Knowledge graph-based heat supply service customer service question and answer method, equipment and medium
By constructing a knowledge graph of heating question-and-answer and combining it with a reinforcement learning network model, the problem of accuracy and flexibility in responding to complex demands in the heating service customer service system was solved, and adaptive and efficient heating service responses were achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- SHANDONG SYNTHESIS ELECTRONICS TECH
- Filing Date
- 2025-12-26
- Publication Date
- 2026-05-08
AI Technical Summary
Existing heating service customer service systems struggle to achieve deep semantic reasoning and implicit demand understanding across multiple rounds of dialogue when faced with complex, dynamic, and highly specialized service requests. They lack self-optimization capabilities, resulting in decreased response accuracy and insufficient service flexibility.
A knowledge graph for heating-related questions and answers is constructed and semantically embedded. Response text is generated through a reinforcement learning network model, and user feedback is used to optimize the knowledge graph, thereby achieving adaptive adjustment of the response strategy.
It improves the accuracy and flexibility of responses, enabling high-quality user interaction during policy changes or service peaks, and achieving autonomous iteration and improvement of service quality.
Smart Images
Figure CN121996752A_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of artificial intelligence customer service technology, specifically to a knowledge graph-based method, equipment, and medium for answering customer service questions in heating services. Background Technology
[0002] In the urban public utility system, heating services cover a vast number of users. With the deepening of digital transformation in the public utility sector, using artificial intelligence (AI) technology to automate and professionally interact with heating service processes has gradually become a core direction for optimizing service efficiency, giving rise to intelligent customer service systems for heating services. These systems represent a significant application of AI technology in the public utility sector, aiming to automate the handling of various service requests from users regarding heating policies, fault reporting, fee inquiries, and business processing.
[0003] However, existing heating service customer service systems are typically built on rule engines or retrieval models, and are gradually incorporating knowledge graph technology to structurally model entities and their relationships in the heating field. These systems fall short when faced with complex, dynamic, and highly specialized service demands. On one hand, heating knowledge involves a large amount of specialized knowledge, dynamic policies, and complex equipment relationships. Traditional knowledge graphs struggle to achieve deep semantic reasoning and understanding of implicit needs across multiple rounds of dialogue, such as locating faults with unclear user descriptions or performing complex queries involving historical fees and tiered policies. On the other hand, existing systems lack the ability to self-optimize based on continuous user interaction feedback, and cannot adaptively adjust knowledge representation and response strategies. This leads to decreased response accuracy, delayed knowledge updates, and insufficient service flexibility when facing seasonal peak repair requests, policy updates, or new types of faults. Summary of the Invention
[0004] To address the aforementioned issues, this application proposes a knowledge graph-based question-and-answer method for heating service customer service, including: Based on historical heating service records, a heating question-and-answer knowledge graph is constructed, and semantic embedding is performed on the heating question-and-answer knowledge graph. The system analyzes user-submitted heating service questions, extracts key question entities, identifies the user's question intent category, generates an enhanced question semantic vector based on the key question entities and the question intent category, and calculates the semantic matching degree between the enhanced question semantic vector and each entity node in the heating question-answering knowledge graph. The entity node corresponding to the highest semantic matching degree is determined as the retrieval starting point. Using the retrieval starting point as the initial node, a multi-hop traversal is performed along the entity relationship edge in the heating question and answer knowledge graph. The end nodes reached by the traversal path are collected as candidate nodes to form a candidate node set. Using a reinforcement learning network model, the expected revenue value of each candidate node in the candidate node set is calculated according to the question intent category. Target response nodes are then selected. Based on the question-and-answer semantic information corresponding to the target response nodes, heating service response text is generated and returned to the user terminal. The system collects interactive feedback data generated by the user in response to the heating service text, calculates a reward signal value based on the interactive feedback data using a reward function, and optimizes the network parameters of the reinforcement learning network model based on the reward signal value.
