Theme extraction method and system for user question text

Abstract

The invention provides a topic extraction method and system for a user question text. The method comprises the following steps: identifying a core intention and an association concept based on a received question text submitted by a user; carrying out mapping processing on a pre-constructed knowledge graph, and determining a theme extension range according to semantic relevance between nodes in the knowledge graph and an implicit relation between the nodes learned through a graph neural network; an adaptive weight distribution mechanism is combined with a link analysis algorithm, and the importance of the nodes in the theme extension range and the centrality of the nodes in the graph neural network are evaluated so as to adjust the weight of each node; according to the connection strength between the nodes obtained through point mutual information calculation, structured description information representing a subject corresponding to the question text is generated; and further extracting and processing the structured description information by using a multi-layer attention mechanism to obtain an extraction result. According to the technical scheme provided by the invention, the efficiency and accuracy of extracting the user question text theme can be improved.

Description

Technical Field

[0001] The embodiments of the present application relate to the technical field of topic extraction, and in particular, to a method and system for extracting topics from user question texts. Background Art

[0002] In today's era of information explosion, users submit question texts through various online platforms to obtain the information or solutions they need. For example, in the fields of customer service, intelligent question-answering systems, search engine optimization, etc., it is crucial to quickly and accurately understand the core intentions and related concepts of user questions.

[0003] At present, common methods for extracting topics from user question texts mainly include rule-based methods, traditional machine learning methods, and deep learning methods. Specific approaches include: rule-based methods rely on predefined grammatical structures and vocabularies; traditional machine learning methods use feature engineering and classification algorithms for topic classification; and deep learning methods such as convolutional neural networks and recurrent neural networks and their variants perform topic modeling by automatically learning text representations. In addition, some studies have also tried to combine knowledge graphs to enhance the ability to understand text.

[0004] Although existing solutions can achieve topic extraction from user question texts to a certain extent, they still have some limitations: rule-based methods and traditional machine learning methods are highly dependent on manually designed rules and feature engineering, and are difficult to adapt to changing language patterns and complex semantic relationships; although many existing deep learning methods can capture the surface features of text, they are not deep enough in deep semantic understanding and implicit relationship mining, especially when dealing with polysemous words and complex contexts; existing solutions usually adopt a fixed weight allocation strategy and cannot dynamically adjust the importance of nodes according to specific problems, resulting in unreasonable resource allocation in some cases, affecting the effect and efficiency of topic extraction. Summary of the invention

[0005] The embodiments of the present application provide a method and system for extracting topics from user question texts, so as to solve the problems of low efficiency and poor accuracy in extracting topics from user question texts in the prior art.

[0006] In a first aspect, an embodiment of the present application provides a method for extracting a topic from a user question text, comprising: Based on the received question text submitted by the user, the question text is parsed to identify the core intent and related concepts; According to the core intent and related concepts, mapping processing is performed on a pre-constructed knowledge graph, and the scope of topic expansion is determined based on the semantic relevance between nodes in the knowledge graph and the implicit relationship between the nodes learned through the graph neural network; Based on the topic extension scope, an adaptive weight allocation mechanism combined with a link analysis algorithm is used to evaluate the importance of nodes within the topic extension scope and the centrality of the nodes in the graph neural network to adjust the weight of each node; Generate structured description information representing the topic corresponding to the question text according to the adjusted node weights and the connection strength between the nodes calculated by point mutual information; The structured description information is further refined using a multi-layer attention mechanism to obtain an extraction result, where the extraction result refers to a summary output result of the topic corresponding to the question text.

[0007] Optionally, according to the adjusted node weights and the connection strength between the nodes calculated by point mutual information, generating structured description information representing the topic corresponding to the question text includes: Based on the adjusted node weights, the nodes in the knowledge graph are prioritized to obtain core topic nodes; Using point mutual information, the semantic association strength between the core topic nodes is quantified to determine the closeness between the core topic nodes; According to the closeness between the core topic nodes, a clustering algorithm or a community discovery algorithm is used to group the core topic nodes to form a topic cluster; Comprehensively analyzing and processing the core topic nodes in the topic cluster to extract topic tags that can summarize the core meaning of the topic cluster; Combining the explicit concepts in the question text with the topic labels, supplementing potential implicit topics through a logical reasoning mechanism, and generating topic content corresponding to the implicit topics; Based on the subject content, the explicit concepts, the subject tags and the implicit subjects are organized to generate structured description information representing the subject corresponding to the question text.

[0008] Optionally, combining the explicit concepts in the question text with the topic labels, supplementing potential implicit topics through a logical reasoning mechanism, and generating topic content corresponding to the implicit topics, including: Constructing a preliminary topic framework based on the explicit concepts in the question text and the topic labels; Using the associations and context information between nodes in the knowledge graph, the subject framework is expanded to identify additional entities or concepts related to the explicit concept and the subject label; Based on rule-based reasoning or machine learning models, the relationships between the additional entities or concepts are analyzed and processed to mine implicit topics; Based on the implicit topic, the explicit concept and the topic label, consistency processing is performed through semantic fusion technology, and the explicit concept, the topic label and the implicit topic are integrated to generate the topic content corresponding to the implicit topic.

[0009] Optionally, based on the topic extension scope, an adaptive weight allocation mechanism combined with a link analysis algorithm is used to evaluate the importance of nodes within the topic extension scope and the centrality of the nodes in the graph neural network to adjust the weight of each node, including: Based on the topic extension scope, the basic weights of the nodes within the topic extension scope are initialized using an adaptive weight allocation mechanism, and the local influence of the basic weight of each node within the topic extension scope is calculated based on the semantic relevance between the nodes in the knowledge graph and the implicit relationship between the nodes learned by the graph neural network to obtain a local influence result; Using the link analysis algorithm, iteratively calculate the scores of the nodes to obtain importance scores; The importance of the nodes is comprehensively evaluated in combination with the local influence result and the importance score, so as to dynamically adjust the weight of each node so that the node obtains a higher weight.

[0010] Optionally, combining the influence result and the importance score, as well as the centrality of the node in the graph neural network, the importance of the node is comprehensively evaluated to dynamically adjust the weight of each node so that the node obtains a higher weight, including: Using the local influence result, the basic influence score of each node is calculated to obtain a basic influence score; Based on the basic influence score and the importance score, combined with a preset functional relationship, the comprehensive evaluation score of the node is calculated to obtain a comprehensive evaluation score reflecting the importance of the node within the scope of the topic extension; Using the comprehensive evaluation score, all nodes are sorted, and a threshold or ranking interval is set to determine nodes with high importance; A gain function is applied to increase the weight of the high-importance node so that the high-importance node obtains a higher weight.

[0011] Optionally, mapping processing is performed on a pre-constructed knowledge graph according to the core intent and related concepts, and the scope of topic expansion is determined according to the semantic relevance between nodes in the knowledge graph and the implicit relationship between the nodes learned through the graph neural network, including: Using a bidirectional encoder representation model and a semantic matching algorithm, based on the core intent and the associated concepts, the concepts in the question text are mapped in a pre-constructed knowledge graph to obtain a topic mapping result; According to the topic mapping results, combined with the semantic relevance between nodes in the knowledge graph, multi-step paths starting from the mapped nodes are explored and processed to identify potential nodes related to the core intent, so as to construct a graph neural network; The graph neural network is combined with a self-attention mechanism to analyze and process the implicit relationship between the potential nodes, and a context-aware model and Bayesian optimization technology are introduced to evaluate the relevance and importance of the potential nodes to the question text, and the relevance and importance scores of the potential nodes are dynamically adjusted to obtain adjustment results; According to the adjustment result, nodes that can be included in the subject extension scope are determined from the potential nodes, and the nodes in the subject extension scope are grouped in combination with a hierarchical clustering algorithm to determine the subject extension scope.

