A text robot application system based on large model

Abstract

The present invention discloses a text robot application system based on a large model, comprising the following modules: a knowledge base and parsing module, which receives and processes business documents, decomposes the content into semantic units and annotates context features; a question and answer generation module, which automatically generates question and answer pairs through deep semantic analysis and stores them in a vector database; a question extension module, which dynamically expands similar questions using natural language generation technology; a semantic intention analysis module, which accurately extracts the core intention of the question through context semantic analysis and maps it to a business path; a text workflow management module, which uses a logical tree model to achieve dynamic routing of multiple business scenarios; a context semantic association module, which optimizes question expression through multi-round dialogue analysis; a bottom-up response module, which provides multi-level answers in combination with a question and answer database, document recall and dynamic generation; and an intelligent learning module, which records unsolved problems and optimizes the question and answer database and semantic rules through a learning algorithm.

Description

Technical Field

[0001] The present invention relates to the technical field of large models, and in particular to a text robot application system based on large models. Background Art

[0002] With the rapid development of artificial intelligence and big data technologies, text customer service systems are being used more and more widely in various business scenarios. However, existing text customer service systems still have significant limitations when facing increasingly complex user needs and diversified problem expressions. The main function of text customer service systems is to provide users with answers through preset rules or manually maintained knowledge bases. These traditional methods are unable to handle complex business scenarios, multi-round conversations, and dynamic optimization.

[0003] In the prior art, traditional text customer service solutions mainly rely on keyword matching and fixed rules to achieve automatic answers to user questions. These methods have the following obvious defects in practical applications:

[0004] 1. Low efficiency in knowledge base construction: Traditional systems require manual organization and maintenance of a large number of keywords and question-answer pairs. The process is cumbersome and error-prone, resulting in low efficiency and accuracy in knowledge base construction.

[0005] 2. Weak semantic matching capability: Existing systems mostly use keyword-based matching methods, which cannot accurately understand the semantic relationship in user questions and lack the ability to handle flexible expressions and complex contexts, which easily leads to mismatches or non-matching situations.

[0006] 3. Lack of contextual understanding ability: Traditional text customer service has difficulty associating contextual information in multiple rounds of conversations, cannot accurately identify the core intent of user questions, and has limited ability to handle continuous conversations.

[0007] 4. Single response mechanism: Existing methods usually provide answers based on fixed rules or preset templates. When faced with questions beyond the scope of the knowledge base, they lack the ability to dynamically generate answers, resulting in a poor user experience.

[0008] 5. Lack of continuous learning capabilities: Traditional customer service systems fail to achieve dynamic optimization. The handling of unresolved issues usually relies on manual analysis and updates. It is impossible to achieve automatic expansion of the knowledge base and continuous optimization of the model through intelligent algorithms.

[0009] Therefore, how to provide a text robot application system based on a large model is an urgent problem that technicians in this field need to solve. Summary of the invention

[0010] One purpose of the present invention is to propose a text robot application system based on a large model. The present invention fully combines semantic parsing algorithms, intelligent learning mechanisms and multi-module collaboration technologies, and describes in detail methods for automated knowledge base construction, contextual semantic association optimization and dynamic answer generation. It has the advantages of accurate response, wide coverage of scenarios, strong continuous optimization capabilities and good user experience.

[0011] A text robot application system based on a large model according to an embodiment of the present invention includes the following steps:

[0012] S1, knowledge base and document parsing module, receives business documents and FAQ forms in various formats provided by users, analyzes the document content using semantic parsing algorithms, decomposes the content into semantically related units, and annotates contextual semantic features. According to needs, it selects and calls precise answer models and generalized answer models;

[0013] S2, the question-answer generation and vectorization module calls the pre-trained large model, performs in-depth analysis on the semantic units generated by the knowledge parsing module, automatically generates question-answer pairs, supports users to view, edit and confirm question-answer pairs through the click-and-click QA extraction function, stores the optimized question-answer pairs in the question-answer vector database, and builds a searchable initial question-answer library;

[0014] S3, the question expansion module, based on the standard questions in the question and answer library, uses the natural language generation capability of the large model to call the AI ​​one-click expansion algorithm to generate diversified questions with similar semantics, dynamically expands the coverage of the question and answer library, and synchronously updates the expanded questions to the question and answer vector database;

[0015] S4, semantic intent analysis module, through the dynamically set prompt, calls the big model to perform deep semantic analysis on the questions input by the user, extracts the core intent features, and maps the question intent to the corresponding business processing path based on the multi-intent parsing rules, thus realizing the diversion of multi-intent scenarios and business automation processing;

[0016] S5, text workflow management module, defines text workflow rules for multiple business scenarios based on the logic tree model, dynamically routes user questions to the corresponding business branches according to the generated business paths, combines business branches, calls the large model knowledge base to generate answers, and supports collaborative processing and dynamic adjustment of multiple business scenarios;

[0017] S6, context semantic association module, by configuring hot questions and calling multi-round dialogue context analysis algorithm, combined with context association switch, performs semantic association analysis on the user's current question and historical dialogue records, sets context association optimization prompt, completes and standardizes user questions, submits the optimized questions to the question-answer vector database for retrieval, and returns the best answer;

[0018] S7, the fallback response module, which first calls the question-answer vector database to match user questions. If no answer is found, the document recall function is enabled to retrieve relevant content from the knowledge base. If there is no result in the document recall, the Spark Big Model is called to generate a dynamic associated answer. If the fixed fallback mechanism is enabled, the answer is generated according to the preset fixed words, and the unresolved questions and their contextual features are recorded in the intelligent learning module;

[0019] S8, intelligent learning and optimization module, for unresolved issues recorded by the catch-all response module, calls on the intelligent learning algorithm to dynamically expand and optimize the question-and-answer database and semantic rules, and retrains the large model to update the question-and-answer database and business logic rules.

