Method for improving reply accuracy of intelligent customer service

Abstract

The invention relates to the technical field of intelligent customer service, in particular to a method for improving reply accuracy of intelligent customer service. The method for improving the reply accuracy of the intelligent customer service comprises the following steps: compressing complex statements; vectorization construction of a standard intention library and vectorization construction of a standard entity library; extracting a customer question intention and identifying a key entity, and identifying the intention and the key entity; matching the question intention of the customer with the task; the customer questions the alignment between the key entity and the standard entity library, so that the accuracy of subsequent service parameter matching is improved; automatic assembly and service search of services, automatic assembly of entities, tasks and background services, and accurate query and analysis of intentions are realized; and feedback and display of a result: intelligent and standardized reply is completed based on a large model and a preset template. According to the method, key intentions can be extracted from the long sentences, redundant parts are removed, and the system is enabled to focus on processing core information.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent customer service, and in particular to a method for improving the accuracy of intelligent customer service responses. Background Art

[0002] With the widespread application of intelligent customer service systems in various industries, various industries have gradually adopted intelligent customer service technology to improve customer service experience and operation and maintenance efficiency. However, the existing intelligent customer service systems still have obvious technical shortcomings when dealing with complex sentences and multi-intent recognition. Especially in scenarios where customers use long sentences and express multiple intents, the system's semantic understanding ability is weak, which often leads to misidentification, missed recognition, and inaccurate execution of intent tasks after recognition, which in turn affects service quality. At present, intelligent customer service systems have certain semantic understanding and dialogue processing capabilities. However, due to the complexity of business scenarios and the diversity of customer demands, the accuracy of existing systems has dropped significantly when processing long sentences, complex sentences, and multi-intent sentences.

[0003] At present, the implementation solutions mainly include the following:

[0004] (I) Semantic understanding solution based on rule matching: It mainly understands the user's semantics through predefined rules, templates and keyword matching. In this system, developers predefine sentence patterns, grammatical rules and keywords for specific business fields. When customers enter sentences, the system identifies the user's intention by matching the preset templates in the rule library. However, if the user expresses a sentence that contains multiple intentions or has a complex structure, it is difficult for the system to effectively identify all the intentions.

[0005] (ii) Semantic understanding solution based on machine learning: Automatically learn the mapping relationship between user input and intent through training on a large amount of labeled data. Common machine learning models include support vector machine (SVM), naive Bayes, decision tree, etc., which can improve the effect of intent recognition to a certain extent. However, this type of solution still has limited processing capabilities for complex sentences, especially when faced with multiple intents or long sentences, the model may not be able to recognize multiple intents at the same time. Limited by the quality of labeled data and the generalization ability of the model, this method often performs unstable when processing more complex sentences with large structural changes.

[0006] (III) Configuration-based service call method is a technical means to dynamically call different services through predefined configuration information. It allows the system to select and execute corresponding services according to the configuration, and has certain service scalability and maintainability. However, this method also has some problems, such as the management and synchronization of configuration information may become complicated, especially the delay of configuration update may lead to call errors; in addition, over-reliance on configuration may also make the system lack of flexibility and increase maintenance costs.

[0007] Existing intelligent customer service systems face significant technical challenges in processing complex sentences and multi-intent recognition, especially when dealing with long sentences and user expressions containing multiple intents, the existing technology shows great limitations. The main reason is that traditional rule matching and machine learning methods can only be based on preset grammatical rules or rely on a single intent recognition mechanism in the training data, and it is difficult to handle sentences with complex structures or lengthy semantics. This technical bottleneck directly leads to the customer service system's inability to effectively extract the core information in the sentence when facing long sentences, resulting in frequent system misidentification and missed recognition, which in turn affects the accuracy and efficiency of the service. When users express multiple needs or intentions in long sentences, the existing customer service system is often disturbed by redundant information because it is difficult to identify the main content. For example, a user may ask about the electricity bill and request a repair of the meter at the same time, and the system may only recognize one of the needs, resulting in the customer's overall needs not being fully processed. The root cause of this situation is that the existing technology lacks the ability to reasonably process long sentences, especially the lack of technical means to extract and compress key information in sentences. The longer the sentence and the more complex the structure, the lower the accuracy of the system's semantic understanding.