[0005] On the other hand, this application also proposes a knowledge graph-based customer service Q&A device for heating services, including: At least one processor; and, A memory communicatively connected to the at least one processor; wherein, The memory stores instructions that can be executed by the at least one processor, which enables the at least one processor to perform, for example, a knowledge graph-based customer service Q&A method for heating services as described in the above example.
[0006] On the other hand, this application also proposes a non-volatile computer storage medium storing computer-executable instructions, wherein the computer-executable instructions are configured as: a knowledge graph-based customer service question-and-answer method for heating services as described in the above example.
[0007] The knowledge graph-based question-and-answer method for heating service customer service proposed in this application can bring the following benefits: By constructing and embedding a semantic heating question-and-answer knowledge graph, we can deeply analyze the intent and key entities in user questions and generate enhanced semantic query vectors. Then, we can perform accurate matching and multi-hop reasoning in the graph, which enables us to accurately locate the answer nodes and logical links most relevant to complex, implicit, or composite heating consultations from a structured professional knowledge network. This results in the generation of accurate and logically complete response texts, overcoming the shortcomings of traditional methods in terms of professionalism and accuracy.
[0008] Furthermore, by introducing a reinforcement learning mechanism, real-time user satisfaction ratings and follow-up questions are quantified into reward signals, which are then used to dynamically optimize the decision-making model and continuously supplement the knowledge graph. This allows the system to break free from dependence on fixed rules, learn autonomously from real-world interactive experiences, and continuously optimize its answer selection strategies and knowledge organization structure. Consequently, it maintains and improves the accuracy of responses and user satisfaction when facing policy changes, new business scenarios, or seasonal service peaks, achieving autonomous iteration and long-term improvement in service quality. Attached Figure Description
[0009] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings: Figure 1 This is a flowchart illustrating a knowledge graph-based customer service question-and-answer method for heating services, as described in an embodiment of this application. Figure 2 This is a schematic diagram of a knowledge graph-based customer service Q&A device for heating services, as described in an embodiment of this application. Detailed Implementation
[0010] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions of this application will be clearly and completely described below in conjunction with specific embodiments and corresponding drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.
[0011] The technical solutions provided by the various embodiments of this application are described in detail below with reference to the accompanying drawings.
[0012] like Figure 1 As shown in the figure, this application provides a knowledge graph-based customer service question-and-answer method for heating services, including: S101. Based on historical heating service records, construct a heating question and answer knowledge graph, and perform semantic embedding on the heating question and answer knowledge graph.
[0013] Specifically, historical heating service records are obtained. In practice, these records are retrieved from the heating service database, log storage system, and document management server. These historical heating service records primarily include, but are not limited to: structured historical work order texts recording user repair requests, inquiries, and processing; unstructured historical customer service dialogue logs storing natural language conversations between customer service representatives and users; and heating policy documents containing heating fee standards, policy terms, equipment specifications, and other relevant information.
[0014] Furthermore, historical heating service records are analyzed to identify and extract heating service object entities and the corresponding service logical relationships between them. A heating question-and-answer knowledge graph is constructed, using heating service object entities as entity nodes and service logical relationships as entity relationship edges. In specific implementation, natural language processing technology is used to deeply analyze the text data, identifying and extracting heating service object entities that represent specific objects of heating services, such as: "user," "heating station," "radiator," "pipeline," "valve," "fee account," and "fault code."
[0015] At the same time, service logic relationships that represent the business logic or service associations between these entities are also extracted, such as: "User-Own-Account", "Device-Located-Site", "Fault-Occurs-Device", "Business-Requires-Materials", etc.
[0016] Each heating service object entity is treated as an entity node in the graph, and each service logic relationship is treated as an entity relationship edge connecting two related entity nodes. Discrete text information is transformed into a networked graph composed of nodes and edges that can clearly express the knowledge structure and business logic of the heating field, namely the heating question-and-answer knowledge graph.
