A multi-dimensional intent recognition and dynamic knowledge fusion question method and device

Abstract

The application discloses a kind of multi-dimension intention recognition and dynamic knowledge fusion question method and device.The method includes S1, obtaining the question of user, pre-processes the input of user, obtains the question after revision;S2, according to the question after revision obtains user intent and extracts entity, then entity is linked with knowledge base;S3, according to user intent and entity, extract the data required from database, then according to the data extracted, user intent and the content of knowledge base linked by entity form prompt word, input into big model to drive its analysis trend, analysis influencing factor, data prediction or generate visual suggestion, form report.Knowledge base and database are dynamically adapted, ensure that knowledge base updates and data changes are mutually induced, enhance system adaptability;Through error correction module and revision module, input deviation is corrected, the accuracy rate of misused character correction and the specification rate of ambiguous expression are improved, and the error rate of intention recognition caused by input deviation is reduced.

Description

Technical Field

[0001] This invention relates to question counting, and in particular to a method and apparatus for multi-dimensional intent recognition and dynamic knowledge fusion question counting. Background Technology

[0002] With the rapid growth of enterprise data volume, query systems (i.e., data retrieval via natural language) have become an important tool for internal enterprise decision-making. Traditional query systems are typically based on rule-based or simple natural language processing (NLP) techniques, directly converting user queries into SQL statements (Text2SQL), and then retrieving corresponding data from the database based on the SQL statements. However, these systems suffer from the following core pain points: Traditional Text2SQL technology is tightly coupled with the underlying table structure, while enterprise domain knowledge is constantly changing; each new data dimension requires a restructuring of the data processing workflow. User queries often involve complex intentions involving multiple dimensions and indicators, which are usually scattered across various business systems, making it impossible to perform interconnected analysis, resulting in low query accuracy and poor adaptability. Furthermore, user input may contain non-standard, erroneous, or ambiguous expressions, further impacting system performance. Existing systems struggle to effectively integrate enterprise-specific knowledge bases and data, failing to provide in-depth data analysis and decision-making recommendations, thus limiting the practical value of query systems.

[0003] Currently, the mainstream query systems mainly adopt the following technical solutions: Traditional Text2SQL methods: directly map user natural language queries into SQL query statements through rule engines or statistical models, such as template matching or sequence-to-sequence Seq2Seq models; Intent recognition and entity extraction: Pre-trained NLP models (such as BERT) are used for intent classification and entity recognition, but they are usually limited to a single dimension or a fixed domain and lack the ability to resolve multi-dimensional intents. Static knowledge integration: Some system integration knowledge bases exist, but these knowledge bases are updated slowly and cannot dynamically adapt to changes in internal enterprise data, resulting in inaccurate data query results.

[0004] While the above solutions have achieved the conversion of natural language into data to a certain extent, they have failed to effectively solve the problems of multi-dimensional querying, dynamic knowledge fusion, and in-depth analysis. Summary of the Invention

[0005] To address the aforementioned problems, this invention provides a method and apparatus for multi-dimensional intent recognition and dynamic knowledge fusion.

[0006] This invention provides the following technical solution: a multi-dimensional intent recognition and dynamic knowledge fusion question-answering method, comprising the following steps: S1. Obtain the user's question, preprocess the user's input, and obtain the rewritten question; S2. Obtain user intent and extract entities based on the rewritten question, and then link the entities to the knowledge base; S3. Based on user intent and entities, extract the required data from the database, and then generate prompt words based on the extracted data, user intent, and knowledge base content linked to the entities. Input these prompt words into the large model to drive it to analyze trends, analyze influencing factors, predict data, or generate visualization suggestions to form a report.

[0007] Furthermore, the preprocessing includes correcting typos in the user's question using an error correction model, and then rewriting the typo-corrected question using a rewriting model to obtain the rewritten question.

[0008] Furthermore, the user intent includes trend analysis, influencing factor analysis, data prediction, and generating visualization suggestions, and the entities include organizations, time, and data indicator names.

