Answer recommendation method and device

Abstract

According to the answer recommendation method and device provided by the invention, the questions of the user are automatically classified, the similar questions are summarized into the same question classification, the corresponding answer recommendation set is generated, and the target user can select the answer recommendation in the answer recommendation set to directly reply, so that the user experience is improved. Therefore, fragmented consultation which needs to be processed one by one originally is converted into an efficient batch decision and response process, the interaction efficiency is remarkably improved, and the use experience of the user and the target user is effectively improved.

Description

Technical Field

[0001] This invention relates to the field of computer technology, and in particular to a method and apparatus for providing answer recommendations. Background Technology

[0002] Currently, interactive platform dialogue systems face the following pain points when providing services to users and target users: target users often face a massive volume of conversations, making it difficult to respond promptly and effectively. Furthermore, the sheer volume of conversations reduces the target users' willingness to respond, which in turn leads to a decrease in user engagement and usage frequency when they frequently do not receive replies. Summary of the Invention

[0003] In view of this, the present invention provides a method and apparatus for answering recommendations, in order to solve the problem of low interaction efficiency in the prior art.

[0004] The first aspect of this invention provides a method for answering recommendations, comprising:

[0005] Receive user problem information;

[0006] Cluster analysis is performed based on the question information and the existing tagged questions in the knowledge base to obtain a list of sentence semantic vectors;

[0007] Based on the statement semantic vector list and the association model, a semantic association information list is generated; wherein, the semantic association information list includes at least one semantic association information, and each semantic association information includes two statement semantic vectors and the degree of association between the two statement semantic vectors;

[0008] Based on the semantic association information list, an undirected association graph is generated; wherein, the nodes in the undirected association graph are generated from the semantic vectors of the statements in the semantic association information, and the edges in the undirected association graph are established based on the degree of association between two semantic vectors of statements;

[0009] The nodes in the associated undirected graph are classified and analyzed to generate classification analysis results; wherein, the classification analysis results include at least one problem classification label for a node partition;

[0010] For each of the aforementioned question categorization tags, a set of recommended answers corresponding to the question categorization tag is generated based on the question associated with the question categorization tag and the historical dialogue records corresponding to the question categorization tag in the knowledge base.

[0011] Optionally, generating a semantic association information list based on the statement semantic vector list and association model includes:

[0012] Input any two statement semantic vectors from the statement semantic vector list into the correlation analysis model to generate the correlation between the two statement semantic vectors; wherein, the correlation analysis model includes a linear transformation layer and a hybrid attention layer, and the hybrid attention layer includes a soft attention mechanism and a self-attention mechanism;

[0013] Semantic association information is generated based on the semantic vectors of the two statements and the degree of correlation between them.

[0014] Optionally, the nodes in the associated undirected graph include template nodes and input nodes, and the step of performing classification analysis on the nodes in the associated undirected graph to generate classification analysis results includes:

[0015] For each template node, starting from the template node, the first target input node belonging to the node partition corresponding to the template node is determined layer by layer using the dynamic correlation threshold and the correlation between nodes, thus obtaining the first classification analysis result; wherein, the dynamic correlation threshold at each level is determined based on the initial correlation threshold and the control factor; the control factor is determined based on the number of nodes in the current different node partitions and the number of nodes that should be included;

[0016] Optionally, the answer recommendation method further includes:

[0017] If the correlation degree between an input node and all the template nodes is less than the minimum correlation degree threshold, then the input node is taken as the target node.

[0018] For each target node, starting from the target node, the second target input node corresponding to the node partition is determined layer by layer using the dynamic correlation threshold and the correlation between nodes, until no input node has a correlation with all template nodes that is less than the minimum correlation threshold, and the second classification analysis result is obtained.

[0019] Based on the first classification analysis result and the second classification analysis result, the final classification analysis result is generated.

[0020] Optionally, after receiving the user's question information, the process further includes:

[0021] The problem information is stored in the problem pool as a task to be processed, and feedback information is generated and sent to the user.

[0022] Optionally, the answer recommendation method further includes:

[0023] When the online event of a target user is detected, the size of the first dynamic virtual congestion window is determined based on the real-time load information of the current pipeline.

[0024] Based on the size of the first dynamic virtual congestion window, after obtaining the question information corresponding to the target user from the question pool, the step of performing cluster analysis based on the question information and the existing tagged questions in the knowledge base to obtain a list of statement semantic vectors is started.

[0025] Optionally, after generating a set of recommended answers for each question categorization tag based on the question associated with the question categorization tag and the historical dialogue records corresponding to the question categorization tag in the knowledge base, the method further includes:

[0026] Receive the target answer information of the target user; wherein, the target answer information is a recommended answer selected by the target user from the answer recommendation set corresponding to the question category tag, or an answer modified by the target user for the recommended answer in the answer recommendation set corresponding to the question category tag;

[0027] The size of the second dynamic virtual congestion window is determined based on the real-time load information of the current pipeline.

[0028] Based on the size of the second dynamic virtual congestion window, the target user's target answer information is fed back to the corresponding user of the corresponding question in the question categorization label.

[0029] Optionally, the answer recommendation method further includes:

[0030] Reward parameters are generated based on the evaluation information of the problem classification tags;

[0031] A state transition process is constructed based on the semantic association information and the reward parameters;

[0032] The correlation analysis model is further optimized using the state transition process to obtain an optimized correlation analysis model.

[0033] Optionally, the answer recommendation method further includes:

[0034] After receiving the question and answer information, the content of the question and answer information is reviewed to obtain the review result; wherein, the question and answer information includes the user's question information and the target answer information;

[0035] Questions and answers deemed risky by the review results will be blocked.

[0036] Optionally, after receiving the user's question information, the process further includes:

[0037] If the language used in the user's question is different from the target user's usual language, the user's question will be translated into the target user's usual language.

[0038] A second aspect of the present invention provides a response recommendation device, comprising:

[0039] The first receiving unit is used to receive the user's problem information;

[0040] The clustering analysis unit is used to perform clustering analysis based on the question information and the existing tagged questions in the knowledge base to obtain a list of sentence semantic vectors;

[0041] The association information generation unit is used to generate a semantic association information list based on the statement semantic vector list and the association model; wherein, the semantic association information list includes at least one semantic association information, and each semantic association information includes two statement semantic vectors and the association degree between the two statement semantic vectors;

[0042] An undirected graph generation unit is used to generate an undirected graph based on the semantic association information list; wherein, the nodes in the undirected graph are generated from the semantic vectors of the statements in the semantic association information, and the edges in the undirected graph are established based on the degree of association between two semantic vectors of statements.

[0043] The classification analysis unit is used to perform classification analysis on the nodes in the associated undirected graph and generate classification analysis results; wherein, the classification analysis results include at least one problem classification label for a node partition;

[0044] The answer recommendation generation unit is used to generate an answer recommendation set corresponding to each question category tag, based on the question associated with the question category tag and the historical dialogue records corresponding to the question category tag in the knowledge base.