[0020] Optionally, the S1 specifically includes:

[0021] S11. receiving business documents and FAQ forms in various formats provided by users, wherein supported document formats include structured data, semi-structured data, and unstructured data;

[0022] S12, calling a semantic parsing algorithm to analyze the received document content, dividing the document into a plurality of semantically related units, each unit including at least one subject, predicate and object, and realizing the logical integrity of the fragmentation according to a predefined semantic fragmentation rule;

[0023] S13, annotating the generated semantic units with contextual semantic features, wherein the annotated content includes the contextual relevance between the semantic units and the keyword weight vector representation, wherein the keyword weight vector is defined as:

[0024] ;

[0025] in, represents the vector of the i-th semantic unit, is the keyword weight distribution of the unit;

[0026] S14. According to user needs, select and call a precise answer model and a generalized answer model. The precise answer model is used to generate a unique and confident answer for the labeled semantic unit, and the generalized answer model is used to generate multiple candidate answers based on semantic similarity.

[0027] S15. Store the generated semantic units and their annotation information into a knowledge base, wherein the knowledge base adopts a multi-level index structure, including a semantic unit index and a context feature index.

[0028] Optionally, the S2 specifically includes:

[0029] S21. Call the pre-trained large model to perform deep semantic analysis on the generated semantic units. The analysis content includes the context relevance of the semantic units, the distribution of keyword features, and the semantic similarity with potential problems:

[0030] ;

[0031] in, represents the semantic similarity between question q and semantic unit u, and are the keyword feature vector components of the question and semantic unit respectively, and is the modulus of the corresponding eigenvector;

[0032] S22. Based on the semantic analysis results, the question-answer generation algorithm is called to generate a question-answer pair for each semantic unit. The question-answer pair includes a question Q and an answer A. The question is generated by expanding the topic and keywords of the semantic unit, and the answer is extracted and formatted from the context content of the semantic unit. The generated question-answer pair is represented in a structured form as follows:

[0033] ;

[0034] Among them, QA is the generated question and answer set, and Respectively represent the i-th question and its corresponding answer;

[0035] S23, through the click-type QA extraction function, the generated question and answer pairs are displayed to the user, allowing the user to edit, optimize and confirm the content of the question and answer, and the optimized question and answer pairs are stored as the final version;

[0036] S24. Call the vectorization algorithm to perform semantic vectorization on the final question-answer pair to generate an embedding vector for the question-answer pair:

[0037] ;

[0038] in, Represents the semantic embedding vector of the question-answer pair, and Semantic vector representations of questions and answers respectively;

[0039] S25. Store the generated question-answer pair embedding vector and its original text into a question-answer vector database to construct a searchable initial question-answer database, wherein the question-answer database adopts a hierarchical vector index structure, including an embedding vector index and an original text index.

[0040] Optionally, the S3 specifically includes:

[0041] S31. Select a standard question set from the question-answer vector database , standard questions are frequently used questions and representative core questions in the question-answering database. Each standard question is represented by a semantic feature vector as follows:

[0042] ;

[0043] in, For standard issues The semantic feature set of Indicates the semantic feature value of the keyword;

[0044] S32, calling the natural language generation module of the large model to generate an extended question according to the semantic features of the standard question. The generation process achieves semantic consistency of the extended question by analyzing the keyword weights and semantic relevance of the standard question;

[0045] S33. Based on the natural language generation results, call the AI ​​one-key expansion algorithm to generate an extended question set :

[0046] ;

[0047] in, To expand the set of questions, represents the i-th extended question, all extended questions are related to the standard question Have semantic similarity;

[0048] S34, the generated extended question The corresponding answer Associate to form an extended question-answer pair set , and store it as a new question-answer pair:

[0049] ;

[0050] Among them, Q is the set of extended question-answer pairs, To expand on the question, For standard issues Relevant answers;

[0051] S35. Call the question and answer vector database update module to synchronize the generated extended question and answer pairs and their semantic embedding vectors to the question and answer vector database, so as to realize the real-time dynamic expansion of the question and answer database and the searchability of the extended questions.

[0052] Optionally, the S4 specifically includes:

[0053] S41. Generate a dynamic Prompt template based on business scenario requirements to guide the big model to perform semantic analysis on user input. The Prompt template contains keywords, context-related prompts, and rule constraints:

[0054] ;

[0055] in, For dynamically generated prompts, Hints for keywords or rules included in the template;

[0056] S42, calling the large model to process user input Perform semantic analysis to extract semantic features of input :

[0057] ;

[0058] in, is the set of semantic features input by the user, Represents semantic feature value;

[0059] S43, based on semantic features , call the multi-intent parsing rules to classify the user input intent and extract the core intent ,The core intent is determined through the similarity scoring function;

[0060] S44. Extract the core intent Map to the corresponding business processing path , the mapping relationship is through the path mapping function Definition, where Represents the business path associated with the core intent;

[0061] S45. Based on the mapped business path , dynamically divert user input to the corresponding business module. The diversion process combines the processing rules of multi-intent scenarios with the path priority configuration to realize the automation and execution of business diversion.

[0062] Optionally, the S5 specifically includes:

[0063] S51. Based on the requirements of multiple business scenarios, a logical tree model of the text workflow is constructed. The logical tree model is composed of a set of business nodes. and a collection of logical paths composition:

[0064] ;

[0065] in, Represents the logical tree model, A collection of business nodes, each of which represents a specific business processing logic. is a set of logical paths;

[0066] S52: Each business node in the logic tree model Configure node rules. Node rules include trigger conditions, associated intents, and response logic. Node rules are defined through mapping functions as follows:

[0067] ;

[0068] in, Represents the i-th node in the logical tree model, represents the rules associated with the node, is the mapping function of the business rules;

[0069] S53. Based on the questions input by the user and node rules , dynamically route user questions to the target node of the logic tree model , the routing rule is expressed by the context matching formula:

[0070] ;

[0071] in, For the problem Is it compliant with the node The judgment result of the condition, is the context matching function;

[0072] S54, when the user question is routed to the target node After that, the large model knowledge base associated with the node is called to generate the answer. The answer generation process includes retrieving candidate answers from the node knowledge base. , and select the best answer through the semantic scoring function :

[0073] ;

[0074] in, is the best answer generated, For user input The semantic similarity score between the semantic features of and the candidate answer a;

[0075] S55, realize the collaborative processing of multiple business scenarios through the logical tree model, when the target node When the conditions are not met, the path backtracking mechanism is used to rematch other nodes. The collaborative processing supports dynamic jumping of user questions between multiple nodes, so that the questions can be properly responded to in complex business scenarios.