[0017] Each entity node and entity relation edge in the heating question-and-answer knowledge graph is vectorized to generate corresponding entity semantic embedding vectors and relation semantic embedding vectors, which are then associated with the corresponding entity nodes and entity relation edges in the heating question-and-answer knowledge graph. In specific implementation, each entity node and each entity relation edge in the heating question-and-answer knowledge graph is vectorized to obtain an entity semantic embedding vector corresponding to each entity node and a relation semantic embedding vector corresponding to each entity relation edge, thus mapping each node and relation to a continuous, low-dimensional vector space. The generated entity semantic embedding vectors and relation semantic embedding vectors are then associated and stored with the original entity nodes and entity relation edges in the heating question-and-answer knowledge graph.
[0018] In this embodiment, historical heating service records are input into the data preprocessing and graph initialization module of the intelligent customer service system. The module parses the historical heating service records into an XMind format mind map, extracts node text content and hierarchical relationships, and uses node content recognition and hierarchical semantic annotation algorithms to initially define the scope of entity and relationship descriptions. Named Entity Recognition (NER) and Relation Extraction (RE) are performed on the parsed text data to automatically identify semantic relationships between entities, constructing an initial knowledge graph KG=(E,R,S), where E is the entity set, R is the relation set, and S is the triple set. The TransE algorithm is used to embed the initial knowledge graph, generating low-dimensional vector representations of entities and relations, providing data support for subsequent semantic matching and reasoning.
[0019] The semantic associations of the knowledge graph are optimized through the semantic enhancement and knowledge completion modules of the intelligent customer service system. Based on the Graph Attention Network (GAT) structure, high-order semantic features are aggregated for nodes in the knowledge graph, and the calculation formula is as follows: ,in For attention weights, As a trainable weight matrix, structured, context-sensitive graph embeddings are generated through multi-layer propagation. The graph embedding results (entity and relation embedding representations z) are mapped to the semantic space of the larger model through an adapter layer. The mapping process is as follows: This approach aligns structured knowledge with natural language semantics. A large model is trained using an instruction-tuning task. The authenticity of triples (h,r,t) is judged, and a semantic consistency score function f(h,r,t) = ||h + rt|| is calculated. The reliability of candidate knowledge links is evaluated based on the Top-K ranking results. If a candidate link score is higher than a threshold, new knowledge triples are automatically generated and added to the knowledge graph. A joint loss function is defined. ,in To optimize the loss for graph embedding, For large model classification loss, , As negative samples, GNN learns structural embedding and large model learns semantic discrimination through joint training, with both sharing the semantic space of entities and relations.
[0020] S102. Analyze the heating service questions raised by the user, extract key question entities, and identify the user's question intent category. Based on the key question entities and the question intent category, generate an enhanced question semantic vector, and calculate the semantic matching degree between the enhanced question semantic vector and each entity node in the heating question-answering knowledge graph.
[0021] By collecting heating service questions raised by users through user terminals, a pre-trained language encoding model, such as a model based on the Transformer architecture, is invoked to semantically encode the text sequence of heating service questions, generating a question semantic context vector, which is a numerical vector representing the entire question context information.
[0022] The semantic context vector of the question is input into a pre-trained intent classification model, such as a classifier based on a fully connected layer. The intent classification model analyzes the semantic context vector of the question, determines the preset business intent category to which the user's question belongs, such as fault reporting, fee inquiry, business consultation, complaint and suggestion, and outputs the corresponding question intent label.
[0023] Simultaneously or sequentially, named entity recognition algorithms, such as Conditional Random Field (CRF) or deep learning sequence labeling models, are used to process the original heating service issue text. The named entity recognition algorithm analyzes sentences word by word, locating and extracting key issue entities that represent elements of the heating business, including location entities, equipment entities, cost and time composite entities, status or fault entities, etc. All extracted entities are organized into a list of key issue entities.
[0024] Furthermore, by querying the intent tag embedding table, the semantic representation of the intent corresponding to the question intent tag is obtained, and the semantic representation of the question corresponding to each key question entity is obtained through the entity encoding model or its pre-trained embedding. Using a feature fusion tool, the question semantic context vector, the intent semantic representation, and the question semantic representation of all key question entities are integrated and transformed, ultimately outputting an enhanced question semantic vector. This enhanced question semantic vector integrates the original semantics of the user's question, the explicit intent, and specific business elements, providing a high-quality query representation for accurate knowledge retrieval.