[0045] Optionally, the association information generation unit includes:

[0046] The correlation degree generation unit is used to input any two statement semantic vectors from the statement semantic vector list into the correlation degree analysis model to generate the correlation degree between the two statement semantic vectors; wherein, the correlation degree analysis model includes a linear transformation layer and a hybrid attention layer, and the hybrid attention layer includes a soft attention mechanism and a self-attention mechanism;

[0047] The association information generation subunit is used to generate semantic association information based on the semantic vectors of the two statements and the degree of association between the two semantic vectors.

[0048] Optionally, the classification analysis unit includes:

[0049] The first classification analysis subunit is used to determine the first target input node belonging to the node partition corresponding to the template node, starting from the template node, by using the dynamic correlation threshold and the correlation between nodes, and obtaining the first classification analysis result; wherein, the dynamic correlation threshold at each level is determined based on the initial correlation threshold and the control factor; the control factor is determined based on the number of nodes in the current different node partitions and the number of nodes that should be included;

[0050] Optionally, the answer recommendation device further includes:

[0051] The target node determination unit is used to determine the input node as the target node if the correlation degree between the current input node and all the template nodes is less than the minimum correlation degree threshold.

[0052] The second classification analysis subunit is used to determine the second target input node corresponding to the node partition of the target node by using the target node as the starting point and the dynamic correlation threshold and the correlation between nodes, until there is no input node whose correlation with all template nodes is less than the minimum correlation threshold, and thus obtain the second classification analysis result.

[0053] The classification analysis result generation unit generates the final classification analysis result based on the first classification analysis result and the second classification analysis result.

[0054] Optionally, the answer recommendation device further includes:

[0055] The storage unit is used to store the problem information as a task to be processed into the problem pool and generate feedback information to provide feedback to the user.

[0056] Optionally, the answer recommendation device further includes:

[0057] The window size determination unit is used to determine the size of the first dynamic virtual congestion window based on the real-time load information of the current pipeline when the online event of the target user is detected.

[0058] The calling unit is used to, based on the size of the first dynamic virtual congestion window, obtain relevant question information corresponding to the target user from the question pool, and then call the clustering analysis unit to start the step of performing clustering analysis based on the question information and existing tagged questions in the knowledge base to obtain a list of statement semantic vectors.

[0059] Optionally, the answer recommendation device further includes:

[0060] The second receiving unit is used to receive the target answer information of the target user; wherein, the target answer information is a recommended answer selected by the target user from the answer recommendation set corresponding to the question classification tag, or an answer modified by the target user for the recommended answer in the answer recommendation set corresponding to the question classification tag;

[0061] The window size determination unit is also used to determine the size of the second dynamic virtual congestion window based on the real-time load information of the current pipeline;

[0062] The feedback unit is used to feed back the target user's target answer information to the corresponding user of the corresponding question in the question classification label, based on the size of the second dynamic virtual congestion window.

[0063] Optionally, the answer recommendation device further includes:

[0064] The reward parameter generation unit is used to generate reward parameters based on the evaluation information of the problem classification tags;

[0065] A state transition process construction unit is used to construct a state transition process based on the semantic association information and the reward parameters;

[0066] An optimization unit is used to further optimize the correlation analysis model using the state transition process to obtain an optimized correlation analysis model.

[0067] Optionally, the answer recommendation device further includes:

[0068] The review unit is used to receive question and answer information, review the content of the question and answer information, and obtain a review result; wherein, the question and answer information includes the user's question information and the target answer information;

[0069] The interception unit is used to intercept question and answer information that is deemed risky by the review results.

[0070] Optionally, the answer recommendation device further includes:

[0071] The translation unit is used to translate the user's question information into the target user's common language type if the language type used in the user's question information is inconsistent with the target user's common language type.

[0072] As can be seen from the above solutions, the present invention provides an answer recommendation method and apparatus. By automatically classifying user questions, similar questions are grouped into the same question category, and a corresponding answer recommendation set is generated. The target user can choose the answer recommendation from the answer recommendation set to reply directly, thereby transforming the fragmented consultation that originally needed to be processed one by one into an efficient batch decision-making and response process, significantly improving interaction efficiency and effectively enhancing the user experience for both the user and the target user. Attached Figure Description

[0073] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0074] Figure 1 A detailed flowchart of a response recommendation method provided in an embodiment of the present invention;

[0075] Figure 2 A diagram illustrating the architecture of a response recommendation system according to another embodiment of the present invention;

[0076] Figure 3 A Q-Network hierarchical structure design diagram is provided for another embodiment of the present invention;

[0077] Figure 4 A structural diagram of an auxiliary intelligent agent system for intelligent induction and response provided in another embodiment of the present invention;

[0078] Figure 5 A structural diagram of a TCP congestion control algorithm implementation scheme provided in another embodiment of the present invention;

[0079] Figure 6 A flowchart of a response recommendation method provided in another embodiment of the present invention;

[0080] Figure 7 This is a schematic diagram of an answer recommendation device provided in another embodiment of the present invention. Detailed Implementation

[0081] 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 embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0082] The term "comprising" and its variations as used herein are open-ended inclusions, meaning "including but not limited to". The term "based on" means "at least partially based on". The term "one embodiment" means "at least one embodiment"; the term "another embodiment" means "at least one additional embodiment"; the term "some embodiments" means "at least some embodiments". Definitions of other terms will be given in the description below.

[0083] It should be noted that the information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, stored data, displayed data, etc.) involved in this invention are all information and data authorized by the user or fully authorized by all parties.

[0084] It should be noted that the concepts of "first" and "second" mentioned in this invention are only used to distinguish different devices, modules or units, and are not used to limit the order of functions performed by these devices, modules or units or their interdependencies.

[0085] It should be noted that the terms "a" and "a plurality of" used in this invention are illustrative rather than restrictive. Those skilled in the art should understand that, unless otherwise expressly indicated in the context, they should be understood as "one or more".

[0086] This invention provides a method for answering recommendations, such as... Figure 1 As shown, the specific steps include:

[0087] S101, Receive user's problem information.

[0088] The user's question information refers to the question raised by the user. For example, the user's question is, "I heard that you, target user A, are going to be the lead actor in XXX. Is that true?"

[0089] In the actual application of this invention, the target users can be internet celebrities, artists, public figures, official accounts, etc., and there is no limitation here.

[0090] It should be noted that a user's question information can be one question, multiple questions, or one question information including multiple sub-questions; there are no restrictions here.

[0091] Furthermore, the user's question information can be text information entered by the user through a keyboard or touch screen, voice information entered by the user through a microphone, or text messages, data links, voice messages, image messages, video messages, etc., entered by the user through an interactive terminal, without any limitation here.

