Intelligent customer service response method and device, electronic equipment, medium and program product

Abstract

The application discloses an intelligent customer service response method and device, electronic equipment, medium and program product. The method comprises the following steps: acquiring a real-time question to be answered, obtaining a corresponding question vector based on the real-time question to be answered; acquiring a semantic entropy value corresponding to each label in a preset dynamic semantic library, obtaining a global semantic entropy value of the preset dynamic semantic library based on the semantic entropy values of all labels; in response to the global semantic entropy value exceeding a preset global entropy threshold, clustering the corresponding label whose semantic entropy value is higher than a preset word entry entropy threshold, and obtaining a clustered label vector; based on the label vector satisfying a preset label density threshold after clustering, generating a target label vector, and saving the target label vector to the preset dynamic semantic library; based on the question vector and the preset dynamic semantic library, similarity comparison is performed to obtain a plurality of target labels satisfying a preset similarity condition, and the plurality of target labels are used to generate a corresponding response result. The response accuracy of the intelligent customer service is improved.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, and in particular to an intelligent customer service response method, device, electronic device, medium, and program product. Background Technology

[0002] With the development of artificial intelligence technology, intelligent customer service has been widely used in customer service scenarios across various industries. To reduce tag redundancy, existing technologies cluster tags in the tag library, merging tags with similar semantics.

[0003] However, existing technologies cannot quantify the degree of uncertainty in the semantics of tags, making it difficult to accurately identify highly ambiguous tags. This leads to overfitting problems in classification models, resulting in insufficient accuracy and stability of intelligent customer service responses. Summary of the Invention

[0004] This application provides an intelligent customer service response method, device, electronic device, medium, and program product, which improves the accuracy of intelligent customer service responses.

[0005] On one hand, this application provides an intelligent customer service response method, including: acquiring real-time questions to be answered, and obtaining corresponding question vectors based on the real-time questions to be answered; acquiring the semantic entropy value corresponding to each tag in a preset dynamic semantic library, and obtaining the global semantic entropy value of the preset dynamic semantic library based on the semantic entropy values ​​of all tags; in response to the global semantic entropy value exceeding a preset global entropy value threshold, clustering the corresponding tags whose semantic entropy values ​​are higher than a preset term entropy value threshold to obtain clustered tag vectors; generating target tag vectors based on the clustered tag vectors that meet a preset tag density threshold, and saving the target tag vectors to the preset dynamic semantic library, wherein the preset tag density threshold is adjusted based on the tag generation time, tag usage frequency, and time decay factor; and performing a similarity comparison between the question vectors and the preset dynamic semantic library to obtain multiple target tags that meet preset similarity conditions, and the multiple target tags are used to generate corresponding response results.

[0006] On the other hand, embodiments of this application provide an intelligent customer service response device, comprising: a first acquisition unit, configured to acquire real-time questions to be answered, and obtain corresponding question vectors based on the real-time questions to be answered; a second acquisition unit, configured to acquire the semantic entropy value corresponding to each tag in a preset dynamic semantic library, and obtain the global semantic entropy value of the preset dynamic semantic library based on the semantic entropy values ​​of all tags; a first processing unit, configured to cluster the corresponding tags whose semantic entropy values ​​are higher than a preset term entropy value threshold in response to the global semantic entropy value exceeding a preset global entropy value threshold, and obtain clustered tag vectors; a second processing unit, configured to generate target tag vectors based on the clustered tag vectors that meet a preset tag density threshold, and save the target tag vectors to the preset dynamic semantic library, wherein the preset tag density threshold is adjusted based on the tag generation time, tag usage frequency, and time decay factor; and a third processing unit, configured to perform similarity comparison between the question vectors and the preset dynamic semantic library, and obtain multiple target tags that meet preset similarity conditions, wherein the multiple target tags are used to generate corresponding response results.

[0007] In another aspect, embodiments of this application provide an electronic device, which includes: a processor and a memory storing computer program instructions; the processor executes the computer program instructions to implement an intelligent customer service response method.

[0008] In another aspect, embodiments of this application provide a computer storage medium on which computer program instructions are stored, and when the computer program instructions are executed by a processor, an intelligent customer service response method is implemented.

[0009] In another aspect, embodiments of this application provide a computer program product in which instructions are executed by the processor of an electronic device, causing the electronic device to perform an intelligent customer service response method.

[0010] The intelligent customer service response method, device, electronic device, medium, and program product of this application embodiment acquires real-time questions to be answered, obtains corresponding question vectors based on the real-time questions to be answered, and transforms unstructured user questions into standardized dense semantic vectors. It acquires the semantic entropy value corresponding to each tag in a preset dynamic semantic library, and obtains the global semantic entropy value of the preset dynamic semantic library based on the semantic entropy values ​​of all tags, quantifying the uncertainty degree of a single tag and the global semantic library. By responding to the global semantic entropy value exceeding a preset global entropy value threshold, it clusters the corresponding tags whose semantic entropy values ​​are higher than a preset term entropy value threshold, which can accurately locate semantic ambiguity and redundancy. Based on the tag vectors that meet the preset tag density threshold after clustering, it generates target tag vectors and saves the target tag vectors to the preset dynamic semantic library, which can continuously maintain the high determinism of the semantic library. Based on the question vector and the preset dynamic semantic library, it performs a similarity comparison to obtain multiple target tags that meet preset similarity conditions, which are used to generate corresponding response results, enabling the output of accurate responses that match the user's true intent. Therefore, it improves the response accuracy of intelligent customer service. Attached Figure Description

[0011] To more clearly illustrate the technical solutions of the embodiments of this application, the accompanying drawings used in the embodiments of this application will be briefly introduced below. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0012] Figure 1 This is a flowchart illustrating an intelligent customer service response method provided in an embodiment of this application; Figure 2 This is an example diagram of the problem tag index provided in the embodiments of this application; Figure 3 This is an example diagram of the tag vector retrieval range provided in the embodiments of this application; Figure 4 This is a schematic diagram illustrating the processing procedure of an intelligent customer service response method provided in an embodiment of this application; Figure 5 This is a schematic diagram of the large model processing procedure provided in the embodiments of this application; Figure 6 This is a flowchart of the problem vector generation and label matching process provided in the embodiments of this application; Figure 7 This is a schematic diagram of the intelligent customer service response device provided in the embodiments of this application; Figure 8 This is a schematic diagram of the structure of the electronic device provided in the embodiments of this application. Detailed Implementation

[0013] The features and exemplary embodiments of various aspects of this application will be described in detail below. To make the objectives, technical solutions, and advantages of this application clearer, the application will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are only intended to explain this application and not to limit it. For those skilled in the art, this application can be implemented without some of these specific details. The following description of the embodiments is merely to provide a better understanding of this application by illustrating examples.

[0014] It should be noted that, in this document, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising..." does not exclude the presence of additional identical elements in the process, method, article, or apparatus that includes the element.

[0015] It should be noted that the acquisition, storage, use, and processing of data in this application embodiment all comply with the relevant provisions of national laws and regulations.

[0016] It should be noted that in the embodiments of this application, certain software, components, models and other existing solutions in the industry may be mentioned. These should be regarded as exemplary and are only intended to illustrate the feasibility of implementing the technical solution of this application. However, it does not mean that the applicant has used or necessarily used the solution.

[0017] To address the aforementioned technical issues, the following detailed description, in conjunction with the accompanying drawings, of a smart customer service response method, apparatus, electronic device, computer storage medium, and computer program product provided by the present application, through specific embodiments and application scenarios, will be provided in detail.

[0018] This application provides an intelligent customer service response method, such as... Figure 1 As shown, the method includes the following steps: S100: Obtain real-time pending questions, and based on the real-time pending questions, obtain the corresponding question vector; In this context, "real-time pending questions" refers to natural language interaction requests submitted by users in real time within an intelligent customer service interaction scenario. These are the direct processing targets of the intelligent customer service system. Specifically, these can be plain text input, text generated from speech-to-text transcription, text extracted from image recognition, text extracted from video subtitles, etc. For example, it could be a user typing "Why hasn't my work order been pushed to the audit department yet?" in a corporate customer service chat box, or it could be the text transcribed from a user's voice question, "How do I change the mobile phone number registered with my bank card?"

[0019] Among them, the question vector refers to the core semantic information used to accurately represent the question and is the basic data for subsequent semantic similarity comparison. Specifically, it can be a bag-of-words vector, TF-IDF vector, Word2Vec word vector, BERT sentence vector, Sentence-BERT context-aware sentence vector, etc. For example, it can be a d-dimensional numerical vector obtained by transforming "Why hasn't my work order been pushed to the audit department yet" through the Sentence-BERT model.