[0025] In this embodiment, the user intent recognition module of the intelligent customer service system uses a pre-trained language model based on the Transformer architecture to semantically encode the natural language statements input by the user, capturing the contextual semantic features of the statements. A classification network is then used to classify the encoded semantic features to identify user intent categories, such as policy consultation, business processing, and problem feedback, outputting intent labels. The Named Entity Recognition (NER) algorithm is used to extract key entities from the statements, and combined with the semantic encoding results to generate a contextual semantic vector.
[0026] After generating the enhanced question semantic vector, each entity node in the heating question-answering knowledge graph is traversed. For the currently traversed target entity node, the entity semantic embedding vector corresponding to that node is obtained through the aforementioned associated storage relationship.
[0027] Calculate the semantic similarity between the semantic vector of the augmented question and the semantic embedding vector of the target node entity. Semantic similarity can be calculated using methods such as cosine similarity, inner product, or the reciprocal of Euclidean distance. The calculated similarity value is taken as the semantic matching degree of the target entity node relative to the current user question.
[0028] Repeat the above process of calculating similarity, traversing all entity nodes in the knowledge graph to obtain the semantic matching degree corresponding to each entity node, thus completing the global semantic matching calculation.
[0029] S103. Determine the entity node corresponding to the highest semantic matching degree as the retrieval starting point. Using the retrieval starting point as the initial node, perform a multi-hop traversal along the entity relationship edge in the heating question and answer knowledge graph, collect the end nodes reached by the traversal path as candidate nodes, and form a candidate node set.
[0030] From all entity nodes and their corresponding semantic matching scores, select the entity node with the highest semantic matching score and officially designate that node as the starting point for in-depth exploration of the knowledge graph in this question-answering task.
[0031] Starting from the retrieval origin, a graph traversal process is initiated in the heating question-and-answer knowledge graph to discover potential answer nodes associated with the initial node. From the initial node, along the entity relationship edges connected to it, the process visits its directly adjacent nodes, i.e., the next-hop nodes. Simultaneously, the number of edges from the initial node to the currently visited node is recorded in real-time, serving as the current path depth of the current traversal path.
[0032] When visiting each node, it is determined whether the current path depth has reached the preset maximum traversal depth threshold.
[0033] If the maximum traversal depth threshold is not reached, the currently visited node is set as the new current node, and the process continues from there, visiting the next hop node along the edges connecting the nodes. Each time the path is explored deeper, the current path depth is updated (e.g., incremented by 1), and the updated depth is re-evaluated to see if it has reached the maximum traversal depth threshold. This process is repeated cyclically to achieve in-depth exploration of the path.
[0034] If the maximum traversal depth threshold has been reached, it is determined that the exploration of the current branch path has reached the preset depth limit, and the traversal along this branch is terminated, and the currently visited node is identified as a candidate end node.
[0035] Continue executing the above traversal loop process to ensure that the traversal covers all reachable nodes starting from the retrieval point and whose current path depth does not exceed the maximum traversal depth threshold. Once all reachable nodes meeting the conditions have been visited and processed, terminate the entire traversal process of the heating question-and-answer knowledge graph.
[0036] All candidate endpoints collected during the traversal are aggregated and integrated to form a candidate node set. It should be noted that the candidate node set represents a series of knowledge graph nodes that may contain answer information, which can be reached through relational reasoning with a finite step size, starting from the most semantically relevant starting point.
[0037] In this embodiment, the knowledge retrieval and reasoning module of the intelligent customer service system uses a semantic similarity calculation model, such as the cosine similarity calculation model, to match the contextual semantic vector output by the user intent recognition module with the entity embedding in the knowledge graph, selecting the most relevant node as the query starting point. A path reasoning algorithm, such as a multi-hop reasoning algorithm, is employed to find relationship links related to the target semantics in the knowledge graph, generating a preliminary set of candidate answers. Combining contextual semantics and historical interaction feedback, a ranking network is used to rank the candidate answers by confidence, selecting the optimal knowledge node and its associated relationships.
