Deep learning-based technology consultation intention accurate identification and response method
By using a deep learning-based technical consulting system to achieve multimodal information fusion, dynamic knowledge adaptation, and personalized responses, the system solves the problems of inaccurate intent recognition, insufficient multimodal adaptation, and lagging knowledge updates in existing technologies, thereby improving the efficiency of technical consulting and user satisfaction.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-23
- Publication Date
- 2026-04-14
AI Technical Summary
Existing technical consultation systems suffer from low accuracy in intent recognition, insufficient multimodal adaptation, inadequate professionalism and relevance in responses, lagging dynamic knowledge updates, and poor continuity of interaction, resulting in low efficiency and poor user satisfaction.
We adopt a fully intelligent processing method based on deep learning, which includes multimodal consultation access, multi-dimensional intent deep recognition, dynamic knowledge graph retrieval and adaptation, personalized response generation and optimization, and multi-round interactive feedback and iteration. Through the synergistic linkage of deep learning models and structured knowledge systems, we achieve full-process intelligent technical consultation.
It improved the accuracy of intent recognition in technical consultations to 95%, and the user problem resolution rate to 88%, ensuring the timeliness and operability of responses, reducing redundant communication, and improving interaction efficiency and user satisfaction.
Smart Images
Figure CN121858745A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of deep learning and technical consulting, specifically a method for accurately identifying and responding to technical consulting intentions based on deep learning. Background Technology
[0002] The core requirement of technical consulting is "accurate understanding of the consulting intent + efficient output of professional responses," but current technical consulting services face many bottlenecks:
[0003] Low accuracy of intent recognition: Technical consultation language is characterized by high professionalism, dense terminology, and flexible expression (such as conversational questions, elliptical sentences, and professional abbreviations). Traditional methods rely on manually preset rules or simple keyword matching, which makes it difficult to understand vague expressions, ambiguous questions, and implicit intents, resulting in an intent recognition accuracy of less than 75%.
[0004] Lack of multimodal consultation adaptation: Existing methods mostly only support text consultation, which cannot handle multimodal inputs commonly found in technical consultations, such as "text + image (e.g., error screenshots, schematic diagrams)," "text + code snippets," and "text + voice." Incomplete information reception leads to response deviations.
[0005] The response lacks professionalism and specificity: Technical consultation needs to provide accurate solutions based on specific technical scenarios, version differences, and environmental constraints. Traditional response systems rely on fixed knowledge base matching, lack adaptation to consultation context and user technical level, and the output responses are mostly general templates, which are difficult to solve personalized technical problems.
[0006] Lagging dynamic knowledge updates: Technical knowledge iterates rapidly (such as programming language version updates, framework upgrades, and vulnerability fixes). The existing system knowledge base updates rely on manual maintenance, making it difficult to synchronize with the latest technical developments in real time, resulting in the failure of old knowledge responses.
[0007] Poor interaction continuity: Users often have needs for multiple rounds of follow-up questions and supplementary explanations. Traditional systems lack the ability to remember and associate the context of multiple rounds of dialogue, and background information needs to be repeated for each round of questions, resulting in low interaction efficiency.
[0008] Existing technologies have failed to achieve full-process intelligentization of "deep understanding of multimodal intent - dynamic knowledge adaptation - personalized professional response - multi-round interaction optimization", resulting in low efficiency and poor user satisfaction in technical consulting services. To address this, a method for accurate identification and response of technical consulting intent based on deep learning is proposed. Summary of the Invention
[0009] In view of this, the present invention provides a method for accurately identifying and responding to technical consultation intentions based on deep learning, so as to solve or alleviate the technical problems existing in the prior art, and at least provide a beneficial option.
[0010] The technical solution of this invention is implemented as follows: a method for accurate identification and response to technical consultation intent based on deep learning, comprising five core processes: multimodal consultation access and preprocessing, multi-dimensional deep identification of technical intent, dynamic knowledge graph retrieval and adaptation, personalized response generation and optimization, and multi-round interactive feedback and iteration. Each process is coordinated and linked through a deep learning model and a structured knowledge system to achieve intelligent processing of the entire technical consultation process. The specific process connection is as follows: multimodal consultation access and preprocessing outputs a structured consultation document, drives multi-dimensional deep identification of technical intent to generate accurate three-level intent, obtains personalized knowledge through dynamic knowledge graph retrieval and adaptation, outputs an adapted response through personalized response generation and optimization, and finally updates the model and knowledge graph through multi-round interactive feedback and iteration.