[0020] Specifically, as an example, when a user enters the question "Why hasn't my work order been pushed to the audit department yet?" through the web interface of the enterprise's internal intelligent customer service, the real-time text of the question awaiting a response is obtained. The text undergoes semantic preprocessing, including word segmentation and stop word removal, resulting in multiple candidate semantic terms such as "work order," "why," "no," "push," and "audit department." The discriminative importance of each candidate semantic term is evaluated based on the TF-IDF algorithm. Noisy semantic terms such as "why" and "no" with discriminative importance below a preset threshold are filtered out, leaving three core semantic terms: "work order," "push," and "audit department." These three core semantic terms are then transformed into d-dimensional question vectors using a pre-trained Sentence-BERT model.

[0021] In practical applications, such as Figure 2 As shown in the diagram, the mapping relationship between user questions and core semantic terms is illustrated. For the user question "I can't log in to the maintenance outsourcing system, what should I do?", two core semantic terms, "maintenance outsourcing" and "login," are extracted. For the user question "The work order for the hidden danger report, which was assigned before August 26th, has not been pushed to the audit system," four core semantic terms, "audit," "work order," "push," and "assignment," are extracted. This question tag indexing mechanism can quickly locate the core semantic elements of the question, providing a foundation for subsequent question vector generation and semantic similarity comparison.

[0022] S110: Obtain the semantic entropy value corresponding to each tag in the preset dynamic semantic library, and obtain the global semantic entropy value of the preset dynamic semantic library based on the semantic entropy values ​​of all tags. Semantic entropy, based on information entropy theory, is a quantitative indicator used to characterize the degree of semantic uncertainty of a corresponding tag. It's important to note that a higher semantic entropy value indicates greater ambiguity and semantic vagueness in the term or tag, while a lower value indicates clearer and more certain semantics. Specifically, it can be a value ranging from 0 to positive infinity, calculated by statistically analyzing the semantic state probability distribution of the text associated with the semantic term. For example, the semantic entropy value of the tag "audit" is 0.2, indicating highly clear semantics, while the semantic entropy value of the tag "system problem" is 3.7, indicating highly ambiguous semantics.

[0023] The global semantic entropy value refers to the overall quantitative indicator calculated from the semantic entropy values ​​of all tags in the preset dynamic semantic library. It is used to characterize the overall semantic uncertainty level and semantic quality of the entire semantic library. The higher the global semantic entropy value, the more ambiguous tags exist in the semantic library, and the more chaotic the semantic distribution. Specifically, it can be the arithmetic mean, weighted average, or sum of the semantic entropy values ​​of all tags. For example, the global semantic entropy value of a company's customer service semantic library is 1.6, which can indicate good semantic quality. When a large number of new business tags are added, the global semantic entropy value rises to 2.8, which indicates that entropy reduction optimization is needed.

[0024] Specifically, as an example, after obtaining the semantic entropy values ​​of all tags, the global semantic entropy value of the preset dynamic semantic library is calculated to be 2.3 by weighted averaging, thereby assessing the overall semantic uncertainty level of the entire semantic library.

[0025] S120: In response to the global semantic entropy value exceeding the preset global entropy value threshold, cluster the corresponding tags whose semantic entropy value is higher than the preset term entropy value threshold to obtain the clustered tag vector; The preset global entropy threshold is a critical value used to determine whether the semantic entropy reduction optimization process needs to be triggered. It serves as a standard for evaluating whether the overall semantic quality of the semantic library meets business requirements. Specifically, it can be a fixed value manually set according to the accuracy requirements of different business scenarios, or it can be an adaptive value dynamically adjusted according to changes in business data. For example, in a bank customer service scenario, to ensure the accuracy of responses, the preset global entropy threshold is set to 2.0.

[0026] Specifically, as an example, the calculated global semantic entropy value of 2.3 is compared with the preset global entropy threshold of 2.0. The global semantic entropy value exceeds the threshold, indicating that there are many ambiguous tags in the semantic database, requiring the initiation of a semantic entropy reduction optimization process. In this process, high-entropy tags with semantic entropy values ​​higher than the preset term entropy threshold of 1.5 are selected; for example, a total of 427 high-entropy tags are obtained. These high-entropy tags can then be clustered using a density-based clustering algorithm.

[0027] S130: Based on the clustered label vectors that meet the preset label density threshold, generate the target label vector and save the target label vector to the preset dynamic semantic library. The preset label density threshold is adjusted based on the label generation time, label usage frequency and time decay factor. The preset tag density threshold is a threshold condition used to filter effective semantic clusters in the clustering results. It can be determined based on at least one of the following: the number of tags within a cluster, the average similarity within a cluster, the tag generation time, the tag usage frequency, and the time decay factor. It is used to determine whether the tag vectors after clustering meet the conditions for generating the target tag vector. This threshold can be adjusted according to the tag generation time, tag usage frequency, and time decay factor. For example, for popular tags generated within one month and used more than 1,000 times, the preset tag density threshold is adjusted to 0.6; for unpopular tags generated more than one year ago and used less than 10 times, the preset tag density threshold is adjusted to 0.8.

[0028] The target tag vector refers to the low-entropy tag vector that meets the preset tag density threshold after clustering high-entropy tags with semantic entropy values ​​higher than the preset term entropy threshold. It is an optimized replacement vector for the original high-entropy tags, keeping the entire semantic database in a low-entropy state. For example, after clustering and merging the three high-entropy tags "work order not pushed", "work order delayed", and "work order abnormal", the resulting low-entropy tag "work order pushed abnormal" is a d-dimensional vector.

[0029] Specifically, as an example, during the clustering process, the preset label density threshold is dynamically adjusted based on the label generation time, usage frequency, and time decay factor. For instance, 82 cluster label vectors that meet the corresponding dynamically adjusted preset label density threshold are selected. These selected cluster label vectors are used as new low-entropy, high-determinism label vectors to replace the original 427 high-entropy label vectors. The updated target label vector set is then saved to the preset dynamic semantic library, completing the entropy reduction optimization of the semantic library and restoring the entire semantic library to a state of full low-entropy, high-determinism.

[0030] S140: Based on the question vector and the preset dynamic semantic library, a similarity comparison is performed to obtain multiple target labels that meet the preset similarity conditions. The multiple target labels are used to generate the corresponding response results.

[0031] The target label refers to a semantically relevant and semantically deterministic label selected by calculating the semantic similarity between the question vector and the label vectors corresponding to all low-entropy labels in a preset dynamic semantic database. This label is used to match the standard response corpus. All low-entropy labels can include the original low-entropy labels that have not been clustered, as well as the new low-entropy labels corresponding to the target label vector generated after clustering. Specifically, it can be a single label or a combination of multiple labels. For example, for the user question "Why hasn't my work order been pushed to the audit department yet?", the target labels selected from the low-entropy labels in the entire database are "work order push" and "audit process".

[0032] Specifically, as an example, the cosine similarity between the d-dimensional question vector generated in step S100 and the tag vectors corresponding to all low-entropy tags in the preset dynamic semantic library is calculated to obtain the semantic relevance between each low-entropy tag and the real-time question to be answered. In practical applications, semantic similarity calculation aims to measure the semantic closeness between two texts, such as the closeness of words, phrases, sentences, and paragraphs in two texts. It requires extracting deep semantic information from the surface text and quantifying the similarity of different language structures. Therefore, this application uses a text vectorization model that is exactly the same as or in the same vector space as the question vector generation to generate tag vectors for all low-entropy tags. Each low-entropy tag text in the preset dynamic semantic library is input into the text vectorization model. After word segmentation, encoding, and pooling, the corresponding d-dimensional tag vector is obtained. The cosine similarity function is used to calculate the similarity score between the question vector and each tag vector. The score ranges from -1 to 1, with the closer to 1 indicating greater semantic similarity.

[0033] like Figure 3 As shown in the figure, the tag vector retrieval range corresponding to the question "Hazard reporting work orders dispatched before August 26th were not pushed to the audit system". The core semantic terms extracted for this question are "audit", "work order", and "dispatch". When performing similarity comparison, the set of related questions corresponding to these three tags will be retrieved first. For example, the questions associated with the tag "audit" include "Login prompt that the account has not been audited", "Work orders were not pushed to the audit system", and "How to determine whether the audit result is passed in the AI ​​audit process of monitoring work orders". The tag vector retrieval mechanism based on core semantic terms can effectively narrow the search range, reducing the computational workload of originally needing to traverse the entire database of tags to only needing to calculate dozens of tags associated with the core semantic terms, which greatly improves the efficiency of semantic matching.

[0034] The semantic similarity calculation results show that the original low-entropy label "work order push" has a similarity of 0.92, the original low-entropy label "audit process" has a similarity of 0.87, and the new low-entropy label "work order push anomaly" generated in this clustering has a similarity of 0.63. All low-entropy labels are sorted from high to low semantic similarity, and the top two target labels, "work order push" and "audit process," are selected as target labels that meet the preset similarity criteria.