[0009] Furthermore, the error correction model is constructed through the following steps: A lightweight BERT model was fine-tuned using the Sighan 2015 Chinese spelling correction data and the First Enterprise dataset. The fine-tuning parameters included the learning rate, the number of training samples per batch, the gradient accumulation steps, the L2 regularization weights, the number of training rounds, and the gradient clipping threshold, to form the error correction model.

[0010] Furthermore, the rewriting model is constructed through the following steps: Based on dialogue-utterance-rewriter, a rewriting model is formed by training on a second enterprise dataset to achieve coreference resolution and information completion.

[0011] Furthermore, the process of obtaining user intent and extracting entities is accomplished using a large model.

[0012] A multi-dimensional intent recognition and dynamic knowledge fusion problem includes at least: The acquisition module is used to acquire user questions, preprocess user input, and obtain rewritten questions; The extraction module is used to obtain user intent and extract entities based on the rewritten question, and then link the entities to the knowledge base; The generation module is used to extract the required data from the database based on user intent and entities. Then, based on the extracted data, user intent, and knowledge base content linked to the entities, it generates prompt words, which are then input into the large model to drive it to analyze trends, analyze influencing factors, predict data, or generate visualization suggestions to form a report.

[0013] An electronic device, comprising: One or more processors; Memory, used to store one or more programs; When the one or more programs are executed by the one or more processors, the one or more processors perform the methods described above.

[0014] A computer-readable storage medium having computer instructions stored thereon, which, when executed by a processor, implement the steps of the method described above.

[0015] The beneficial effects of this invention are as follows: (1) Multi-dimensional intent recognition: accurately analyze the multi-dimensional and multi-indicator intent in user queries to improve query accuracy; (2) Dynamic knowledge integration: The knowledge base and the database are dynamically adapted to ensure that knowledge base updates and data changes drive each other and enhance the system's adaptability; (3) Intelligent decision output: Based on data query, integrate large models to perform data analysis, chart generation, report writing and even business suggestions, to achieve full automation of query, data and suggestions; (4) By correcting input deviations through the error correction module and the rewriting module, the accuracy of typo correction and the standardization rate of fuzzy expression are improved, and the error rate of intent recognition caused by input deviations is reduced; (5) Modular architecture: Construct a three-layer collaborative mechanism of algorithm, data and big model, with each layer iterating independently, supporting rapid expansion of new business scenarios and improving response speed and flexibility. Attached Figure Description

[0016] Figure 1 This is a flowchart of the multi-dimensional intent recognition and dynamic knowledge fusion question-answering method of the present invention; Figure 2 This is a block diagram of the multi-dimensional intent recognition and dynamic knowledge fusion question-answering device of the present invention. Detailed Implementation

[0017] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the following detailed description of the specific implementation methods, structures, features, and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided below.

[0018] By correcting input biases through error correction and rewriting modules, the accuracy of typo correction and the standardization of ambiguous expressions are improved, reducing the intention recognition error rate caused by input biases. The knowledge base and database are dynamically adapted to ensure that knowledge base updates and data changes are mutually reinforcing, enhancing the system's adaptability. Finally, a report is generated by analyzing trends, influencing factors, predicting or generating visual suggestions through a large model.

[0019] The embodiments of the present invention will be further described below with reference to several examples.

[0020] Example 1 This invention adopts a three-layer collaborative architecture: an algorithm layer, a data layer, and a large model layer. Its core is to transform fuzzy user natural language queries into structured data query instructions through layer-by-layer refinement and enhancement, and finally output insights and suggestions with decision-making value.

[0021] like Figure 1 A multi-dimensional intent recognition and dynamic knowledge fusion question-answering method includes the following steps: S1. Obtain the user's question, preprocess the user's input, and obtain the rewritten question; Preprocessing includes correcting typos in the user's question using an error correction model, and then rewriting it using a rewriting model to obtain the rewritten question.

[0022] The error correction model is constructed through the following steps: A lightweight BERT model was fine-tuned using the Sighan 2015 Chinese spelling correction data and the First Enterprise dataset. The fine-tuning parameters included the learning rate, batch size, gradient accumulation steps, L2 regularization weights, training epochs, and gradient clipping threshold, which formed the error correction model.