[0076] Optionally, the S6 specifically includes:

[0077] S61. Construct a collection of hot issues , used to quickly associate the user's current question. Each hot question is represented by keywords and semantic feature vectors:

[0078] ;

[0079] in, A collection of hot issues. The semantic feature vector representing the hot issues, n is the total number of hot issues;

[0080] S62: Call the multi-round dialogue context analysis algorithm and combine it with the user's current question And historical conversation records , generate contextual features , used to optimize semantic analysis and question completion:

[0081] ;

[0082] in, Represents contextual features, and are the semantic vectors of the user’s current question and historical conversation records, respectively, and f is the context fusion function;

[0083] S63, based on contextual features and Hot Issues Collection , optimize the user's current problem , generate completed and normalized questions ,The optimization process combines the context and hot issue features;

[0084] S64. Generate standardized questions Submit to the question-answer vector database and retrieve the candidate answer set based on semantic matching rules , and select the best answer :

[0085] ;

[0086] in, For the best answer, Represents the similarity score between the normalized question semantic features and the candidate answer semantic features.

[0087] Optionally, the S7 specifically includes:

[0088] S71. Questions about receiving user input , call the question-answer vector database to perform semantic matching on the question, based on the semantic feature vector of the user's question and the answer feature vector in the question-answer database The similarity score of is used to calculate the matching answer;

[0089] S72: If no answer is matched, the document recall module is activated to retrieve relevant document sets from the knowledge base. , the recall process is based on the keywords of the user's question and semantic features , filter relevant documents by semantic similarity:

[0090] ;

[0091] in, represents the collection of recalled documents, is a collection of knowledge base documents. is the similarity threshold, Score the semantic similarity between the question and the document;

[0092] S73. If no document is recalled, the Spark model is called to generate a dynamic associated answer , dynamic answer generation based on user input questions Semantic features of and contextual features :

[0093] ;

[0094] Among them, G is the dynamic answer generation function;

[0095] S74: If there is no result from S71 to S73, if the fixed fallback mechanism is enabled, the answer will be generated according to the preset fixed script. ,Fixed script content is generated based on the standard template configured for business;

[0096] S75: Issues that were not resolved through S71 to S74 Its contextual features Recorded to the intelligent learning module, including the semantic features of the question , context features , Document recall results and fallback response types.

[0097] Optionally, the S8 specifically includes:

[0098] S81. A collection of unresolved issues recorded in the fallback response module Extract questions and their context features, and record the semantic features of user questions , context features , Recall Results and fallback response types , the set of unsolved problems is represented as:

[0099] ;

[0100] S82, call the intelligent learning algorithm to solve the set of unsolved problems Perform semantic analysis to generate optimized data sets , optimize the dataset to generate new question-answer pairs by semantically associating unsolved questions with existing question-answer libraries:

[0101] ;

[0102] in, represents the generated answer, is the semantic feature set of the answers in the question-answering database, and gen is the generating function;

[0103] S83. Optimize the data set The question-answer pairs generated in Add to Q&A , and generate embedded vectors through vectorization algorithms, and update the index structure of the question-answering database vector database;

[0104] S84, based on the set of unsolved problems The analysis results are used to dynamically optimize the existing semantic parsing rules. The optimization rules adjust the semantic feature extraction and matching strategies through intelligent learning algorithms:

[0105] ;

[0106] in, is the optimized semantic rule, h is the optimization function;

[0107] S85, based on the optimized question-answer database and semantic rules , call the large model training module to retrain the model and update the model parameters , optimizing the model’s ability to understand and handle extended question-answering scenarios.

[0108] The beneficial effects of the present invention are:

[0109] (1) The present invention combines large model technology, semantic parsing algorithms and intelligent learning mechanisms to achieve deep semantic understanding and dynamic adaptability of user questions, enabling the system to accurately match user needs and efficiently handle multiple rounds of complex dialogue scenarios, thereby overcoming the limitations of traditional text customer service systems in semantic understanding and context association. In particular, it demonstrates excellent processing capabilities in multi-intent parsing and dynamic business diversion.

[0110] (2) The present invention significantly improves the efficiency of knowledge base construction and the coverage of questions and answers through the automated knowledge base construction module and question-answer vectorization technology, combined with the question-answer expansion algorithm, and reduces the need for manual intervention. The question-answer database can be expanded and optimized in real time, making the system highly flexible and scalable when facing diverse user expressions and new business needs.

[0111] (3) The present invention uses a contextual semantic association module and a dynamically optimized prompt template, combined with multi-round dialogue context analysis and hot question configuration, to achieve the completion and standardization of question statements. The system can accurately extract the question intent and generate high-quality standardized questions, greatly improving the processing efficiency and answer accuracy of user questions.

[0112] (4) The present invention uses a comprehensive response module, combined with a multi-level response mechanism that includes question-answer vector database matching, document recall, and dynamically generated answers, to ensure that users can obtain reasonable answers when faced with complex and unknown questions, and to record unresolved questions in the intelligent learning module to support subsequent optimization and continuous improvement.