[0035] In practical applications, semantic similarity calculations can accurately pinpoint the most relevant content from multiple candidate problem descriptions to match the user's problem. For example, the problem description to be matched is "Hazard reporting work orders dispatched before August 26th were not pushed to the audit system." The system needs to identify the semantically closest description from seven candidate options: "Login prompt: Account not audited," "Work order not pushed to the audit system," "How to determine if the audit result is passed in the AI ​​audit stage of monitoring work orders," "Inconsistent VLANs between inner and outer layers of the work order," "Emergency platform unable to dispatch emergency drill work orders," "Inbound customer order dispatch product number lacks resource CID," and "Residential customer order dispatch shows executor execution failure." By calculating the semantic similarity between the problem to be matched and each candidate option, the option with the highest semantic similarity, "Work order not pushed to the audit system," is identified as the most relevant description. This precise matching and locking mechanism effectively improves the relevance of the response content, ensuring that the output response accurately meets the user's actual needs. Based on these two target tags, the system matches the corresponding standard response content in the pre-set internal customer service response corpus and outputs the response result to the user: "Hello, it usually takes 1-2 business days for the work order to be reviewed by the audit department. If it has not been pushed after the time limit, please provide the specific work order number, and we will immediately check the work order status and follow up on the processing."

[0036] In this embodiment, real-time questions awaiting response are acquired, and corresponding question vectors are obtained based on these questions, transforming unstructured user questions into standardized dense semantic vectors. The semantic entropy value corresponding to each tag in a preset dynamic semantic library is acquired, and the global semantic entropy value of the preset dynamic semantic library is obtained based on the semantic entropy values ​​of all tags, quantifying the uncertainty of individual tags relative to the global semantic library. In response to the global semantic entropy value exceeding a preset global entropy threshold, corresponding tags with semantic entropy values ​​higher than a preset term entropy threshold are clustered, accurately locating semantic ambiguity and redundancy. Target tag vectors are generated based on the clustered tag vectors that meet a preset tag density threshold, and these target tag vectors are saved to the preset dynamic semantic library, continuously maintaining a high degree of determinism in the semantic library. A similarity comparison is performed between the question vector and the preset dynamic semantic library to obtain multiple target tags that meet preset similarity conditions, which are used to generate corresponding response results, outputting accurate responses that match the user's true intent. This improves the accuracy of intelligent customer service responses.

[0037] In any of the above embodiments, optionally, before obtaining the semantic entropy value corresponding to each tag in the preset dynamic semantic library, the method further includes: S1101: Based on historical business data, obtain the question text, and build a user question corpus based on the question text; Historical business data refers to the collection of raw data related to historical business operations. Specifically, it can include customer service dialogue logs, work order system records, internal business knowledge base documents, user feedback, product user manuals, and frequently asked questions documents. For example, it can include customer service chat logs, historical work order data, and business knowledge base documents.

[0038] Among them, the question text refers to the natural language question content extracted from historical business data, which is the smallest processing unit for semantic analysis and tag generation. Specifically, it can be the user question part in customer service dialogue, the work order title and problem description in the work order system, the problem statement in user feedback, the question part in the FAQ, etc. For example, it can be "Why hasn't my work order been pushed to the audit department yet?", "How do I reset my system login password?", "What documents are required for issuing an invoice?"

[0039] Among them, the user question corpus refers to a structured data warehouse that centrally stores all cleaned and standardized question texts. Specifically, it can be a collection of question texts classified by business line, a historical question database with timestamps and user identifiers, or a distributed question corpus cluster.

[0040] In practical applications, data acquisition builds a dynamic semantic database by integrating heterogeneous data from multiple sources. Raw data is obtained from various channels, including real-time dialogue logs, work order records, structured knowledge graphs, and publicly available external data. For example, historical business data from the past 12 months can be collected from an enterprise's customer service system, work order system, and internal knowledge base, including 870,000 customer service chat records, 320,000 historical work order records, and 1,500 internal business knowledge base documents. In practical applications, data cleaning is the core step in the preprocessing layer. It involves in-depth analysis and processing of the data involved in the large model, performing cleaning, deduplication, normalization, and standardization operations, properly handling outliers and missing values, and finely filtering and correcting based on domain-specific knowledge bases or rule sets to improve data quality. Data cleaning technology plays a crucial role in the data processing workflow. Its core purpose is to discover and correct errors, inconsistencies, or incompleteness in the dataset. With the help of data cleaning, erroneous data can be effectively removed, missing data values ​​filled, and outlier data handled, thereby improving the accuracy and reliability of the data.

[0041] In practical applications, the Apache NiFi method can also be used to integrate multi-source data, including text, language, and images, unifying heterogeneous data formats and establishing data pipelines. The specific implementation process is as follows: Import the data flow builder tool DataFlowBuilder; create a new data flow instance; add a Kafka data source, subscribing to the raw customer service data on the topic "customer_service_raw"; add a character encoding conversion processor to convert the raw data's encoding format from GBK to UTF-8, resolving Chinese character encoding issues; add a JSON field extraction processor to extract the "dialog_content" field from the raw data, representing the user's actual dialogue content; add an HDFS data receiver to save the cleaned data to the " / cleaned / phase1" path in HDFS; execute the constructed data flow to complete batch data cleaning. The above standardized preprocessing was applied to the collected historical business data to remove duplicate, invalid, and sensitive information, extracting all the actual user-submitted question texts, resulting in 650,000 standardized question texts. Based on these question texts, an internal user question corpus can be built for the enterprise and stored in categories according to business lines such as customer service, operations and maintenance, finance, and human resources.

[0042] S1102: Based on the semantic information of each question text in the user question corpus, obtain the corresponding semantic terms, and based on the semantic terms, obtain the corresponding question tags; The semantic information of the question text refers to the content information contained in the question text that can express the user's true intention. Specifically, it can include the topic of the question, core entities, actions, attributes, relationships, time, location, etc. For example, for the question text "Why hasn't my work order been pushed to the audit department yet?", its semantic information includes the topic "work order push", the core entities "work order" and "audit department", the action "push", and the question point "why it hasn't been completed".

[0043] Semantic terms refer to words or phrases extracted from the semantic information of the question text that can independently express complete semantic units, serving as an intermediary connecting the question text and the question tag. Specifically, they can be nouns, verbs, noun phrases, verb phrases, proper nouns, etc. For example, they could be terms extracted from the aforementioned question text such as "work order," "push notification," "audit department," "login password," and "invoice issuance."

[0044] Specifically, as an example, semantic analysis is performed on each question text in the user question corpus to extract the core semantic information and corresponding semantic terms contained in each question text, and an automated tag generation operation is performed.

[0045] In practical applications, a label semantic entropy reduction technique based on vector models is adopted. The problem text is encoded using BERT, and similar problems are clustered using K-means or DBSCAN algorithms. For example, "login failed" and "unable to log in" are clustered into "login abnormal". Each cluster center generates a core label, and additional mandatory labels are set for domain-specific terms such as "VoLTE" and "IoT card" to avoid model omissions.

[0046] During the issue tagging process, the length of each tag should preferably be 2-6 Chinese characters or a business phrase, and the use of synonymous summaries is allowed to ensure the conciseness of the tags; the tag content should closely revolve around the core content of the issue description to ensure the accuracy and relevance of the tags; the tag content should be included in the description of the issue; the number of tags generated for one issue should preferably be controlled within 5.

[0047] In practical applications, for question 1 "What is deep learning?", after extracting the core semantic information, the corresponding tags are generated as "machine learning" and "neural network". For question 2 "How to train a neural network?", the corresponding tags are generated as "neural network" and "training". The generated tags strictly follow the tag generation principles. The length of each tag is preferably 2-6 Chinese characters or a business phrase. Synonymous summarization tags are allowed. The tag content closely revolves around the core content of the question description. The number of tags generated for each question is controlled within 5, which can accurately and concisely represent the core semantics of the question.

[0048] S1103: Establish a question tag mapping relationship based on the semantic information of each question text, the corresponding semantic terms, and the corresponding question tags. Based on the question tag mapping relationship, generate a preset dynamic semantic library.

[0049] The issue tag mapping relationship refers to the many-to-many association established between issue text, semantic terms, and issue tags. It represents the semantic relationships between different data and is the core data structure of the pre-defined dynamic semantic database. Specifically, it can be a relational table in a relational database, a node and edge connection in a graph database, or a mapping relationship in a key-value pair store. For example, it could be the association between the issue text "Why hasn't my work order been pushed to the audit department yet?" and the semantic terms "work order," "push," and "audit department," as well as the issue tags "work order push" and "audit process."