[0038] S104. Using a reinforcement learning network model, calculate the expected return value of each candidate node in the candidate node set according to the question intent category, filter the target response node, generate the heating service response text based on the question and answer semantic information corresponding to the target response node, and return it to the user terminal.
[0039] Based on the obtained query intent tags, a fixed-dimensional intent semantic vector is generated by querying a predefined intent embedding table or using a lightweight mapping network to numerically represent the user's query intent. Simultaneously, the semantic embedding vector of the candidate entity corresponding to each candidate node in the candidate node set is retrieved one by one from the associated stored data.
[0040] For each candidate node in the candidate node set, the corresponding candidate entity semantic embedding vector and intent semantic vector are concatenated to form a candidate feature vector, thus fusing specific node information with global user intent. The candidate feature vector corresponding to each candidate node is then sequentially input into a pre-trained reinforcement learning network model.
[0041] By performing nonlinear transformations and feature abstraction on candidate feature vectors through one or more fully connected layers within the policy network of a reinforcement learning network model, a high-level feature representation is obtained that characterizes the high-level semantics and decision value of the candidate node under a given intent. Based on this high-level feature representation, a scalar value is calculated and output through the output layer of the policy network; this scalar value represents the expected reward of the candidate node in the current decision state. This process is repeated for all candidate nodes in the set to obtain the expected reward value for each node.
[0042] The candidate node with the highest expected return value is selected and determined as the target response node for this question-and-answer session. Based on the identifier of the target response node, predefined question-and-answer semantic information bound to that node is retrieved from the graph database or its associated text library, such as standard answer text, business processing steps, and fault handling instructions. Simultaneously, the current dialogue history is obtained as context.
[0043] The semantic information of the question and answer from the target response node, the relevant knowledge context, and the current user question text are all input into the natural language generation model to generate a logically complete and fluent natural language text that conforms to the expression standards of the heating industry—that is, the heating service response text. The generated heating service response text is then pushed and displayed on the user's terminal interface that initiated the query via an application programming interface (API) or message queue, completing the service loop for this intelligent response.
[0044] In this embodiment, the intelligent response generation module of the intelligent customer service system generates initial text content based on a language generation model with a Transformer-Decoder architecture, combined with candidate answers output by the knowledge retrieval and reasoning module. Beam Search and Top-k sampling strategies are employed to control the diversity and fluency of the output text, avoiding the generation of repetitive or semantically incoherent content. The tone and expression of the response are automatically adjusted according to the scenario category. The logical completeness of the response is verified by a semantic consistency detection module, ensuring that the response is consistent with the knowledge graph content and semantically coherent with the context. The final response is output after successful verification.
[0045] S105. Collect the interactive feedback data generated by the user in response to the heating service text, calculate the reward signal value based on the interactive feedback data using a reward function, and optimize the network parameters of the reinforcement learning network model based on the reward signal value.
[0046] By integrating functionality into the front-end interface and back-end session tracking module, the system collects real-time user interaction feedback data related to each response. Specific data items collected include: user satisfaction rating (the rating explicitly submitted by the user through the interface); follow-up question count (the number of times the user asks related questions again within the same session regarding the current response); and conversation interruption flag (an identifier used to indicate whether the user ended or abandoned the current session without conducting effective follow-up questions after the response).
[0047] Based on a pre-defined interruption level mapping table, different interruption scenarios are mapped to corresponding specific penalty values. All data are normalized or standardized to ensure they fall within a weighted range. According to pre-defined reward function weights, the user satisfaction score, number of follow-up questions, and penalty value are weighted and summed to obtain the reward signal value for this interaction.
[0048] Furthermore, the set of candidate nodes corresponding to this question-and-answer session and the expected payoff values of each node previously calculated by the policy network are retrieved from historical decision records. Using a normalization function, such as the Softmax function, the original selection probability of each candidate node in its current state is calculated based on the expected payoff values, thus obtaining the original selection probability distribution of the entire set.