[0113] (5) The present invention uses an intelligent learning and optimization module to dynamically expand the question-answer library and optimize semantic rules using an intelligent learning algorithm, and retrains the large model based on the optimized data and rules, thereby achieving the ability to automatically learn and continuously improve unresolved problems, thereby enabling the system to continuously improve its answer coverage and semantic understanding accuracy, significantly enhancing user experience and business processing capabilities. BRIEF DESCRIPTION OF THE DRAWINGS

[0114] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings:

[0115] Figure 1 This is the overall framework diagram of a large model-based text robot application system proposed by the present invention. DETAILED DESCRIPTION

[0116] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, which only illustrate the basic structure of the present invention in a schematic manner, and therefore only show the components related to the present invention.

[0117] refer to Figure 1 , a text robot application system based on a large model, comprising the following steps:

[0118] S1, knowledge base and document parsing module, receives business documents and FAQ forms in various formats provided by users, analyzes the document content using semantic parsing algorithms, decomposes the content into semantically related units, and annotates contextual semantic features. According to needs, it selects and calls precise answer models and generalized answer models;

[0119] S2, the question-answer generation and vectorization module calls the pre-trained large model, performs in-depth analysis on the semantic units generated by the knowledge parsing module, automatically generates question-answer pairs, supports users to view, edit and confirm question-answer pairs through the click-and-click QA extraction function, stores the optimized question-answer pairs in the question-answer vector database, and builds a searchable initial question-answer library;

[0120] S3, the question expansion module, based on the standard questions in the question and answer library, uses the natural language generation capability of the large model to call the AI ​​one-click expansion algorithm to generate diversified questions with similar semantics, dynamically expands the coverage of the question and answer library, and synchronously updates the expanded questions to the question and answer vector database;

[0121] S4, semantic intent analysis module, through the dynamically set prompt, calls the big model to perform deep semantic analysis on the questions input by the user, extracts the core intent features, and maps the question intent to the corresponding business processing path based on the multi-intent parsing rules, thus realizing the diversion of multi-intent scenarios and business automation processing;

[0122] S5, text workflow management module, defines text workflow rules for multiple business scenarios based on the logic tree model, dynamically routes user questions to the corresponding business branches according to the generated business paths, combines business branches, calls the large model knowledge base to generate answers, and supports collaborative processing and dynamic adjustment of multiple business scenarios;

[0123] S6, context semantic association module, by configuring hot questions and calling multi-round dialogue context analysis algorithm, combined with context association switch, performs semantic association analysis on the user's current question and historical dialogue records, sets context association optimization prompt, completes and standardizes user questions, submits the optimized questions to the question-answer vector database for retrieval, and returns the best answer;

[0124] S7, the fallback response module, which first calls the question-answer vector database to match user questions. If no answer is found, the document recall function is enabled to retrieve relevant content from the knowledge base. If there is no result in the document recall, the Spark Big Model is called to generate a dynamic associated answer. If the fixed fallback mechanism is enabled, the answer is generated according to the preset fixed words, and the unresolved questions and their contextual features are recorded in the intelligent learning module;

[0125] S8, intelligent learning and optimization module, for unresolved issues recorded by the catch-all response module, calls on the intelligent learning algorithm to dynamically expand and optimize the question-and-answer database and semantic rules, and retrains the large model to update the question-and-answer database and business logic rules.

[0126] In this implementation, S1 specifically includes:

[0127] S11. receiving business documents and FAQ forms in various formats provided by users, wherein supported document formats include structured data, semi-structured data, and unstructured data;

[0128] S12, calling a semantic parsing algorithm to analyze the received document content, dividing the document into a plurality of semantically related units, each unit including at least one subject, predicate and object, and realizing the logical integrity of the fragmentation according to a predefined semantic fragmentation rule;

[0129] S13, annotating the generated semantic units with contextual semantic features, wherein the annotated content includes the contextual relevance between the semantic units and the keyword weight vector representation, wherein the keyword weight vector is defined as:

[0130] ;

[0131] in, represents the vector of the i-th semantic unit, is the keyword weight distribution of the unit;

[0132] S14. According to user needs, select and call a precise answer model and a generalized answer model. The precise answer model is used to generate a unique and confident answer for the labeled semantic unit, and the generalized answer model is used to generate multiple candidate answers based on semantic similarity.

[0133] S15. Store the generated semantic units and their annotation information into a knowledge base, wherein the knowledge base adopts a multi-level index structure, including a semantic unit index and a context feature index.

[0134] In this implementation, S2 specifically includes:

[0135] S21. Call the pre-trained large model to perform deep semantic analysis on the generated semantic units. The analysis content includes the context relevance of the semantic units, the distribution of keyword features, and the semantic similarity with potential problems:

[0136] ;

[0137] in, represents the semantic similarity between question q and semantic unit u, and are the keyword feature vector components of the question and semantic unit respectively, and is the modulus of the corresponding eigenvector;

[0138] S22. Based on the semantic analysis results, the question-answer generation algorithm is called to generate a question-answer pair for each semantic unit. The question-answer pair includes a question Q and an answer A. The question is generated by expanding the topic and keywords of the semantic unit, and the answer is extracted and formatted from the context content of the semantic unit. The generated question-answer pair is represented in a structured form as follows:

[0139] ;

[0140] Among them, QA is the generated question and answer set, and Respectively represent the i-th question and its corresponding answer;

[0141] S23, through the click-type QA extraction function, the generated question and answer pairs are displayed to the user, allowing the user to edit, optimize and confirm the content of the question and answer, and the optimized question and answer pairs are stored as the final version;

[0142] S24. Call the vectorization algorithm to perform semantic vectorization on the final question-answer pair to generate an embedding vector for the question-answer pair:

[0143] ;

[0144] in, Represents the semantic embedding vector of the question-answer pair, and Semantic vector representations of questions and answers respectively;

[0145] S25. Store the generated question-answer pair embedding vector and its original text into a question-answer vector database to construct a searchable initial question-answer database, wherein the question-answer database adopts a hierarchical vector index structure, including an embedding vector index and an original text index.