[0012] Optionally, the structured description information is further refined by using a multi-layer attention mechanism to obtain an extraction result, including: Using the constructed multi-layer attention network model, encoding the obtained structured description information to obtain the structured description information after encoding; Based on the structured description information after the encoding process, the structured description information after the encoding process is preliminarily filtered and condensed by using the first-layer attention mechanism, the elements directly related to the core intent are identified and strengthened, and the irrelevant structured description information after the encoding process is weakened or ignored, so as to obtain a preliminary extraction result; According to the preliminary extraction result, in each subsequent layer of the attention mechanism, based on the output of the previous layer of the attention mechanism, the structured description information is subjected to a layer-by-layer in-depth analysis and processing to obtain the multi-layer attention mechanism screening result; Based on the multi-layer attention mechanism screening results, the output of the attention mechanism of the final layer is processed to obtain the extraction result.

[0013] In a second aspect, an embodiment of the present application provides a topic extraction system for a user question text, comprising: A parsing module, based on the received question text submitted by the user, parses the question text to identify the core intent and related concepts; A mapping module performs mapping processing on a pre-constructed knowledge graph according to the core intent and related concepts, and determines the scope of topic expansion according to the semantic relevance between nodes in the knowledge graph and the implicit relationship between the nodes learned through the graph neural network; An adjustment module, based on the topic extension scope, adopts an adaptive weight allocation mechanism combined with a link analysis algorithm to evaluate the importance of nodes within the topic extension scope and the centrality of the nodes in the graph neural network to adjust the weight of each node; A generation module, generating structured description information representing a topic corresponding to the question text according to the adjusted node weights and the connection strength between the nodes calculated by point mutual information; The processing module further refines the structured description information using a multi-layer attention mechanism to obtain an extraction result, where the extraction result refers to a summary output result of the topic corresponding to the question text.

[0014] In an embodiment of the present application, based on the question text received and submitted by the user, the question text is parsed to identify the core intent and related concepts; according to the core intent and related concepts, mapping is performed on a pre-constructed knowledge graph, and the topic extension range is determined based on the semantic relevance between nodes in the knowledge graph and the implicit relationship between the nodes learned through the graph neural network; based on the topic extension range, an adaptive weight allocation mechanism is combined with a link analysis algorithm to evaluate the importance of nodes within the topic extension range and the centrality of the nodes in the graph neural network to adjust the weight of each node; according to the adjusted node weights and the connection strength between the nodes calculated by point mutual information, structured description information representing the topic corresponding to the question text is generated; the structured description information is further refined using a multi-layer attention mechanism to obtain an extraction result, which refers to the summary output result of the topic corresponding to the question text.

[0015] The technical solution of this application has the following beneficial effects: This application can effectively identify the core intent and related concepts in the question text by parsing the question text submitted by the received user. It ensures a more accurate and comprehensive understanding of the user's questions, providing a solid foundation for subsequent steps. Using the pre-built knowledge graph for mapping processing, it can not only capture the explicit semantic relevance, but also learn the implicit relationship between nodes through the graph neural network, so as to determine a wider and more accurate range of topic extension. By combining the adaptive weight allocation mechanism and the link analysis algorithm, the importance and centrality of the nodes within the topic extension range are evaluated and processed, and the weight of each node is adjusted. By quantifying the importance of the nodes, the model can focus on the most critical information points, optimize resource allocation, and improve information processing efficiency. By generating structured description information representing the corresponding topic of the question text based on the adjusted node weights and the connection strength between nodes calculated by point mutual information, it is ensured that the scattered information is integrated into a coherent whole, which is convenient for further analysis and refinement, while retaining the core intent and related concepts of the original text.

[0016] Furthermore, by prioritizing the nodes in the knowledge graph, the core topic nodes can be accurately identified, ensuring that the model focuses on the most important and relevant nodes, thereby improving the accuracy of topic extraction. By quantifying the semantic association strength between core topic nodes through point mutual information, the implicit relationship between nodes can be more deeply explored, enhancing the ability to understand complex contexts. By determining the closeness between core topic nodes, the generated topic description information is ensured to have higher coherence and logical consistency. The core topic nodes are grouped through clustering algorithms or community discovery algorithms to form topic clusters. Ensure that scattered information is integrated into multiple meaningful topic sets for subsequent analysis. By organizing explicit concepts, topic tags, and implicit topics, structured description information is generated, providing users with a comprehensive and detailed perspective, ensuring that users can quickly grasp the core points of the problem.

[0017] Furthermore, through the explicit concepts and topic tags in the question text, a preliminary topic framework is constructed, which provides a clear direction and foundation for the subsequent expansion processing and ensures the pertinence and effectiveness of the subsequent steps. By identifying additional entities or concepts, the model can capture more information implicit in the question text, further improving the ability to understand user intent. By analyzing the relationship between additional entities or concepts, it ensures that the mined implicit topics are highly relevant to the original question text, avoids the introduction of irrelevant information, and improves the quality of topic extraction. Through the consistency processing of implicit topics and explicit concepts and topic tags through semantic fusion technology, it ensures the semantic consistency and logical coherence of the generated topic content.

[0018] These and other aspects of the present application will become more clearly understood in the description of the following embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, a brief introduction will be given below to the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0020] Figure 1 A flowchart of a method for extracting topics from user question text provided in an embodiment of the present application; Figure 2 A schematic diagram of the structure of a topic extraction system for user question text provided in an embodiment of the present application; Figure 3 A schematic diagram of the structure of a computing device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0021] In order to enable those skilled in the art to better understand the solution of the present application, the technical solution in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application.

[0022] In some of the processes described in the specification and claims of this application and the above-mentioned figures, multiple operations that appear in a specific order are included, but it should be clearly understood that these operations may not be executed in the order in which they appear in this article or executed in parallel. The serial numbers of the operations, such as 101, 102, etc., are only used to distinguish between different operations, and the serial numbers themselves do not represent any execution order. In addition, these processes may include more or fewer operations, and these operations may be executed in sequence or in parallel. It should be noted that the descriptions of "first", "second", etc. in this article are used to distinguish different messages, devices, modules, etc., do not represent the order of precedence, and do not limit the "first" and "second" to be different types.

[0023] The following will be combined with the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative work are within the scope of protection of this application.

[0024] Figure 1 A flowchart of a method for extracting a topic from a user question text is provided for an embodiment of the present application. Figure 1 As shown, the method includes: 101. Based on the received question text submitted by the user, the question text is parsed to identify the core intent and related concepts; In this step, the question text submitted by the user is parsed to identify the core intent and related concepts. Parsing includes natural language processing technologies such as word segmentation, part-of-speech tagging, named entity recognition, syntactic analysis, etc., which are used to decompose the question text into structured data. These data not only contain vocabulary information, but also cover grammatical structure and semantic features, which are the basis for understanding user intent. The core intent refers to the main purpose or need of the user's question, while the related concepts refer to other important information points closely related to the core intent.

[0025] In the application embodiment, in the customer service system, when a user asks "How do I check the status of my order", the system first parses the question text into "how", "query", "I", "of", "order", and "status". Through part-of-speech tagging and named entity recognition, "order" is marked as a key entity, and "query" is regarded as an action word that expresses the user's intention. Syntactic analysis further confirms that "query" is the main action and "order status" is the target object. In the end, the system recognizes that the user's core intention is "query the order status", and related concepts include "order number", "logistics information", etc. This parsing result provides a clear direction for subsequent steps, ensuring that the system can accurately map to relevant nodes in the knowledge graph.