[0049] The calculated reward signal value is used as the evaluation criterion for the quality of this decision. Combined with the original selection probability distribution, the policy gradient algorithm is applied to calculate the update direction of the policy network parameters, i.e., the policy gradient.
[0050] Based on the calculated policy gradient, a policy optimization algorithm is used to make a small adjustment to the network parameters of the policy network. By updating the parameters, when encountering similar question-and-answer states in the future, i.e., similar intentions and candidate node features, the policy network can output higher expected reward values for candidate nodes that have the potential to obtain higher reward signal values, thereby guiding the reinforcement learning model to make better response choices.
[0051] In this embodiment, the large model reinforcement learning fine-tuning layer in the reinforcement learning optimization module of the intelligent customer service system uses knowledge graph entity relationship embedding as input and employs a reinforcement learning mechanism to fine-tune the large model. The reward signal comes from the triple authenticity score and the semantic consistency of the generated content. The large model parameters are optimized through a policy gradient algorithm to make the generated knowledge links more consistent with the real graph structure, thereby improving knowledge completion and semantic reasoning capabilities.
[0052] By optimizing the user feedback self-learning layer in the reinforcement learning module of the intelligent customer service system, the interaction process with users is modeled as a reinforcement learning process. This involves defining a state space S (current dialogue semantics and context vector), an action space A (model-generated response content or retrieval strategy), and a reward signal R (user feedback, including satisfaction rating, number of follow-up questions, dialogue interruption markers, etc.). The reward signal is quantified, and a reward function is defined. Where accuracy is the accuracy of the answer matching standard knowledge, relevance is the relevance of the answer to the context semantics, user_feedback is the user feedback score, and w is the weight coefficient. Real negative samples are automatically identified and labeled, and a negative sample penalty term is introduced into the reward function; the policy parameters are updated using the policy gradient method, as shown in the formula: ,in To improve the learning rate, we reinforce correct response paths and suppress incorrect patterns.
[0053] This application constructs and embeds a semantic heating question-and-answer knowledge graph to deeply analyze the intent and key entities in user questions and generate enhanced semantic query vectors. Then, it performs accurate matching and multi-hop reasoning in the graph, which enables the accurate identification of the answer nodes and logical links most relevant to complex, implicit, or composite heating consultations from a structured professional knowledge network. This results in the generation of accurate and logically complete response texts, overcoming the shortcomings of traditional methods in terms of professionalism and accuracy.
[0054] Furthermore, by introducing a reinforcement learning mechanism, real-time user satisfaction ratings and follow-up questions are quantified into reward signals, which are then used to dynamically optimize the decision-making model and continuously supplement the knowledge graph. This allows the system to break free from dependence on fixed rules, learn autonomously from real-world interactive experiences, and continuously optimize its answer selection strategies and knowledge organization structure. Consequently, it maintains and improves the accuracy of responses and user satisfaction when facing policy changes, new business scenarios, or seasonal service peaks, achieving autonomous iteration and long-term improvement in service quality.
[0055] like Figure 2 As shown in the embodiments of this application, a knowledge graph-based customer service Q&A device for heating services is also proposed, including: At least one processor; and, A memory communicatively connected to the at least one processor; wherein, The memory stores instructions that can be executed by the at least one processor, which, when executed by the at least one processor, enables the at least one processor to perform a knowledge graph-based customer service Q&A method for heating services, as described in any of the above embodiments.
[0056] This application also provides a non-volatile computer storage medium storing computer-executable instructions, wherein the computer-executable instructions are configured as: a knowledge graph-based customer service question-and-answer method for heating services as described in any of the above embodiments.
[0057] The various embodiments in this application are described in a progressive manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the device and medium embodiments are basically similar to the method embodiments, so the description is relatively simple; relevant parts can be referred to the description of the method embodiments.
[0058] The devices and media provided in this application are one-to-one with the methods. Therefore, the devices and media also have similar beneficial technical effects as their corresponding methods. Since the beneficial technical effects of the methods have been described in detail above, the beneficial technical effects of the devices and media will not be repeated here.
[0059] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.