[0008] Secondly, during the single-intent recognition process, the system may fail to fully extract the key entity information in the question, and the extracted information cannot be automatically aligned with the parameters in the service. As a result, even if the intent recognition is accurate, it will not be able to give a correct response due to the missing key information, reducing the accuracy of the service and user experience.

[0009] Therefore, the existing technical solutions have low overall accuracy when processing the intent recognition of long sentences and cannot meet the needs of customers to solve problems quickly. Summary of the invention

[0010] The present invention proposes a method for improving the accuracy of intelligent customer service responses based on complex sentence compression and methods, multi-stage entity alignment, and service automatic assembly technology. The specific technical solution is as follows:

[0011] A method for improving the accuracy of intelligent customer service responses includes the following process:

[0012] S100: Complex sentence compression; specifically includes:

[0013] S110: Sentence analysis and recognition; Specific process: For long sentences expressed by users, the sentence structure is first parsed through natural language processing technology to identify the basic components of the subject, predicate, and object in the sentence, and analyze their subordinate relationships and sentence structures; the sentence is structurally analyzed using a syntactic tree or dependency graph; when multiple clauses are identified in a sentence, the system divides it into independent short sentences so that each short sentence can be processed separately;

[0014] S120: semantically compress each clause; specifically, the process includes the following:

[0015] S121: Eliminate redundant information in the sentence and only keep the core semantic parts directly related to customer needs;

[0016] S122: Based on the dual-channel attention mechanism, through the deep learning model, Chinese word segmentation and part-of-speech tagging are regarded as joint tasks. The dual-channel attention mechanism weights the automatically acquired context features and syntactic knowledge respectively, predicts the word segmentation and part-of-speech tag of each word, and effectively eliminates the word segmentation "noise";

[0017] S123: Extract the core needs of users by analyzing keywords and key phrases in sentences;

[0018] S200: vectorized construction of standard intent library and standard entity library;

[0019] S300: Extracting the customer's question intention and identifying key entities, identifying the intention and key entities; specifically, first predefine a set of entity types, such as business type, place name, and equipment type; then use the labeled intention classification data to train a large power customer service model to complete intention extraction and named entity recognition;

[0020] S400: Matching customer question intention with task;

[0021] S500: Align key entities in customer questions with the standard entity library to improve the accuracy of subsequent service parameter matching;

[0022] S600: Automatic assembly and search of services, automatic assembly of entities, tasks and background services, and accurate query and analysis of intent;

[0023] S700: Feedback and display of results, based on large models and preset templates, to complete intelligent and standardized responses.

[0024] The present invention proposes a technical method based on sentence compression. By introducing automatic grammatical analysis and context understanding algorithms, the present invention can extract key intents from long sentences, remove redundant parts, and enable the system to focus on processing core information. Through this sentence compression method, the system can more accurately understand the real needs of users, thereby reducing misidentification and missed identification. Compared with traditional rule matching and single intent recognition methods, sentence compression technology can significantly improve the system's processing efficiency for long sentences, allowing the customer service system to respond to customer needs in complex dialogue scenarios more quickly and accurately; in addition, by adopting multi-stage entity alignment and service automatic assembly technology, it aims to ensure that the key parameter entity information in the question can be fully extracted during the single intent recognition process, and automatically aligned with the parameters in the service, so as to achieve accurate query and reply of intent. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] Figure 1 It is a flow chart of a method for improving the accuracy of intelligent customer service responses according to the present invention.