[0050] In practical applications, dynamic semantic libraries can continuously learn from user dialogues and industry terminology, dynamically updating semantic representations and contextual associations. This enables accurate identification of user intent, such as clarifying whether the ambiguous expression "apple" refers to a brand or a fruit, and supports semantic coherence in multi-turn conversations, such as tracing back to the order number mentioned above. This improves response accuracy, reduces human intervention, and optimizes service strategies through sentiment analysis, providing users with a more intelligent, flexible, and up-to-date interactive experience. By establishing a mapping relationship between each question text and its corresponding semantic terms and question tags, and storing all question texts, semantic terms, question tags, and their mapping relationships in a pre-defined dynamic semantic library, a foundational data support is provided for subsequent semantic entropy calculations and intelligent customer service responses.

[0051] In this embodiment, historical business data is collected to obtain question text, and a user question corpus is constructed based on the question text, providing comprehensive and reliable basic corpus support for subsequent semantic analysis. Automated tag generation based on vector models replaces traditional manual annotation methods, significantly improving tag generation efficiency and semantic consistency, while ensuring the core semantic representativeness and business adaptability of the tags. A multi-dimensional association mapping relationship is established between question text, semantic terms, and question tags. The resulting dynamic semantic library can continuously learn business knowledge and dynamically update semantic representations, providing an accurate semantic benchmark for subsequent semantic similarity matching and intelligent response, and overall improving the intent recognition accuracy and business scenario generalization capability of the intelligent customer service system.

[0052] In any of the above embodiments, optionally, obtaining the semantic entropy value corresponding to each tag in a preset dynamic semantic library includes: S1104: For each semantic term corresponding to a preset tag, count the set of semantic states of all question texts associated with the semantic term in the preset dynamic semantic library, and the probability of each semantic state in the set of semantic states appearing in the associated question texts; The semantic state set refers to the complete set of all different semantic meanings expressed by the same semantic term in different contexts and business scenarios, used to comprehensively characterize the semantic ambiguity range of the semantic term. Specifically, it can be the set of all meanings of a polysemous word, the set of different meanings of the same professional term in different business processes, the set of different referents of the same phrase in different contexts, etc. For example, the semantic state set of the semantic term "system" can include IT system, financial system, work order system, and human resources system.

[0053] Semantic state refers to a specific and clear semantic meaning expressed by the same semantic term in a particular context and business scenario. It is the basic unit that constitutes the set of semantic states. Specifically, it can be an independent meaning of a polysemous word, a specific business meaning of a professional term, or the specific reference of the same phrase in a particular context. For example, in the problem text "System login failed", the semantic state corresponding to the semantic term "system" is "IT system".

[0054] The probability of a semantic state appearing in associated question texts refers to the proportion of the frequency of a particular semantic state appearing in all question texts associated with that semantic term out of the total frequency of associations for that semantic term. This quantifies the prevalence of the semantic state. Specifically, it can be based on a statistically obtained value between 0 and 1. A higher value indicates a more common semantic state, while a lower value indicates a rarer semantic state. For example, if the semantic term "system" is associated with 1000 question texts, and 600 of them correspond to the semantic state "IT system," then the probability of this semantic state appearing is 0.6.

[0055] S1105: Based on the number of semantic states and the probability of occurrence of each semantic state, obtain the semantic entropy value corresponding to the semantic term, and obtain the semantic entropy value corresponding to each tag based on the semantic entropy value corresponding to the semantic term.

[0056] The number of semantic state sets refers to the total number of different semantic states corresponding to the same semantic term, which directly determines the basic semantic ambiguity of the semantic term. Specifically, it can be a positive integer greater than or equal to 1. The more states there are, the stronger the ambiguity of the semantic term. A number of 1 indicates that the semantic term is unambiguous. For example, the semantic term "audit" has a semantic state set of 1, meaning that this semantic term only refers to business auditing, while the semantic term "system" has a semantic state set of 4.

[0057] In practical applications, a semantic entropy calculation model is constructed based on information entropy theory. For a single semantic term, the specific formula for calculating its information entropy based on its feature information is as follows: H(X) = i=1 n p(x i log2p(x) i ); Where X represents a semantic term, n is the number of possible semantic states of the term, and p(x) i Let be the probability of the i-th semantic state occurring. For the entire semantic database, the overall entropy value of the semantic database is obtained by comprehensively calculating the entropy value of each term, which reflects the degree of uncertainty of the information in the semantic database.

[0058] Based on the semantic entropy calculation model described above, the semantic entropy value of each semantic term is calculated one by one. For example, the semantic entropy value of the semantic term "system" is calculated to be 1.57, while the semantic entropy value of the semantic term "audit" is 0. The semantic entropy values ​​of all tags are calculated and the results are saved to a preset dynamic semantic library. The entropy values ​​of all tags are then calculated using a weighted average method to obtain the global semantic entropy value of the preset dynamic semantic library. This value reflects the overall uncertainty of the information in the current semantic library, providing a basis for subsequent judgments on whether to trigger the semantic entropy reduction optimization process.

[0059] In this embodiment, by statistically analyzing the set of semantic states of the question text associated with each semantic term and the probability of occurrence of each semantic state, it is possible to comprehensively and systematically capture all semantic meanings of the same semantic term in different business scenarios and contexts, fully covering the semantic ambiguity range of semantic terms, and providing comprehensive and accurate basic data for the quantitative assessment of semantic uncertainty. Based on information entropy theory, a semantic entropy value calculation model is constructed, transforming the abstract degree of semantic ambiguity into a quantifiable objective indicator, achieving a precise assessment of the semantic certainty of each tag. This quantitative assessment method solves the problem that traditional semantic libraries cannot objectively measure semantic quality, accurately locating highly ambiguous tags, providing a clear goal and basis for the subsequent semantic entropy reduction optimization process, effectively improving the overall semantic certainty of the dynamic semantic library, and enhancing the intent recognition accuracy and response reliability of the intelligent customer service system.

[0060] In any of the above embodiments, optionally, after obtaining the semantic entropy value corresponding to each tag in the preset dynamic semantic library, and obtaining the global semantic entropy value of the preset dynamic semantic library based on the semantic entropy values ​​of all tags, the method further includes: S1106: When the preset dynamic semantic library meets the preset update triggering conditions, determine the semantic entries affected by the update; Among them, semantic terms affected by updates refer to the set of objects whose semantic entropy values ​​need to be recalculated when any form of update operation occurs in the preset dynamic semantic library. Specifically, these can be newly added semantic terms, semantic terms whose content has been modified, semantic terms whose related question text quantity or content has changed, semantic terms whose semantic relationship connection weight has been adjusted, and deleted semantic terms, etc. For example, it could be the newly added semantic term "electronic invoice", the semantic term "work order push" whose related question text has increased by 500, or the semantic term "system" whose connection weight with "IT system" has been adjusted.

[0061] In practical applications, the preset dynamic semantic library is continuously updated. When the preset dynamic semantic library is updated, the semantic entropy value of the affected terms and the global semantic entropy value are recalculated. If the updated global semantic entropy value exceeds the preset threshold, the high-entropy tag clustering process is restarted to maintain the low-entropy and high-determinism state of the semantic library and ensure the long-term accuracy and stability of intelligent customer service responses.

[0062] S1107: Recalculate the semantic entropy value corresponding to the affected semantic terms. Based on the updated semantic entropy values ​​of all semantic terms in the preset dynamic semantic library, recalculate the global semantic entropy value of the preset dynamic semantic library. Specifically, as an example, the preset dynamic semantic library can be triggered by incremental data updates or by semantic relationship optimization. Taking the incremental data update scenario as an example, the semantic entropy value of each affected semantic term is recalculated according to the semantic entropy calculation method in S1104-S1105. For example, the semantic entropy value of the newly added semantic term "electronic invoice" is 0.3, indicating clear semantics; the semantic entropy value of "online reimbursement" is 2.1, indicating ambiguous semantics; the semantic entropy value of the existing semantic term "invoice issuance" increases from 0.8 to 1.2; the semantic entropy value of "work order push" increases from 0.5 to 0.7; and the semantic entropy value of "password reset" remains unchanged at 0.4. Based on the updated semantic entropy values ​​of all semantic terms, the global semantic entropy value of the preset dynamic semantic library is recalculated to be 2.1. Comparing this updated global semantic entropy value with the preset global entropy value threshold of 2.0, it is found that the global semantic entropy value has exceeded the threshold, and the semantic entropy reduction optimization process needs to be initiated.

[0063] S1108: In response to the updated global semantic entropy value exceeding the preset global entropy value threshold, cluster the corresponding tags whose updated semantic entropy value is higher than the preset term entropy value threshold to obtain the updated target tag vector.

[0064] In practical applications, a density-based clustering algorithm is used to cluster regions with high semantic entropy values ​​based on similar semantics. The similarity between semantic terms is calculated using their feature vectors, and terms with similarity exceeding a set threshold are grouped into the same cluster. For each cluster, terms with similar semantics are merged, retaining representative semantic expressions and eliminating semantic redundancy. For example, terms with similar semantics such as "login failed," "unable to log in," and "login error" are merged into "login error," reducing the entropy value of the semantic database.