[0026] 102. According to the core intent and related concepts, mapping processing is performed on a pre-constructed knowledge graph, and the scope of topic expansion is determined according to the semantic relevance between nodes in the knowledge graph and the implicit relationship between the nodes learned through the graph neural network; In this step, mapping processing is performed on the pre-built knowledge graph, and the scope of topic expansion is determined based on the semantic relevance between nodes in the knowledge graph and the implicit relationship between nodes learned through the graph neural network. The knowledge graph is a semantic network composed of entity nodes and the relationship edges between them, which can represent complex knowledge structures. Mapping processing refers to mapping the parsed text elements to the corresponding nodes in the knowledge graph, thereby establishing a connection between the text and the knowledge base. Topic expansion is based on these mapping results, using the deep-level associations between nodes captured by the graph neural network to expand the scope of understanding of the original problem.

[0027] In the application embodiment, the customer service system maps "order status" to the "order management" field in the knowledge graph, and finds that the related nodes are "order details", "payment status", "delivery progress", etc. Through the learning of the graph neural network, the system recognizes that "logistics information" and "return policy" are also important related concepts. Therefore, the scope of subject expansion is not limited to the query of "order status", but also includes other factors that may affect order processing, such as logistics delays, return procedures, etc. This expansion ensures that the system can provide comprehensive and targeted answers to meet the diverse needs of users.

[0028] 103. Based on the topic extension scope, an adaptive weight allocation mechanism combined with a link analysis algorithm is used to evaluate the importance of nodes within the topic extension scope and the centrality of the nodes in the graph neural network to adjust the weight of each node; In this step, an adaptive weight allocation mechanism combined with a link analysis algorithm is used to evaluate the importance and centrality of nodes within the topic extension range and adjust the weight of each node. Adaptive weight allocation dynamically adjusts the weight of a node based on its position and connection in the knowledge graph to ensure efficient allocation of resources. The link analysis algorithm is used to quantify the strength of connections between nodes and evaluate the relative importance of nodes in the network. This process aims to highlight the nodes that are most critical to understanding and solving user problems while optimizing overall processing efficiency.

[0029] In the application embodiment, in the customer service case, the system calculated that the "order status" node has a high centrality because it is closely connected to multiple other important nodes (such as "order details" and "payment status"). Through link analysis, it was found that there is a strong connection between "logistics information" and "order status", indicating that the two concepts are closely related. Therefore, the system gives "order status" and "logistics information" higher weights to ensure that they occupy a more important position in subsequent topic extraction. At the same time, the system also appropriately reduces the weight of "return policy" because although it is also an associated concept, it is not the primary focus in the current query. This step ensures the effective allocation of resources and improves processing efficiency.

[0030] 104. Generate structured description information representing a topic corresponding to the question text according to the adjusted node weights and the connection strength between the nodes calculated by point mutual information; In this step, structured description information representing the topic corresponding to the question text is generated based on the adjusted node weights and the connection strength between nodes calculated by point mutual information. Structured description information is a high-level summary of the nodes and their relationships within the extended scope of the topic, aiming to provide a clear and coherent topic framework. Point mutual information is used to measure the probability of co-occurrence of two nodes, thereby quantifying the strength of the semantic association between them, ensuring that the generated information is both concise and complete.

[0031] In the application embodiment, the customer service system selected "order status" and "logistics information" as core nodes, and constructed structured description information based on the high connection strength between them. This information not only outlines the order status that users care about, but also details the logistics progress, providing a complete solution. In addition, the system also supplements the "return policy" as additional information, because although its weight is slightly lower, it may also be the focus of user attention in some cases, enhancing the comprehensiveness and practicality of the answer. This structured description information provides a solid foundation for the subsequent multi-layer attention mechanism refinement, ensuring the integrity and logic of the information.

[0032] 105. The structured description information is further refined by using a multi-layer attention mechanism to obtain an extraction result, where the extraction result refers to a summary output result of the topic corresponding to the question text.

[0033] In this step, the structured description information is further refined using a multi-layer attention mechanism to obtain the final summary output result. The multi-layer attention mechanism allows the model to understand information from different levels and angles, focusing on the most critical parts, thereby generating a highly summarized topic summary that closely revolves around the core intent. This mechanism improves the model's adaptability to different types of input data, ensuring that the output results are both accurate and easy to understand.

[0034] In the application embodiment, in the customer service scenario, the system uses a multi-layer attention mechanism to refine the previously generated structured description information. The first layer of attention emphasizes the two core concepts of "order status" and "logistics information", and the second layer delves into the specific connections between them, such as "reasons for logistics delays" and "estimated delivery time". In the end, the system generates a concise and clear summary: "Your order has been shipped and is currently in transit. It is expected to arrive within the next three days. If you have any questions, please refer to our return policy." This summary not only answers the user's questions, but also provides additional useful information, greatly improving the user experience. Through the continuous implementation of five steps, the system realizes the automation of the entire process from question parsing to final answer generation, ensuring efficient and accurate service response.

[0035] In summary, through the above steps 101-105, the present invention can achieve efficient and accurate topic extraction and content generation in the fields of customer service, and provide users with high-quality help and support. Each step not only works independently, but also closely connects to form a complete solution chain, ensuring seamless connection from problem understanding to answer provision.

[0036] Optionally, in step 104, structured description information representing the topic corresponding to the question text is generated according to the adjusted node weights and the connection strength between the nodes calculated by point mutual information, including: Based on the adjusted node weights, the nodes in the knowledge graph are prioritized to obtain core topic nodes; the semantic association strength between the core topic nodes is quantified using point mutual information to determine the closeness between the core topic nodes; according to the closeness between the core topic nodes, a clustering algorithm or a community discovery algorithm is used to group the core topic nodes to form a topic cluster; a comprehensive analysis is performed on the core topic nodes in the topic cluster to extract topic labels that can summarize the core meaning of the topic cluster; the explicit concepts in the question text and the topic labels are combined to supplement potential implicit topics through a logical reasoning mechanism, and the topic content corresponding to the implicit topic is generated; based on the topic content, the explicit concepts, the topic labels and the implicit topics are organized to generate structured description information representing the topic corresponding to the question text.

[0037] Optionally, combining the explicit concepts in the question text with the topic labels, supplementing potential implicit topics through a logical reasoning mechanism, and generating topic content corresponding to the implicit topics, including: Based on the explicit concepts and the topic labels in the question text, a preliminary topic framework is constructed; the topic framework is expanded by utilizing the correlation between nodes in the knowledge graph and contextual information, and additional entities or concepts related to the explicit concepts and the topic labels are identified; based on rule-based reasoning or machine learning models, the relationships between the additional entities or concepts are analyzed and processed to mine implicit topics; based on the implicit topics, the explicit concepts and the topic labels, consistency processing is performed through semantic fusion technology, and the explicit concepts, the topic labels and the implicit topics are integrated to generate topic content corresponding to the implicit topics.

[0038] In this step, in the knowledge graph, each node represents an entity or concept. Through the adjusted node weights, the system can identify which nodes are most critical to the current problem. Prioritization processing is to sort the nodes according to these weights to filter out the most core topic nodes. It ensures that subsequent analysis focuses on the most important information, improving processing efficiency and accuracy. Point mutual information is a statistical method to measure the probability of co-occurrence of two nodes, which is used to quantify the strength of semantic association between them. Clustering algorithms or community discovery algorithms are used to group core topic nodes with similar characteristics to form multiple topic clusters. Each cluster represents a group of highly related concepts, which helps to understand the topic of the question text from different perspectives. This grouping method not only simplifies the complex information structure, but also facilitates subsequent comprehensive analysis and refinement. The system integrates explicit concepts, topic labels, and implicit topics to generate a comprehensive and coherent structured description information, ensuring that the output results cover both the core content of the original question text and the extended background information, providing a more complete and in-depth answer.