[0026] Figure 2 It is a schematic diagram of the results of the sentence analysis and recognition example in S110 of the embodiment of the present invention. DETAILED DESCRIPTION

[0027] A method for improving the accuracy of intelligent customer service responses includes the following process:

[0028] S100: Complex sentence compression; specifically includes:

[0029] S110: Sentence analysis and recognition; Specific process: For long sentences expressed by users, firstly parse the sentence structure through natural language processing (NLP) technology, identify the basic components of subject, predicate and object in the sentence, and analyze their subordinate relationship and sentence structure; use syntax tree or dependency graph (Dependency Parsing) to perform structured analysis on the sentence; for example, for the sentence "I need to check the electricity bill and also want to report the repair of the electricity meter", the system will parse that the sentence contains two clauses: "I need to check the electricity bill" and "I also want to report the repair of the electricity meter"; when it is recognized that there are multiple clauses in the sentence, the system divides it into independent short sentences, so that each short sentence can be processed separately; this step ensures that even complex sentences can be decomposed into multiple simple semantic units by the system.

[0030] S120: semantically compress each clause; specifically, the process includes the following:

[0031] S121: Eliminate redundant information in sentences and keep only the core semantic parts directly related to customer needs; in actual conversations, users' expressions often contain a lot of non-critical information, such as descriptive words, additional instructions, polite expressions, etc.

[0032] S122: Based on the dual-channel attention mechanism, through deep learning models such as the BERT model, Chinese word segmentation and part-of-speech tagging are regarded as joint tasks. The dual-channel attention mechanism weights the automatically acquired context features and syntactic knowledge respectively, predicts the word segmentation and part-of-speech tag of each word, and effectively removes the "noise" of word segmentation; for example, in the sentence "Can you help me check my electricity bill for last month?", polite expressions such as "can you help me" and "a while" are removed, and only the core information "check last month's electricity bill" is retained;

[0033] S123: extracting the core needs of the user by analyzing the keywords and key phrases in the sentence; for example, for the sentence "I want to check the electricity bill for the last three months", the system can extract the core information "check the electricity bill" and "last three months";

[0034] Dynamic sentence compression strategy: The present invention introduces a dynamic adjustment strategy in the sentence compression process, which can flexibly adjust the compression intensity and method according to the length of the sentence, business complexity and importance, and user interaction behavior; for long sentences exceeding a certain length (such as 100 words), the system will give priority to compression to ensure that its core semantics can be quickly extracted and processed; for different user expressions, the system will dynamically adjust the compression strategy according to the specific context of the conversation; for lengthy and repetitive expressions, the system will more aggressively remove redundant information; and for shorter sentences, the system will be more cautious to avoid excessive compression and information loss; the specific steps are as follows:

[0035] S124: Feature extraction, specifically: for each input sentence, extract its length, business importance, complexity, user interaction behavior and context environment features, and automatically extract and quantify them through BERT or a large model;

[0036] S125: weight allocation, specifically: assigning a corresponding weight to each factor according to its importance in semantic compression;

[0037] S126: weighted calculation, specifically: multiplying the value of each feature by its corresponding weight to obtain a weighted score for each factor; summing the weighted scores of all factors to obtain a comprehensive score;

[0038] First, sort the quantitative values ​​of each evaluation index from large to small, and apply the C-OWA weight function to calculate the comprehensive score based on the sorted values. Assuming that we use a linear weight function and the sorted evaluation index values ​​are x1, x2, x3, x4 (where x1 is the maximum value and x4 is the minimum value), the comprehensive score S can be calculated as:

[0039] S=wl'x1+w2'x2+w3'x3+w4'x4;

[0040] S127: strategy selection, specifically: selecting the most appropriate strategy from a predefined semantic compression strategy library based on the comprehensive score; these strategies can be preconfigured, including different compression strengths, methods, and retained information types;

[0041] S128: Model optimization, specifically: continuously optimize the weight distribution and strategy selection in the multi-factor weighted model by collecting user feedback and system performance data; use online learning algorithms to update model parameters in real time to adapt to changing user needs.