[0023] Specifically: In this embodiment, the lightweight BERT model is Macbert, with learning_rate of 2e-5, batch_size of 64, gradient_accumulation_steps of 2, weight_decay of 0.01, num_train_epochs of 3, and max_grad_norm of 1.0.

[0024] The first enterprise dataset includes commonly used enterprise information, terminology, and business information, which can be constructed by those skilled in the art according to the actual situation of different enterprises.

[0025] The rewritten model is constructed through the following steps: Based on dialogue-utterance-rewriter, a rewriting model is formed by training on a second enterprise dataset to achieve coreference resolution and information completion.

[0026] The rewriting model uses dialogue-utterance-rewriter, obtained from https: / / github.com / LIJY0402 / dialogue-utterance-rewriter. It is fine-tuned using the second enterprise dataset. After fine-tuning, the model achieves coreference resolution (ensuring that the rewritten statement contains the correct referent entities) and information completion (filling in the missing key parts of the dialogue).

[0027] The fine-tuning parameters include: the maximum length of the source sequence (max_length_source) is 64, the maximum length of the target sequence (max_length_target) is 64, the vocabulary size (vocab_size) is 5000, the hidden layer dimension (hidden_size) is 256, the number of hidden layers (num_hidden_layers) is 2, the number of attention heads (num_heads) is 4, and the feedforward network dimension (filter_size) is 1024.

[0028] The second enterprise dataset includes commonly used enterprise information, terminology, and business information, which can be constructed by those skilled in the art according to the actual situation of different enterprises.

[0029] The goal of this step is to optimize user input, laying the foundation for accurate intent recognition in the future. This approach employs a dual-module collaborative strategy: In this embodiment, the first enterprise dataset and the second enterprise dataset are only examples. Those skilled in the art can construct the first enterprise dataset and the second enterprise dataset as needed.

[0030] S2. Obtain user intent and extract entities based on the rewritten question, and then link the entities to the knowledge base; User intent includes trend analysis, influencing factor analysis, data prediction, and generating visualization suggestions. Entities include organizations, time periods, and data indicator names.

[0031] The process of obtaining user intent and extracting entities is accomplished using a large model.

[0032] Specifically: The large model uses a locally deployed qwen3 large model combined with prompts to obtain user intent and extract entities. The prompts are as follows: You are a professional enterprise data analytics assistant, responsible for intent recognition and entity extraction from user queries. Please perform your tasks strictly according to the following requirements: 1. Identify a unique idea graph from user questions, choosing from only four given intentions; 2. Extract all entities of the specified type from the user query; return an empty array if none are found. 3. The output must be in strict JSON format and cannot contain any additional explanations, descriptions, or Markdown; 4. If the user's question is unclear or does not belong to any given intent, the intent field should be set to "unknown".

[0033] Intent definition (only one of the following 4 can be selected): Trend analysis: Inquire about the changes, trends, rises and falls, year-on-year / month-on-month changes, etc. of a certain indicator over a period of time; Influencing factor analysis: Inquire about the reasons for changes in a certain indicator, including influencing factors, driving factors, and limiting factors; Data prediction: Inquiring about the future value, trend, prediction, and estimate of a certain indicator; Generate visualization suggestions: Ask what charts should be used to display the data, how to visualize it, and recommend chart types, etc. Entity definition (extract all content that matches the following types): Organizational structure: Names of companies, departments, branches, subsidiaries, institutions, teams, etc. Time: Year, quarter, month, date, time period, cycle, etc. (e.g., "2024", "first quarter", "last three months", "last year's period"); Data metric names: Names of various business metrics, statistical metrics, and data items (such as "sales revenue", "profit", "number of users", "conversion rate", "market share"); Output format: { "intent": "intent name" “entities”: { “organizations”: [“Entity 1”, “Entity 2”, ...], "times": ["Entity 1", "Entity 2", ...], “metrics”: [“Entity 1”, “Entity 2”, ……] }, "confidence": 0.95 }; Example: Input: What is the sales trend of Transfar Logistics in the first quarter of 2024? Output: { "intent": "trend analysis" “entities”: { “organizations”: [“logistics”], "times": ["First quarter of 2024"], “metrics”: [“Sales”] }, "confidence": 0.98 }; Input: What are the main factors affecting the growth of Hangzhou Port's throughput? Output: { "intent": "Analysis of influencing factors" “entities”: { “organizations”: [“Hangzhou Port”], "times": [], “metrics”: [“throughput”] }, "confidence": 0.97 }; Input: What is the predicted net profit of Transfar Smart Logistics in 2025? Output: { "intent": "data prediction" “entities”: { “organizations”: [“Company X”], "times": ["2025"], “metrics”: [“Net Profit”] }, "confidence": 0.96 }; Input: What chart should be used to display the distribution of freight volume at highway ports across the country? Output: { "intent": "Generate visualization suggestions" “entities”: { “organizations”: [“National Highway Ports”], "times": [], “metrics”: [“freight volume”] }, "confidence": 0.95 }; Please now address the following user questions: {{user_query}}.