[0039] In an embodiment of the present application, based on the adjusted node weights, the system prioritizes the nodes in the knowledge graph and selects the most important batch of nodes as core topic nodes. Using point mutual information, the system evaluates the strength of semantic associations between core topic nodes and determines the closeness between them. Using a clustering algorithm or a community discovery algorithm, the system groups the core topic nodes to form multiple topic clusters, each cluster representing a group of highly related concepts. A comprehensive analysis is performed on the core topic nodes within each topic cluster to extract topic labels that can summarize the core meaning of the cluster. Combining the explicit concepts and topic labels in the question text, potential implicit topics are mined through a logical reasoning mechanism, and corresponding topic content is generated. Explicit concepts, topic labels, and implicit topics are integrated together to generate a comprehensive and coherent structured description information for users.

[0040] The system first identifies "order status", "logistics information", "shipping" and other nodes as core topic nodes based on the adjusted node weights. Through point mutual information calculation, there is a very strong semantic association between "order status" and "logistics information", indicating that the two concepts are closely related. The system uses a clustering algorithm to classify nodes such as "order status", "logistics information", and "shipping" into the same topic cluster, named "order transportation". At the same time, another topic cluster "after-sales service" is identified, including nodes such as "return policy" and "customer service contact information". A comprehensive analysis of the "order transportation" cluster is performed to extract the topic label "logistics delay". This label accurately summarizes the main issue that users are concerned about, that is, the reason why logistics information is not updated in time. The system generates a structured description message: "Your order has been shipped, but the logistics information has not been updated. The logistics delay may be due to improper selection of the logistics company or the impact of holidays. It is recommended that you contact the express company to inquire about the specific reasons, or check the latest logistics progress on our website. If you have any further questions, please refer to our after-sales service policy." Optionally, in step 103, based on the topic extension scope, an adaptive weight allocation mechanism combined with a link analysis algorithm is used to evaluate the importance of nodes within the topic extension scope and the centrality of the nodes in the graph neural network to adjust the weight of each node, including: Based on the topic extension scope, the basic weights of the nodes within the topic extension scope are initialized using an adaptive weight allocation mechanism, and the local influence of the basic weight of each node within the topic extension scope is calculated based on the semantic relevance between nodes in the knowledge graph and the implicit relationship between nodes learned through the graph neural network to obtain a local influence result; the link analysis algorithm is used to iteratively calculate the score of the node to obtain an importance score; the importance of the node is comprehensively evaluated by combining the local influence result and the importance score, as well as the centrality of the node in the graph neural network, so as to dynamically adjust the weight of each node so that the node obtains a higher weight.

[0041] Optionally, the importance of the nodes is comprehensively evaluated in combination with the influence result and the importance score to dynamically adjust the weight of each node so that the node obtains a higher weight, including: using the local influence result to calculate the basic influence score of each node to obtain a basic influence score; based on the basic influence score and the importance score, combined with a preset functional relationship, the comprehensive evaluation score of the node is calculated to obtain a comprehensive evaluation score that reflects the importance of the node within the scope of the topic extension; using the comprehensive evaluation score to sort all nodes, and set a threshold or ranking range to determine high-importance nodes; applying a gain function to increase the weight of the high-importance nodes so that the high-importance nodes obtain a higher weight.

[0042] In this step, the adaptive weight allocation mechanism is a method of dynamically adjusting node weights, aiming to optimize resource allocation according to the position and connection of nodes in the knowledge graph. Local influence refers to the degree of influence of a node within the scope of a specific topic extension. By combining the semantic relevance between nodes in the knowledge graph and the implicit relationship learned by the graph neural network, the system can quantify the local influence of each node within the scope, making the evaluation more comprehensive and accurate. Comprehensive evaluation is the process of combining the local influence results with the importance score to comprehensively evaluate the importance of the node. Through the preset functional relationship, the system calculates the comprehensive evaluation score of each node, and then determines its actual importance within the scope of the topic extension. Dynamic weight adjustment is to increase the weight of high-importance nodes based on the comprehensive evaluation results to ensure that they occupy a more important position in subsequent processing. The gain function is used to further increase the weight of high-importance nodes so that they receive greater attention in the overall structure. The design of the gain function can be adjusted according to the specific application scenario to achieve the best resource allocation effect.

[0043] In an embodiment of the present application, the system first assigns a basic weight to each node. This weight is based on the basic properties of the node and serves as the basis for subsequent calculations. Through the semantic relevance between nodes in the knowledge graph and the implicit relationship learned by the graph neural network, the system calculates the local influence of each node within the scope of the topic extension to ensure the comprehensiveness of the evaluation. Combining the local influence results and importance scores, the system conducts a comprehensive assessment of the importance of the nodes. Through a preset functional relationship, the comprehensive evaluation score of each node is calculated to reflect its actual importance within the scope of the topic extension. The system applies a gain function to enhance the weights of high-importance nodes. In this way, it is ensured that these nodes occupy a more important position in the subsequent topic extraction and content generation process, and the response efficiency and accuracy of the model are improved.

[0044] Suppose a user asks: "My order status shows that it has been shipped, but the logistics information has not been updated. What should I do?" The system first assigns a basic weight to each node in the knowledge graph (such as "order status", "logistics information", "shipping", "logistics company", "customer service contact information", etc.). Using the semantic relevance between nodes in the knowledge graph and the implicit relationship learned by the graph neural network, the system calculates the local influence of each node in the extended range of the "order status has not been updated in time" problem. Combining the local influence results and the importance score, the system conducts a comprehensive evaluation of the importance of each node. Through the preset functional relationship, the comprehensive evaluation score of each node is calculated. The results show that "logistics information" and "logistics company" perform well in these two dimensions, so they obtain a higher comprehensive evaluation score. The system applies a gain function to increase the weight of "logistics information" and "logistics company". In this way, in the subsequent topic extraction and content generation process, these two nodes will receive more attention, ensuring that the system can provide more accurate service recommendations.

[0045] Optionally, in step 102, mapping processing is performed on a pre-constructed knowledge graph according to the core intent and the associated concepts, and the scope of topic expansion is determined according to the semantic relevance between nodes in the knowledge graph and the implicit relationship between the nodes learned through the graph neural network, including: Using a bidirectional encoder representation model and a semantic matching algorithm, the concepts in the question text are mapped in a pre-constructed knowledge graph to obtain a topic mapping result; based on the topic mapping result, the multi-step paths starting from the mapped nodes are explored to identify potential nodes related to the core intent to construct a graph neural network; using the graph neural network in combination with a self-attention mechanism, the implicit relationship between the potential nodes is analyzed and processed, and a context-aware model and Bayesian optimization technology are introduced to evaluate the relevance and importance of the potential nodes to the question text, and the relevance and importance scores of the potential nodes are dynamically adjusted to obtain an adjustment result; based on the adjustment result and in combination with a hierarchical clustering algorithm, the nodes within the topic extension range are grouped to determine the topic extension range.

[0046] In this step, the semantic matching algorithm is used to compare the concepts in the question text with the nodes in the knowledge graph, find the most similar or relevant mapping results, and accurately map the concepts in the question text to the corresponding nodes in the knowledge graph. The topic mapping result refers to the corresponding nodes in the knowledge graph of the concepts in the question text obtained by the semantic matching algorithm. Ensure that the core intent and related concepts of the question text can be accurately mapped to the relevant parts of the knowledge graph. Graph neural network is a neural network model specially designed for processing graph structured data, which can capture the complex relationships between nodes. The context-aware model is used to consider the specific background information of the question text and improve the accuracy of mapping and analysis. Bayesian optimization technology is used to evaluate the relevance and importance of potential nodes to the question text, and dynamically adjust the score to ensure that the final selected node best meets user needs.