[0042] S200: vectorized construction of standard intent library and standard entity library; specifically includes:

[0043] S210: Standard intent library vectorization construction; specifically includes:

[0044] S211: Maintain the basic intent library; specifically: establish an initial intent library, including common customer questions and corresponding intent tags;

[0045] S212: Identify the intention based on the large model; specifically: use the intention recognition module to identify the intention of the customer's question;

[0046] S213: Automatically collect intents. For responses with high customer ratings, automatically collect the corresponding intents into the standard intent library to enrich and optimize the intent library. The specific process includes:

[0047] S213a: Evaluation threshold setting: Setting an evaluation threshold, such as when the satisfaction score reaches a certain score, is used to determine which responses are worth being included in the standard intent library;

[0048] S213b: Intent aggregation: For replies with evaluations above the threshold, their corresponding intent labels and question-reply pairs are automatically aggregated into the standard intent library;

[0049] S213c: Intent library vectorization: Convert each intent into a vector using the Word2Vec model

[0050] S220: vectorized construction of standard entity library; specifically includes:

[0051] S221: Data collection and preprocessing; Collect data extensively from text data sources related to the power industry, including knowledge base technical documents, user manuals, industry standards, and academic papers; Preprocess the collected data, including removing irrelevant characters, word segmentation, removing stop words, stemming or lemmatization to ensure data quality;

[0052] S222: Word2Vec model training; constructing a training data set based on the preprocessed text data and the extracted entities; selecting a suitable Word2Vec model and setting training parameters, the parameters including one or more of vector dimension, context window size, and number of iterations; training the Word2Vec model using the training data set until the model converges or reaches a preset training round, and obtaining a word vector representation of the entity;

[0053] S223: extracting entities from the preprocessed text data, using the pre-trained power customer service model to complete named entity recognition, and identifying key entities such as power equipment, technical terms, and concepts;

[0054] S224: Entity similarity calculation and merging; specifically includes the following process:

[0055] S224a: Calculate the cosine similarity between any two entity word vectors as the similarity measure between them;

[0056] S224b: Setting a similarity threshold, merging entity pairs whose similarity exceeds the threshold; selecting the most representative entity name, vector representation, merging descriptive information, and updating entity relationships for merging;

[0057] S224c: repeating steps S224a and S224b until all entities with high similarity are merged to form a preliminary electric power industry standard entity library;

[0058] S225: Entity library optimization and verification; the process specifically includes:

[0059] S225a: Manually review the initially constructed power industry standard entity library to ensure that the merged entities are accurate and reasonable;

[0060] S225b: According to the audit results, make necessary adjustments and optimizations to the entity database, including correcting incorrect merges, adding missing entities, and updating entity descriptions;

[0061] S225c: Use the validation dataset to validate the optimized entity library and evaluate its practicality and accuracy;

[0062] S300: Extracting the customer’s question intention and identifying key entities, identifying the intention and key entities; specifically including:

[0063] S331: First, a group of entity types are predefined, such as entity types including business types, place names, and equipment types;

[0064] S332: Then, use the labeled intent classification data to train a large power customer service model to complete intent extraction and named entity recognition. For example, "Help me check my September electricity bill and report a repair to the meter" identifies that "check the electricity bill" and "report a repair to the meter" are two independent intents, and have time and month entities.

[0065] S400: Matching the customer's question intention with the task; specifically includes the following process:

[0066] S410: Intent matching; specifically: according to the intent identified by the intent recognition module, vector similarity calculation is performed with the intent in the standard intent library by using the cosine similarity algorithm to obtain the top 5 intents with the highest similarity, and the intent with the highest intent similarity is preferentially selected for matching;

[0067] S420: Determine the configuration item according to the task; specifically: search for the corresponding task according to the matched intent, and once the task is determined, retrieve the corresponding service list from the configuration item based on the task;

[0068] S430: screening and sorting the service list; specifically, screening and sorting the retrieved service list according to the priority of the task and factors related to the availability of the service;

[0069] S500: Align key entities in customer questions with the standard entity library to improve the accuracy of subsequent service parameter matching; specifically includes the following processes:

[0070] S510: Extract entity key-value pairs from the question and answer statements; specifically: use the pre-trained power customer service model to accurately identify and extract multiple entity key-value pairs E from the customer's questions.<k,v> ;

[0071] S520: matching of service names; specifically, comparing key information in the extracted entity key-value pair with the retrieved service list to find the best matching service;

[0072] S530: matching and checking parameter values; then, the system needs to verify whether the parameter values ​​in the entity key-value pair meet the parameter requirements of the service, and perform necessary checks and adjustments;

[0073] S600: Automatic assembly and search of services, automatic assembly of entities, tasks and background services to achieve accurate query and analysis of intent; specifically includes the following processes:

[0074] S610: Calling and executing the service; specifically, after successfully matching the service and confirming that the parameters are correct, the system will call and execute the service;

[0075] S620: Monitoring and recording of execution results; specifically, the system needs to monitor the execution process of the service in real time and accurately record its execution results;

[0076] S700: Feedback and display of results, based on large models and preset templates, to complete intelligent and standardized responses; specifically includes the following processes:

[0077] S710: Integrate the multi-intent query results; specifically including:

[0078] S711: Result collection; specifically: collecting all relevant results from queries with different intents, where the results come from different services, data sources, or query paths;

[0079] S712: Deduplication and sorting; specifically: deduplication of the collected results to avoid duplicate information; and sorting the results according to their relevance and priority;

[0080] S720: Result assembly and integration; specifically including:

[0081] S721: content analysis; specifically: using the pre-trained power customer service model to perform content analysis on the collected results and identify key information, themes or trends;

[0082] S722: Logical association; specifically: based on the analysis results, establish logical associations between different results, where the associations include causal relationships, time sequence, or parallel relationships;

[0083] S723: Assembly and integration; specifically: using the text reorganization and summarization capabilities of the pre-trained power customer service model, assemble multiple results into a coherent and orderly whole based on logical associations;

[0084] S730: Generate standardized responses; based on the results of intent classification, select the most appropriate response template to complete content filling.

[0085] Explanation of relevant terms:

[0086] Intent recognition: is a natural language processing technique used to determine the intent or purpose of a user when conversing with a computer.

[0087] Sentence compression: It is a language processing technology that aims to simplify a longer sentence into a shorter and more concise sentence while retaining the important information of the original sentence. The core of sentence compression is to remove redundant information in the sentence while ensuring that the simplified sentence still conforms to grammatical norms and can accurately express the core content of the original sentence.

[0088] Big Model: Translated into English as Big Model or Large Model, it is an important technology for processing large-scale data and complex tasks in the fields of machine learning and artificial intelligence. It usually refers to deep learning models with tens of millions or even hundreds of millions of parameters. These models are built by deep neural networks and have billions or even hundreds of billions of parameters.

[0089] Word2vec: Word2vec is a set of related models used to generate word vectors. These models are shallow and two-layer neural networks that are trained to reconstruct linguistic word texts. In simple terms, Word2vec is a technology that converts words or features into vectors, also known as word embedding.

Claims

1. A method for improving the accuracy of intelligent customer service responses, characterized in that: The process includes the following: S100: Complex sentence compression; specifically includes: S110: Sentence analysis and recognition; Specific process: For long sentences expressed by users, the sentence structure is first parsed through natural language processing (NLP) technology to identify the basic components of the subject, predicate, and object in the sentence, and analyze their subordinate relationships and sentence structures; the sentence is structurally analyzed using a syntax tree or dependency graph (Dependency Parsing); when multiple clauses are identified in a sentence, the system divides it into independent short sentences so that each short sentence can be processed separately; S120: semantically compress each clause; specifically, the process includes the following: S121: Eliminate redundant information in the sentence and only keep the core semantic parts directly related to customer needs; S122: Based on the dual-channel attention mechanism, through the deep learning model, Chinese word segmentation and part-of-speech tagging are regarded as joint tasks. The dual-channel attention mechanism weights the automatically acquired context features and syntactic knowledge respectively, predicts the word segmentation and part-of-speech tag of each word, and effectively eliminates the word segmentation "noise"; S123: Extract the core needs of users by analyzing keywords and key phrases in sentences; S200: vectorized construction of standard intent library and standard entity library; S300: Extracting the customer's question intention and identifying key entities, identifying the intention and key entities; specifically, first predefine a set of entity types, such as business type, place name, and equipment type; then use the labeled intention classification data to train a large power customer service model to complete intention extraction and named entity recognition; S400: Matching customer question intention with task; S500: Align key entities in customer questions with the standard entity library to improve the accuracy of subsequent service parameter matching; S600: Automatic assembly and search of services, automatic assembly of entities, tasks and background services, and accurate query and analysis of intent; S700: Feedback and display of results, based on large models and preset templates, to complete intelligent and standardized responses.