[0034] The above are the prompt words.

[0035] In this embodiment, {{user_query}} represents the modified query.

[0036] The knowledge base is a structured text knowledge base based on vector retrieval, deployed on the underlying Milvus vector database. It is used to store vectorized representations and associated attribute information of text data such as news, regulations, and laws. After collecting text file data such as news, regulations, and laws, the text is structured and parsed to extract the core attribute information of each text, obtaining preprocessed text content, which includes, but is not limited to: file name, file upload timestamp, and complete file text content. The BGE-M3 embedding model is used to perform vector transformation on the preprocessed text content, generating text content vectors with semantic representation capabilities. The text content vectors are then written into the Milvus vector database, completing the construction of the knowledge base.

[0037] This step aims to gain a deeper understanding of the "motivation" and "components" of user queries.

[0038] Dynamic knowledge fusion and entity linking: The extracted entities are linked to the knowledge base. Cosine similarity is used to calculate the semantic similarity between the entity and the text content vector in the knowledge base. The links are sorted according to the similarity scores, thereby enabling the retrieval of knowledge base content.

[0039] S3. Based on user intent and linked entities, extract the required data from the database, and then generate prompts based on the extracted data, user intent, and text content of the knowledge base. Input these prompts into the large model to drive it to analyze trends, analyze influencing factors, predict data, or generate visualization suggestions to form a report.

[0040] In this embodiment, the large model is qwen3, but LLaMA or other models can also be used.

[0041] The database is an enterprise-level structured operational database. Its core purpose is to store and manage structured business data generated by various internal departments (such as marketing, sales, R&D, and finance) during their daily operations. This includes, but is not limited to, the following categories: Departmental business operation data: business order data, project progress data, customer communication data, and task execution data for each department; Departmental resource allocation data: Human resource allocation data for each department (number of personnel, job distribution, attendance data), material requisition data (office equipment, material consumption), and budget usage data (expenditure, budget surplus, cost allocation). Departmental process operation data: Business process flow data of each department (approval process nodes, process time), cross-departmental collaboration data (collaboration task allocation, collaboration completion status), and operational anomaly data (fault records, complaint handling data).

[0042] This step no longer simply returns data, but leverages the analytical and reasoning capabilities of large models to provide value-added services.