[0060] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.
[0061] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.
[0062] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.
[0063] In a typical configuration, a computing device includes one or more processors (CPU), input / output interfaces, network interfaces, and memory.
[0064] Memory may include non-persistent storage in computer-readable media, such as random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash RAM. Memory is an example of computer-readable media.
[0065] Computer-readable media include both permanent and non-permanent, removable and non-removable media that can store information using any method or technology. Information can be computer-readable instructions, data structures, modules of programs, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic disk storage or other magnetic storage devices, or any other non-transferable medium that can be used to store information accessible by a computing device. As defined herein, computer-readable media does not include transient computer-readable media, such as modulated data signals and carrier waves.
[0066] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.
[0067] The above description is merely an embodiment of this application and is not intended to limit the scope of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of the claims of this application.
Claims
1. A knowledge graph-based question-and-answer method for heating service customer service, characterized in that, include: Based on historical heating service records, a heating question-and-answer knowledge graph is constructed, and semantic embedding is performed on the heating question-and-answer knowledge graph. The system analyzes user-submitted heating service questions, extracts key question entities, identifies the user's question intent category, generates an enhanced question semantic vector based on the key question entities and the question intent category, and calculates the semantic matching degree between the enhanced question semantic vector and each entity node in the heating question-answering knowledge graph. The entity node corresponding to the highest semantic matching degree is determined as the retrieval starting point. Using the retrieval starting point as the initial node, a multi-hop traversal is performed along the entity relationship edge in the heating question and answer knowledge graph. The end nodes reached by the traversal path are collected as candidate nodes to form a candidate node set. Using a reinforcement learning network model, the expected revenue value of each candidate node in the candidate node set is calculated according to the question intent category. Target response nodes are then selected. Based on the question-and-answer semantic information corresponding to the target response nodes, heating service response text is generated and returned to the user terminal. The system collects interactive feedback data generated by the user in response to the heating service text, calculates a reward signal value based on the interactive feedback data using a reward function, and optimizes the network parameters of the reinforcement learning network model based on the reward signal value.
2. The knowledge graph-based customer service question-and-answer method for heating services according to claim 1, characterized in that, The process of constructing a heating question-and-answer knowledge graph based on historical heating service records and semantically embedding the knowledge graph specifically includes: Obtain historical heating service records; these historical service records include historical work order texts, historical customer service dialogue logs, and heating policy documents. The historical heating service records are parsed to identify and extract heating service object entities and the corresponding service logic relationships between entities. A heating question-and-answer knowledge graph is constructed with the heating service object entities as entity nodes and the service logic relationships as entity relationship edges. Each entity node and each entity relationship edge in the heating question-and-answer knowledge graph are vectorized to generate corresponding entity semantic embedding vectors and relationship semantic embedding vectors, which are then associated with the corresponding entity nodes and entity relationship edges in the heating question-and-answer knowledge graph.
3. The knowledge graph-based customer service question-and-answer method for heating services according to claim 2, characterized in that, The process involves parsing user-submitted heating service questions, extracting key question entities, identifying the user's question intent category, and generating an enhanced question semantic vector based on the key question entities and the question intent category. Specifically, this includes: The system collects heating service questions raised by users through user terminals, performs semantic encoding on the heating service questions, and generates question semantic context vectors. Based on the question semantic context vector, the user's question intent category is identified through an intent classification model, and the corresponding question intent label is output. And by using a named entity recognition algorithm, the key problem entities that characterize the elements of heating services in the heating service problem are extracted; An enhanced question semantic vector is generated based on the intent semantic representation of the question intent tag and the question semantic representation of the key question entity.
4. The knowledge graph-based customer service question-and-answer method for heating services according to claim 3, characterized in that, The calculation of the semantic matching degree between the enhanced question semantic vector and each entity node in the heating question-answering knowledge graph specifically includes: For each entity node in the heating question-and-answer knowledge graph, obtain the corresponding entity semantic embedding vector; Calculate the semantic similarity between the augmented question semantic vector and the entity semantic embedding vector to obtain the semantic matching degree corresponding to each entity node.