[0092] S102. Based on the question information and the existing tagged questions in the knowledge base, perform cluster analysis to obtain a list of sentence semantic vectors.

[0093] It should be noted that the knowledge base is not simply a list of terms. It includes at least an industry terminology database and a private knowledge base for the target user. Taking artists as an example, it can contain a wealth of structured contextual information such as standardized terms in the entertainment industry, information on the target user's works, commonly used nicknames, recent hot topics, and cultural background annotations. No limit is set here.

[0094] In the practical application of this invention, textual data such as publicly available industry discussions, technical documents, and news information can be actively obtained by calling the API interfaces provided by social media platforms. Subsequently, it uses natural language processing technology to perform in-depth analysis on this data. The core tasks include identifying and extracting frequently used basic terms, emerging technology concepts, and hot industry events, which are not limited here.

[0095] Furthermore, this invention also performs targeted monitoring of target users. It queries and analyzes discussions, fan interactions, and trending topics related to target users from social media platforms. Its advanced capability lies in its ability to intelligently identify specific popular expressions created by fan groups and understand the semantics and applicable scenarios of these expressions. Specific identification methods may include, but are not limited to, obtaining comments from the target user's official account on mainstream social media platforms by the operations team, inputting the comment content + prompt into an intelligent agent, and having the intelligent agent output the expression. For example, if a target user participates in a work and gains a nickname that spreads widely, the intelligent agent will be aware of this and the information will be stored in a knowledge base.

[0096] Of course, in the practical application of this invention, an information enhancement mechanism that can be intervened by operators can also be designed. Operators can proactively push and inject industry dynamics and target user-specific information collected from other channels (such as third-party data services, public networks, or internal reports) into the data processing pipeline through a dedicated operations management backend. After the system deduplication, cleaning, and structuring, this external information can be dynamically integrated as incremental knowledge into the industry terminology database and the target user's private knowledge base.

[0097] Regarding the aforementioned knowledge base, this invention can be implemented based on an information statistics intelligent agent, such as... Figure 2 As shown, the information statistics intelligent agent, acting as the data perception and processing center of the system, can automatically and continuously collect, clean, and analyze multi-source heterogeneous data from external public networks (especially various social media platforms). This multi-source heterogeneous data includes, but is not limited to, API query interfaces provided by external social media platforms, publicly available information, internal company data, and data obtained by operations personnel.

[0098] It should be noted that the system architecture of this invention can use the Model Context Protocol (MCP) as the core interconnection framework. This architecture integrates the Browser-Use open-source toolkit as the browser automation execution engine and deploys a custom MCP server, which is not limited here.

[0099] This invention utilizes a specialized information statistics agent to collect data from various data sources via an MCP server and outputs it to a knowledge base. The content in the database is dynamically linked by a pipeline agent to provide background knowledge. Simultaneously, after a target user answers a question, that type of question is also recorded in the knowledge base for future question answering.

[0100] Understandably, the knowledge base in this invention acts as a continuously updated long-term memory module. It works in conjunction with the agent's short-term working memory (dialogue context) and toolset. When an agent in the pipeline needs to handle professional advice, it proactively retrieves relevant information fragments from the knowledge base and integrates them with the current dialogue context to form a basis for decision-making. For example, it stores the target user's nickname, the name of a character in a work, and its corresponding scene in the knowledge base to help other agents make decisions. This mechanism significantly improves the agent's cognitive ability in the professional field and the reliability of its services.

[0101] In the practical application of this invention, natural language processing technology can be used to automatically identify the semantics of question information and existing tagged questions in the knowledge base, and a pre-trained semantic encoding model can be used to map the input statement / tag into a high-dimensional vector representation. The specific implementation steps are as follows (processes A1 to A3):

[0102] Process A1: The system first executes two data acquisition and preprocessing processes in parallel: real-time acquisition of massive amounts of user-submitted question information streams, and dynamic extraction of a list of frequently occurring user question modules from the knowledge base.

[0103] Understandably, to support subsequent intelligent analysis, the system will assign unique identifiers to these two types of data and inject structured metadata. Injecting structured metadata refers to metadata that forms a standard structure, meaning each piece of information is uniquely identified to aid in subsequent processes. For example:

{"uq_key":"","msg_raw":"I heard you're going to star in xxx?"","...":"..."}

[0104] Specifically, a globally unique Query_ID is generated for each user's question. Simultaneously, a Tag_ID is generated for each tag. This lays the data foundation for subsequently constructing an undirected graph linking questions and tags.

[0105] Process A2: Deep learning-based semantic encoding models (such as BERT) first perform word segmentation, dividing the sentence into n units. Named entity recognition and dependency parsing then help understand key features such as modifiers and nouns. For example, it can identify that "AX" is an alias for the target user "AAA," and that "released last week" in "new song released last week" modifies "new song."

[0106] Named entity recognition (NER) aims to identify entities, or words or phrases with specific meanings, from text. For example, in the sentence "AAA starred in XXX," the goal of NER is to identify the entities "AAA" and "XXX."

[0107] Dependency parsing helps in understanding the grammatical roles and syntactic structure of words in a sentence. For example, in the sentence “AAA starred in XXX.”, “AAA” is the subject, “starred in” is the predicate, and “XXX” is the object.

[0108] Process A3: Utilize the deep semantic understanding capabilities of deep learning-based semantic coding models (such as BERT) to map word segmentation expressions to specific entities and works in the knowledge base, thereby enabling the exploration of specific semantics, emotional expressions, and other content in the input statement.

[0109] The specific methods for mapping word segmentation expressions to concrete entities and works in the knowledge base include: First, performing an entity query to find specific information about a particular entity or work (such as definition, historical background, and the work itself). Next, performing a relationship query, that is, querying the relationships between the identified entities. For example, other works by a particular author, or historical events that occurred at a particular location. Finally, performing contextual understanding, that is, using contextual information to understand the user's needs. For example, when a user mentions "Luo Yunxi," the agent may need to determine from the context whether the user is asking about Luo Yunxi himself or the background of a specific work.

[0110] It is important to emphasize that this invention uses a semantic encoding model to transform text into high-dimensional vectors, namely, a list of sentence semantic vectors. Instead of mapping to the final text content.

[0111] S103. Generate a list of semantic association information based on the list of semantic vectors of statements and the association model.

[0112] The semantic association information list includes at least one semantic association information, and each semantic association information includes two statement semantic vectors and the degree of association between the two statement semantic vectors.

[0113] Optionally, in another embodiment of the present invention, one implementation of step S103 may specifically include the following steps (processes B1 to B2):

[0114] Process B1: Input any two statement semantic vectors from the statement semantic vector list into the correlation analysis model to generate the correlation between the two statement semantic vectors.

[0115] The correlation analysis model includes a linear transformation layer and a hybrid attention layer, which includes soft attention mechanism and self-attention mechanism.