[0043] Specifically: Based on user intent and entities, the required data is extracted from the database. Then, based on the extracted data, user intent, and the knowledge base content linked to the entities, prompt words are generated and input into the large model to drive it to perform one or more of the following tasks: The specific prompt words are: You are a professional enterprise data analytics expert, skilled at generating accurate, clear, and valuable analytical content based on structured data and industry knowledge. Please strictly adhere to the following requirements when performing this task: Core Ironclad Rules (Violation of any one of these rules will result in an unsatisfactory answer): 1. Zero Illusion Principle: All content must be 100% derived from the "database" and "knowledge base," and fabricating any data, cases, reasons, or conclusions is strictly prohibited. 2. Absolute Intent Match: Responses must be completely focused on the identified "user intent" and must not deviate from the topic or answer irrelevant questions. 3. Data Priority Principle: All conclusions must be supported by concrete data. Databases should be used preferentially; knowledge bases should only be used to explain background, provide reasons, and supplement context. 4. Principle of objectivity and neutrality: Use professional and objective language, and avoid subjective assumptions and emotional expressions. 5. Clear Structure Principle: Employ a standardized analytical structure, using subheadings and bullet points to organize content and improve readability. Input information: Original user question: {{user_query}}; Identified user intent: {{intent}}; Extracted entity information: {{entities}}; Structured data retrieved from the database: {{database_data}}; The relevant content retrieved from the knowledge base is: {{knowledge_base_content}}; Five key intentions and specific response requirements: 1. Trend Analysis Core objective: To clearly demonstrate the changing patterns of indicators over time; Answer structure: Key conclusion: Summarize the overall trend (upward / downward / fluctuating / stable) and magnitude of the indicator in one sentence; Detailed data: List the specific values ​​for key time points and calculate the year-on-year and month-on-month growth rates; Trend characteristics: Extract the main features of the trend (such as continuous growth, rise followed by fall, seasonal fluctuations, cyclical changes, etc.). Brief description: Based on the knowledge base content, provide reasonable explanations for obvious trend changes.

[0044] 2. Analysis of Influencing Factors Core objective: To identify and quantify the key factors influencing changes in indicators; Answer structure: Key conclusion: Clearly identify 1-2 of the most important factors affecting this indicator; Factor list: List all relevant factors in descending order of their influence; Quantitative analysis: Using data to explain the specific degree of influence of each factor on the indicator; Causal explanation: Explaining how each factor affects the target indicator.

[0045] 3. Data Prediction Core objective: To provide reasonable future predictions based on historical data; Answer structure: Forecast results: Clearly provide the core values ​​or ranges for the forecast; Basis for prediction: Explain the range of historical data used and the prediction method; Confidence assessment: Provide the confidence level (high / medium / low) of the prediction results and the reasons; Risk warning: This section identifies uncertainties that may affect the accuracy of the forecast.

[0046] 4. Visualization suggestions Core objective: To recommend the most suitable chart types and design schemes for displaying data; Answer structure: Core Recommendation: We recommend one main chart that best showcases the core conclusions; Detailed design: Chart Types: Explicitly recommended chart types (line chart / bar chart / pie chart / scatter plot / heat map, etc.); Coordinate axes: These indicate the data dimensions displayed by the X and Y axes, respectively. Data Series: List the data series that need to be displayed; Key annotations: Indicate data points or trends that require special annotation; Additional suggestion: We recommend using 1-2 supplementary charts to display detailed information.

[0047] 5. Generate report Core objective: To generate a complete, structured, and comprehensive data analysis report; the report structure follows the "five key intent definitions and specific response requirements" as follows: I. Report Summary Report Topic: Clearly define the subject matter and time frame of the report's analysis; Key findings: Summarize the most important findings of this analysis in 3-5 sentences; Data source: Explain the source of the data and the statistical period.

[0048] II. Trend Analysis The discussion unfolds according to the "Trend Analysis" structure within the "Five Intent Definitions and Specific Response Requirements".

[0049] III. Analysis of Influencing Factors The discussion unfolds according to the "Analysis of Influencing Factors" structure within the "Five Intents Definition and Specific Response Requirements".

[0050] IV. Data Forecasting The discussion unfolds according to the "Data Prediction" structure within the "Five Intents Definition and Specific Response Requirements".

[0051] V. Visualization Solution The presentation follows the structure of "Visual Suggestions" within the "Five Intents Definition and Specific Response Requirements".

[0052] VI. Conclusions and Recommendations Main conclusions: Summarize the core findings of this analysis; Action recommendations: Provide 2-3 specific and actionable business recommendations; Output requirements: The answer should be in Chinese, concise, professional, and avoid redundancy. Important data and conclusions are marked in bold; The table data should retain its original format to ensure neat alignment. Response length: 300-800 words for a single intent, 1000-2000 words for the generated report; You must not output any content unrelated to the question, and you must not add any additional explanations.