[0047] In the embodiment of the present application, through the semantic matching algorithm, the system maps the concepts in the question text to the corresponding nodes in the pre-built knowledge graph, obtains the topic mapping result, and ensures that the core intent and associated concepts of the question text can be accurately mapped to the relevant parts of the knowledge graph. According to the topic mapping result, the system starts from the mapped nodes, performs multi-step path exploration in the knowledge graph, and identifies the potential nodes related to the core intent. By constructing a graph neural network and combining the self-attention mechanism, the system conducts an in-depth analysis of the implicit relationship between potential nodes. The self-attention mechanism allows the model to dynamically adjust its weight according to the importance of the node, so as to better understand the relationship between the nodes. Introducing the context-aware model, the system considers the specific background information of the question text to improve the accuracy of mapping and analysis. At the same time, using the Bayesian optimization technology, the system evaluates the relevance and importance of the potential nodes to the question text. According to the adjusted relevance and importance scores, the system determines the nodes that can be included in the scope of the topic extension from the potential nodes. Then, in combination with the hierarchical clustering algorithm, these nodes are grouped and processed to form multiple topic clusters, and the scope of the topic extension is finally determined.

[0048] The customer service system first uses a semantic matching algorithm to map concepts such as "order status" and "logistics information" in the question text to the corresponding nodes in the pre-built knowledge graph. For example, "order status" is mapped to the relevant nodes in the field of "order management", and "logistics information" is mapped to the nodes in the field of "logistics services". By constructing a graph neural network and combining it with the self-attention mechanism, the system conducts an in-depth analysis of the implicit relationships between the above potential nodes. The self-attention mechanism allows the model to dynamically adjust its weights according to the importance of the nodes, thereby better understanding the relationship between the nodes. Then, combined with the hierarchical clustering algorithm, these nodes are grouped and processed to form multiple topic clusters, and finally the scope of topic expansion is determined. For example, "logistics company" and "transportation progress" are classified into the same cluster, named "reasons for logistics delays", while "customer service contact information" is separately classified into another cluster, named "help channels".

[0049] Optionally, the structured description information is further refined by using a multi-layer attention mechanism in step 105 to obtain an extraction result, including: The obtained structured description information is encoded using the constructed multi-layer attention network model to obtain the structured description information after encoding; based on the structured description information after encoding, the first layer of attention mechanism is used to perform preliminary filtering and concentration on the structured description information after encoding, identify and strengthen elements directly related to the core intent, weaken or ignore irrelevant structured description information after encoding, and obtain preliminary extraction results; according to the preliminary extraction results, in each subsequent layer of the attention mechanism, based on the output of the attention mechanism of the previous layer, the structured description information is analyzed and processed in depth layer by layer to obtain the screening results of the multi-layer attention mechanism; based on the screening results of the multi-layer attention mechanism, the output of the attention mechanism of the final layer is processed to obtain the extraction results.

[0050] In this step, the multi-layer attention network model is a deep learning architecture that can understand information from different levels and angles. Each layer of attention mechanism focuses on a specific level of information to ensure that the final output is both accurate and easy to understand. The initial filtering and concentration process uses the first layer of attention mechanism to perform the initial screening of the structured description information after encoding. In this way, the system can quickly lock on key information points and improve processing efficiency. The layer-by-layer in-depth analysis and processing means that in each subsequent layer of attention mechanism, the understanding of information is further refined and deepened based on the output of the previous layer, ensuring the comprehensiveness and accuracy of information processing. The extraction result refers to the final topic summary or key information after processing by the multi-layer attention mechanism. These results are closely centered around the user's core intent, provide concise and clear answers, and may contain additional relevant information to enhance the user experience.

[0051] In an embodiment of the present application, based on the structured description information, the system uses the first layer of attention mechanism to perform preliminary filtering and concentration processing. The preliminary extraction results focus on reflecting the most critical parts of the question text, thereby improving the efficiency of information processing. According to the preliminary extraction results, in each subsequent layer of attention mechanism, the system performs a layer-by-layer in-depth analysis and processing of the structured description information based on the output of the previous layer. Each layer of attention mechanism focuses on different information levels, gradually digging out deeper relationships and implicit topics. Through a progressive approach layer by layer, the system can capture more complex information structures to ensure the comprehensiveness and accuracy of the final output. Based on the screening results of the multi-layer attention mechanism, the system processes the output of the final layer of attention mechanism to obtain the extraction results. These results are closely centered around the user's core intent, provide concise and clear answers, and may contain additional relevant information to enhance the user experience.

[0052] The customer service system uses the first layer of attention mechanism to perform preliminary filtering and condensation based on structured description information. For example, identify and strengthen elements directly related to "order status" and "logistics information", and weaken or ignore irrelevant information such as "return policy". According to the preliminary extraction results, in each subsequent layer of attention mechanism, the system conducts a layer-by-layer in-depth analysis and processing of the structured description information based on the output of the previous layer. For example, the second layer of attention mechanism may find that "improper selection of logistics companies" or "holiday impact" are important reasons for logistics delays. Based on the screening results of multiple layers of attention mechanisms, the system processes the output of the final layer of attention mechanism to obtain the extraction results. Finally, a concise and clear summary is generated: "Your order has been shipped, but the logistics information has not been updated. The delay may be due to problems with the logistics company. It is recommended that you contact the courier company to inquire about the specific reasons, or check the latest logistics progress on our website. If you have further questions, please refer to our after-sales service policy or contact customer service for help." In the process of topic extraction for user question text, it is crucial to determine the importance and weight of nodes within the scope of topic extension. Traditional weight assignment methods are usually based on static features such as frequency or predefined rules, but this method is difficult to adapt to complex and changing contexts and dynamic information needs. Therefore, an adaptive weight assignment mechanism combined with a link analysis algorithm is introduced to dynamically adjust the node weights according to the centrality of the node in the graph neural network and other relevant factors.

[0053] Optionally, based on the topic extension scope, an adaptive weight allocation mechanism combined with a link analysis algorithm is used to evaluate the importance of nodes within the topic extension scope and the centrality of the nodes in the graph neural network to adjust the weight of each node, including: Based on the topic extension scope and the obtained local influence results, the link analysis algorithm is used to iteratively calculate the scores of the nodes to obtain the importance scores of the nodes within the topic extension scope; wherein the importance Evaluation calculation formula: ; in, Indicates the node in the extended scope of the topic The centrality of Indicates the node in the extended scope of the topic Link analysis score, is a balancing factor between 0 and 1, and are the exponential factors applied to the centrality and the linkage analysis scores, respectively, Representation Node The variability score reflects the stability of the node. is the coefficient used to adjust the variability score, Represents the node Activity score, is a coefficient used to adjust the activity score, is the importance; Represents the nodes in the graph neural network The centrality is calculated using the PageRank algorithm. The centrality calculation formula is: ; in, is the damping coefficient, is the total number of nodes, is pointing to the node The set of all nodes of Is a node , Is a node The number of outgoing links, is an additional exponential factor, Is a distance decay function used to determine the node and the node The shortest path distance between , , is to control the decay rate, Represents the nodes in the graph neural network centrality; Combine the local influence result and the importance score, as well as the centrality of the node in the graph neural network, and perform a comprehensive evaluation on the importance of the node to dynamically adjust the weight of each node and adjust the node Weight ; Weight calculation formula: ; in, is the node The original weight of is a mixing ratio factor between 0 and 1, used to control the mixing ratio between the calculated importance and the original weight, is an exponential factor for the importance, is the importance, is the node The weight of .

[0054] The overall formula introduces node importance evaluation and weight calculation, aiming to build an efficient, flexible and intelligent node importance evaluation and weight adjustment framework to improve the system's understanding of user problems and service quality. The introduction of centrality calculation reflects the relative importance of nodes in the network. The overall formula design fully considers multiple aspects such as multi-dimensional evaluation, dynamic adaptability, balance adjustment, enhanced differentiation, smooth transition, context sensitivity and wide application, so that the system can better serve the accurate understanding and answering of user problems.

[0055] The following is a brief explanation of the design reasons for each item in the formula: Importance Reasons for designing the evaluation formula: Importance The evaluation calculation formula can strike a balance between multiple factors, providing a comprehensive and dynamic evaluation method to ensure that the system can accurately identify and highlight the most important nodes and provide users with more accurate service recommendations.