2. A method for improving the accuracy of intelligent customer service responses according to claim 1, characterized in that: A dynamic adjustment strategy is introduced in the S100 complex statement compression process, specifically: The system flexibly adjusts the intensity and method of compression based on the length of the sentence, business complexity and importance, and user interaction behavior. For long sentences exceeding a certain length (such as 100 words), the system will give priority to compression to ensure that their core semantics can be quickly extracted and processed. For different user expressions, the system will dynamically adjust the compression strategy based on the specific context of the conversation. For lengthy and repetitive expressions, the system will more aggressively remove redundant information. For shorter sentences, the system will be more cautious to avoid information loss due to excessive compression.

3. A method for improving the accuracy of intelligent customer service responses according to claim 2, characterized in that: The specific steps are as follows: S124: Feature extraction, specifically: for each input sentence, extract its length, business importance, complexity, user interaction behavior and context environment features, and automatically extract and quantify them through BERT or a large model; S125: weight allocation, specifically: assigning a corresponding weight to each factor according to its importance in semantic compression; S126: weighted calculation, specifically: multiplying the value of each feature by its corresponding weight to obtain a weighted score for each factor; summing the weighted scores of all factors to obtain a comprehensive score; S127: strategy selection, specifically: selecting the most appropriate strategy from a predefined semantic compression strategy library based on the comprehensive score; these strategies can be preconfigured, including different compression strengths, methods, and retained information types; S128: Model optimization, specifically: continuously optimizing the weight allocation and strategy selection in the multi-factor weighted model by collecting user feedback and system performance data.

4. A method for improving the accuracy of intelligent customer service responses according to claim 1, characterized in that: S200 specifically includes the following processes: S210: Standard intent library vectorization construction; specifically includes: S211: Maintain the basic intent library; specifically: establish an initial intent library, including common customer questions and corresponding intent tags; S212: Identify the intention based on the large model; specifically: use the intention recognition module to identify the intention of the customer's question; S213: Automatically collect intents: For replies with high customer ratings, automatically collect the corresponding intents into the standard intent library to enrich and optimize the intent library; S220: vectorized construction of standard entity library; specifically includes: S221: Data collection and preprocessing; Collect data extensively from text data sources related to the power industry, including knowledge base technical documents, user manuals, industry standards, and academic papers; Preprocess the collected data, including removing irrelevant characters, word segmentation, removing stop words, stemming or lemmatization to ensure data quality; S222: Word2Vec model training; constructing a training data set based on the preprocessed text data and the extracted entities; selecting a suitable Word2Vec model and setting training parameters, the parameters including one or more of vector dimension, context window size, and number of iterations; training the Word2Vec model using the training data set until the model converges or reaches a preset training round, and obtaining a word vector representation of the entity; S223: extracting entities from the preprocessed text data, using the pre-trained power customer service model to complete named entity recognition, and identifying key entities such as power equipment, technical terms, and concepts; S224: Entity similarity calculation and merging; S225: Entity library optimization and verification.