[0053] Special case handling: If the database is empty: explicitly state "No relevant data was found, analysis cannot be performed"; If the knowledge base is empty: This means "There is no relevant background knowledge available, and the analysis is based solely on the database data"; If the data is insufficient to answer the question: clearly state, "The existing data is insufficient to answer this question. We suggest supplementing the following data:..."; If the intent is unclear: "We cannot clarify your analysis needs; please provide further details." Now, please generate an answer based on the above requirements.

[0054] The above are the prompt words.

[0055] The technical principles of this invention are based on the following core pillars: The principle of hierarchical collaboration between task-specific models and general models: Explanation of the principle: Following the principle of "letting specialized models do specialized tasks," lightweight, targeted error correction and rewriting models are used for well-defined tasks requiring high accuracy and low latency (such as typo correction and rewriting). For tasks requiring generalization, complex reasoning, and content generation (such as user intent recognition, entity extraction, and report generation), more powerful general-purpose models are invoked. This layered architecture achieves an optimal balance between cost and effectiveness while ensuring performance.

[0056] In practical applications, users ask, "Based on the 2025 data on vehicle parking revenue at Hangzhou Highway Port, analyze the changing trends and influencing factors." The error correction model retrieved the question for typo correction: "Based on the parking revenue data of Hangzhou Highway Port in 2025, analyze the changing trends and influencing factors." The model was rewritten to obtain the revised question as "Based on the parking revenue data of Hangzhou Highway Port in 2025, analyze the changing trends and influencing factors", which is the aforementioned {{user_query}}; The large model (qwen3) uses prompt words (as described in step S2 of this embodiment) to identify the user's intent as trend analysis and influencing factor analysis based on the rewritten question. It obtains the entity as: organization - Hangzhou Highway Port, time - 2025, and data indicator name - parking revenue. It links the entity and the knowledge base to obtain the knowledge base content, and then queries the database. Combining the powerful analytical capabilities of the large model, it performs trend analysis on historical data and influencing factor analysis based on the content in the knowledge base. Finally, it outputs the analysis results and provides business suggestions.

[0057] In some embodiments, different models or hybrid techniques may be used for each step. The error correction model, rewriting model, joint model, and large model can adopt a single ultra-large model core architecture. That is, a powerful large model is used to directly complete all tasks from input, error correction, intent recognition, knowledge retrieval, data query to analysis report generation by designing prompt words and context learning. The system layer only interfaces with the knowledge base, database, and prompt word templates. This approach simplifies the system complexity but places extremely high demands on the capabilities of the large model and the context window.

[0058] In some embodiments, a hybrid architecture combining traditional machine learning and rules is used. For scenarios with limited resources or high cost constraints, it is not necessary to rely entirely on large models. An alternative is to adopt a hybrid architecture combining traditional machine learning and rule engines. For example, the joint model uses traditional BERT to obtain user intent, and then combines rule templates for fine-grained slot filling to extract entities. When generating a report, a template filling method is used to fill the extracted data, user intent, and knowledge base content linked to the entities into a preset report template to form the report.

[0059] In some embodiments, alternatives to intent recognition and entity extraction include retraining or fine-tuning a smaller, more specialized model for an enterprise-specific domain, instead of relying on a general, pre-trained large model. This model is specifically designed to identify domain-specific metrics, dimensions, and entities, and may be more accurate and faster on specific tasks.

[0060] In some embodiments, alternatives to linking entities to the knowledge base include static indexing with periodic updates, or near real-time or batch update modes.

[0061] Example 2 like Figure 2 A multi-dimensional intent recognition and dynamic knowledge fusion problem, including at least: Module 1 is used to acquire user questions, preprocess user input, and obtain rewritten questions; Extraction module 2 is used to obtain user intent and extract entities based on the rewritten question, and then link the entities to the knowledge base; Module 3 is used to extract the required data from the database based on user intent and entities, and then generate prompt words based on the extracted data, user intent, and knowledge base content linked to the entities. These prompt words are then input into the large model to drive it to analyze trends, analyze influencing factors, predict data, or generate visualization suggestions to form a report.