5. A knowledge graph-based customer service question-and-answer method for heating services according to claim 3, characterized in that, The step of using a reinforcement learning network model to calculate the expected reward value of each candidate node in the candidate node set based on the question intent category, and then selecting the target response node, specifically includes: Based on the intent semantic representation of the question intent tag, an intent semantic vector is generated, and the candidate entity semantic embedding vector corresponding to each candidate node in the candidate node set is obtained. For each candidate node, the intent semantic vector and the candidate entity semantic embedding vector are fused to obtain the corresponding candidate feature vector; The candidate feature vector is input into a pre-trained reinforcement learning network model. The candidate feature vector is then subjected to a nonlinear transformation through the fully connected layer of the policy network in the reinforcement learning network model to obtain the high-level feature representation of the candidate node. Based on the high-level feature representation, the expected return value corresponding to the candidate node is calculated through the output layer of the policy network, and the candidate node corresponding to the highest value among the expected return values is selected as the target response node.
6. The knowledge graph-based customer service question-and-answer method for heating services according to claim 1, characterized in that, The process involves using the retrieval starting point as the initial node, performing a multi-hop traversal along the entity relationship edges in the heating question-and-answer knowledge graph, and collecting the terminal nodes reached by the traversal path as candidate nodes to form a candidate node set. Specifically, this includes: Using the retrieval starting point as the initial node, in the heating question and answer knowledge graph, the next hop node is accessed along the entity relationship edge connected to the initial node; Record the current path depth of the access traversal path in real time, determine whether the current path depth has reached the preset maximum traversal depth threshold, and start the traversal loop. If the maximum traversal depth threshold is not reached, then the next hop node is taken as the current node, and the new next hop node is visited along the entity relationship edge connected to the current node. The current path depth of the traversal path is updated, and it is re-determined whether the updated current path depth has reached the maximum traversal depth threshold. If the maximum traversal depth threshold has been reached, the traversal of the currently visited branch path is determined to be terminated, and the currently visited node is collected as a candidate end node. After all reachable nodes have been traversed, the traversal loop is terminated, and all candidate end nodes are counted to form a candidate node set. The reachable node is an entity node in the heating question-and-answer knowledge graph that takes the retrieval starting point as the initial node and whose current path depth has not reached the maximum traversal depth threshold.
7. The knowledge graph-based customer service question-and-answer method for heating services according to claim 1, characterized in that, The mutual feedback data includes user satisfaction ratings, number of follow-up questions, and dialogue interruption indicators. The step of calculating the reward signal value based on the interaction feedback data using a reward function specifically includes: Based on the preset interruption level mapping relationship, determine the penalty value corresponding to the dialogue interruption identifier; Based on the preset reward function weights, the user satisfaction score, the number of follow-up questions, and the penalty value are weighted to generate a reward signal value.
8. A knowledge graph-based customer service question-and-answer method for heating services according to claim 7, characterized in that, The optimization of the network parameters of the reinforcement learning network model based on the reward signal value specifically includes: Based on the expected return value of each candidate node in the candidate node set, the original selection probability of the corresponding candidate node is generated through a normalization function, thus obtaining the original selection probability distribution of the candidate node set. Based on the reward signal value and the original selection probability distribution, calculate the policy gradient of the policy network in the reinforcement learning network model; Based on the policy gradient, the network parameters of the policy network are adjusted by a policy optimization algorithm to increase the expected return value of candidate nodes in similar states.
9. A knowledge graph-based customer service Q&A device for heating services, characterized in that, include: At least one processor; as well as, A memory communicatively connected to the at least one processor; wherein, The memory stores instructions that can be executed by the at least one processor, which, when executed by the at least one processor, enables the at least one processor to perform a knowledge graph-based customer service Q&A method for heating services as described in any one of claims 1 to 8.
10. A non-volatile computer storage medium storing computer-executable instructions, characterized in that, The computer-executable instructions are configured to execute a knowledge graph-based customer service Q&A method for heating services as described in any one of claims 1 to 8.