[0011] The multimodal consultation access and preprocessing process supports unified access of multimodal consultation carriers such as text, images, voice, and code files. It extracts the core information of each modality through technical dictionary word segmentation algorithm, improved YOLOv8+OCR, and syntax parser, and integrates them into a structured consultation document with "consultation subject, core question, auxiliary information, and environmental constraints".
[0012] The multi-dimensional technical intent deep recognition process constructs a three-level classification system of "first-level intent - second-level intent - third-level intent", uses "BERT-ConvBERT hybrid pre-trained model" for semantic encoding and coarse classification, mines implicit intent through GNN, combines attention mechanism to dissolve ambiguity, and outputs accurate intent after confidence verification;
[0013] The dynamic knowledge graph retrieval and adaptation process constructs a dynamic knowledge graph containing technical entities and relational edges. It automatically updates knowledge through real-time crawling, and uses a dual retrieval mechanism of "semantic similarity matching + knowledge association path mining" to obtain candidate knowledge. It also performs personalized adaptation based on the user's technical level and environmental constraints.
[0014] The personalized response generation and optimization process is organized according to the following structure: "Problem Analysis - Core Solution - Implementation Steps - Precautions - Alternative Solutions". It generates multi-style responses through the Transformer model and outputs them after technical accuracy and completeness verification and format optimization.
[0015] The multi-round interactive feedback and iteration process uses LSTM to build a context memory module, collects and classifies user feedback, and combines high-quality cases to incrementally train the core model and update the dynamic knowledge graph in sync.
[0016] In a further preferred embodiment, the three-level technical intent classification system includes first-level intents such as problem investigation, technology selection, operation guidance, knowledge Q&A, and optimization and upgrade. Each first-level intent is further subdivided into 5-8 second-level intents, and the second-level intents are further subdivided into specific third-level intents, forming a complete intent classification tree.
[0017] In a further preferred embodiment, the three-level technical intent classification system includes first-level intents such as problem investigation, technology selection, operation guidance, knowledge Q&A, and optimization and upgrade. Each first-level intent is further subdivided into 5-8 second-level intents, and the second-level intents are further subdivided into specific third-level intents, forming a complete intent classification tree.
[0018] Further preferably, the BERT-ConvBERT hybrid pre-trained model incorporates incremental training with 10 million+ technical field corpora on the basis of general pre-training, achieving a coarse classification accuracy of ≥92%; the intent confidence verification standard is ≥85%, and if it is lower than this value, a guided question is generated to obtain supplementary information before re-identification.
[0019] More preferably, the entity nodes of the dynamic knowledge graph include technical terms, problem types, solutions, version information, environmental parameters, and user roles, and the relationship edges include "problem-solution", "technology-applicable version", and "solution-dependency conditions", with the knowledge update cycle as short as 1 hour.
[0020] In a further preferred embodiment, in the dynamic knowledge graph retrieval and adaptation process, the semantic similarity threshold of the dual retrieval mechanism is set to ≥70%, and the user's technical level is divided into beginner, intermediate and advanced levels, respectively pushing knowledge content of corresponding depth, while filtering out knowledge that is not suitable for environmental constraints.
[0021] Furthermore, the personalized response generation and optimization process supports switching between three styles: professional and rigorous, concise and efficient, and easy to understand. Format optimization includes code syntax highlighting, step numbering, and bolding of key information to improve response readability.
[0022] In a further preferred embodiment, in the multi-round interactive feedback and iteration process, user feedback is divided into three categories: "solved problem", "partially solved", and "unsolved", which respectively trigger case storage, incremental model training, intent re-identification or knowledge supplementation operations to ensure continuous optimization of the method.
[0023] Further preferred embodiments of the method include a technical consultation intent recognition accuracy of ≥95%, a user problem resolution rate of ≥88%, a knowledge update cycle of at least 1 hour, and adaptability to multiple scenarios such as IT technical support, engineering technical consultation, and scientific research technical Q&A.
[0024] The embodiments of the present invention have the following advantages due to the adoption of the above technical solutions:
[0025] I. This invention achieves a technical consultation intent recognition accuracy of ≥95% through multimodal information fusion, pre-trained models in the technical field, and implicit intent mining, which is significantly better than traditional methods. It effectively avoids response deviations caused by misjudgment of intent, supports multimodal input such as text, images, voice, and code, and fully receives technical consultation information. It solves the problem of information loss caused by single text input and adapts to diverse expression scenarios of technical consultation.
