Query request processing method and system based on scatter neural network

Abstract

The invention discloses a query request processing method and system based on a scatter neural network, and relates to the technical field of artificial intelligence and knowledge engineering. According to the method, an original knowledge source of a target domain is obtained, ternary array knowledge scatter points containing a core object, association logic and domain connotation are generated through structured analysis, and ternary array and dense vector double views are adopted for representation and storage; and responding to a query request, performing adaptive multi-path retrieval on the knowledge scatter points based on the semantic type and other characteristics of the query request, and generating a target answer after confidence evaluation and optimization. According to the method, the problems of low analysis efficiency of an original knowledge source in the financial field, poor semantic matching of a single retrieval strategy, insufficient understanding of general and special knowledge and low question and answer accuracy caused by large model illusion can be solved, and the accuracy and efficiency of vertical field query request processing and the dynamic knowledge strengthening capability are improved.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and knowledge engineering technology, and in particular to a query request processing method and system based on a scattered neural network. Background Technology

[0002] Driven by the need for refined management of expense discrepancies and comprehensive new business value in the financial and insurance industry, the digital transformation of business and financial data integration and knowledge management has become crucial for industry development. Intelligent question-answering systems, as the core carrier for efficient knowledge retrieval, are widely used in scenarios such as policy inquiries and decision support. The accuracy of their question-answering directly impacts the efficiency of expense discrepancy management and the rationality of resource allocation. However, knowledge in the financial and accounting field is characterized by high specialization, complex terminology, and significant differences in cognition across various scenarios. Traditional intelligent question-answering technologies struggle to meet the high-precision knowledge retrieval and semantic matching requirements of this field, thus limiting the practicality of knowledge services.

[0003] To address this challenge, existing technologies generally employ a whole-segment segmentation, single dense vector indexing, and single-recall mode based on the RAG framework. By simply segmenting knowledge sources such as institutional documents, a basic vector knowledge base is constructed. Then, a single strategy of sparse retrieval or dense retrieval is combined to obtain relevant knowledge fragments. Finally, question-and-answer results are generated based on a large model, thereby realizing online storage and intelligent retrieval of knowledge.

[0004] However, this existing technology has a core flaw: insufficient adaptability of knowledge representation and recall strategies, making it unable to accurately solve the problems of understanding and efficiently retrieving specialized knowledge in the financial and accounting vertical fields. Its simple block-based model makes it difficult to form refined knowledge units, and the single recall strategy cannot adapt to different types of query needs. Furthermore, it lacks dynamic response to user feedback and knowledge reinforcement mechanisms, resulting in low semantic matching accuracy for professional terms and vague expressions, insufficient recall accuracy, and ultimately leading to a high illusion rate in large models. This makes it difficult to meet the core requirements of accuracy and reliability for knowledge-based question answering in cost difference management scenarios. Summary of the Invention

[0005] This invention provides a query request processing method and system based on a scattered neural network, which can solve the problems of low efficiency in parsing original knowledge sources in the financial / accounting field, poor semantic matching of single retrieval strategies, insufficient understanding of general and proprietary knowledge, and low accuracy of question answering due to the illusion of large models. It improves the accuracy, efficiency and dynamic knowledge reinforcement capabilities of query request processing in vertical fields.

[0006] To achieve the above objectives, the embodiments of the present invention adopt the following technical solutions: Firstly, a query request processing method based on a scatter neural network is provided, comprising: acquiring an original knowledge source containing target domain knowledge; performing structured parsing on the original knowledge source to generate knowledge scatter points of the smallest semantic unit; the knowledge scatter points are triple arrays including core objects, association logic, and domain connotations, each triple array independently carrying a complete set of domain knowledge; encoding the knowledge scatter points using a dual-view representation method combining the triple arrays and dense vectors to form standardized knowledge data and storing it in a knowledge base; responding to a query request input by a user, adaptively selecting a multi-path retrieval strategy based on the semantic type, keyword attributes, and domain relevance features of the query request, and retrieving the knowledge scatter points corresponding to the query request from the knowledge base; performing confidence evaluation and optimization processing on the knowledge scatter points corresponding to the query request, and generating and displaying the target answer corresponding to the query request based on the optimized knowledge scatter points.