[0146] In this implementation, S3 specifically includes:

[0147] S31. Select a standard question set from the question-answer vector database , standard questions are frequently used questions and representative core questions in the question-answering database. Each standard question is represented by a semantic feature vector as follows:

[0148] ;

[0149] in, For standard issues The semantic feature set of Indicates the semantic feature value of the keyword;

[0150] S32, calling the natural language generation module of the large model to generate an extended question according to the semantic features of the standard question. The generation process achieves semantic consistency of the extended question by analyzing the keyword weights and semantic relevance of the standard question;

[0151] S33. Based on the natural language generation results, call the AI ​​one-key expansion algorithm to generate an extended question set :

[0152] ;

[0153] in, To expand the set of questions, represents the i-th extended question, all extended questions are related to the standard question Have semantic similarity;

[0154] S34, the generated extended question The corresponding answer Associate to form an extended question-answer pair set , and store it as a new question-answer pair:

[0155] ;

[0156] Among them, Q is the set of extended question-answer pairs, To expand on the question, For standard issues Relevant answers;

[0157] S35. Call the question and answer vector database update module to synchronize the generated extended question and answer pairs and their semantic embedding vectors to the question and answer vector database, so as to realize the real-time dynamic expansion of the question and answer database and the searchability of the extended questions.

[0158] In this implementation, S4 specifically includes:

[0159] S41. Generate a dynamic Prompt template based on business scenario requirements to guide the big model to perform semantic analysis on user input. The Prompt template contains keywords, context-related prompts, and rule constraints:

[0160] ;

[0161] in, For dynamically generated prompts, Hints for keywords or rules included in the template;

[0162] S42, calling the large model to process user input Perform semantic analysis to extract semantic features of input :

[0163] ;

[0164] in, is the set of semantic features input by the user, Represents semantic feature value;

[0165] S43, based on semantic features , call the multi-intent parsing rules to classify the user input intent and extract the core intent ,The core intent is determined through a similarity scoring function;

[0166] S44. Extract the core intent Map to the corresponding business processing path , the mapping relationship is through the path mapping function Definition, where Represents the business path associated with the core intent;

[0167] S45. Based on the mapped business path , dynamically divert user input to the corresponding business module. The diversion process combines the processing rules of multi-intent scenarios with the path priority configuration to realize the automation and execution of business diversion.

[0168] In this implementation manner, S5 specifically includes:

[0169] S51. Based on the requirements of multiple business scenarios, a logical tree model of the text workflow is constructed. The logical tree model is composed of a set of business nodes. and a collection of logical paths composition:

[0170] ;

[0171] in, Represents the logical tree model, A collection of business nodes, each of which represents a specific business processing logic. is a set of logical paths;

[0172] S52: Each business node in the logic tree model Configure node rules. Node rules include trigger conditions, associated intents, and response logic. Node rules are defined through mapping functions as follows:

[0173] ;

[0174] in, Represents the i-th node in the logical tree model, represents the rules associated with the node, is the mapping function of the business rules;

[0175] S53. Based on the questions input by the user and node rules , dynamically route user questions to the target node of the logic tree model , the routing rule is expressed by the context matching formula:

[0176] ;

[0177] in, For the problem Is it compliant with the node The judgment result of the condition, is the context matching function;

[0178] S54, when the user question is routed to the target node After that, the large model knowledge base associated with the node is called to generate the answer. The answer generation process includes retrieving candidate answers from the node knowledge base. , and select the best answer through the semantic scoring function :

[0179] ;

[0180] in, is the best answer generated, For user input The semantic similarity score between the semantic features of and the candidate answer a;

[0181] S55, realize the collaborative processing of multiple business scenarios through the logical tree model, when the target node When the conditions are not met, the path backtracking mechanism is used to rematch other nodes. The collaborative processing supports dynamic jumping of user questions between multiple nodes, so that the questions can be properly responded to in complex business scenarios.

[0182] In this implementation manner, S6 specifically includes:

[0183] S61. Construct a collection of hot issues , used to quickly associate the user's current question. Each hot question is represented by keywords and semantic feature vectors:

[0184] ;

[0185] in, A collection of hot issues. The semantic feature vector representing the hot issues, n is the total number of hot issues;

[0186] S62: Call the multi-round dialogue context analysis algorithm and combine it with the user's current question And historical conversation records , generate contextual features , used to optimize semantic analysis and question completion:

[0187] ;

[0188] in, Represents contextual features, and are the semantic vectors of the user’s current question and historical conversation records, respectively, and f is the context fusion function;

[0189] S63, based on contextual features and Hot Issues Collection , optimize the user's current problem , generate completed and normalized questions ,The optimization process combines the context and hot issue features;

[0190] S64. Generate standardized questions Submit to the question-answer vector database and retrieve the candidate answer set based on semantic matching rules , and select the best answer :

[0191] ;

[0192] in, For the best answer, Represents the similarity score between the normalized question semantic features and the candidate answer semantic features.

[0193] In this implementation manner, the S7 specifically includes:

[0194] S71. Questions about receiving user input , call the question-answer vector database to perform semantic matching on the question, based on the semantic feature vector of the user's question and the answer feature vector in the question-answer database The similarity score of is used to calculate the matching answer;

[0195] S72: If no answer is matched, the document recall module is activated to retrieve relevant document sets from the knowledge base. , the recall process is based on the keywords of the user's question and semantic features , filter relevant documents by semantic similarity:

[0196] ;

[0197] in, represents the collection of recalled documents, is a collection of knowledge base documents. is the similarity threshold, Score the semantic similarity between the question and the document;

[0198] S73. If no document is recalled, the Spark model is called to generate a dynamic associated answer , dynamic answer generation based on user input questions Semantic features of and contextual features :

[0199] ;

[0200] Among them, G is the dynamic answer generation function;

[0201] S74: If there is no result from S71 to S73, if the fixed fallback mechanism is enabled, the answer will be generated according to the preset fixed script. ,Fixed script content is generated based on the standard template configured for business;

[0202] S75: Issues that were not resolved through S71 to S74 Its contextual features Recorded to the intelligent learning module, including the semantic features of the question , context features , Document recall results and fallback response types.