[0026] Second, this invention provides real-time updated professional knowledge based on a dynamic knowledge graph, and is personalized to users' technical level and environmental constraints. The responses are highly targeted and operable, with a user problem resolution rate of ≥88%. It automatically captures and updates the latest knowledge in the technical field, solving the problem of knowledge lag in traditional systems and ensuring the timeliness and effectiveness of the responses. Through the context memory module, it associates multi-turn dialogue information and continuously optimizes based on user feedback, reducing the cost of repeated communication and improving interaction efficiency and user satisfaction.
[0027] The above overview is for illustrative purposes only and is not intended to be limiting in any way. In addition to the illustrative aspects, embodiments, and features described above, further aspects, embodiments, and features of the invention will become readily apparent from the accompanying drawings and the following detailed description. Attached Figure Description
[0028] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0029] Figure 1 This is an overall flowchart of the method of the present invention;
[0030] Figure 2 This is a flowchart of the multi-dimensional technical intent deep recognition sub-process of the present invention;
[0031] Figure 3 This is a flowchart of the personalized response generation and optimization sub-process of the present invention. Detailed Implementation
[0032] In the following description, only certain exemplary embodiments are briefly described. As those skilled in the art will recognize, the described embodiments can be modified in various ways without departing from the spirit or scope of the invention. Therefore, the drawings and description are considered to be exemplary in nature and not restrictive.
[0033] The embodiments of the present invention will now be described in detail with reference to the accompanying drawings.
[0034] like Figure 1-3 As shown, this invention provides a method for accurate identification and response to technical consultation intent based on deep learning. It includes five core processes: multimodal consultation access and preprocessing, multi-dimensional deep technical intent identification, dynamic knowledge graph retrieval and adaptation, personalized response generation and optimization, and multi-round interactive feedback and iteration. Each process works in synergy with a deep learning model and a structured knowledge system to achieve intelligent processing of the entire technical consultation process. Specifically, the process flow is as follows: multimodal consultation access and preprocessing outputs a structured consultation document, driving multi-dimensional deep technical intent identification to generate accurate three-level intents; personalized knowledge is obtained through dynamic knowledge graph retrieval and adaptation; adapted responses are output through personalized response generation and optimization; and finally, the model and knowledge graph are updated through multi-round interactive feedback and iteration.
[0035] The multimodal consultation access and preprocessing process supports unified access to multimodal consultation carriers such as text, images, voice, and code files. It extracts the core information of each modality through technical dictionary word segmentation algorithm, improved YOLOv8+OCR, and syntax parser, and integrates them into a structured consultation document with "consultation subject, core question, auxiliary information, and environmental constraints".
[0036] The multi-dimensional technical intent deep recognition process constructs a three-level classification system of "first-level intent - second-level intent - third-level intent", uses "BERT-ConvBERT hybrid pre-trained model" for semantic encoding and coarse classification, mines implicit intent through GNN, combines attention mechanism to dissolve ambiguity, and outputs accurate intent after confidence verification;
[0037] The dynamic knowledge graph retrieval and adaptation process constructs a dynamic knowledge graph containing technical entities and relational edges. It automatically updates knowledge through real-time crawling and uses a dual retrieval mechanism of "semantic similarity matching + knowledge association path mining" to obtain candidate knowledge. It also performs personalized adaptation based on the user's technical level and environmental constraints.
[0038] The personalized response generation and optimization process is organized according to the following steps: "Problem Analysis - Core Solution - Implementation Steps - Precautions - Alternative Solutions". It generates multi-style responses through the Transformer model and outputs them after technical accuracy and completeness verification and format optimization.
[0039] The multi-round interactive feedback and iteration process uses LSTM to build a context memory module, collects and classifies user feedback, and combines high-quality cases to incrementally train the core model and update the dynamic knowledge graph in sync.
[0040] In one embodiment, in the three-level technical intent classification system, the first-level intents include problem investigation, technology selection, operation guidance, knowledge Q&A, and optimization and upgrade. Each first-level intent is further subdivided into 5-8 second-level intents, and the second-level intents are further subdivided into specific third-level intents, forming a complete intent classification tree.
[0041] In one embodiment, in the three-level technical intent classification system, the first-level intents include problem investigation, technology selection, operation guidance, knowledge Q&A, and optimization and upgrade. Each first-level intent is further subdivided into 5-8 second-level intents, and the second-level intents are further subdivided into specific third-level intents, forming a complete intent classification tree.