[0007] The method provided by this invention acquires the original knowledge source of the target domain and performs structured parsing to generate knowledge scatter points in the form of triple arrays, consisting of core objects, related logic, and domain connotations. Each knowledge scatter point independently carries complete domain knowledge, realizing the smallest semantic unit decomposition of domain knowledge and ensuring the fineness and completeness of knowledge granularity. At the same time, a dual-view representation method combining triple arrays and dense vectors is used to encode and store the knowledge scatter points, which not only retains the structured semantic information of the knowledge but also has the ability to express semantic features in a vectorized manner, providing dual data support for subsequent retrieval. By adaptively selecting a multi-path retrieval strategy based on the semantic type, keyword attributes, and domain relevance features of the query request, it can accurately match retrieval needs and improve retrieval targeting. After confidence evaluation and optimization processing, the target answer is generated, ensuring the accuracy and reliability of the answer. Overall, it realizes efficient processing of the entire process from knowledge parsing, encoding and storage to retrieval response, greatly improving the accuracy and efficiency of query request processing.

[0008] In one possible implementation of the first aspect, the original knowledge source includes at least one of PDF files, Word files, web page data, and database data, and the target domain includes the financial or accounting domain. The step of acquiring the original knowledge source containing target domain knowledge, performing structured parsing on the original knowledge source, and generating knowledge scatter points of the smallest semantic unit includes: performing hierarchical processing on the original knowledge source using a parallel pipeline parsing mechanism with a Map-Reduce architecture, wherein in the Map phase, basic data fragments are split according to preset units, and in the Reduce phase, the basic data fragments are syntactically segmented to obtain a set of sentences; a domain-fine-tuned entity recognition model is invoked to extract a triple array from the sentence set, including a core object, association logic, and domain connotation. The core object is a specific concept or object within the target domain, the association logic is the association type between core objects, and the domain connotation is the target domain knowledge meaning corresponding to the association logic. The core object, association logic, and domain connotation constitute the smallest indivisible semantic unit.

[0009] The method provided by this invention clearly defines the original knowledge sources as encompassing various types, including PDF files, Word files, web page data, and database data, and is adaptable to target domains such as finance or accounting, thus broadening the range of knowledge acquisition sources and meeting the knowledge processing needs of specific domains. By employing a parallel pipeline parsing mechanism based on a Map-Reduce architecture, the original knowledge sources are split into basic data fragments in the Map phase, and the basic data fragments are syntactically segmented into sentence sets in the Reduce phase, achieving efficient parallel processing of large-scale original knowledge sources and improving the efficiency of knowledge parsing. The domain-fine-tuned entity recognition model extracts a triple array composed of core objects, related logic, and domain connotations from the sentence set, and this triple array is the smallest indivisible semantic unit, ensuring that the extracted knowledge points not only conform to the characteristics of the target domain but also possess semantic integrity and independence, laying a high-quality data foundation for subsequent knowledge encoding and retrieval.

[0010] In one possible implementation of the first aspect, the method further includes: adding identification information to each of the knowledge scatter points, the identification information including a source identifier, a version number and a domain classification label, wherein the source identifier is used to record the origin of the original knowledge source, the version number is used to trace the generation and update records of the knowledge scatter points, and the domain classification label is used to distinguish the sub-domain category to which the knowledge scatter points belong.

[0011] The method provided by this invention adds source identifiers, version numbers, and domain classification tags to each knowledge point. The source identifier clearly records the origin of the original knowledge, facilitating the tracing of the original basis of the knowledge and improving the credibility of the knowledge. The version number accurately traces the generation and update records of the knowledge point, supports the full lifecycle management of knowledge, and facilitates subsequent updates, maintenance, and verification of knowledge. The domain classification tag effectively distinguishes the sub-domain category to which the knowledge point belongs, making the knowledge classification in the knowledge base clearer. During subsequent retrieval, knowledge points in sub-domains can be quickly located, improving retrieval efficiency and accuracy.