[0116] Each vector in the statement semantic vector list is combined with any different vector to form a tuple. , > is used to represent a potential semantic relationship between two semantic segments. The number of combinations is This combination method uses a structured representation of pairwise semantic relationships as the reinforcement learning state, providing a learnable feature representation space for subsequent association degree decisions.

[0117] This invention treats semantic vector pairs as "states" through a reinforcement learning framework, enabling trainable state modeling of the semantic space, rather than static similarity calculation in traditional methods, thus laying the foundation for adaptive learning of semantic relations.

[0118] In the practical application of this invention, the correlation analysis model can be, but is not limited to, an information Q-Network structure based on a linear transformation (LL) layer and a hybrid attention (AL) layer, which is used to enhance the ability of Deep Q-Network in semantic deep feature modeling.

[0119] It should be noted that the LL layer maps the input vector to a new high-dimensional semantic feature space through a linear transformation, expressed as: rFF(X) = relu(XW + b). X is the input, W and b are the learning parameters, and relu is the activation function. The AL layer employs an adaptive multi-head attention mechanism, combining soft-attention and self-attention to jointly model the multidimensional dependencies in the input vector pairs. This structure can capture complex association patterns of semantic units in different contextual dimensions in parallel, achieving a deep feature representation capability that traditional DQN cannot achieve in semantic association tasks.

[0120] This invention introduces a hybrid attention structure into DQN for the first time to handle semantic association tasks, enabling Q-Networks to possess "semantic context awareness." Compared to ordinary MLP-based Q-Networks, this structure can learn dependencies across semantic segments, achieving higher accuracy in semantic association estimation.

[0121] In the practical application of this invention, any two statement semantic vectors from the statement semantic vector list are used as feature combinations and input into the information Q-Network structure based on the linear transformation (LL) layer and the hybrid attention (AL) layer to output the Q value of the corresponding action.

[0122] It should be noted that the semantic modification of the action selection process in Q-learning in this invention is as follows: Q-Network forward propagation: outputs the Q-value estimates of all possible actions. Policy selection (e.g., ε-greedy): selects the optimal action in the current state.

[0123] It is important to emphasize that, in this invention, the Q-value no longer represents the expected reward in the traditional sense, but is redefined as the degree of correlation. That is, vector , A quantitative metric for the strength of semantic association between elements. The Q-function learned during the training process of Q-Network essentially becomes a learnable semantic association mapping model, whose output can be dynamically adjusted according to the context, rather than fixed rules or static distance metrics.

[0124] This invention redefines the Q-value of DQN as semantic relevance, representing a breakthrough application of reinforcement learning frameworks in NLP semantic relation modeling. The system can continuously optimize the semantic relevance model through multiple training rounds, making the relevance calculation learnable, iterative, and context-adaptive.

[0125] like Figure 3 The diagram shown illustrates a Q-Network hierarchical structure design according to an embodiment of the present invention. After the feature vector is input, it first undergoes a linear transformation through the LL layer, then passes through the Self-attention and Soft-attention parts in the AL layer to capture the complex association patterns of semantic units under different contextual dimensions, and then undergoes another linear transformation through the LL layer, finally outputting the Q value.

[0126] Process B2: Generate semantic association information based on the semantic vectors of the two statements and the degree of association between the two semantic vectors.

[0127] Continuing with the above example, through the above mechanism, we can finally obtain the semantic association triple: < , , >, which refers to semantic association information, and its data is shown in Table 1.

[0128] Table 1

[0129] Statement semantic vector i Statement semantic vector j Vector correlation degree (Q-value) {0.5，...，0.97} {0.66，...，0.57} 0.45 {0.41，...，0.99} {0.3，...，0.14} 0.97

[0130] S104. Generate an undirected graph of association based on the list of semantically related information.

[0131] In this context, nodes in the undirected graph of association are generated from the semantic vectors of statements in the semantic association information, and edges in the undirected graph of association are established based on the degree of association between two statement semantic vectors. Each node corresponds to a semantic unit, and the edges between nodes are established according to the degree of association. The distance of the edges is assigned as a weight attribute to the edges, thereby quantifying the strength of the association or semantic similarity between nodes.

[0132] S105. Perform classification analysis on the nodes in the associated undirected graph and generate classification analysis results.

[0133] The classification analysis results include problem classification labels for at least one node partition.

[0134] It should be noted that, in order to distinguish the source and nature of nodes, in the actual application of this invention, nodes can be divided into two types: template nodes and input nodes.

[0135] Template nodes represent vectors corresponding to predefined, approved tag templates or standard question-and-answer pairs in the system, forming a stable knowledge skeleton of the graph.

[0136] The input node represents the vector formed by encoding the user's real-time input statement, serving as a dynamic extension of the graph.

[0137] Optionally, in another embodiment of the present invention, one implementation of step S105 may be as follows (processes C1 to C3):

[0138] Process C1: For each template node, starting from the template node, the first target input node belonging to the node partition corresponding to the template node is determined layer by layer using the dynamic correlation threshold and the correlation between nodes, and the first classification analysis result is obtained.

[0139] The dynamic correlation threshold for each level is determined based on the initial correlation threshold and the control factor; the control factor is determined based on the number of nodes in different node partitions and the number of nodes that should be included.

[0140] like Figure 4 As shown in the undirected graph of association, starting from the template node, input nodes with a value greater than the dynamic association threshold are assigned to the node partition to which the template node belongs as the first target input node. The search and matching continues with the first target data node. Output nodes with a value greater than the dynamic association threshold are assigned to the first target data node of the next level, and so on, until there are no output nodes with a value greater than the dynamic association threshold.

[0141] In the practical application of this invention, nodes with a dynamic correlation threshold can be connected with solid lines, and nodes with a correlation threshold less than the dynamic correlation threshold can be connected with dashed lines; this is not a limitation.

[0142] In the practical application of this invention, the search matching method can be, but is not limited to, breadth-first search (BFS), that is, starting from the template node, the input nodes are traversed in a breadth-first manner, and the input nodes most related to the template node are preferentially added to the category. This is not limited here.

[0143] It is important to emphasize that the dynamic correlation threshold in this invention is dynamically adjusted according to the local semantic density, thereby forming an automatic stratification in different semantic clusters. Unlike the traditional BFS, this algorithm guides the clustering direction through "template nodes", forming a "semantic core-driven" graph clustering mechanism. This avoids the cross-class misconnection problem that is prone to occur in traditional BFS clustering in this problem.

[0144] Specifically, when traversing from template nodes or new tag classification nodes, the system sets an initial semantic association strength threshold T (initial association threshold). The core function of this initial association threshold T is to define a boundary for a local search.

