved, including past activity participation records, social interaction-related behaviors, request expression and execution records, etc. Activity records related to the request type, scene attributes, and key elements involved in the second user's wishes are filtered to determine whether there are any valid historical behavioral traces. The results from various dimensions are used to determine whether the conditions for triggering the couple's wish recommendation are met.

[0082] In this embodiment, as shown in the appendix Figure 2As shown, an information recommendation system based on emotional semantics is also disclosed, including: an information receiving module 1, an intent determination module 2, and a matching feedback module 3; wherein, the information receiving module 1 is configured to receive real-time input information from a first user, fuse the real-time input information with the historical dialogue content of the current session to construct a complete dialogue context; input the complete dialogue context into a trained intent recognition model to determine the core interaction intent of the first user; and input the real-time input information into a text vector representation model for semantic encoding to generate a query feature vector; the intent determination module 2 is configured to, when the interaction intent is determined to be a couple's wish-related intent, obtain the second user's intent that is associated with the first user as a couple. The system includes identity information, where the second user has a romantic relationship tag associated with the first user; based on the second user's identity information, it searches for the second user's user profile data from a vector database, the user profile data containing the user's wish list feature vector set; the wish list feature vector set is configured to be limited to query matching only with the user to which it belongs and users with a romantic relationship tag associated with the user; the matching feedback module 3 is configured to perform similarity matching calculations between the query feature vector and each wish feature vector in the wish list feature vector set, obtain the appropriate wish feature vector and convert it into recommended wish information, and generate corresponding response content based on the recommended wish information to feed back to the first user.

[0083] In this embodiment, a doll device is also disclosed, including an audio acquisition device, a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the steps of the information recommendation method based on emotional semantics as described above.

[0084] The doll device receives user input information through a built-in interactive interface, completes data processing and logical judgment based on various functional modules, and outputs corresponding response feedback to the user through a communication interface.

[0085] It should be noted that the various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. Regarding the sentiment-based information recommendation system disclosed in the embodiments, since it corresponds to the sentiment-based information recommendation method disclosed in the embodiments, its description can be simplified, and its specific implementation can be referred to the foregoing method embodiments.

[0086] In this embodiment, a computer-readable storage medium is also disclosed, which, when executed by a processor, enables the implementation of the steps of the information recommendation method based on sentiment semantics as described in the preceding embodiment.

[0087] If the sentiment-based semantic information recommendation method is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, all or part of the processes in the above embodiments of the present invention can also be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can include any entity or device capable of carrying the computer program code, a recording medium, a USB flash drive, a portable hard drive, a magnetic disk, an optical disk, a computer memory, etc.

[0088] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

[0089] In summary, the above description is only a preferred embodiment of the present invention. All equivalent changes and modifications made within the scope of the claims of the present invention should be covered by the present invention.

Claims

1. An information recommendation method based on sentiment semantics, characterized in that, Includes the following steps: S1, Receive real-time input information from the first user, and merge the real-time input information with the historical dialogue content of the current session to construct a complete dialogue context; The complete dialogue context is fed into the trained intent recognition model to determine the core interaction intent of the first user; The real-time input information is then input into a text vector representation model for semantic encoding to generate a query feature vector. S2, when the interaction intent is determined to be a couple's wish association intent, obtain the identity information of the second user who is a couple associated with the first user, and the second user has a couple relationship tag associated with the first user; Based on the identity information of the second user, the user profile data of the second user is retrieved from the vector database. The user profile data contains the user's wish list feature vector set. The wish list feature vector set is configured to be limited to user queries and matching only with the user and users with a romantic relationship tag associated with the user. S3, perform similarity matching calculations between the query feature vector and each wish feature vector in the wish list feature vector set, obtain the appropriate wish feature vector and convert it into recommended wish information, and generate corresponding reply content based on the recommended wish information and feed it back to the first user.

2. The information recommendation method based on sentiment semantics according to claim 1, characterized in that: The wish list feature vector set is configured to be generated based on the wish information pre-input by the second user, after semantic encoding through a text vector representation model.

3. The information recommendation method based on sentiment semantics according to claim 2, characterized in that: The wish list feature vector set is further configured to generate new wish feature vectors by semantically encoding the interaction content containing wishes collected by the second user during the dialogue interaction with the terminal through a text vector representation model, and then supplementing the user's original wish list feature vector set.

4. The information recommendation method based on sentiment semantics according to claim 3, characterized in that: The user profile data also includes the user's interest vector, preference vector, and behavioral feature vector; The interest vector is generated based on the user's pre-input information through a text vector representation model after semantic encoding, and is configured to be used for query matching by the current user and users with corresponding couple relationship tags, while rejecting queries from other users; The preference vector is generated based on the personal preference information extracted by the user during the dialogue and interaction with the terminal, and is configured to be available only for the user to query and match and to reject queries from other users. The behavioral feature vector is generated based on the historical dynamic behavioral features extracted by the user during the dialogue and interaction with the terminal, and is configured to be available only for the user to query and match and to reject queries from other users.

5. The information recommendation method based on sentiment semantics according to claim 4, characterized in that: The query feature vector is matched with the behavioral feature vector, preference vector, and interest vector of the first user, as well as the interest vector and the couple's wish vector of the second user for similarity. Based on the matching results, the most similar wish feature vectors of a preset number are obtained and converted into recommended wish information. Based on the recommended wish information, the corresponding reply content is generated and fed back to the first user.

6. The information recommendation method based on sentiment semantics according to claim 5, characterized in that, Step S2 further includes: When the couple's wish association intent includes the requirement for related recommendations, the second user's wish list feature vector set is retrieved based on the couple relationship tag. The current time node, the realization conditions of each item in the wish list, and the first user's historical behavior data are combined to determine whether the recommendation triggering conditions are met. If they are met, the corresponding reply content is generated based on the recommended wish and the adapted recommended content and fed back to the first user.

7. The information recommendation method based on sentiment semantics according to claim 6, characterized in that: The current time node includes statutory holidays, couples' anniversaries, the second user's birthday, and the second user's custom time node in the wish list. The real-time time is obtained through the time calibration module and accurately matched with the time node.

8. An information recommendation system based on sentiment semantics, characterized in that, include: The information receiving module is configured to receive real-time input information from the first user and merge the real-time input information with the historical dialogue content of the current session to construct a complete dialogue context. The complete dialogue context is fed into the trained intent recognition model to determine the core interaction intent of the first user; The real-time input information is then input into a text vector representation model for semantic encoding to generate a query feature vector. The intent determination module is configured to, when the interaction intent is determined to be a couple's wish-related intent, obtain the identity information of a second user who is a couple associated with the first user, and the second user has a couple relationship tag associated with the first user; Based on the identity information of the second user, the user profile data of the second user is retrieved from the vector database. The user profile data contains the user's wish list feature vector set. The wish list feature vector set is configured to be limited to user queries and matching only with the user and users with a romantic relationship tag associated with the user. The matching feedback module is configured to perform similarity matching calculations between the query feature vector and each wish feature vector in the wish list feature vector set, obtain the appropriate wish feature vector and convert it into recommended wish information, and generate corresponding reply content based on the recommended wish information to feed back to the first user.

9. A doll device, comprising an audio acquisition device, a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the computer program, it implements the steps of the method as described in any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, When the executable computer program in the storage medium is executed by a processor, it can implement the method as described in any one of claims 1-7.

Images