[0012] In one possible implementation of the first aspect, the method of encoding the knowledge scatter points using a dual-view representation combining the triple array and dense vectors to form standardized knowledge data and store it in a knowledge base includes: semantic vector encoding of the triple arrays corresponding to the knowledge scatter points to generate dense vectors that reflect the core semantic features of the triple arrays; structurally integrating the triple arrays, dense vectors, and identification information of the knowledge scatter points to form standardized knowledge data containing knowledge content, semantic features, and management attributes; and batch writing the standardized knowledge data into a vector database, wherein the vector database supports precise retrieval based on triple array keywords and semantic similarity retrieval based on dense vectors.

[0013] The method provided by this invention performs semantic vector encoding on triple arrays of knowledge points to generate dense vectors. These dense vectors accurately reflect the core semantic features of the triple arrays, providing crucial support for semantic similarity retrieval. The method also structurally integrates the triple arrays, dense vectors, and identifier information to form standardized knowledge data containing knowledge content, semantic features, and management attributes. This makes the knowledge data structure more unified and the information more complete, facilitating subsequent storage and retrieval. Furthermore, the standardized knowledge data is written into a vector database that supports both precise keyword retrieval based on triple arrays and semantic similarity retrieval based on dense vectors. This achieves compatibility between the two retrieval methods, satisfying both the need for precise keyword matching and the need for fuzzy retrieval based on semantic similarity, thus broadening the retrieval scenarios and improving the flexibility and comprehensiveness of the retrieval process.

[0014] In one possible implementation of the first aspect, the adaptive selection of a multi-path retrieval strategy based on the semantic type, keyword attributes, and domain relevance features of the query request, and the retrieval of knowledge points corresponding to the query request from the knowledge base, includes: performing feature extraction on the query request to obtain the semantic type, keyword attributes, and domain relevance features corresponding to the query request; The retrieval channels are adaptively matched based on the semantic type, keyword attributes, and domain relevance features. The retrieval channels include a sparse retrieval channel using the BM25 algorithm, a dense retrieval channel using a vector matching algorithm, and a knowledge graph-based relevance retrieval channel. The retrieval operations are executed in parallel through the matched retrieval channels to obtain preliminary retrieval results for each channel. Based on the fit between the query request and the retrieval channel, a dynamic weighting coefficient is determined, and the preliminary retrieval results are dynamically weighted and fused to obtain the knowledge scatter points corresponding to the query request.

[0015] The method provided by this invention extracts features from query requests to obtain semantic type, keyword attributes, and domain relevance features. Based on these features, it adaptively matches a combination of retrieval channels, including sparse, dense, and related retrieval channels, achieving precise matching between the retrieval channels and the query request. Parallel execution of retrieval operations through the matched channels improves retrieval efficiency. Dynamic weighting coefficients are determined based on the fit between the query request and the retrieval channels, and the preliminary retrieval results are dynamically weighted and fused. This integrates the retrieval advantages of each channel, effectively filters invalid retrieval results, and improves the accuracy and comprehensiveness of the final retrieval results.

[0016] In one possible implementation of the first aspect, determining a dynamic weighting coefficient based on the fit between the query request and the retrieval channel, and dynamically weighting and fusing the preliminary retrieval results to obtain the knowledge scatter points corresponding to the query request, includes: pre-constructing a fit score matrix between the query request and the retrieval channel, wherein the fit score matrix records the fit score between the query request and each retrieval channel, wherein the fit score between the query request and the sparse retrieval channel is a first value, the fit score between the query request and the dense retrieval channel is a second value, and the fit score between the query request and the associated retrieval channel is a third value, wherein the second value is less than or equal to the first value and greater than or equal to the third value; and retrieving the fit score matrix from the retrieval channel... Extract the corresponding matching scores for each retrieval channel, normalize the matching scores to obtain the dynamic weighting coefficients for each retrieval channel, and ensure that the sum of the weighting coefficients for all retrieval channels is 1. Obtain the preliminary retrieval results for each retrieval channel, assign the dynamic weighting coefficient of the retrieval channel to each knowledge point in the preliminary retrieval results, and calculate the weighted score for each knowledge point. The weighted score is the product of the semantic matching score of the knowledge point and the dynamic weighting coefficient. Sort all knowledge points in descending order of weighted score, and select a predetermined number of knowledge points with the highest weighted scores as the knowledge points corresponding to the query request.