[0145] Once the traversal process reaches the second-level nodes from the template node, the system will activate the threshold adaptive adjustment mechanism. The core of this mechanism lies in dynamically increasing the current threshold T. The new threshold T' is calculated using the formula T' = T + (1-T) × α. In this formula, α is a configurable control factor, whose value is strictly defined as (0, 1). The physical meaning of this factor is to control the magnitude and rate of threshold increase: the larger the value of α, the greater the threshold increase in each iteration. This reflects the algorithm's strategy of tending to converge to a more precise local region faster, thus intelligently balancing the breadth and depth of the search.

[0146] When the threshold is increased for the third time at the template node, the value is... And so on, at the Nth layer, the threshold calculation formula is:

[0147] ;

[0148] Therefore, the higher the level n traversed by the algorithm, the higher the threshold required to connect the (n-1)th and nth layers, thus avoiding the cross-class misconnection problem that is prone to occur in the traditional BFS clustering algorithm in the scenario presented in this paper.

[0149] However, in practical applications, the semantic relevance between user input questions can dynamically change due to external environmental factors (such as recent hot topics related to the target user and fluctuations in public opinion intensity). In this situation, if a fixed control factor is still used for semantic relevance judgment, it can easily lead to limited local search boundaries, resulting in problems such as multiple or few cross-class connections. To solve the generalization deficiencies caused by local optima and threshold fixation, this invention proposes a dynamic adaptive adjustment mechanism for the control factor α.

[0150] Specifically, the system randomly selects N classification results for evaluation. For a certain category, suppose there are already x semantic vectors classified into that category, and according to the actual annotation or statistical rules, the category should contain y semantic vectors.

[0151] If x > y, it indicates the presence of (x-y) cross-class multi-link phenomena, meaning the current value of α is too conservative, causing the system to loosen its judgment on associations. α needs to be appropriately increased to enhance its discriminative ability. If y > x, it indicates the presence of (y-x) cross-class few-link phenomena, meaning the value of α is too aggressive, causing the system to excessively suppress potential associations. α needs to be appropriately decreased to improve recall.

[0152] The above evaluation process is repeated N times across different categories. The system comprehensively analyzes the deviation between multiple connections across categories and fewer connections across categories, automatically calculates the update magnitude of α, and achieves dynamic adjustment of α. The formula for calculating the update magnitude is as follows:

[0153] ;

[0154] This invention achieves an intelligent trade-off between search breadth and depth through the above-described process. In the shallower layers of traversal (when N is small), the threshold increase is relatively large, allowing the algorithm to expand rapidly and conduct extensive "exploration." However, as the algorithm delves deeper (when N increases), the threshold increase significantly narrows. This makes the algorithm extremely "conservative" in deeper traversals, ensuring that only nodes with extremely high relevance are included in subsequent steps. This effectively suppresses excessive branching of the search tree and avoids consuming excessive computational resources on paths with weak semantic connections.

[0155] In this invention, the control factor α, as a dynamically adjustable parameter, acts as a knob for adjusting the "aggressive" or "conservative" nature of the search. A larger α value results in a faster threshold increase and a more rapid convergence of the search range; conversely, a smaller α value leads to a more detailed and persistent search. This invention enables the association judgment threshold to adaptively change with user input features and external environmental variations, thereby effectively improving classification accuracy and the system's robustness to semantic changes.

[0156] In the application scenarios of this invention, experimental results show that the proposed algorithm achieves significant improvements in key performance indicators compared to the traditional BFS clustering algorithm. Specifically, the cross-class false linking error rate is reduced by more than 29%, while the data processing or clustering exploration efficiency is improved by more than 45%.

[0157] In the actual application of this invention, after the search and matching of all template nodes is completed, process C2 is executed to further classify the new problems found.

[0158] Process C2: If the correlation between an input node and all template nodes is less than the minimum correlation threshold, then the input node is taken as the target node (representing a new category dynamically identified in this session), which is the starting point of the new question classification label.

[0159] Process C3: For each target node, starting from the target node, the second target input node corresponding to the node partition of the target node is determined layer by layer using the dynamic correlation threshold and the correlation between nodes, until the correlation between the current input node and all template nodes is less than the minimum correlation threshold (e.g., 0.4), and the second classification analysis result is obtained.

[0160] Specifically, the traversal and search process for the target node can be referred to the implementation method corresponding to process C1, which will not be repeated here.

[0161] Each node partition represents the same semantic category. It can be understood that if a partition contains a template node, the category is considered an existing category; if it does not contain any existing template nodes, the category is considered a new category.

[0162] Finally, the results of the first and second classification analyses can be combined to generate the final classification analysis result.

[0163] In the practical application of this invention, after each input node is classified into a category corresponding to the semantic text, if there is an existing category, the previous answer history of the target user can be obtained from the knowledge base as supplementary content, which is not limited here.

[0164] Optionally, in another embodiment of the present invention, in order to improve the correlation between Q-Network computation vectors... To ensure accuracy, according to the DQN algorithm, the Q-values ​​of all actions need to be evaluated to train DQN. The specific steps are as follows (processes D1 to D3):

[0165] D1: Generate reward parameters based on the evaluation information of the problem classification labels.

[0166] In this robust practical application, operators can extract the output node categories (problem classification tags) and evaluate their categorization to determine if they meet expectations. If they do, the state assessment is considered reasonable and receives a high rating; otherwise, it is considered unreasonable and receives a low rating. The evaluation is then calculated. The formula is as follows:

[0167] ;

[0168] D2: Construct the state transition process based on semantic association information and reward parameters.

[0169] Specifically, the future state is the pair of vertices j and j+1 associated with the previous vertex i. , Therefore, the state transition process of the Q-Network can be constructed, and the corresponding relationship is shown in Table 2. The successful state transition process is as follows: (< , >, ,1,< , >), its transfer failure affects the transfer process as follows: (< , >, ,0,< , >).

[0170] Table 2

[0171] Q-learning algorithm transfer process This invention mapping State Action Next state (future state) Reward

[0172] D3: The correlation analysis model is further optimized using the state transition process to obtain the optimized correlation analysis model.

[0173] Specifically, the state transitions are placed into the training pool (memorypool) of the correlation analysis model (such as Q-Network), and the data in the training pool is used for training Q-Network.

[0174] S106. For each question category tag, based on the questions associated with the question category tag and the historical dialogue records corresponding to the question category tag in the knowledge base, generate a set of recommended answers corresponding to the question category tag.

[0175] Among them, the questions associated with the question classification tags are user input question sets formed by aggregating approved tag templates, that is, standardized question expressions associated with the question classification tags; the historical dialogue records corresponding to the question classification tags in the knowledge base include, but are not limited to, the target user's historical response corpus that has been manually reviewed and standardized, etc., which are not limited here.

[0176] This invention constructs a semantically enhanced dual-source conditional input expression based on an approval template by using a structured association of "tag template - user input - knowledge base corpus". This ensures that the context received by the model is not only highly consistent in semantics, but also controllable in terms of quality, security and target user style.