[0065] Specifically, as an example, high-entropy tags with updated semantic entropy values ​​higher than a preset term entropy threshold of 1.5 are selected. A density-based clustering algorithm is then used to cluster these high-entropy tags. During clustering, clustering can be based on the cosine similarity between tag vectors, and semantic clusters meeting the conditions are selected according to a preset tag density threshold. For example, when the average intra-cluster similarity of a cluster is higher than 0.7, and the number of tags within the cluster meets a preset quantity condition, the cluster is determined as a valid semantic cluster. For each semantic cluster, all tags with similar semantics are merged, and the most representative semantic expression is retained as a new low-entropy tag. For example, four semantically similar high-entropy tags, "work order not pushed," "work order delayed," "work order flow abnormal," and "work order stuck," are merged into one semantic cluster, retaining "work order push abnormal" as the representative tag of this cluster. After clustering, these new low-entropy target tag vectors are saved to a preset dynamic semantic library, replacing the original high-entropy tag vectors. After the update is completed, the system recalculates the global semantic entropy value. When the global semantic entropy value drops below the preset global entropy value threshold, the semantic database optimization after this data incremental update is completed, which effectively eliminates semantic redundancy and improves the overall determinism and retrieval efficiency of the semantic database.

[0066] In this embodiment, by identifying the semantic terms affected by the update when the preset dynamic semantic library meets the preset update trigger conditions, computational overhead is reduced, ensuring the real-time and targeted nature of the update. By recalculating the semantic entropy values ​​corresponding to the affected semantic terms, the semantic quality changes of the semantic library can be perceived in real time, accurately reflecting the overall uncertainty level of the semantic library after the update. When the global semantic quality declines, an entropy reduction optimization process can be automatically initiated to eliminate semantic redundancy and ambiguity, continuously maintaining a low-entropy, high-determinism state of the semantic library. This solves the problems of traditional static semantic libraries being unable to adapt to business iterations and the accumulation of semantic ambiguity leading to a decrease in response accuracy, thus improving the long-term stability and business adaptability of the intelligent customer service system.

[0067] In any of the above embodiments, optionally, when the preset dynamic semantic library meets the preset update triggering conditions, the semantic terms affected by the update are determined, including: S1106a: When incremental historical business data is detected imported into the user question corpus, the newly added question text is obtained based on the incremental historical business data, and the semantic terms corresponding to the newly added question text are identified as the semantic terms affected by the update. Incremental historical business data refers to various business data sets that have not yet been included in the user question corpus within a preset time period. It is the data source that triggers incremental updates to the preset dynamic semantic database. Specifically, it can include newly generated customer service dialogue logs, work order system records, user feedback, product update documents, newly released business rules, etc., which may be 12,000 new customer service chat records generated by a company in the past week.

[0068] The newly added question text refers to user question content extracted from incremental historical business data that was not previously present in the user question corpus. Specifically, this can include new customer service user questions, question descriptions in new work orders, and question expressions in new user feedback that were not previously included.

[0069] S1106b: Obtain a preset domain knowledge graph and analyze the semantic network structure between semantic entries in the preset dynamic semantic library based on the domain knowledge graph; The pre-defined domain knowledge graph refers to a structured knowledge network used to represent all core concepts, entities, and their standard semantic relationships within a specific business domain. It serves as the basis for judging the rationality of the semantic network structure. Specifically, it can be an enterprise's internal business knowledge graph, an industry-wide general knowledge graph, or a vertical domain's professional knowledge graph. For example, it could be an enterprise's pre-built internal business knowledge graph, which contains thousands of business entities such as "work orders," "audits," and "invoices," as well as the standard relationships between them.

[0070] The semantic network structure refers to a mesh data structure composed of all semantic terms in a pre-defined dynamic semantic database as nodes and the semantic relationships between semantic terms as edges, used to represent the semantic connections between different semantic terms. For example, it can be a semantic network composed of semantic term nodes such as "work order", "push", and "audit department" and related edges such as "work order-push" and "push-audit department".

[0071] S1106c: Identify conflicting semantic relations in the semantic network structure and determine the semantic terms corresponding to the conflicting semantic relations as the semantic terms affected by the update.

[0072] Conflicting semantic relationships refer to unreasonable semantic relationships in the semantic network structure of the pre-defined dynamic semantic library that are inconsistent with the pre-defined domain knowledge graph, logically contradictory, semantically overlapping, or incorrectly associated. Specifically, these can be incorrect hierarchical relationships, incorrect synonymous relationships, contradictory causal relationships, or duplicate associations. For example, "electronic invoice" could be incorrectly set as a hierarchical term of "paper invoice," or "system problem" and "network failure" could be incorrectly marked as synonyms.

[0073] In this embodiment, by detecting the import of incremental historical business data and extracting semantic terms corresponding to newly added problem texts, newly generated business data can be promptly incorporated into the semantic database system, adapting to the dynamic iteration of business and the expansion of emerging scenarios. By introducing a pre-defined domain knowledge graph to comprehensively analyze the semantic network structure, and by identifying conflicting semantic relationships and locating corresponding semantic terms, logical contradictions and association errors in the semantic database can be accurately discovered and corrected. This covers both data increment and semantic structure update needs, and only affected semantic terms are processed subsequently, effectively avoiding the accumulation of semantic ambiguity and ensuring the semantic consistency and business adaptability of the dynamic semantic database.

[0074] In any of the above embodiments, optionally, real-time pending questions are obtained, and based on the real-time pending questions, a corresponding question vector is obtained, including: S1001: Based on the real-time unanswered question, multiple candidate semantic terms are obtained; Among them, candidate semantic terms refer to the set of all words or phrases that can independently express semantic units, extracted from the preliminary word segmentation of the real-time question text. This is the original candidate set for selecting core semantic features. It can include core semantic terms that accurately express the user's true intent, as well as noisy terms such as stop words, modal particles, and conjunctions that contribute little to semantic expression. Specifically, it can include single Chinese characters, two-character words, multi-character words, phrases, proper nouns, etc.

[0075] S1002: Determine the importance of each candidate semantic term, filter out candidate semantic terms whose importance is lower than the preset importance threshold, and obtain the filtered semantic terms. The importance is positively correlated with the number of times the candidate semantic term appears in the real-time question to be answered, and negatively correlated with the frequency of the candidate semantic term in the user question corpus. The importance of candidate semantic terms refers to their contribution to expressing the user's true intent. A higher importance indicates that the semantic term better represents the core semantics of the question, while a lower importance indicates a lower semantic contribution and is considered noise. Specifically, the importance can be a value between 0 and 1. The importance is positively correlated with the number of times the candidate semantic term appears in the real-time pending questions and negatively correlated with the frequency of the candidate semantic term in the user question corpus. For example, the importance of the candidate semantic term "work order" is 0.85, and the importance of the candidate semantic term "why" is 0.12.

[0076] In practical applications, the TF-IDF (Term Frequency-Inverse Document Frequency) algorithm is mainly used to evaluate the importance of words in a question to a corpus. The importance of a word increases with its frequency in the question, but decreases with its frequency in the corpus. Then, the TF-IDF statistical algorithm is applied to each tag phrase for compression, thereby extracting the most suitable phrase tags. The specific formula for calculating TF-IDF is: ; in, This represents the final importance score of term t in the question text d; This represents the term frequency, which is the number of times term t appears in the question text d. Document Frequency (N) represents the number of question texts containing term t in the user question corpus; N represents the total number of question texts in the user question corpus.

[0077] The final formula for calculating importance is: ; in, It consists of two parts. The first part is Term Frequency (TF), which is the number of times a given word appears in the question. Its calculation formula is as follows: ; in, TF indicates the number of times a term appears in a question. TF is a commonly used metric in information retrieval and text mining to assess the importance of a word. The TF value is usually normalized to avoid bias towards longer questions.

[0078] The second part is Inverse Document Frequency (IDF). Document frequency refers to the prevalence of a specific word. It can be obtained by dividing the total number of questions by the number of questions containing that word, and then taking the logarithm to the base 10 of the quotient. The formula is as follows: ; in, Indicates the number of questions. Indicates that the term is included. The number of questions. If it contains entries. The fewer problems, The larger the value, the more significant the entry. It has a strong ability to distinguish between categories. If the number of questions in a certain category C containing the term t is m, and the total number of questions in other categories containing t is k, then the total number of questions containing t is obviously greater than or equal to the total number of questions containing t. When m is large, n is also large, and the IDF value obtained according to the IDF formula will be small, which means that the t tag of the term has weak distinguishing ability.