[0017] The method provided by this invention pre-constructs a matching score matrix between query requests and retrieval channels, clarifying the matching score between query requests and each retrieval channel. By reasonably setting the score relationship, a scientific basis is provided for determining the dynamic weighting coefficient. The dynamic weighting coefficient is obtained by normalizing the matching score, ensuring the rationality of weighted fusion. A corresponding dynamic weighting coefficient is assigned to the preliminary retrieval results of each retrieval channel. By calculating the weighted score of knowledge points (the product of semantic matching score and the dynamic weighting coefficient), and sorting and selecting a preset number of knowledge points according to the score, the weight of retrieval results from retrieval channels with high matching scores can be highlighted, while also taking into account the effective information from other channels, further improving the accuracy of retrieval results and ensuring that the selected knowledge points have the optimal matching degree with the query request.

[0018] In one possible implementation of the first aspect, the step of performing confidence assessment and optimization processing on the knowledge scatter points corresponding to the query request, and generating and displaying the target answer corresponding to the query request based on the optimized knowledge scatter points, includes: calculating the overall confidence of the knowledge scatter points corresponding to the query request, wherein the confidence is based on a comprehensive judgment of the semantic matching degree between the knowledge scatter points and the query request, the historical call accuracy of the knowledge scatter points, and the version validity in the identification information; if the overall confidence is less than a preset confidence threshold, then calling a text expansion model to expand the query request with synonyms, generating at least two query variants, and performing multi-path retrieval again based on the query variants to obtain supplementary retrieval results; merging the preliminary retrieval results and the supplementary retrieval results to remove duplicates, sorting the deduplicated knowledge scatter points according to the objective function of minimizing evidence conflict, and filtering out the top N knowledge scatter points, where N is a positive integer; concatenating the top N knowledge scatter points with the query request to form a prompt text, inputting it into a large language model to generate the target answer, and synchronously displaying the target answer on the user interaction interface.

[0019] The method provided by this invention comprehensively determines the overall confidence level based on the semantic matching degree between knowledge points and query requests, historical call accuracy, and version validity, enabling a comprehensive assessment of the reliability of retrieved knowledge points. When the overall confidence level is lower than a preset confidence threshold, a query variant is generated through a text expansion model, and a multi-path retrieval is performed again, thus supplementing and improving the retrieval results and avoiding inaccurate answers due to insufficient initial retrieval information. After merging and deduplicating the preliminary and supplementary retrieval results, N knowledge points are sorted and selected according to the objective function of minimizing evidence conflict, ensuring that the selected knowledge points have minimal conflict and the highest consistency. These knowledge points are concatenated with the query request to form a prompt text, which is then input into a large language model to generate the target answer and display it. This method not only fully utilizes accurate knowledge points to ensure the accuracy of the answer but also leverages the natural language generation capabilities of the large language model to make the target answer more in line with the user's reading habits, thereby improving the user experience.

[0020] In one possible implementation of the first aspect, the method further includes: obtaining user feedback information on the target answer, the feedback information being generated through user interaction and including a binary evaluation label for the accuracy of the target answer; based on the binary evaluation label, locating and generating all knowledge scatter points on which the target answer depends, and calling a preset referee model to re-estimate the confidence of the knowledge scatter points, the preset referee model being used to perform secondary calibration of the confidence score based on the historical call frequency, matching accuracy, and domain relevance of the knowledge scatter points; if the binary evaluation label is a positive evaluation, increasing the confidence score of the corresponding knowledge scatter point and recording the number of valid calls; if it is a negative evaluation, decreasing the confidence score of the corresponding knowledge scatter point and marking it as pending review; based on the confidence re-estimated results of the knowledge scatter points, correcting the vector parameters of the low-confidence knowledge scatter points to obtain corrected knowledge scatter points, and structurally organizing the knowledge content corresponding to the high-confidence knowledge scatter points to form new knowledge scatter points; writing the corrected knowledge scatter points and the new knowledge scatter points into a knowledge base, and synchronously updating the version number and associated knowledge scatter point mapping relationship.