[0177] In the practical application of this invention, the questions associated with the question classification tags and the historical dialogue records corresponding to the question classification tags in the knowledge base can be input into the pre-trained recommendation model to output the answer recommendation set corresponding to the question classification tags.

[0178] The pre-trained recommendation model can be, but is not limited to, autoregressive generative models based on the Transformer decoder architecture, large language models, etc., and is not limited here. This model utilizes enhanced conditional language modeling capabilities to simultaneously model the semantic specifications of tag templates, the contextual details of user questions, and the stylistic features of corresponding historical corpora, achieving deep alignment with the target user's fixed style, personality traits, and tone habits.

[0179] Specifically, if the Transformer is used, it can utilize a self-attention mechanism to understand dependencies within the context, and then progressively generate new words until the entire text is complete. This approach allows the generated text to be natural, fluent, and semantically meaningful, adaptable to different input content and contexts.

[0180] Unlike traditional generation methods that rely on static templates or a single input source, this invention introduces a combined mechanism of "review template constraints + enhancement of high-quality historical corpus + autoregressive conditional generation," which significantly improves the consistency, semantic accuracy, personalization, and content controllability of response generation, providing target users with efficient and low-risk intelligent response decision support.

[0181] Optionally, in another embodiment of the invention, such as Figure 5 As shown, when a user raises a question, the system does not send it immediately. Instead, it stores the question request as a pending task in a dynamic question queue (i.e., the "question pool") for temporary storage and immediately sends feedback to the user, such as: "The question has been collected, please wait for a reply."

[0182] When the online event of the target user is detected, the size of the first dynamic virtual congestion window is determined based on the real-time load information of the current pipeline. Based on the size of the first dynamic virtual congestion window, the relevant issue information of the target user is obtained from the issue pool, and then step S102 is executed.

[0183] It's important to note that when the processing flow is smooth, this window increases linearly, allowing more messages to be aggregated and submitted in batches, aiming to improve overall processing throughput. When the system detects increased processing latency or potential congestion, it automatically reduces the window size, switching to smaller batches and higher frequency submissions to prioritize processing real-time performance and system stability. This implements a negative feedback control loop, enabling the system to intelligently balance the trade-off between throughput and latency.

[0184] In the practical application of this invention, the method for determining the dynamic virtual congestion window size based on the real-time load information of the current pipeline can be calculated using the following formula:

[0185] ;

[0186] in, Let be the size of the dynamic virtual congestion window at time t, C be the scaling factor that controls the agility of queue depth growth, t be the time elapsed since the last processing timeout (i.e., the "congestion event"), and K be the time the algorithm recovers from the current window to the current window. The estimated time required; where K can be calculated using the following formula:

[0187] ;

[0188] Here, β is the window reduction factor. Once a timeout occurs, the current window will be immediately multiplied by β to reduce it (e.g., to 0.7 or 0.8 of the original value). β and C together affect K.

[0189] The window reduction rule is as follows: ;

[0190] In the actual application of the present invention, the problem information corresponding to the target user is obtained from the problem pool based on the size of the first dynamic virtual congestion window. This can be, but is not limited to, obtaining the problem information corresponding to the target user from the problem pool based on the maximum quantization of the first dynamic virtual congestion window. No limitation is made here.

[0191] Of course, when receiving the target user's target answer information, it is also necessary to use a dynamic virtual congestion window, that is, to determine the size of the second dynamic virtual congestion window based on the real-time load information of the current pipeline; then, based on the size of the second dynamic virtual congestion window, the target user's target answer information is fed back to the corresponding user of the corresponding question in the question classification tag.

[0192] It should be noted that the specific implementation method for determining the size of the second dynamic virtual congestion window based on the real-time load information of the current pipeline can refer to the specific implementation method for determining the size of the first dynamic virtual congestion window based on the real-time load information of the current pipeline, and will not be repeated here.

[0193] In the actual application of this invention, based on the size of the second dynamic virtual congestion window, the target user's target answer information is fed back to the corresponding user of the corresponding question in the question classification label. This can be, but is not limited to, feeding back to the corresponding user of the corresponding question in the question classification label based on the maximum quantization of the second dynamic virtual congestion window. No limitation is made here.

[0194] Optionally, to ensure the security of interaction between the target user and the user, in another embodiment of the present invention, one implementation of the answer recommendation method further includes:

[0195] After receiving the Q&A information, the content of the Q&A information is reviewed and the review results are obtained; Q&A information that is deemed to be risky is blocked.

[0196] The question-and-answer information includes the user's question information and the target answer information.

[0197] This invention does not rely on simple keyword filtering, but rather deeply integrates a dynamically updated multi-dimensional knowledge base (such as a sensitive word database, risk topic model, compliance strategy rules, etc.).

[0198] In the practical application of this invention, the information statistics intelligent agent can continuously update and maintain the target user's knowledge base (and also supports manual modification by operators). This is not limited here. Therefore, this invention dynamically obtains new data through the knowledge base, thereby avoiding information lag.

[0199] Optionally, in another embodiment of the present invention, one implementation of the answer recommendation method further includes:

[0200] If the language used in the user's question is different from the target user's usual language, the user's question will be translated into the target user's usual language.

[0201] In the practical application of this invention, a knowledge base is actively queried and deeply integrated before performing a translation task. This knowledge base is not a simple list of terms, but contains rich structured contextual information such as standardized terms in the entertainment industry, information on the target user's works, common nicknames, recent hot topics, and cultural background annotations.

[0202] like Figure 6The flowchart shown is a method for recommending answers provided in an embodiment of the present invention. In the process of generating answer recommendations, it can not only solve the problem of expression distortion in the dialogue translation process of international interactive platforms, but also greatly improve the speed at which target users reply to questions, significantly improve the efficiency of interaction, greatly protect the security of interaction between target users and users, and effectively identify hidden and variant illegal content.