[0056] Formula expression: It is used to comprehensively evaluate the structural characteristics and stability of nodes in the network, where: Is a node The centrality of Adjust its impact strength, reflects the variability score of the node, Representation Node The variability score reflects the stability of the node and is adjusted by the coefficient control its impact; Used to comprehensively evaluate the connection strength and activity of nodes in the network. Reflects the link analysis score of the node , and through the exponential factor Adjust its impact strength, Reflects the activity score of the node , and by adjusting the coefficient Control its impact, Through the balance factor , ensuring an appropriate balance between link analysis scores and activity scores.

[0057] Centrality Reason for designing the calculation formula: Centrality Reflecting the relative importance of nodes in the network, nodes with high centrality often play a key role in the network and are important hubs for information dissemination.

[0058] Formula expression: Centrality In the calculation formula What is calculated is the basic probability of each node, that is, the initial weight given to each node at the beginning of each iteration, the damping coefficient It is a parameter between 0 and 1, usually set to around 0.85, which represents the probability that a user will continue to click on a link while browsing a web page or moving in the network. Indicates the total number of nodes in the network; Mainly used for computing nodes The weights obtained from other nodes pointing to it are transferred to accurately calculate the node The centrality score of is pointing to the node The set of all nodes of is the node The number of outgoing links, is an additional exponential factor, Is a distance decay function used to determine the node and the node The shortest path distance between , , It controls the decay rate Reasons for designing the weight calculation formula: The weight formula plays a key role in adjusting node weights. It achieves smooth and dynamic adjustment of node weights by combining the newly evaluated importance and the original weights. This not only improves the flexibility and adaptability of the system, but also ensures the stability and continuity of the results, enabling the system to better serve complex and changing application scenarios.

[0059] Formula expression: Used to determine the importance of nodes Calculate the new weighted contribution, is a mixing ratio factor between 0 and 1, used to control the mixing ratio between the calculated importance and the original weight, is an exponential factor for the importance, is the importance; Used to preserve the original weights of the nodes To ensure the stability and continuity of the system, is the node The weight of .

[0060] The following is a brief explanation of how to obtain the parameters of the formula: : These parameters can be tuned through experiments and techniques such as Bayesian optimization to find the best values ​​that best suit a specific application scenario; :These scores can be automatically calculated using historical data, expert knowledge, or machine learning models. Calculated using the PageRank algorithm, Calculated by link analysis algorithm, Calculated based on the frequency of node attribute changes, Calculated based on the activity of the node over the past period of time.

[0061] In the embodiment of this application, suppose we have a customer service scenario, and the user asks: "My order status shows that it has been shipped, but the logistics information has not been updated, what should I do?" Initial settings: The nodes involved in the knowledge graph include "order status", "logistics information", "logistics company", and "customer service contact information". The initial weight of each node Set to 0.5, Parameter settings: .

[0062] Calculation process: Calculating centrality : Use the PageRank algorithm to calculate the centrality of each node, for example, the centrality of "order status" Probably higher because it is a core concept; Evaluate importance : For the "Order Status" node (assuming it is node 1), substitute the formula to calculate the importance: ; Assumptions ,but: ; Adjust weights : For the "Order Status" node, substitute the formula to calculate the new weight: ; The system determines that the importance score of the "Order Status" node is approximately 0.916, and adjusts its weight to approximately 0.695, so that the "Order Status" node will receive more attention in the subsequent topic refinement and content generation process, thereby ensuring that the system can provide more accurate service recommendations.

[0063] In the process of topic extraction from user question texts, generating structured description information requires comprehensive consideration of node weights, connection strengths between nodes, and dynamic features of nodes. Traditional methods are usually based on static features, but this method is difficult to adapt to complex and changing contexts and dynamic information needs. Therefore, point mutual information is introduced to calculate the connection strength between nodes, and combined with adjusted node weights and other dynamic features to more accurately evaluate the importance of nodes, and finally generate structured description information representing the corresponding topic of the question text.

[0064] Optionally, generating structured description information representing a topic corresponding to the question text according to the adjusted node weights and the connection strength between the nodes calculated by point mutual information, further comprising: Generate the connection strength between the nodes according to the adjusted node weights and point mutual information; Connection strength Calculation formula: ; in, is the node and the node The probability of simultaneous occurrence, and The nodes are and the node The probability of a single occurrence, is an enhancement factor used to strengthen the effect of high co-occurrence frequency on the connection strength. is an exponential factor used to adjust the influence of the high co-occurrence frequency. is the connection strength, is the node The weight of A comprehensive score is used based on the node weight, the connection strength between the nodes, the variability and activity of the nodes. Evaluate the node The importance of the structured description information, Comprehensive score calculation formula: ; in, Indicates that the node The set of directly connected nodes, is an exponential factor applied to the mutual information of the points, is an exponential factor applied to the connection strength, is a distance decay function, the node and the node The shortest path distance between , Controls the decay rate, is a similarity enhancement function based on the node and the node The connection strength between ,in is an index that controls the influence of the semantic similarity. is a coefficient that balances the strength of the node connections, is the node The variability score, is the exponential factor for the variability score, is the node Activity score, is the exponential factor of the activity score, is the node The popularity rating of is the exponential factor of the popularity score, and are coefficients for adjusting the influence of the variability score, the activity score and the popularity score, respectively. is the composite score; Based on the comprehensive score , selecting the node with the highest score and its connection relationship to form structured description information representing the topic corresponding to the question text; Calculation formula: ; in, is a preset connection strength threshold, used to filter out weak connection strengths, is the node and the node connection strength.

[0065] By combining the strength of and comprehensive score The overall design of is to provide a comprehensive and dynamic method to evaluate the importance of nodes. By combining the static structural characteristics of nodes (such as weight, connection strength) and dynamic characteristics (such as variability, activity, popularity), the formula can strike a balance between multiple factors to ensure the comprehensiveness and accuracy of the evaluation results. In addition, by introducing parameters such as exponential factors, adjustment coefficients, and distance decay functions, the system is ensured to be highly flexible and adaptable, and suitable for complex application scenarios in different fields.

[0066] The following is a brief explanation of the design reasons for each item in the formula: Connection strength Design reason: Connection strength measures the co-occurrence relationship between nodes through point mutual information, reflects the association strength between nodes, ensures that nodes with high co-occurrence frequency contribute more to the connection strength, and highlights strong associations.

[0067] Formula expression: Representation Node and nodes The ratio of the probability of simultaneous occurrence to the product of their individual occurrence probabilities measures the co-occurrence relationship between nodes through point mutual information, reflecting the strength of association between nodes; Used to enhance the impact of high co-occurrence frequency on connection strength by introducing an enhancement coefficient and exponential factor ,The formula can highlight those nodes that frequently co-occur and are closely related while keeping the value range reasonable, making the evaluation results more comprehensive and accurate, is the node and the node The probability of simultaneous occurrence, and The nodes are and the node The probability of a single occurrence, is an enhancement factor used to strengthen the effect of high co-occurrence frequency on the connection strength. is an exponential factor used to adjust the influence of the high co-occurrence frequency. is the connection strength; The reason for designing the comprehensive score calculation formula: The comprehensive score calculation formula combines node weights, connection strengths, and dynamic features to comprehensively evaluate the importance of nodes, strike an appropriate balance between the static structural features and dynamic features of nodes, ensure a smooth transition between new and old information, and avoid drastic fluctuations in results due to frequent updates.