[0203] In this implementation manner, S8 specifically includes:

[0204] S81. A collection of unresolved issues recorded in the fallback response module Extract questions and their context features, and record the semantic features of user questions , context features , Recall Results and fallback response types , the set of unsolved problems is represented as:

[0205] ;

[0206] S82, call the intelligent learning algorithm to solve the set of unsolved problems Perform semantic analysis to generate optimized data sets , optimize the dataset to generate new question-answer pairs by semantically associating unsolved questions with existing question-answer libraries:

[0207] ;

[0208] in, represents the generated answer, is the semantic feature set of the answers in the question-answering database, and gen is the generating function;

[0209] S83. Optimize the data set The question-answer pairs generated in Add to Q&A , and generate embedded vectors through vectorization algorithms, and update the index structure of the question-answering database vector database;

[0210] S84, based on the set of unsolved problems The analysis results are used to dynamically optimize the existing semantic parsing rules. The optimization rules adjust the semantic feature extraction and matching strategies through intelligent learning algorithms:

[0211] ;

[0212] in, is the optimized semantic rule, h is the optimization function;

[0213] S85, based on the optimized question-answer database and semantic rules , call the large model training module to retrain the model and update the model parameters , optimizing the model’s understanding and processing capabilities for extended question-answering scenarios.

[0214] Embodiment 1:

[0215] In order to verify the feasibility and effectiveness of the present invention, the present invention is applied to the online customer service system of a large e-commerce platform in a certain city. The platform has an average daily user visit volume of more than one million, and the user consultation content covers multiple scenarios such as order inquiry, return and exchange, logistics tracking, etc. The types of questions are complex and the language expressions are diverse. Traditional text customer service systems usually respond by keyword matching and fixed rules, but they are unable to cope with complex multi-round dialogues, flexible question expressions, and multi-scenario business processing, resulting in low response efficiency and insufficient answer accuracy, which cannot meet the user's demand for efficient services.

[0216] In this application scenario, the system of the present invention achieves accurate answers to user questions and dynamic optimization of the system by deploying a knowledge base construction module, a question and answer generation module, a contextual semantic association module, a bottom-up response module, and an intelligent learning module. The system automatically parses platform business documents, converts document content into semantic fragments, and generates high-quality question and answer pairs. At the same time, it optimizes user question statements in combination with contextual information and uses semantic vectorization technology for efficient retrieval. When user questions exceed the coverage of the knowledge base, the system can provide relevant answers through document recall and dynamic generation functions, and unresolved questions are recorded in the intelligent learning module. The system capabilities are enhanced through subsequent optimization to expand the question and answer library and update model parameters.

[0217] During a big promotion, the e-commerce platform fully launched the system of the present invention, aiming to improve the processing efficiency and service quality of user consultation questions. In actual application, the system received a question from the user: "When will the pair of sneakers I bought arrive?" The system analyzes the historical conversation records through the context semantic association module, identifies the object of the user's question as "sneakers", and completes the question statement as "When will the sneakers of my order 123456789 arrive" in combination with the user's order information. The optimized question is submitted to the question-answer vector database for matching, and the system quickly retrieves the answer: "According to the logistics information, the product is expected to be delivered tomorrow." Then the user continues to ask: "Can the coupon still be used?" The system combines the context information and determines that the order the user asked is the previously mentioned order 123456789, and optimizes the question statement as "Can the coupon of order 123456789 still be used?" Finally, the answer is returned: "The coupon has expired and cannot be used." This multi-round dialogue processing capability significantly improves the user experience.

[0218] When faced with questions that are beyond the scope of the knowledge base, such as when a user asks, "How long does it take for same-city delivery to arrive?" the system activates the fallback response module. First, the document recall function is tried, but when no relevant documents are retrieved from the knowledge base, the Spark Big Model is called to dynamically generate an answer: "According to the same-city delivery rules, it is expected to be delivered in 1-2 days." This answer is generated by combining platform rules and contextual information, ensuring the effective resolution of user questions while accumulating data for future optimization of the Q&A library. In other cases, such as when a user asks, "Where is the installation video for my coffee machine?" Since neither the knowledge base nor the fallback module can generate an accurate answer, the question is recorded in the intelligent learning module. In subsequent optimizations, the system adds the question and the generated standard answer to the Q&A library to ensure that similar questions can be accurately answered in the future.

[0219] In order to further verify the system performance, a detailed analysis of the data within 30 days of going online was conducted, and the results showed that the system of the present invention is superior to the traditional text customer service system in many aspects. The data is shown in Table 1.

[0220] Table 1 Performance comparison of e-commerce platform customer service systems

[0221]

[0222] From the data, we can see that the average response time of the system of the present invention has been shortened from 7.5 seconds of the traditional system to 2.1 seconds, the accuracy of answers has been significantly improved to 96.7%, and the success rate of multi-round dialogues has been increased from 65.3% to 92.5%. User satisfaction has reached 94.8%, an increase of 16.6 percentage points over the traditional system, and the efficiency of knowledge base expansion has increased significantly, with the number of new question-answer pairs per day increasing from 15 to 230, fully demonstrating the advantages of the system in automation and dynamic optimization capabilities.

[0223] In an actual application, a user bought a coffee machine due to the "Double Eleven" promotion and asked, "When will the coffee machine I bought be shipped?" The system completed the question through semantic analysis to "What is the logistics status of the coffee machine in order 987654321?" and quickly retrieved the answer in the question and answer vector library: "The product has been shipped and is expected to arrive within 3 days." After that, the user asked, "How can I find the installation video?" Since there was no matching answer in the knowledge base, the bottom-up module tried document recall but still could not find relevant content, and finally dynamically generated the answer: "Please go to the official website to search for the product model and view the relevant installation video." This answer solved the user's problem and accumulated data records of unresolved problems for the system. With the optimization of the intelligent learning module, the question was expanded into a new question and answer pair and added to the question and answer library. In the future, the system will be able to directly answer similar questions.