[0203] First, in step ①, the user submits a question they wish to ask the target user, such as their current schedule or plans to release a new work. Step ② indicates that the system will push the questions to the dialogue system's processing pipeline in batches. This invention determines the size of a dynamic virtual congestion window based on the real-time load information of the current pipeline, intelligently balancing the throughput and latency in user submissions and target user responses using this dynamic virtual congestion window. As shown in step ③, after the user's question is pushed to the pipeline in step ②, it first passes through a security and risk management intelligent agent. This agent will intercept inappropriate statements, ensuring the safety of the target user's interaction with other users and identifying implicit or variant violations. During this process, a knowledge base (including an industry terminology database and the target user's private knowledge base) will be dynamically linked. As shown in step ④, after step ③ is completed, the pipeline will translate the content sent by the user to ensure consistency with the target user's language. Again, this process will involve dynamically linking to the knowledge base. As shown in step ⑤, this module first performs real-time cluster analysis on the massive amount of user questions, automatically identifying semantically similar question groups using natural language processing technology and categorizing them into several standardized topics. Subsequently, based on a pre-built target user knowledge base and historical high-quality dialogue data, the system dynamically generates a multi-dimensional, selectable list of response suggestions for each standardized topic. These suggested responses not only accurately match the user's intent but also fully consider the target user's personal style and the current context. Finally, the system presents the summarized question groups and their corresponding intelligent response suggestions to the target user in a batch push manner. The target user can then quickly reply with one click or make lightweight edits before sending, thus transforming fragmented inquiries that originally required processing one by one into an efficient batch decision-making and response process. As shown in step ⑥, after the large model aggregates and generates answer recommendations in step ⑤, the target user will edit or select content to reply. Compared to relying on purely manual replies from the target user, this large model aggregation and answer generation can greatly reduce the target user's time cost and increase the number of times users can reply. As shown in step ⑦, after the target user replies, security and risk management are implemented to avoid public opinion crises.

[0204] like Figure 4 As shown, this invention provides an intelligent induction and response auxiliary agent system structure, which is essentially a classification + text generation system based on deep reinforcement learning (DQN), greatly improving the speed at which target users respond to questions and significantly enhancing interaction efficiency.

[0205] like Figure 4 As shown in step 1, a deep learning-based semantic encoding model (BERT model in this example) obtains the question text sent by the user and retrieves the current target user's background information and the tag templates from the knowledge base for the past month. These are used for semantic analysis to generate semantic vectors. Figure 4 As shown in step 2, the semantic vector is input into the Q-Network as the state of the DQN algorithm. Figure 4 Step 3, as shown, indicates that according to the DQN algorithm, we need to generate Q-values ​​for all states and map them to calculate the correlation between vectors. For example... Figure 4 As shown in step 4, semantic classification is achieved using the depth-first clustering algorithm based on semantic association graphs, which employs class-based depth-first search (BFS) proposed in this invention. Figure 4 As shown in step 5, the autoregressive generative model based on the Transformer decoder architecture (the Mango Big Model in this example) dynamically generates a set of multi-dimensional, personalized candidate response suggestions that match the target user's style and tone, providing efficient reference and decision-making for the target user. Figure 4 As shown in steps 6 and 7, operations personnel extract and categorize a portion of the issues, then evaluate and judge them. These evaluations are then input into the DQN as the reward portion. Figure 4 The transfer process shown in step 8 is put into the training pool (memory pool) of Q-Network to train Q-Network.

[0206] As can be seen from the above solutions, the present invention provides an answer recommendation method. By automatically classifying user questions, similar questions are grouped into the same question category, and a corresponding set of recommended answers is generated. Target users can choose the recommended answers from the set to reply directly, thereby transforming fragmented inquiries that originally required processing one by one into an efficient batch decision-making and response process, significantly improving interaction efficiency and effectively enhancing the user experience for both users and target users.

[0207] Another embodiment of the present invention provides a response recommendation device, such as... Figure 7 As shown, it specifically includes:

[0208] The first receiving unit 701 is used to receive user problem information.

[0209] Clustering analysis unit 702 is used to perform clustering analysis based on question information and existing tagged questions in the knowledge base to obtain a list of sentence semantic vectors.

[0210] The association information generation unit 703 is used to generate a list of semantic association information based on the list of statement semantic vectors and the association model.

[0211] The semantic association information list includes at least one semantic association information, and each semantic association information includes two statement semantic vectors and the degree of association between the two statement semantic vectors.

[0212] The undirected graph generation unit 704 is used to generate an undirected graph based on a list of semantic association information.

[0213] In this context, the nodes in the associated undirected graph are generated from the semantic vectors of the statements in the semantic association information, and the edges in the associated undirected graph are established based on the degree of association between the semantic vectors of two statements.

[0214] The classification analysis unit 705 is used to perform classification analysis on nodes in the associated undirected graph and generate classification analysis results.

[0215] The classification analysis results include problem classification labels for at least one node partition.

[0216] The answer recommendation generation unit 706 is used to generate a set of answer recommendations corresponding to each question category tag, based on the questions associated with the question category tag and the historical dialogue records corresponding to the question category tag in the knowledge base.

[0217] For details on the specific operation of the units disclosed in the above embodiments of the present invention, please refer to the corresponding method embodiments, such as... Figure 1 As shown, it will not be elaborated further here.

[0218] As can be seen from the above solutions, the present invention provides an answer recommendation device that automatically categorizes user questions, aggregates similar questions into the same question category, and generates a corresponding set of recommended answers. Target users can choose the recommended answers from the set to reply directly, thereby transforming fragmented inquiries that originally required processing one by one into an efficient batch decision-making and response process, significantly improving interaction efficiency and effectively enhancing the user experience for both users and target users.

[0219] The functions described above in this document can be performed at least in part by one or more hardware logic components. For example, exemplary types of hardware logic components that can be used, without limitation, include: field-programmable gate arrays (FPGAs), application-specific integrated circuits (ASICs), application-specific standard products (ASSPs), system-on-a-chip (SoCs), complex programmable logic devices (CPLDs), and so on.

[0220] Another embodiment of the present invention provides an electronic device, comprising:

[0221] One or more processors.

[0222] A storage device on which one or more programs are stored.

[0223] When the one or more programs are executed by the one or more processors, the one or more processors implement the answer recommendation method as described in the above embodiments.

[0224] Another embodiment of the present invention provides a computer storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the answer recommendation method as described in the above embodiments.

[0225] In the context of this invention, a machine-readable medium can be a tangible medium that may contain or store a program for use by or in conjunction with an instruction execution system, apparatus, or device. A machine-readable medium can be a machine-readable signal medium or a machine-readable storage medium. Machine-readable media can include, but are not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any suitable combination of the foregoing. More specific examples of machine-readable storage media include electrical connections based on one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fibers, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination of the foregoing.

[0226] It should be noted that the computer-readable medium described above in this invention can be a computer-readable signal medium, a computer-readable storage medium, or any combination thereof. A computer-readable storage medium can be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of a computer-readable storage medium may include, but are not limited to: an electrical connection having one or more wires, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof. In this invention, a computer-readable storage medium can be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device. In this invention, a computer-readable signal medium can include a data signal propagated in baseband or as part of a carrier wave, carrying computer-readable program code. Such propagated data signals can take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. A computer-readable signal medium can be any computer-readable medium other than a computer-readable storage medium, which can send, propagate, or transmit a program for use by or in connection with an instruction execution system, apparatus, or device. The program code contained on the computer-readable medium can be transmitted using any suitable medium, including but not limited to: wires, optical fibers, RF (radio frequency), etc., or any suitable combination thereof.

[0227] The aforementioned computer-readable medium may be included in the aforementioned electronic device; or it may exist independently and not assembled into the electronic device.