[0021] The method provided by this invention obtains user feedback information in the form of binary evaluation tags for target answers, enabling direct acquisition of user evaluations of the accuracy of the answers and providing a real user feedback basis for knowledge base optimization. Based on the binary evaluation tags, the method locates corresponding knowledge points and calls a preset referee model to re-estimate the confidence of these knowledge points based on historical call frequency, matching accuracy, and domain relevance, achieving dynamic calibration of the knowledge point confidence. For positive evaluations, the confidence of the corresponding knowledge points is increased and the number of valid calls is recorded; for negative evaluations, the confidence is decreased and the points are marked as pending review, accurately distinguishing between high-quality knowledge points and those requiring optimization. Based on the confidence re-estimation results, the method corrects the vector parameters of low-confidence knowledge points, organizes high-confidence knowledge points to form new knowledge points, and updates them to the knowledge base, achieving dynamic iterative optimization of the knowledge base. This continuously improves the quality of knowledge in the knowledge base, thereby ensuring the accuracy and reliability of subsequent query request processing.

[0022] Secondly, this invention provides a query request processing system based on a scatter neural network, comprising: a knowledge scatter generation module, used to acquire an original knowledge source containing target domain knowledge, perform structured parsing on the original knowledge source, and generate knowledge scatter points of the smallest semantic unit; the knowledge scatter points are triple arrays including core objects, association logic, and domain connotations, each triple array independently carrying a complete set of domain knowledge; a knowledge scatter writing module, used to encode the knowledge scatter points using a dual-view representation method combining the triple arrays and dense vectors, forming standardized knowledge data and storing it in a knowledge base; a knowledge scatter retrieval module, used to respond to a query request input by a user, adaptively select a multi-path retrieval strategy based on the semantic type, keyword attributes, and domain relevance features of the query request, and retrieve the knowledge scatter points corresponding to the query request from the knowledge base; and a target answer generation module, used to perform confidence evaluation and optimization processing on the knowledge scatter points corresponding to the query request, and generate and display the target answer corresponding to the query request based on the optimized knowledge scatter points.

[0023] Thirdly, an electronic device is provided, the electronic device including a memory and one or more processors; the memory is coupled to the processors; wherein the memory stores computer program code, the computer program code including computer instructions, which, when executed by the processor, cause the electronic device to perform the method as described in any implementation of the first aspect.

[0024] Fourthly, a computer-readable storage medium is provided, including computer instructions that, when executed on an electronic device, cause the electronic device to perform a method as described in any implementation of the first aspect.

[0025] Fifthly, a computer program product is provided that, when run on a computer, causes the computer to perform the method in any implementation of the first aspect.

[0026] Understandably, the beneficial effects achieved by the system of the second aspect, the electronic device of the third aspect, the computer-readable storage medium of the fourth aspect, and the computer program product of the fifth aspect provided above can be referred to with reference to the beneficial effects of the first aspect and any of its possible design embodiments, which will not be repeated here. Attached Figure Description

[0027] Figure 1 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention; Figure 2 A flowchart illustrating a query request processing method based on a scattered neural network, provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of the structure of a query request processing system provided in an embodiment of the present invention. Detailed Implementation

[0028] The technical solutions of the embodiments of the present invention will be described below with reference to the accompanying drawings. In the description of the present invention, unless otherwise stated, " / " indicates that the objects before and after are in an "or" relationship. For example, A / B can represent A or B. The "or" in the present invention is merely a description of the relationship between the related objects, indicating that three relationships can exist. For example, A or B can represent: A alone, A and B simultaneously, and B alone. A and B can be singular or plural. Furthermore, in the description of the present invention, unless otherwise stated, "multiple" refers to two or more. "At least one of the following" or similar expressions refer to any combination of these items, including any combination of single or plural items.