[0224] To sum up, the large model-based text robot application system of the present invention effectively solves the deficiencies of traditional customer service systems in semantic understanding, context processing and dynamic optimization capabilities, greatly improves response efficiency and service quality, and at the same time achieves continuous improvement of the system through intelligent learning, providing an efficient and intelligent solution for text customer service in complex scenarios.

[0225] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can make equivalent replacements or changes according to the technical scheme and inventive concept of the present invention within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.

Claims

1. A text robot application system based on a large model, characterized in that: The steps include: S1, knowledge base and document parsing module, receives business documents in various formats provided by users and Table, use semantic parsing algorithm to analyze the document content, decompose the content into semantically related units, and mark the contextual semantic features. According to the needs, select and call the precise answer model and generalized answer model; S2, question-answer generation and vectorization module, calls the pre-trained large model, performs in-depth analysis on the semantically associated units generated by the knowledge parsing module, automatically generates question-answer pairs, stores the optimized question-answer pairs in the question-answer vector database, and builds a searchable initial question-answer library; S3, the question expansion module, based on the standard questions in the question and answer library, uses the natural language generation capability of the large model to call the AI ​​one-click expansion algorithm to generate diversified questions with similar semantics, and synchronously updates the expanded questions to the question and answer vector database; S4, semantic intent analysis module, through the dynamic setting template, The template contains keywords, context-related prompts, and rule constraints. It calls the big model to perform in-depth semantic analysis on the questions input by the user, extract the core intent features, and map the question intent to the corresponding business processing path based on multi-intent parsing rules. S5, text workflow management module, defines text workflow rules for multiple business scenarios based on the logic tree model, dynamically routes user questions to corresponding business branches according to the generated business paths, and calls the large model knowledge base to generate answers based on the business branches; S6, context semantic association module, by configuring hot issues and calling multiple rounds of dialogue context analysis algorithm, combined with context association switch, performs semantic association analysis on the user's current question and historical dialogue records, and sets context association optimization , complete and standardize user questions, submit the optimized questions to the question-answer vector database for retrieval, and return the best answer; S7, the fallback response module, which first calls the question-answer vector database to match the user's question. If no answer is matched, the document recall function is enabled to retrieve relevant content from the knowledge base; S8, intelligent learning and optimization module, for unresolved issues recorded by the catch-all response module, calls on the intelligent learning algorithm to dynamically expand and optimize the question-and-answer database and semantic rules, and retrains the large model to update the question-and-answer database and business logic rules.

2. A text robot application system based on a large model according to claim 1, characterized in that: The S1 specifically includes: S11. Receive business documents in various formats provided by users and Tables, which support document formats including structured data, semi-structured data, and unstructured data; S12, calling a semantic parsing algorithm to analyze the received document content, dividing the document into a plurality of semantically related units, each unit including at least one subject, predicate and object, and realizing the logical integrity of the fragmentation according to a predefined semantic fragmentation rule; S13, annotating the generated semantic units with contextual semantic features, wherein the annotated content includes the contextual relevance between the semantic units and the keyword weight vector representation, wherein the keyword weight vector is defined as: ； in, represents the vector of the i-th semantic unit, is the keyword weight distribution of this unit; S14. According to user needs, select and call a precise answer model and a generalized answer model. The precise answer model is used to generate a unique and confident answer for the labeled semantic unit, and the generalized answer model is used to generate multiple candidate answers based on semantic similarity. S15. Store the generated semantic units and their annotation information into a knowledge base, wherein the knowledge base adopts a multi-level index structure, including a semantic unit index and a context feature index.

3. A text robot application system based on a large model according to claim 1, characterized in that: The S2 specifically includes: S21. Call the pre-trained large model to perform deep semantic analysis on the generated semantic units. The analysis content includes the context relevance of the semantic units, the distribution of keyword features, and the semantic similarity with potential problems: ； in, represents the semantic similarity between question q and semantic unit u, and are the keyword feature vector components of the question and semantic unit respectively, and is the modulus of the corresponding eigenvector; S22. Based on the semantic analysis results, the question-answer generation algorithm is called to generate a question-answer pair for each semantic unit. The question-answer pair includes a question Q and an answer A. The question is generated by expanding the topic and keywords of the semantic unit, and the answer is extracted and formatted from the context content of the semantic unit. The generated question-answer pair is represented in a structured form as follows: ； Among them, QA is the generated question and answer set, and Respectively represent the i-th question and its corresponding answer; S23, through the click-type QA extraction function, the generated question and answer pairs are displayed to the user, allowing the user to edit, optimize and confirm the content of the question and answer, and the optimized question and answer pairs are stored as the final version; S24. Call the vectorization algorithm to perform semantic vectorization on the final question-answer pair to generate an embedding vector for the question-answer pair: ； in, Represents the semantic embedding vector of the question-answer pair, and Semantic vector representations of questions and answers respectively; S25. Store the generated question-answer pair embedding vector and its original text into a question-answer vector database to construct a searchable initial question-answer database, wherein the question-answer database adopts a hierarchical vector index structure, including an embedding vector index and an original text index.

4. A text robot application system based on a large model according to claim 1, characterized in that: The S3 specifically includes: S31. Select a standard question set from the question-answer vector database , standard questions are frequently used questions and representative core questions in the question-answering database. Each standard question is represented by a semantic feature vector as follows: ； in, For standard issues The semantic feature set of Indicates the semantic feature value of the keyword; S32, calling the natural language generation module of the large model to generate an extended question according to the semantic features of the standard question. The generation process achieves semantic consistency of the extended question by analyzing the keyword weights and semantic relevance of the standard question; S33. Based on the natural language generation results, call the AI ​​one-key expansion algorithm to generate an extended question set : ； in, To expand the set of questions, represents the i-th extended question, all extended questions are related to the standard question Have semantic similarity; S34, the generated extended question The corresponding answer Associate to form an extended question-answer pair set , and store it as a new question-answer pair: ； Among them, Q is the set of extended question-answer pairs, To expand on the question, For standard issues Relevant answers; S35. Call the question and answer vector database update module to synchronize the generated extended question and answer pairs and their semantic embedding vectors to the question and answer vector database, so as to realize the real-time dynamic expansion of the question and answer database and the searchability of the extended questions.