[0068] Formula expression: For evaluating nodes The connection strength and its contribution to the comprehensive score are comprehensively evaluated by combining node weights, connection strength and dynamic features (such as distance decay and semantic similarity). The importance of Exponential Factor and The formula can strike a balance between multiple factors to ensure the comprehensiveness and accuracy of the evaluation results. Used to accumulate all nodes Directly connected nodes The connection strength of the network is calculated, and the influence of distance decay and semantic similarity is considered; By combining the dynamic characteristics of nodes, the importance of nodes is comprehensively evaluated, and the balance factor is introduced and adjustment coefficient , , The formula can strike a balance between multiple factors to ensure the comprehensiveness and accuracy of the evaluation results. Indicates that the node The set of directly connected nodes, is an exponential factor applied to the mutual information of the points, is an exponential factor applied to the strength of the connection, is a distance decay function, the node and the node The shortest path distance between , Controls the decay rate, is a similarity enhancement function based on the node and the node The connection strength between ,in is an index that controls the influence of the semantic similarity. is a coefficient that balances the strength of the node connections, is the node The variability score, is the exponential factor for the variability score, is the node Activity score, is the exponential factor of the activity score, is the node The popularity rating of is the exponential factor of the popularity score, and are coefficients for adjusting the influence of the variability score, the activity score and the popularity score, respectively. is the composite score; Reasons for designing structured description information: Through the computational design of structured description information, the system can effectively remove weak connections that are irrelevant to the problem, thereby generating a more concise and useful structured description information.

[0069] Formula expression: If , the connection is retained ,in, is a preset connection strength threshold, which is set by , can effectively remove those weak connections that have little impact on the overall structure, thereby focusing on important node relationships, which helps to improve the efficiency of the system and the accuracy of the results, is the node and the node The connection strength of the structured description information represents that the system can effectively remove those weak connections that are irrelevant to the problem, thereby generating a more concise and useful structured description information.

[0070] The following is a brief explanation of how to obtain the parameters of the formula: Node Weight : Calculated by combining the adaptive weight distribution mechanism with the link analysis algorithm, Connection strength : Calculated by point mutual information, Variability score , Activity score and popularity score : Through historical data analysis, Exponential Factor and adjustment coefficient , , : can be tuned through experiments and techniques such as Bayesian optimization, Enhancement factor and exponential factor : Used to adjust the impact of high co-occurrence frequency to ensure the rationality of the evaluation results. Distance decay parameter and similarity enhancement index : Used to adjust the impact of distance decay and semantic similarity to ensure the rationality of the evaluation results.

[0071] In the embodiment of the present application, a logistics network consisting of multiple logistics centers is considered. Each logistics center has its own weight. , reflecting its importance or capacity; and the strength of its connections with other logistics centers , calculated based on historical data. Our goal is to evaluate the importance of each logistics center based on this information and select the most effective connections to form an optimized logistics network.

[0072] Parameter settings and assumptions: Set the weights of logistics centers 1 to 5 to , and 0.5.

[0073] Setting the Variability Score They are , and 0.3.

[0074] Activity score They are , and 0.4.

[0075] Popularity Rating They are , and 0.6.

[0076] Connection strength The specific value of will be calculated based on the actual co-occurrence probability and other factors. Here we simplify and directly give some example values, such as the connection strength between logistics centers 1 and 2 is 0.8, 1 to 3 is 0.6, and so on. Set the connection strength threshold , that is, only connections with connection strength greater than 0.5 are retained; , the remaining parameters are adjusted according to the actual situation. To simplify the calculation, we will use the default values ​​or empirical estimates.

[0077] Calculate overall score : For logistics center 1, we first calculate the sum of the weighted connection strengths between it and all other logistics centers, and then calculate according to the given comprehensive scoring formula . Assume that: ; ; Distance decay function (Assuming that distance decay is not significant); Similarity Enhancement Function (assuming linearity); ; here It indicates the node weight that may have undergone some adjustments. Other index factors and coefficients also need to be determined based on actual conditions.

[0078] After calculating the comprehensive score of each logistics center through the above formula, we can sort them according to the score, select the logistics center with the highest score as the key node, and retain those whose connection strength exceeds the threshold. The resulting structured description information can help decision makers understand which logistics centers and routes are most important, and then guide practical operations such as resource allocation and route planning. The calculation results show that logistics center 4 has the highest comprehensive score, and strong connections are mainly concentrated on logistics centers 1 and 2, which indicates that establishing an efficient logistics channel between these three logistics centers may be the best choice.

[0079] Figure 2 A structural diagram of a topic extraction system for a user question text is provided for an embodiment of the present application, such as Figure 2 As shown, the device comprises: The parsing module 21 parses the question text based on the received question text submitted by the user to identify the core intention and related concepts; A mapping module 22 performs mapping processing on a pre-constructed knowledge graph according to the core intent and related concepts, and determines the scope of topic expansion according to the semantic relevance between nodes in the knowledge graph and the implicit relationship between the nodes learned through the graph neural network; The adjustment module 23 uses an adaptive weight allocation mechanism combined with a link analysis algorithm based on the topic extension scope to evaluate the importance of nodes within the topic extension scope and the centrality of the nodes in the graph neural network to adjust the weight of each node; A generating module 24 generates structured description information representing a topic corresponding to the question text according to the adjusted node weights and the connection strength between the nodes calculated by point mutual information; Processing module 25 further refines the structured description information using a multi-layer attention mechanism to obtain an extraction result, where the extraction result refers to a summary output result of the topic corresponding to the question text.

[0080] Figure 2 The topic extraction system of a user question text can be performed Figure 1 The implementation principle and technical effect of the method for extracting the topic of a user question text described in the embodiment shown will not be described in detail. The specific way in which each module and unit performs operations in the topic extraction system for a user question text in the above embodiment has been described in detail in the embodiment of the method, and will not be described in detail here.

[0081] In one possible design, Figure 2 A topic extraction system for user question text in the illustrated embodiment can be implemented as a computing device, such as Figure 3 As shown, the computing device may include a storage component 31 and a processing component 32; The storage component 31 stores one or more computer instructions, wherein the one or more computer instructions are called and executed by the processing component 32 .

[0082] The processing component 32 is used to: parse the question text based on the received question text submitted by the user, and identify the core intention and related concepts; According to the core intent and related concepts, mapping processing is performed on a pre-constructed knowledge graph, and the scope of topic expansion is determined based on the semantic relevance between nodes in the knowledge graph and the implicit relationship between the nodes learned through the graph neural network; Based on the topic extension scope, an adaptive weight allocation mechanism combined with a link analysis algorithm is used to evaluate the importance of nodes within the topic extension scope and the centrality of the nodes in the graph neural network to adjust the weight of each node; Generate structured description information representing the topic corresponding to the question text according to the adjusted node weights and the connection strength between the nodes calculated by point mutual information; The structured description information is further refined using a multi-layer attention mechanism to obtain an extraction result, where the extraction result refers to a summary output result of the topic corresponding to the question text.

[0083] The processing component 32 may include one or more processors to execute computer instructions to complete all or part of the steps in the above method. Of course, the processing component may also be implemented by one or more application-specific integrated circuits (ASICs), digital signal processors (DSPs), digital signal processing devices (DSPDs), programmable logic devices (PLDs), field programmable gate arrays (FPGAs), controllers, microcontrollers, microprocessors or other electronic components to perform the above method.

[0084] The storage component 31 is configured to store various types of data to support operations at the terminal. The storage component can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic memory, flash memory, magnetic disk or optical disk.

[0085] Of course, the computing device may also include other components, such as input / output interfaces, display components, communication components, etc.

[0086] The input / output interface provides an interface between the processing component and the peripheral interface module, which may be an output device, an input device, etc.

[0087] The communication component is configured to facilitate, among other things, wired or wireless communications between the computing device and other devices.

[0088] Among them, the computing device can be a physical device or an elastic computing host provided by a cloud computing platform, etc. In this case, the computing device can refer to a cloud server, and the above-mentioned processing components, storage components, etc. can be basic server resources rented or purchased from the cloud computing platform.

[0089] The present application also provides a computer storage medium storing a computer program, wherein the computer program can achieve the above-mentioned Figure 1 The illustrated embodiment is a method for extracting topics from a user question text.