[0029] Furthermore, to facilitate a clear description of the technical solutions of the embodiments of the present invention, the terms "first" and "second" are used in the embodiments of the present invention to distinguish identical or similar items with essentially the same function and effect. Those skilled in the art will understand that the terms "first" and "second" do not limit the quantity or execution order, and that the terms "first" and "second" are not necessarily different.

[0030] In this embodiment of the invention, the terms "exemplary" or "for example" are used to indicate that something is an example, illustration, or description. Any embodiment or design described as "exemplary" or "for example" in this embodiment of the invention should not be construed as superior or more advantageous than other embodiments or designs. Specifically, the use of terms such as "exemplary" or "for example" is intended to present the relevant concepts in a concrete manner for ease of understanding.

[0031] Driven by the need for refined management of expense discrepancies and comprehensive new business value in the financial and insurance industry, the digital transformation of business and financial data integration and knowledge management has become crucial for industry development. Intelligent question-answering systems, as the core carrier for efficient knowledge retrieval, are widely used in scenarios such as policy inquiries and decision support. The accuracy of their question-answering directly impacts the efficiency of expense discrepancy management and the rationality of resource allocation. However, knowledge in the financial and accounting field is characterized by high specialization, complex terminology, and significant differences in cognition across various scenarios. Traditional intelligent question-answering technologies struggle to meet the high-precision knowledge retrieval and semantic matching requirements of this field, thus limiting the practicality of knowledge services.

[0032] To address this challenge, existing technologies generally employ a whole-segment segmentation, single dense vector indexing, and single-recall mode based on the RAG framework. By simply segmenting knowledge sources such as institutional documents, a basic vector knowledge base is constructed. Then, a single strategy of sparse retrieval or dense retrieval is combined to obtain relevant knowledge fragments. Finally, question-and-answer results are generated based on a large model, thereby realizing online storage and intelligent retrieval of knowledge.

[0033] However, this existing technology has a core flaw: insufficient adaptability of knowledge representation and recall strategies, making it unable to accurately solve the problems of understanding and efficiently retrieving specialized knowledge in the financial and accounting vertical fields. Its simple block-based model makes it difficult to form refined knowledge units, and the single recall strategy cannot adapt to different types of query needs. Furthermore, it lacks dynamic response to user feedback and knowledge reinforcement mechanisms, resulting in low semantic matching accuracy for professional terms and vague expressions, insufficient recall accuracy, and ultimately leading to a high illusion rate in large models. This makes it difficult to meet the core requirements of accuracy and reliability for knowledge-based question answering in cost difference management scenarios.

[0034] In view of this, embodiments of the present invention provide a query request processing method and system based on a scatter neural network. The method includes: acquiring an original knowledge source containing target domain knowledge; performing structured parsing on the original knowledge source to generate knowledge scatter points of the smallest semantic unit; the knowledge scatter points are triple arrays including core objects, association logic, and domain connotations, each triple array independently carrying a complete set of domain knowledge; encoding the knowledge scatter points using a dual-view representation method combining the triple arrays and dense vectors to form standardized knowledge data and storing it in a knowledge base; responding to a query request input by a user, adaptively selecting a multi-path retrieval strategy based on the semantic type, keyword attributes, and domain relevance features of the query request, and retrieving the knowledge scatter points corresponding to the query request from the knowledge base; performing confidence evaluation and optimization processing on the knowledge scatter points corresponding to the query request, and generating and displaying the target answer corresponding to the query request based on the optimized knowledge scatter points.