5. A text robot application system based on a large model according to claim 1, characterized in that: The S4 specifically includes: S41. Generate a dynamic prompt template based on business scenario requirements to guide the big model to perform semantic analysis on user input: ； in, For dynamically generated prompts, Hints for keywords or rules included in the template; S42, calling the large model to process user input Perform semantic analysis to extract semantic features of input : ； in, is the set of semantic features input by the user, Represents semantic feature value; S43, based on semantic features , call the multi-intent parsing rules to classify the user input intent and extract the core intent ,The core intent is determined through the similarity scoring function; S44. Extract the core intent Map to the corresponding business processing path , the mapping relationship is through the path mapping function Definition, where Represents the business path associated with the core intent; S45. Based on the mapped business path , dynamically divert user input to the corresponding business module. The diversion process combines the processing rules of multi-intent scenarios with the path priority configuration to realize the automation and execution of business diversion.

6. A text robot application system based on a large model according to claim 1, characterized in that: The S5 specifically includes: S51. Based on the requirements of multiple business scenarios, a logical tree model of the text workflow is constructed. The logical tree model is composed of a set of business nodes. and a collection of logical paths composition: ； in, Represents the logical tree model, A collection of business nodes, each of which represents a specific business processing logic. is a set of logical paths; S52: Each business node in the logic tree model Configure node rules. Node rules include trigger conditions, associated intents, and response logic. Node rules are defined through mapping functions as follows: ； in, Represents the i-th node in the logical tree model, represents the rules associated with the node, is the mapping function of the business rules; S53. Based on the questions input by the user and node rules , dynamically route user questions to the target node of the logic tree model , the routing rule is expressed by the context matching formula: ； in, For the problem Is it compliant with the node The judgment result of the condition, is the context matching function; S54, when the user question is routed to the target node After that, the large model knowledge base associated with the node is called to generate the answer. The answer generation process includes retrieving candidate answers from the node knowledge base. , and select the best answer through the semantic scoring function : ； in, is the best answer generated, For user input The semantic similarity score between the semantic features of and the candidate answer a; S55, realize the collaborative processing of multiple business scenarios through the logical tree model, when the target node When the conditions are not met, the path backtracking mechanism is used to rematch other nodes. The collaborative processing supports dynamic jumping of user questions between multiple nodes, so that the questions can be properly responded to in complex business scenarios.

7. A text robot application system based on a large model according to claim 1, characterized in that: The S6 specifically includes: S61. Construct a collection of hot issues , used to quickly associate the user's current question. Each hot question is represented by keywords and semantic feature vectors: ； in, A collection of hot issues. The semantic feature vector representing the hot issues, n is the total number of hot issues; S62: Call the multi-round dialogue context analysis algorithm and combine it with the user's current question And historical conversation records , generate contextual features , used to optimize semantic analysis and question completion: ； in, Represents contextual features, and are the semantic vectors of the user’s current question and historical conversation records, respectively, and f is the context fusion function; S63, based on contextual features and Hot Issues Collection , optimize the user's current problem , generate completed and normalized questions ,The optimization process combines the context and hot issue features; S64. Generate standardized questions Submit to the question-answer vector database and retrieve the candidate answer set based on semantic matching rules , and select the best answer : ； in, For the best answer, Represents the similarity score between the normalized question semantic features and the candidate answer semantic features.

8. A text robot application system based on a large model according to claim 1, characterized in that: The S7 specifically includes: S71. Questions about receiving user input , call the question-answer vector database to perform semantic matching on the question, based on the semantic feature vector of the user's question and the answer feature vector in the question-answer database The similarity score of is used to calculate the matching answer; S72: If no answer is found, the document recall module is enabled to retrieve relevant document sets from the knowledge base. , the recall process is based on the keywords of the user's question and semantic features , filter relevant documents by semantic similarity: ； in, represents the collection of recalled documents, is a collection of knowledge base documents. is the similarity threshold, Score the semantic similarity between the question and the document; S73. If no document is recalled, the Spark model is called to generate a dynamic associated answer , dynamic answer generation based on user input questions Semantic features of and contextual features : ； Among them, G is the dynamic answer generation function; S74: If there is no result from S71 to S73, if the fixed fallback mechanism is enabled, the answer will be generated according to the preset fixed script. ,Fixed script content is generated based on the standard template configured for business; S75: Issues that were not resolved through S71 to S74 Its contextual features Recorded to the intelligent learning module, including the semantic features of the question , context features , Document recall results and fallback response types.

9. A text robot application system based on a large model according to claim 1, characterized in that: The S8 specifically includes: S81. A collection of unresolved issues recorded in the fallback response module Extract questions and their context features, and record the semantic features of user questions , context features , Recall Results and fallback response types , the set of unsolved problems is represented as: ； S82, call the intelligent learning algorithm to solve the set of unsolved problems Perform semantic analysis to generate optimized data sets , optimize the dataset to generate new question-answer pairs by semantically associating unsolved questions with existing question-answer libraries: ； in, represents the generated answer, is the semantic feature set of the answers in the question-answering database, and gen is the generating function; S83. Optimize the data set The question-answer pairs generated in Add to Q&A , and generate embedded vectors through vectorization algorithms, and update the index structure of the question-answering database vector database; S84, based on the set of unsolved problems The analysis results are used to dynamically optimize the existing semantic parsing rules. The optimization rules adjust the semantic feature extraction and matching strategies through intelligent learning algorithms: ； in, is the optimized semantic rule, h is the optimization function; S85, based on the optimized question-answer database and semantic rules , call the large model training module to retrain the model and update the model parameters , optimizing the model’s understanding and processing capabilities for extended question-answering scenarios.

Images