[0062] For the device embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to in the description of the method embodiments. The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this application according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0063] Accordingly, this application also provides an electronic device, including: One or more processors; Memory, used to store one or more programs; When the one or more programs are executed by the one or more processors, the one or more processors perform the methods described above.

[0064] Accordingly, this application also provides a computer-readable storage medium having computer instructions stored thereon, which, when executed by a processor, implement the steps of the method described above.

[0065] In the embodiments provided in this application, it should be understood that the disclosed methods and embodiments can also be implemented in other ways. The methods and embodiments described above are merely illustrative. For example, the flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of methods and computer program products according to various embodiments of this application. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions marked in the blocks may occur in a different order than those marked in the drawings. For example, two consecutive blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in a block diagram and / or flowchart, and combinations of blocks in block diagrams and / or flowcharts, can be implemented using dedicated hardware-based implementations that perform the specified functions or actions, or can be implemented using a combination of dedicated hardware and computer instructions.

[0066] In addition, the functional modules in the various embodiments of this application can be integrated together to form an independent part, or each module can exist independently, or two or more modules can be integrated to form an independent part.

[0067] On the other hand, a computer-readable storage medium stores computer instructions thereon, which, when executed by a processor, implement the steps of the above-described method. When the computer program is executed by the processor, it implements the method as described in any of the first aspects above. If the function is implemented as a software functional module and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as a USB flash drive, a portable hard drive, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk.

[0068] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A method for multi-dimensional intent recognition and dynamic knowledge fusion in question analysis, characterized in that, Includes the following steps: S1. Obtain the user's question, preprocess the user's input, and obtain the rewritten question; S2. Obtain user intent and extract entities based on the rewritten question, and then link the entities to the knowledge base; S3. Based on user intent and entities, extract the required data from the database, and then generate prompt words based on the extracted data, user intent, and knowledge base content linked to the entities. Input these prompt words into the large model to drive it to analyze trends, analyze influencing factors, predict data, or generate visualization suggestions to form a report.

2. The method according to claim 1, characterized in that, The preprocessing includes correcting typos in the user's question using an error correction model, and then rewriting the typo-corrected question using a rewriting model to obtain the rewritten question.

3. The method according to claim 1, characterized in that, The user intent includes trend analysis, influencing factor analysis, data prediction, and generation of visualization suggestions. Entities include organizational structure, time, and data indicator names.

4. The method according to claim 1, characterized in that, The error correction model is constructed through the following steps: A lightweight BERT model was fine-tuned using the Sighan 2015 Chinese spelling correction data and the First Enterprise dataset. The fine-tuning parameters included the learning rate, the number of training samples per batch, the gradient accumulation steps, the L2 regularization weights, the number of training rounds, and the gradient clipping threshold, to form the error correction model.

5. The method according to claim 1, characterized in that, The rewriting model is constructed through the following steps: Based on dialogue-utterance-rewriter, a rewriting model is formed by training on a second enterprise dataset to achieve coreference resolution and information completion.

6. The method according to claim 1, characterized in that, The process of obtaining user intent and extracting entities is accomplished using a large model.

7. A multi-dimensional intent recognition and dynamic knowledge fusion question, characterized in that, At least including: The acquisition module is used to acquire user questions, preprocess user input, and obtain rewritten questions; The extraction module is used to obtain user intent and extract entities based on the rewritten question, and then link the entities to the knowledge base; The generation module is used to extract the required data from the database based on user intent and entities. Then, based on the extracted data, user intent, and knowledge base content linked to the entities, it generates prompt words, which are then input into the large model to drive it to analyze trends, analyze influencing factors, predict data, or generate visualization suggestions to form a report.

8. An electronic device, characterized in that, include: One or more processors; Memory, used to store one or more programs; When the one or more programs are executed by the one or more processors, the one or more processors implement the method as described in any one of claims 1-6.

9. A computer-readable storage medium storing computer instructions thereon, characterized in that, When executed by the processor, this instruction implements the steps of the method as described in any one of claims 1-6.

Images