5. A method for improving the accuracy of intelligent customer service responses according to claim 1, characterized in that: The specific process of automatically collecting intentions in S213 includes: S213a: Evaluation threshold setting: Setting an evaluation threshold to determine which responses are worth being included in the standard intent library; S213b: Intent aggregation: For replies with evaluations above the threshold, their corresponding intent labels and question-reply pairs are automatically aggregated into the standard intent library; S213c: Intent library vectorization: Convert each intent into a vector using the Word2Vec model.

6. A method for improving the accuracy of intelligent customer service responses according to claim 1, characterized in that: S224 entity similarity calculation and merging specifically include the following process: S224a: Calculate the cosine similarity between any two entity word vectors as the similarity measure between them; S224b: Setting a similarity threshold, merging entity pairs whose similarity exceeds the threshold; selecting the most representative entity name, vector representation, merging descriptive information, and updating entity relationships for merging; S224c: repeating steps S224a and S224b until all entities with high similarity are merged to form a preliminary electric power industry standard entity library; The S225 entity library optimization and verification process specifically includes: S225a: Manually review the initially constructed power industry standard entity library to ensure that the merged entities are accurate and reasonable; S225b: According to the audit results, make necessary adjustments and optimizations to the entity database, including correcting incorrect merges, adding missing entities, and updating entity descriptions; S225c: Use the validation dataset to validate the optimized entity library and evaluate its practicality and accuracy.

7. A method for improving the accuracy of intelligent customer service responses according to claim 1, characterized in that: S400 specifically includes the following processes: S410: Intent matching; specifically: according to the intent identified by the intent recognition module, vector similarity calculation is performed with the intent in the standard intent library by using the cosine similarity algorithm to obtain the top 5 intents with the highest similarity, and the intent with the highest intent similarity is preferentially selected for matching; S420: Determine the configuration item according to the task; specifically: search for the corresponding task according to the matched intent, and once the task is determined, retrieve the corresponding service list from the configuration item based on the task; S430: Screening and sorting the service list; specifically, screening and sorting the retrieved service list according to the priority of the task and factors related to the availability of the service.

8. A method for improving the accuracy of intelligent customer service responses according to claim 1, characterized in that: S500 specifically includes the following processes: S510: Extract entity key-value pairs from the question and answer statements; specifically: use the pre-trained power customer service model to accurately identify and extract multiple entity key-value pairs E from the customer's questions.<k,v> ; S520: matching of service names; specifically, comparing key information in the extracted entity key-value pair with the retrieved service list to find the best matching service; S530: Matching and verifying parameter values; Next, the system needs to verify whether the parameter values ​​in the entity key-value pair meet the parameter requirements of the service, and perform necessary verification and adjustments.

9. A method for improving the accuracy of intelligent customer service responses according to claim 1, characterized in that: S600 specifically includes the following processes: S610: Calling and executing the service; specifically, after successfully matching the service and confirming that the parameters are correct, the system will call and execute the service; S620: Monitoring and recording of execution results; specifically: the system needs to monitor the execution process of the service in real time and accurately record its execution results.

10. A method for improving the accuracy of intelligent customer service responses according to claim 1, characterized in that: S700 specifically includes the following processes: S710: Integrate the multi-intent query results; specifically including: S711: Result collection; specifically: collecting all relevant results from queries with different intents, where the results come from different services, data sources, or query paths; S712: Deduplication and sorting; specifically: deduplication of the collected results to avoid duplicate information; and sorting the results according to their relevance and priority; S720: Result assembly and integration; specifically including: S721: content analysis; specifically: using the pre-trained power customer service model to perform content analysis on the collected results and identify key information, themes or trends; S722: Logical association; specifically: based on the analysis results, establish logical associations between different results, where the associations include causal relationships, time sequence, or parallel relationships; S723: Assembly and integration: Specifically: using the text reorganization and summarization capabilities of the pre-trained power customer service model, assemble multiple results into a coherent and orderly whole based on logical associations; S730: Generate standardized responses; based on the results of intent classification, select the most appropriate response template to complete content filling.

Images