[0042] In one embodiment, the BERT-ConvBERT hybrid pre-trained model incorporates incremental training with 10 million+ technical corpora on the basis of general pre-training, achieving a coarse classification accuracy of ≥92%; the intent confidence verification standard is ≥85%, and if it is lower than this value, a guided question is generated to obtain supplementary information before re-identification.
[0043] In one embodiment, the entity nodes of the dynamic knowledge graph include technical terms, problem types, solutions, version information, environmental parameters, and user roles, while the relationship edges include "problem-solution", "technology-applicable version", and "solution-dependencies". The knowledge update cycle can be as short as 1 hour.
[0044] In one embodiment, in the dynamic knowledge graph retrieval and adaptation process, the semantic similarity threshold of the dual retrieval mechanism is set to ≥70%, and the user's technical level is divided into beginner, intermediate and advanced, respectively, and knowledge content of corresponding depth is pushed, while knowledge that is not suitable for environmental constraints is filtered out.
[0045] In one embodiment, the personalized response generation and optimization process supports switching between three styles: professional and rigorous, concise and efficient, and easy to understand. Format optimization includes code syntax highlighting, step numbering, and bolding of key information to improve response readability.
[0046] In one embodiment, in a multi-round interactive feedback and iteration process, user feedback is divided into three categories: "solved problem", "partially solved", and "unsolved". These categories trigger case storage, incremental model training, intent re-identification, or knowledge supplementation operations, respectively, to ensure continuous optimization of the method.
[0047] In one embodiment, the method achieves a technical consultation intent recognition accuracy of ≥95%, a user problem resolution rate of ≥88%, and a knowledge update cycle of at least 1 hour, adapting to multiple scenarios such as IT technical support, engineering technical consultation, and scientific research technical Q&A.
[0048] In operation, this invention works as follows: Taking "Python code execution error consultation" as an example, the user inputs "Python code execution error, screenshot below + error screenshot + code snippet". After the system connects, it uses OCR to extract the error text "ModuleNotFoundError:Nomodulenamed'pandas'" from the screenshot, and uses a Python syntax parser to extract the "import pandas" statement from the code. This is then integrated into a structured consultation document: "Consultation subject: Python code error; Core problem: Failure to import pandas module; Auxiliary information: Error screenshot text, code snippet; Environmental constraints: System version and Python version not specified". The system loads a technical intent classification system, encodes the structured document using the BERT-ConvBERT model, and outputs the primary intent "problem investigation" and the secondary intent "code error". It also uses GNN to mine the implicit intent "missing dependency package", combined with "Pyt Keywords such as "python" and "pandas" are used to dispel ambiguity, outputting a level-three intent: "Python dependency package missing, causing module import failure." With a confidence level of 90% (≥85%), the process proceeds to the next step. The dynamic knowledge graph is searched for related paths related to "Python dependency package missing - solutions," filtering out candidate knowledge such as "pip installation," "conda installation," and "mirror source configuration." Considering the user's lack of clarity regarding their Python version, knowledge about "differences in installation commands for different Python versions" is added. The user is identified as a novice developer (based on the question's wording), and step-by-step operation guides are prioritized. The content is organized according to "problem analysis - core solution - implementation steps - precautions," generating an easy-to-understand response: "Problem analysis: The error is due to the pandas dependency package not being installed; Core solution: Install via pip command; Implementation steps: 1. Open the command line; 2. Enter pip install pandas (for Python 3, enter pip3 install)." (pandas); When a user reports "Installation successful, problem solved", the system stores the case in the high-quality case library; For subsequent similar inquiries, the model automatically associates the case to optimize response efficiency; The knowledge graph is synchronized in real time with the latest version of pandas installation notes to ensure timely responses.
[0049] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any person skilled in the art can easily conceive of various variations or substitutions within the technical scope disclosed in the present invention, and these should all be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.