[0079] TF-IDF tends to filter out common words and retain important words. The system calculates the importance of each candidate semantic term according to the above formula, compares the importance of each candidate semantic term with the preset importance threshold of 0.1, and filters out noisy semantic terms with an importance lower than 0.1 such as "my", "why", "still", "not yet", and "to". It retains core semantic terms with an importance higher than the threshold such as "work order", "push", and "audit department", and completes the compression of the prepared tag phrases to extract the most suitable core word phrase tags.

[0080] S1003: Based on the filtered semantic terms, obtain the question vector corresponding to the real-time question to be answered.

[0081] In this embodiment, by segmenting the real-time questions to be answered, all possible semantic units are comprehensively extracted to form a set of candidate semantic terms, ensuring that no potential core semantic information is missed. The semantic contribution of each candidate semantic term is quantified based on the correlation characteristics of word frequency and inverse document frequency. This effectively filters out noise information such as stop words and interjections that contribute little to expressing the user's intent, accurately retains semantic terms that reflect the core of the question, and transforms the filtered core semantic terms into a question vector of a unified dimension. This ensures the integrity of semantic information and effectively removes redundant interference. The generated question vector can accurately represent the user's true intent, providing a reliable semantic foundation for subsequent tag retrieval and intelligent response.

[0082] Figure 4 This is a schematic diagram illustrating the processing procedure of an intelligent customer service response method provided in an embodiment of this application, as detailed below: S300: User input; S310: Semantic preprocessing layer; In this step, the semantic preprocessing layer transforms the unstructured raw text input by the user into structured numerical features that can be recognized and computed by a computer, including word segmentation, stop word removal, TF-IDF importance evaluation, core semantic term extraction, and semantic encoding.

[0083] S320: Tag Clustering and Entropy Reduction Module; In this step, the label clustering and entropy reduction module maintains the fully low-entropy, highly deterministic state of the preset dynamic semantic library in real time, including semantic entropy value calculation, global semantic quality assessment, high-entropy label clustering, and target label vector update. At the same time, it receives input from the feedback closed-loop module to continuously optimize the semantic library structure.

[0084] S330: Hybrid search engine; In this step, the hybrid retrieval engine compares the question vector generated by the semantic preprocessing layer with the low-entropy label vector maintained by the label clustering and entropy reduction modules to select the most relevant target labels. It adopts an indexing mechanism based on core semantic terms and a cosine similarity calculation method, which greatly improves retrieval efficiency while ensuring retrieval accuracy.

[0085] S340: Adaptive Learning Engine; In this step, the adaptive learning engine dynamically adjusts various parameters based on historical interaction data and user feedback, enabling it to adapt to changes in different business scenarios and user needs. Its core functions include retrieval weight adjustment, similarity threshold optimization, and self-learning of tag clustering parameters.

[0086] S350: Response generation module; In this step, the response generation module generates and outputs the final response result based on the target label output by the adaptive learning engine. Its core functions include standard response matching, context fusion, and personalized response generation.

[0087] S360: Feedback closed-loop module.

[0088] In this step, the feedback loop module collects user feedback on their responses and passes the feedback data to the label clustering and entropy reduction module and the adaptive learning engine for continuous optimization.

[0089] Figure 5 This is a schematic diagram of the large model processing procedure provided in the embodiments of this application, which is described in detail below: S1: The user submits a question, and the system performs text preprocessing on the received question; In this step, the user can input a question through an intelligent customer service interface, such as a webpage, mobile application, or chatbot. After receiving the question, the system will preprocess it using natural language processing technology.

[0090] S2: The large model extracts the labels of the problem by understanding the problem and performs corpus matching; This step may involve accessing a dynamic corpus database containing a large amount of tagged data. The question intent and key elements output by the large model are used as search criteria, and a tag semantic entropy subtraction method is employed for matching. This method reduces the search space and improves retrieval efficiency through intelligent aggregation and entropy reduction of tags.

[0091] S3: Further identify the user's intent through large model intent recognition and match it in the candidate standard corpus; In this step, the candidate question set matched by the corpus in S2 can be input into a pre-trained large model, and the language integration capability of the large model can be used to perform semantic analysis to find the question that best matches the user's intent.

[0092] S4: Output the answer.

[0093] In this step, specifically, after matching the corresponding question, the answer is extracted from the corpus database and post-processed, such as format adjustment and information supplementation, to ensure that the answer is accurate and easy to understand, and finally displayed to the user through the user interface.

[0094] Figure 6 This is a flowchart of the problem vector generation and label matching process provided in an embodiment of this application, which is described in detail below: S400: Start; S401: Enter the question text; S402: Text Vectorization Model; In this step, a text vectorization model can be used to convert the input question text sequence into a fixed-length dense vector representation. The text vectorization model can be a Transformer-based pre-trained language model, a Sentence-BERT model, or other sequence-to-vector models capable of mapping text sequences to fixed-dimensional vectors. This model is used to extract contextual semantic features from the question text, providing a vector basis for subsequent similarity calculations.

[0095] S403: Embedding layer: Convert to dense vector representation; In this step, the embedding layer is used to convert discrete words, subwords, or character units in the input question text into continuous dense vector representations, enabling the model to perform numerical processing of natural language text. The embedding layer can use pre-trained word vectors, subword embeddings, or embedding representations from pre-trained language models.

[0096] S404: Encoding layer: Transformer-based context semantic encoding; In this step, the encoding layer is used to perform contextual semantic encoding on the vector sequence output by the embedding layer, capturing the relationships between semantic units, word order information, and overall semantic intent in the question text. Specifically, a Transformer-based encoding structure or a sequence encoding structure such as a bidirectional recurrent neural network can be used to generate a hidden state sequence containing contextual semantic features.

[0097] S405: Pooling layer: generates fixed-length sentence vectors; In this step, the pooling layer aggregates the hidden state sequence output by the encoding layer to generate a fixed-length question vector. Specifically, average pooling, max pooling, [CLS] labeled vectors, attention pooling, etc., can be used. The dimension of the generated question vector is determined by the text vectorization model employed.

[0098] S406: Label representation: Generate label vectors; In this step, a text vectorization model with the same or consistent vector space as the question text is used to convert the tag text in the pre-set dynamic semantic library into corresponding tag vectors. The tag vectors and question vectors are in the same semantic vector space to ensure the effectiveness and accuracy of subsequent similarity calculations.

[0099] S407: Calculate semantic similarity; In this step, the semantic similarity between the question vector and each tag vector in the preset dynamic semantic library is calculated to measure the semantic similarity between the real-time question to be answered and each tag. Specifically, cosine similarity, inner product similarity, Euclidean distance, or other vector similarity calculation methods can be used. Taking cosine similarity as an example, its value ranges from -1 to 1, and the closer the value is to 1, the more semantically similar the question vector and the tag vector are.

[0100] S408: Tag Retrieval: Sort and filter target tags; In this step, candidate tags are sorted based on the calculated semantic similarity, and target tags are filtered according to preset similarity conditions. These preset similarity conditions may include a similarity score higher than a preset similarity threshold, a similarity score ranking within the top N positions, or both a similarity threshold and a ranking condition being met simultaneously. The value of N can be adjusted according to the business scenario and response accuracy requirements, for example, set to 2 to 5.

[0101] S409: Output target label; In this step, the selected target tags are output as input for subsequent response generation. These target tags characterize the core semantics of the real-time question to be answered and are used to match the corresponding response content in a standard response corpus, knowledge base, or business rule base.

[0102] S410: End.

[0103] Before executing the tag retrieval process described above, the text vectorization model needs to be trained first. The specific training process is as follows: Training Parameter Settings: In the training parameter setting stage, the loss function must first be selected based on the task type of the label retrieval vector model. For the classification task, the cross-entropy loss algorithm is used. This algorithm effectively measures the difference between the model's predicted probability distribution and the true label distribution, helping the model learn the accurate classification boundary. Next, the Adam optimizer is chosen because it combines the advantages of momentum and adaptive learning rate adjustment, resulting in fast convergence and the ability to adaptively adjust the learning rate for different parameters. Then, the learning rate is determined, as it determines the step size of model parameter updates and has a significant impact on the model's convergence speed and final performance. The initial learning rate is set to 1e-3, but in actual training, due to differences in the training set, it needs to be adjusted based on the decrease in the loss function and the model's performance on the validation set. If the loss decreases too slowly, the learning rate can be increased appropriately; if the loss fluctuates or fails to converge, the learning rate may need to be decreased. The batch size refers to the number of samples input into the model during each training iteration; the initial value for this system is 16. If the training set is large, setting a larger batch size allows for more data information to be used for parameter updates, potentially speeding up convergence. However, this consumes more memory resources and may cause the model to get stuck in local optima. Generally, the batch size can be set in doubling order, such as 32, 64, 128, etc., increasing incrementally. The specific choice should be made based on hardware resources and data characteristics. Next is the number of training epochs, which is the number of times the entire training dataset is traversed. Too few epochs may prevent the model from fully learning the patterns in the data, leading to underfitting; too many epochs may cause the model to overfit the training data, resulting in decreased performance on the validation and test sets. Therefore, it is usually necessary to conduct multiple trials, observing changes in metrics such as accuracy and recall on the validation set. When the metrics no longer show significant improvement or signs of overfitting appear, the appropriate number of training epochs is determined. This system uses a training epoch set of 500 epochs.