[0228] Another embodiment of the present invention provides a computer program product, which, when executed, is used to perform the above-described answer recommendation method.

[0229] In particular, according to embodiments of the present invention, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments of the present invention include a computer program product comprising a computer program carried on a non-transitory computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via a communication device, or installed from a storage device, or installed from a ROM. When the computer program is executed by a processing device, it performs the functions defined in the methods of the embodiments of the present invention.

[0230] Although the subject matter has been described using language specific to structural features and / or methodological logic, it should be understood that the subject matter defined in this invention is not necessarily limited to the specific features or actions described above. Rather, the specific features and actions described above are merely exemplary forms for implementing the invention.

[0231] While several specific implementation details are included in the foregoing discussion, these should not be construed as limiting the scope of the invention. Certain features described in the context of individual embodiments may also be implemented in combination in a single embodiment. Conversely, various features described in the context of a single embodiment may also be implemented individually or in any suitable sub-combination in multiple embodiments.

[0232] The above description is merely a preferred embodiment of the present invention and an explanation of the technical principles employed. Those skilled in the art should understand that the scope of the invention is not limited to the specific combination of the above-described technical features, but also includes other technical solutions formed by arbitrary combinations of the above-described technical features or their equivalents without departing from the inventive concept. For example, technical solutions formed by substituting the above-described features with technical features of the present invention (but not limited to) that have similar functions.

Claims

1. A method for recommending answers, characterized in that, include: Receive user problem information; Cluster analysis is performed based on the question information and the existing tagged questions in the knowledge base to obtain a list of sentence semantic vectors; Based on the statement semantic vector list and the association model, a semantic association information list is generated; wherein, the semantic association information list includes at least one semantic association information, and each semantic association information includes two statement semantic vectors and the degree of association between the two statement semantic vectors; Based on the semantic association information list, an undirected association graph is generated; wherein, the nodes in the undirected association graph are generated from the semantic vectors of the statements in the semantic association information, and the edges in the undirected association graph are established based on the degree of association between two semantic vectors of statements; The nodes in the associated undirected graph are classified and analyzed to generate classification analysis results; wherein, the classification analysis results include at least one problem classification label for a node partition; For each of the aforementioned question categorization tags, a set of recommended answers corresponding to the question categorization tag is generated based on the question associated with the question categorization tag and the historical dialogue records corresponding to the question categorization tag in the knowledge base.

2. The answer recommendation method according to claim 1, characterized in that, The step of generating a semantic association information list based on the statement semantic vector list and association model includes: Input any two statement semantic vectors from the statement semantic vector list into the correlation analysis model to generate the correlation between the two statement semantic vectors; wherein, the correlation analysis model includes a linear transformation layer and a hybrid attention layer; Semantic association information is generated based on the semantic vectors of the two statements and the degree of correlation between them.

3. The answer recommendation method according to claim 1, characterized in that, The nodes in the associated undirected graph include template nodes and input nodes. The process of classifying and analyzing the nodes in the associated undirected graph to generate classification analysis results includes: For each template node, starting from the template node, the first target input node belonging to the node partition corresponding to the template node is determined layer by layer using the dynamic correlation threshold and the correlation between nodes, thus obtaining the first classification analysis result; wherein, the dynamic correlation threshold of each level is determined based on the initial correlation threshold and the control factor; the control factor is determined based on the number of nodes in the current different node partitions and the number of nodes that should be included.

4. The answer recommendation method according to claim 3, characterized in that, Also includes: If the correlation degree between an input node and all the template nodes is less than the minimum correlation degree threshold, then the input node is taken as the target node. For each target node, starting from the target node, the second target input node corresponding to the node partition is determined layer by layer using the dynamic correlation threshold and the correlation between nodes, until no input node has a correlation with all template nodes that is less than the minimum correlation threshold, and the second classification analysis result is obtained. Based on the first classification analysis result and the second classification analysis result, the final classification analysis result is generated.

5. The answer recommendation method according to claim 1, characterized in that, Also includes: When the online event of a target user is detected, the size of the first dynamic virtual congestion window is determined based on the real-time load information of the current pipeline; Based on the size of the first dynamic virtual congestion window, after obtaining the question information corresponding to the target user from the question pool, the step of performing cluster analysis based on the question information and the existing tagged questions in the knowledge base to obtain a list of statement semantic vectors is started.

6. The answer recommendation method according to claim 5, characterized in that, After generating a set of recommended answers for each question category tag based on the question associated with the question category tag and the historical dialogue records corresponding to the question category tag in the knowledge base, the method further includes: Receive the target answer information of the target user; wherein, the target answer information is a recommended answer selected by the target user from the answer recommendation set corresponding to the question category tag, or an answer modified by the target user for the recommended answer in the answer recommendation set corresponding to the question category tag; The size of the second dynamic virtual congestion window is determined based on the real-time load information of the current pipeline. Based on the size of the second dynamic virtual congestion window, the target user's target answer information is fed back to the corresponding user of the corresponding question in the question categorization label.

7. The answer recommendation method according to claim 2, characterized in that, Also includes: Reward parameters are generated based on the evaluation information of the problem classification tags; A state transition process is constructed based on the semantic association information and the reward parameters; The correlation analysis model is further optimized using the state transition process to obtain an optimized correlation analysis model.

8. The answer recommendation method according to claim 6, characterized in that, Also includes: After receiving the question and answer information, the content of the question and answer information is reviewed to obtain the review result; wherein, the question and answer information includes the user's question information and the target answer information; Questions and answers that are deemed risky by the review results will be blocked.

9. The answer recommendation method according to claim 1, characterized in that, After receiving the user's question information, the process also includes: If the language used in the user's question is different from the target user's usual language, the user's question will be translated into the target user's usual language.

10. A response recommendation device, characterized in that, include: The first receiving unit is used to receive the user's problem information; The clustering analysis unit is used to perform clustering analysis based on the question information and the existing tagged questions in the knowledge base to obtain a list of sentence semantic vectors; The association information generation unit is used to generate a semantic association information list based on the statement semantic vector list and the association model; wherein, the semantic association information list includes at least one semantic association information, and each semantic association information includes two statement semantic vectors and the association degree between the two statement semantic vectors; An undirected graph generation unit is used to generate an undirected graph based on the semantic association information list; wherein, the nodes in the undirected graph are generated from the semantic vectors of the statements in the semantic association information, and the edges in the undirected graph are established based on the degree of association between two semantic vectors of statements. The classification analysis unit is used to perform classification analysis on the nodes in the associated undirected graph and generate classification analysis results; wherein, the classification analysis results include at least one problem classification label for a node partition; The answer recommendation generation unit is used to generate an answer recommendation set corresponding to each question category tag, based on the question associated with the question category tag and the historical dialogue records corresponding to the question category tag in the knowledge base.

Images