Claims
1. A method for accurately identifying and responding to technical consultation intent based on deep learning, characterized by: It includes five core processes: multimodal consultation access and preprocessing, multi-dimensional technical intent deep recognition, dynamic knowledge graph retrieval and adaptation, personalized response generation and optimization, and multi-round interactive feedback and iteration. Each process works together through deep learning models and structured knowledge systems to achieve intelligent processing of the entire technical consultation process. The specific process connection is as follows: multimodal consultation access and preprocessing outputs structured consultation documents, drives multi-dimensional technical intent deep recognition to generate accurate three-level intents, obtains personalized knowledge through dynamic knowledge graph retrieval and adaptation, outputs adapted responses through personalized response generation and optimization, and finally updates the model and knowledge graph through multi-round interactive feedback and iteration. The multimodal consultation access and preprocessing process supports unified access to multimodal consultation carriers such as text, images, voice, and code files. It extracts the core information of each modality through technical dictionary word segmentation algorithm, improved YOLOv8+OCR, and syntax parser, and integrates them into a structured consultation document with "consultation subject, core question, auxiliary information, and environmental constraints". The multi-dimensional technical intent deep recognition process constructs a three-level classification system of "first-level intent - second-level intent - third-level intent", uses "BERT-ConvBERT hybrid pre-trained model" for semantic encoding and coarse classification, mines implicit intent through GNN, combines attention mechanism to dissolve ambiguity, and outputs accurate intent after confidence verification; The dynamic knowledge graph retrieval and adaptation process constructs a dynamic knowledge graph containing technical entities and relational edges. It automatically updates knowledge through real-time crawling, and uses a dual retrieval mechanism of "semantic similarity matching + knowledge association path mining" to obtain candidate knowledge. It also performs personalized adaptation based on the user's technical level and environmental constraints. The personalized response generation and optimization process is organized according to the following structure: "Problem Analysis - Core Solution - Implementation Steps - Precautions - Alternative Solutions". It generates multi-style responses through the Transformer model and outputs them after technical accuracy and completeness verification and format optimization. The multi-round interactive feedback and iteration process uses LSTM to build a context memory module, collects and classifies user feedback, and combines high-quality cases to incrementally train the core model and update the dynamic knowledge graph in sync.
2. The method for accurately identifying and responding to technical consultation intentions based on deep learning according to claim 1, characterized in that: In the three-level technical intent classification system, the first-level intents include problem investigation, technology selection, operation guidance, knowledge Q&A, and optimization and upgrade. Each first-level intent is further subdivided into 5-8 second-level intents, and the second-level intents are further subdivided into specific third-level intents, forming a complete intent classification tree.
3. The method for accurately identifying and responding to technical consultation intentions based on deep learning according to claim 1, characterized in that: In the three-level technical intent classification system, the first-level intents include problem investigation, technology selection, operation guidance, knowledge Q&A, and optimization and upgrade. Each first-level intent is further subdivided into 5-8 second-level intents, and the second-level intents are further subdivided into specific third-level intents, forming a complete intent classification tree.
4. The method for accurately identifying and responding to technical consultation intentions based on deep learning according to claim 1, characterized in that: The BERT-ConvBERT hybrid pre-trained model incorporates incremental training with over 10 million technical corpora on top of general pre-training, achieving a coarse classification accuracy of ≥92%; the intent confidence verification standard is ≥85%, and if it falls below this value, a guided question is generated to obtain supplementary information before re-identification.
5. The method for accurately identifying and responding to technical consultation intentions based on deep learning according to claim 1, characterized in that: The entity nodes of the dynamic knowledge graph include technical terms, problem types, solutions, version information, environmental parameters, and user roles. The relationship edges include "problem-solution", "technology-applicable version", and "solution-dependency conditions". The knowledge update cycle can be as short as 1 hour.
6. The method for accurately identifying and responding to technical consultation intentions based on deep learning according to claim 1, characterized in that: In the dynamic knowledge graph retrieval and adaptation process, the semantic similarity threshold of the dual retrieval mechanism is set to ≥70%. The user's technical level is divided into beginner, intermediate and advanced levels, and knowledge content of corresponding depth is pushed to them respectively, while knowledge that is not suitable for environmental constraints is filtered out.
7. The method for accurately identifying and responding to technical consultation intentions based on deep learning according to claim 1, characterized in that: The personalized response generation and optimization process supports switching between three styles: professional and rigorous, concise and efficient, and easy to understand. Format optimization includes code syntax highlighting, step numbering, and bolding of key information to improve response readability.
8. The method for accurately identifying and responding to technical consultation intentions based on deep learning according to claim 1, characterized in that: In the multi-round interactive feedback and iteration process, user feedback is divided into three categories: "solved problem", "partially solved", and "unsolved". These categories trigger case storage, incremental model training, intent re-identification, or knowledge supplementation operations, respectively, to ensure continuous optimization of the method.
9. The method for accurately identifying and responding to technical consultation intentions based on deep learning according to any one of claims 1-8, characterized in that: The method achieves a technical consultation intent recognition accuracy of ≥95%, a user problem resolution rate of ≥88%, and a knowledge update cycle of at least 1 hour, making it suitable for various scenarios such as IT technical support, engineering technical consultation, and scientific research technical Q&A.