[0104] Training Data Partitioning and Loading: After setting the training parameters, the next step is data partitioning and loading. The primary task at this stage is to rationally partition the preprocessed dataset. The dataset is divided into training, validation, and test sets according to a certain ratio; in this system, the ratio is 8:1:1. The training set is mainly used by the model to learn the features and patterns of the data during training, prompting the model to continuously adjust its parameters to fit the data. The validation set is used to monitor the model's performance during training, evaluate its generalization ability, assist in adjusting hyperparameters, and avoid overfitting or underfitting. The test set is used after model training is complete to objectively evaluate the model's final performance and test its performance on unknown data.

[0105] When partitioning the data, 80% of the questions for each type of user question are extracted from the pre-organized training set as training data. To avoid bias during data partitioning, a random sampling method is used to shuffle the data before partitioning, ensuring that each subset covers data samples of all types. Simultaneously, data balance must be considered, especially for classification tasks; in practice, attention must be paid to the differences in the number of samples in each category. If the differences between categories are too large, the model may be biased towards the category with the larger number of samples, thus affecting model performance. A stratified sampling method is used, allocating samples to each subset according to the proportion of each category, ensuring a relatively balanced distribution of each category across the subsets. For example, for a dataset of 650,000 labeled user questions within an enterprise, after partitioning in an 8:1:1 ratio, 520,000 training questions, 65,000 validation questions, and 65,000 test questions are obtained, with the distribution ratio of questions for each business category in the three subsets remaining consistent with the original dataset.

[0106] After data partitioning, the partitioned data needs to be loaded into the model training environment. In deep learning frameworks, this is achieved using a DataLoader. The DataLoader can group data according to a set batch size, allowing the model to read data in batches during training, thus improving training efficiency. Furthermore, enabling the DataLoader supports multi-threaded data loading, further accelerating data reading speed.

[0107] Model Iterative Training: After data partitioning and loading, the model training and iteration phase begins. First, according to the pre-defined training parameters, the partitioned training set data is input into the text vectorization-based model in batches according to the set batch size, using a data loader. Once a batch of data is input into the model, it undergoes a forward propagation process. During this process, the input text sequence is calculated and transformed at various layers of the model. The text vectorization model encodes the input sequence, extracts its semantic features, and maps them to a label vector space, thereby generating a label vector prediction result corresponding to the input text. Subsequently, the model's prediction result is compared with the true labels of this batch of data, and the difference between the two is calculated using the previously selected loss function, i.e., the loss value. This loss value reflects the gap between the model's current predictive ability and the actual situation, and is an important basis for subsequent parameter updates.

[0108] In the backpropagation phase, based on the calculated loss value, the selected optimizer calculates the gradients of each model parameter. The gradient represents the rate of change of the loss function relative to each parameter, indicating the direction and magnitude of the parameter updates. The optimizer updates the model parameters according to the set learning rate based on these gradients, gradually reducing the value of the loss function, thus making the model's predictions closer to the true labels. After training on a batch of data, the next batch of data is loaded, and the process of forward propagation, loss calculation, backpropagation, and parameter updates is repeated until the entire training set has been traversed, completing one round of training. Typically, to allow the model to fully learn the features and patterns in the data, multiple rounds of training are required, with the model's parameters further optimized after each round.

[0109] Training process monitoring: During model training and iteration, monitoring the training process is crucial to ensuring model performance and stability. The training progress is monitored by continuously observing the loss function, which measures the difference between the model's predicted values ​​and the true labels. The training set loss should generally decrease with each training epoch. If it stops decreasing or even increases, it indicates a possible problem of an excessively high learning rate or model overfitting. When the validation set loss stops decreasing and starts increasing, while the training set loss continues to decrease, this is a clear sign of overfitting. In this case, it is necessary to introduce regularization parameters, reduce model complexity, or terminate training prematurely.

[0110] Simultaneously, appropriate evaluation metrics should be selected for monitoring based on the specific task. This system performs a classification task, using accuracy, precision, recall, and F1 score; for a regression task, mean squared error, root mean square error, and mean absolute error are used. These metrics should be calculated periodically on the validation set, and their changes over training epochs should be observed. If the metrics do not show significant improvement over a prolonged period, it may be necessary to adjust the model architecture or perform more in-depth feature engineering. The learning rate metric should also be monitored, especially when using learning rate scheduling strategies. It's crucial to ensure that it adjusts as expected, as a learning rate decrease that is too rapid or too slow can negatively impact the model's training performance.

[0111] In addition, it's crucial to monitor hardware resource usage during training. Excessive GPU usage can slow down training and even cause program crashes; insufficient memory can lead to slow or unloaded data. Monitoring allows for adjustments to batch size and optimization of data loading methods, ensuring training stability and efficiency.

[0112] like Figure 7 As shown in the figure, this application embodiment provides an intelligent customer service response device 200, the device 200 including: The first acquisition unit 202 is used to acquire real-time questions to be answered and to obtain the corresponding question vector based on the real-time questions to be answered. The second acquisition unit 204 is used to acquire the semantic entropy value corresponding to each tag in the preset dynamic semantic library, and obtain the global semantic entropy value of the preset dynamic semantic library based on the semantic entropy values ​​of all tags. The first processing unit 206 is used to cluster the corresponding tags whose semantic entropy values ​​are higher than the preset global entropy value threshold in response to the global semantic entropy value exceeding the preset term entropy value threshold, and obtain the clustered tag vector. The second processing unit 208 is used to generate a target label vector based on the label vectors that meet the preset label density threshold after clustering, and save the target label vector to a preset dynamic semantic library. The preset label density threshold is adjusted based on the label generation time, label usage frequency and time decay factor. The third processing unit 210 is used to perform similarity comparison between the question vector and the preset dynamic semantic library to obtain multiple target labels that meet the preset similarity conditions. The multiple target labels are used to generate corresponding response results.

[0113] The intelligent customer service response device 200 provided in this application embodiment acquires real-time questions awaiting response, obtains corresponding question vectors based on these questions, and transforms unstructured user questions into standardized dense semantic vectors. It acquires the semantic entropy value corresponding to each tag in a preset dynamic semantic library, and obtains the global semantic entropy value of the preset dynamic semantic library based on the semantic entropy values ​​of all tags, quantifying the uncertainty level of a single tag relative to the global semantic library. By clustering tags whose semantic entropy values ​​exceed a preset global entropy threshold in response to a global entropy value exceeding a preset term entropy threshold, it can accurately locate semantic ambiguity and redundancy. Based on the clustered tag vectors that meet a preset tag density threshold, it generates target tag vectors and saves these target tag vectors to the preset dynamic semantic library, continuously maintaining a high deterministic state of the semantic library. Based on a similarity comparison between the question vectors and the preset dynamic semantic library, it obtains multiple target tags that meet preset similarity conditions to generate corresponding response results, outputting accurate responses that match the user's true intent. This improves the accuracy of intelligent customer service responses.

[0114] Figure 8 A schematic diagram of the hardware structure of the electronic device provided in an embodiment of this application is shown.

[0115] The electronic device may include a processor 601 and a memory 602 storing computer program instructions.

[0116] Specifically, the processor 601 may include a central processing unit (CPU), an application-specific integrated circuit (ASIC), or one or more integrated circuits that can be configured to implement the embodiments of this application.

[0117] Memory 602 may include mass storage for data or instructions. For example, and not limitingly, memory 602 may include a hard disk drive (HDD), floppy disk drive, flash memory, optical disk, magneto-optical disk, magnetic tape, or Universal Serial Bus (USB) drive, or a combination of two or more of these. Where appropriate, memory 602 may include removable or non-removable (or fixed) media. Where appropriate, memory 602 may be internal or external to the integrated gateway. In a particular embodiment, memory 602 is non-volatile solid-state memory.

[0118] Memory may include read-only memory (ROM), random access memory (RAM), disk storage media devices, optical storage media devices, flash memory devices, and electrical, optical, or other physical / tangible memory storage devices. Therefore, typically, memory includes one or more tangible (non-transitory) computer-readable storage media (e.g., memory devices) encoded with software including computer-executable instructions, and when the software is executed (e.g., by one or more processors), it is operable to perform the operations described with reference to the methods according to one aspect of this disclosure.

[0119] The processor 601 reads and executes computer program instructions stored in the memory 602 to implement any of the intelligent customer service response methods in the above embodiments.