[0090] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working processes of the systems, devices and units described above can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0091] The device embodiments described above are merely illustrative, wherein the units described as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they may be located in one place, or they may be distributed on multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the scheme of this embodiment. Ordinary technicians in this field can understand and implement it without paying creative labor.

[0092] Through the description of the above implementation methods, those skilled in the art can clearly understand that each implementation method can be implemented by means of software plus a necessary general hardware platform, and of course, can also be implemented by hardware. Based on this understanding, the above technical solution is essentially or the part that contributes to the prior art can be embodied in the form of a software product, and the computer software product can be stored in a computer-readable storage medium, such as a disk, an optical disk, etc., including a number of instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in each embodiment or some parts of the embodiments.

[0093] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit it. Although the present application has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A topic extraction method for user question text, characterized in that: include: Based on the received question text submitted by the user, the question text is parsed to identify the core intent and related concepts; According to the core intent and related concepts, mapping processing is performed on a pre-constructed knowledge graph, and the scope of topic expansion is determined based on the semantic relevance between nodes in the knowledge graph and the implicit relationship between the nodes learned through the graph neural network; Based on the topic extension scope, an adaptive weight allocation mechanism combined with a link analysis algorithm is used to evaluate the importance of nodes within the topic extension scope and the centrality of the nodes in the graph neural network to adjust the weight of each node; Generate structured description information representing the topic corresponding to the question text according to the adjusted node weights and the connection strength between the nodes calculated by point mutual information; The structured description information is further refined using a multi-layer attention mechanism to obtain an extraction result, where the extraction result refers to a summary output result of the topic corresponding to the question text.

2. The method according to claim 1, characterized in that According to the adjusted node weights and the connection strengths between the nodes calculated by point mutual information, structured description information representing the topic corresponding to the question text is generated, including: Based on the adjusted node weights, the nodes in the knowledge graph are prioritized to obtain core topic nodes; Using point mutual information, the semantic association strength between the core topic nodes is quantified to determine the closeness between the core topic nodes; According to the closeness between the core topic nodes, a clustering algorithm or a community discovery algorithm is used to group the core topic nodes to form a topic cluster; Comprehensively analyzing and processing the core topic nodes in the topic cluster to extract topic tags that can summarize the core meaning of the topic cluster; Combining the explicit concepts in the question text with the topic labels, supplementing potential implicit topics through a logical reasoning mechanism, and generating topic content corresponding to the implicit topics; Based on the subject content, the explicit concepts, the subject tags and the implicit subjects are organized to generate structured description information representing the subject corresponding to the question text.

3. The method according to claim 2, characterized in that Combining the explicit concepts in the question text with the topic labels, supplementing potential implicit topics through a logical reasoning mechanism, and generating topic content corresponding to the implicit topics, including: Constructing a preliminary topic framework based on the explicit concepts in the question text and the topic labels; Using the associations and context information between nodes in the knowledge graph, the subject framework is expanded to identify additional entities or concepts related to the explicit concept and the subject label; Based on rule-based reasoning or machine learning models, the relationships between the additional entities or concepts are analyzed and processed to mine implicit topics; Based on the implicit topic, the explicit concept and the topic label, consistency processing is performed through semantic fusion technology, and the explicit concept, the topic label and the implicit topic are integrated to generate the topic content corresponding to the implicit topic.

4. The method according to claim 1, characterized in that: Based on the topic extension scope, an adaptive weight allocation mechanism combined with a link analysis algorithm is used to evaluate the importance of nodes within the topic extension scope and the centrality of the nodes in the graph neural network to adjust the weight of each node, including: Based on the topic extension scope, the basic weights of the nodes within the topic extension scope are initialized using an adaptive weight allocation mechanism, and the local influence of the basic weight of each node within the topic extension scope is calculated based on the semantic relevance between the nodes in the knowledge graph and the implicit relationship between the nodes learned by the graph neural network to obtain a local influence result; Using the link analysis algorithm, iteratively calculate the scores of the nodes to obtain importance scores; The importance of the nodes is comprehensively evaluated by combining the local influence results and the importance scores, as well as the centrality of the nodes in the graph neural network, so as to dynamically adjust the weights of the nodes so that the nodes obtain higher weights.

5. The method according to claim 4, characterized in that Combining the influence result and the importance score, a comprehensive evaluation process is performed on the importance of the node to dynamically adjust the weight of each node so that the node obtains a higher weight, including: Using the local influence result, the basic influence score of each node is calculated to obtain a basic influence score; Based on the basic influence score and the importance score, combined with a preset functional relationship, the comprehensive evaluation score of the node is calculated to obtain a comprehensive evaluation score reflecting the importance of the node within the scope of the topic extension; Using the comprehensive evaluation score, all nodes are sorted, and a threshold or ranking interval is set to determine nodes with high importance; A gain function is applied to increase the weight of the high-importance node so that the high-importance node obtains a higher weight.

6. The method according to claim 1, characterized in that According to the core intent and related concepts, mapping processing is performed on the pre-constructed knowledge graph, and the scope of topic expansion is determined according to the semantic relevance between nodes in the knowledge graph and the implicit relationship between the nodes learned through the graph neural network, including: Using a bidirectional encoder representation model and a semantic matching algorithm, based on the core intent and the associated concepts, the concepts in the question text are mapped in a pre-constructed knowledge graph to obtain a topic mapping result; According to the topic mapping results, combined with the semantic relevance between nodes in the knowledge graph, multi-step paths starting from the mapped nodes are explored and processed to identify potential nodes related to the core intent, so as to construct a graph neural network; The graph neural network is combined with a self-attention mechanism to analyze and process the implicit relationship between the potential nodes, and a context-aware model and Bayesian optimization technology are introduced to evaluate the relevance and importance of the potential nodes to the question text, and the relevance and importance scores of the potential nodes are dynamically adjusted to obtain adjustment results; According to the adjustment result, nodes that can be included in the subject extension scope are determined from the potential nodes, and the nodes in the subject extension scope are grouped in combination with a hierarchical clustering algorithm to determine the subject extension scope.

7. The method according to claim 1, characterized in that The structured description information is further refined by using a multi-layer attention mechanism to obtain an extraction result, including: Using the constructed multi-layer attention network model, encoding the obtained structured description information to obtain the structured description information after encoding; Based on the structured description information after the encoding process, the structured description information after the encoding process is preliminarily filtered and condensed by using the first-layer attention mechanism, the elements directly related to the core intent are identified and strengthened, and the irrelevant structured description information after the encoding process is weakened or ignored, so as to obtain a preliminary extraction result; According to the preliminary extraction result, in each subsequent layer of the attention mechanism, based on the output of the previous layer of the attention mechanism, the structured description information is subjected to a layer-by-layer in-depth analysis and processing to obtain the multi-layer attention mechanism screening result; Based on the multi-layer attention mechanism screening results, the output of the attention mechanism of the final layer is processed to obtain the extraction result.

8. A method for extracting topics from user question text, characterized in that: include: A parsing module, based on the received question text submitted by the user, parses the question text to identify the core intent and related concepts; A mapping module performs mapping processing on a pre-constructed knowledge graph according to the core intent and related concepts, and determines the scope of topic expansion according to the semantic relevance between nodes in the knowledge graph and the implicit relationship between the nodes learned through the graph neural network; An adjustment module, based on the topic extension scope, adopts an adaptive weight allocation mechanism combined with a link analysis algorithm to evaluate the importance of nodes within the topic extension scope and the centrality of the nodes in the graph neural network to adjust the weight of each node; A generation module, generating structured description information representing a topic corresponding to the question text according to the adjusted node weights and the connection strength between the nodes calculated by point mutual information; The processing module further refines the structured description information using a multi-layer attention mechanism to obtain an extraction result, where the extraction result refers to a summary output result of the topic corresponding to the question text.

Images