[0120] In one example, the electronic device may also include a communication interface 603 and a bus 610. For example, Figure 8 As shown, the processor 601, memory 602, and communication interface 603 are connected through bus 610 and complete communication with each other.

[0121] The communication interface 603 is mainly used to realize communication between various modules, devices, units and / or equipment in the embodiments of this application.

[0122] Bus 610 includes hardware, software, or both, that couples components of an electronic device together. For example, and not limitingly, the bus may include an Accelerated Graphics Port (AGP) or other graphics bus, an Enhanced Industry Standard Architecture (EISA) bus, a Front Side Bus (FSB), HyperTransport (HT) interconnect, an Industry Standard Architecture (ISA) bus, an Infinite Bandwidth Interconnect, a Low Pin Count (LPC) bus, a memory bus, a Microchannel Architecture (MCA) bus, a Peripheral Component Interconnect (PCI) bus, a PCI-Express (PCI-X) bus, a Serial Advanced Technology Attachment (SATA) bus, a Video Electronics Standards Association Local (VLB) bus, or other suitable buses, or combinations of two or more of these. Where appropriate, bus 610 may include one or more buses. Although specific buses are described and illustrated in embodiments of this application, this application contemplates any suitable bus or interconnect.

[0123] Furthermore, in conjunction with the intelligent customer service response methods in the above embodiments, this application embodiment can provide a computer storage medium for implementation. The computer storage medium stores computer program instructions; when these computer program instructions are executed by a processor, they implement any of the intelligent customer service response methods in the above embodiments.

[0124] This application also provides a computer program product, including a computer program that, when executed by a processor, implements any of the intelligent customer service response methods described in the above embodiments.

[0125] It should be clarified that this application is not limited to the specific configurations and processes described above and shown in the figures. For the sake of brevity, detailed descriptions of known methods are omitted here. In the above embodiments, several specific steps are described and shown as examples. However, the method process of this application is not limited to the specific steps described and shown. Those skilled in the art can make various changes, modifications, and additions, or change the order of steps, after understanding the spirit of this application.

[0126] The functional blocks shown in the above block diagram can be implemented as hardware, software, firmware, or a combination thereof. When implemented in hardware, they can be, for example, electronic circuits, application-specific integrated circuits (ASICs), appropriate firmware, plug-ins, function cards, etc. When implemented in software, the elements of this application are programs or code segments used to perform the required tasks. Programs or code segments can be stored on a machine-readable medium or transmitted over a transmission medium or communication link via data signals carried on a carrier wave. "Machine-readable medium" can include any medium capable of storing or transmitting information. Examples of machine-readable media include electronic circuits, semiconductor memory devices, ROM, flash memory, erasable ROM (EROM), floppy disks, CD-ROMs, optical disks, hard disks, fiber optic media, radio frequency (RF) links, etc. Code segments can be downloaded via computer networks such as the Internet, intranets, etc.

[0127] It should also be noted that the exemplary embodiments mentioned in this application describe methods or systems based on a series of steps or apparatus. However, this application is not limited to the order of the above steps; that is, the steps can be performed in the order mentioned in the embodiments, or in a different order, or several steps can be performed simultaneously.

[0128] The above are merely specific embodiments of this application. Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems, modules, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here. It should be understood that the protection scope of this application is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in this application, and these modifications or substitutions should all be covered within the protection scope of this application.

Claims

1. An intelligent customer service response method, characterized in that, include: Obtain real-time pending questions, and based on the real-time pending questions, obtain the corresponding question vector; Obtain the semantic entropy value corresponding to each tag in the preset dynamic semantic library, and obtain the global semantic entropy value of the preset dynamic semantic library based on the semantic entropy values ​​of all tags; In response to the global semantic entropy value exceeding a preset global entropy value threshold, the corresponding tags with semantic entropy values ​​higher than the preset term entropy value threshold are clustered to obtain a clustered tag vector; Based on the clustered tag vectors that meet the preset tag density threshold, a target tag vector is generated and the target tag vector is saved to the preset dynamic semantic library. The preset tag density threshold is adjusted based on the tag generation time, tag usage frequency, and time decay factor. Based on the question vector and the preset dynamic semantic library, a similarity comparison is performed to obtain multiple target tags that meet the preset similarity conditions. The multiple target tags are used to generate corresponding response results.

2. The method according to claim 1, characterized in that, Before obtaining the semantic entropy value corresponding to each tag in the preset dynamic semantic library, the method further includes: Based on historical business data, question texts are obtained, and a user question corpus is constructed based on these question texts. Based on the semantic information of each question text in the user question corpus, corresponding semantic terms are obtained, and corresponding question tags are obtained based on the semantic terms; A question tag mapping relationship is established based on each question text, its corresponding semantic term, and its corresponding question tag. Based on the question tag mapping relationship, the preset dynamic semantic library is generated.

3. The method according to claim 2, characterized in that, The step of obtaining the semantic entropy value corresponding to each tag in the preset dynamic semantic library includes: For each semantic term corresponding to a preset tag, the semantic state set of all question texts associated with the semantic term in the preset dynamic semantic library is counted, as well as the probability of each semantic state in the semantic state set appearing in the associated question text; Based on the number of semantic states in the set and the probability of occurrence of each semantic state, the semantic entropy value corresponding to the semantic term is obtained, and the semantic entropy value corresponding to each tag is obtained based on the semantic entropy value corresponding to the semantic term.

4. The method according to claim 3, characterized in that, After obtaining the semantic entropy value corresponding to each tag in the preset dynamic semantic library, and obtaining the global semantic entropy value of the preset dynamic semantic library based on the semantic entropy values ​​of all tags, the method further includes: When the preset dynamic semantic library meets the preset update triggering conditions, the semantic entries affected by the update are determined; Recalculate the semantic entropy value corresponding to the affected semantic terms, and recalculate the global semantic entropy value of the preset dynamic semantic library based on the updated semantic entropy values ​​of all semantic terms in the preset dynamic semantic library. In response to the updated global semantic entropy value exceeding a preset global entropy threshold, the corresponding tags whose updated semantic entropy value is higher than the preset term entropy threshold are clustered to obtain the updated target tag vector.

5. The method according to claim 4, characterized in that, The step of determining the semantic terms affected by the update when the preset dynamic semantic library meets the preset update triggering conditions includes: If incremental historical business data is detected imported into the user question corpus, new question text is obtained based on the incremental historical business data, and the semantic term corresponding to the new question text is determined as the semantic term affected by the update; or, Obtain a preset domain knowledge graph, and analyze the semantic network structure between semantic entries in the preset dynamic semantic library based on the domain knowledge graph; Identify conflicting semantic relationships in the semantic network structure, and determine the semantic terms corresponding to the conflicting semantic relationships as the semantic terms affected by the update.

6. The method according to claim 1, characterized in that, The process of obtaining real-time pending questions, and based on these questions, deriving a corresponding question vector, includes: Based on the real-time unanswered question, multiple candidate semantic terms are obtained; The importance of each candidate semantic term is determined, and candidate semantic terms with an importance lower than a preset importance threshold are filtered out to obtain the filtered semantic terms. The importance is positively correlated with the number of times the candidate semantic term appears in the real-time question to be answered, and negatively correlated with the frequency of the candidate semantic term in the user question corpus. Based on the filtered semantic terms, the question vector corresponding to the real-time question to be answered is obtained.

7. An intelligent customer service response device, characterized in that, The device includes: The first acquisition unit is used to acquire real-time unanswered questions and obtain the corresponding question vector based on the real-time unanswered questions; The second acquisition unit is used to acquire the semantic entropy value corresponding to each tag in the preset dynamic semantic library, and to obtain the global semantic entropy value of the preset dynamic semantic library based on the semantic entropy values ​​of all tags. The first processing unit is configured to cluster the corresponding tags whose semantic entropy values ​​are higher than the preset global entropy value threshold in response to the global semantic entropy value exceeding the preset term entropy value threshold, and obtain the clustered tag vector. The second processing unit is used to generate a target label vector based on the label vectors that meet the preset label density threshold after clustering, and save the target label vector to the preset dynamic semantic library. The preset label density threshold is adjusted based on the label generation time, label usage frequency and time decay factor. The third processing unit is used to perform a similarity comparison between the question vector and the preset dynamic semantic library to obtain multiple target tags that meet preset similarity conditions. The multiple target tags are used to generate corresponding response results.

8. An electronic device, characterized in that, The electronic device includes: a processor and a memory storing computer program instructions; When the processor executes the computer program instructions, it implements the intelligent customer service response method as described in any one of claims 1 to 6.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer program instructions, which, when executed by a processor, implement the intelligent customer service response method as described in any one of claims 1 to 6.

10. A computer program product, characterized in that, When the instructions in the computer program product are executed by the processor of the electronic device, the electronic device causes the electronic device to perform the intelligent customer service response method as described in any one of claims 1 to 6.

Images