[0035] The method provided by this invention acquires the original knowledge source of the target domain and performs structured parsing to generate knowledge scatter points in the form of triple arrays, consisting of core objects, related logic, and domain connotations. Each knowledge scatter point independently carries complete domain knowledge, realizing the smallest semantic unit decomposition of domain knowledge and ensuring the fineness and completeness of knowledge granularity. At the same time, a dual-view representation method combining triple arrays and dense vectors is used to encode and store the knowledge scatter points, which not only retains the structured semantic information of the knowledge but also has the ability to express semantic features in a vectorized manner, providing dual data support for subsequent retrieval. By adaptively selecting a multi-path retrieval strategy based on the semantic type, keyword attributes, and domain relevance features of the query request, it can accurately match retrieval needs and improve retrieval targeting. After confidence evaluation and optimization processing, the target answer is generated, ensuring the accuracy and reliability of the answer. Overall, it realizes efficient processing of the entire process from knowledge parsing, encoding and storage to retrieval response, greatly improving the accuracy and efficiency of query request processing.

[0036] In some embodiments, the query request processing method based on a scatter neural network provided in this invention can be executed by a query request processing system 100 based on a scatter neural network (hereinafter referred to as query request processing system 100).

[0037] As an example, the query request processing system 100 can be any electronic device 200 with data processing capabilities, such as a general-purpose computer, personal computer, laptop computer, switch, or tablet computer. The specific implementation of the query request processing system 100 is not limited here.

[0038] Figure 1 A schematic diagram of the hardware structure of an electronic device provided in an embodiment of the present invention is shown. The electronic device 200 includes a processor 210, a memory 220, and a communication interface 230.

[0039] Processor 210 may include one or more processing cores. Processor 210 connects to various parts within electronic device 200 using various interfaces and lines, and performs various functions and processes data of electronic device 200 by running or executing instructions, programs, code sets, or instruction sets stored in memory 220, and by calling data stored in memory 220. Optionally, processor 210 may be implemented using at least one of the following hardware forms: Central Processing Unit (CPU), Graphics Processing Unit (GPU), Digital Signal Processing (DSP), Field-Programmable Gate Array (FPGA), and Programmable Logic Array (PLA).

[0040] The memory 220 may include random access memory (RAM) or read-only memory (ROM). Optionally, the memory 220 may include a non-transitory computer-readable storage medium. The memory 220 may be used to store instructions, programs, code, code sets, or instruction sets. The memory 220 may include a program storage area. This program storage area may store instructions for implementing an operating system, instructions for implementing at least one function, instructions for implementing the various method embodiments described above, etc.

[0041] Communication interface 230 is used to communicate with other devices, equipment or communication networks, such as data storage devices, image processing devices or Ethernet, wireless access network (RAN), wireless local area network (WLAN), etc.

[0042] In terms of physical implementation, the aforementioned devices (such as processor 210, memory 220, and communication interface 230) can each be devices within the same device (such as a laptop computer). Alternatively, at least two of these devices can be located within the same device, i.e., as different devices within the same device, similar to the deployment of devices or components in a distributed system.

[0043] It is understood that the structure illustrated in this embodiment does not constitute a specific limitation on the electronic device 200. In other embodiments of the present invention, the electronic device 200 may include more or fewer components than illustrated, or combine some components, or split some components, or have different component arrangements. The illustrated components may be implemented in hardware, software, or a combination of software and hardware.

[0044] The following description, in conjunction with the accompanying drawings, illustrates a query request processing method based on a scattered neural network provided by an embodiment of the present invention.

[0045] Figure 2 This is a flowchart illustrating a query request processing method based on a scattered neural network, provided as an embodiment of the present invention. Optionally, this method can be... Figure 1 The illustrated electronic device 200 performs this method, which includes the following steps: S1. Obtain the original knowledge source containing knowledge of the target domain, perform structured parsing on the original knowledge source, and generate knowledge scatter points of the smallest semantic unit.

[0046] Specifically, the knowledge scatter plots are triple arrays comprising core objects, relational logic, and domain connotations, with each triple array independently carrying a complete set of domain knowledge; the original knowledge sources include at least one of PDF files, Word files, web page data, and database data, and the target domain includes the financial or accounting field.

[0047] Furthermore, the aforementioned S1 specifically includes: The original knowledge source is processed hierarchically using a parallel pipeline parsing mechanism based on the Map-Reduce architecture. In the Map phase, basic data fragments are obtained by splitting the data into preset units. In the Reduce phase, the basic data fragments are syntactically segmented to obtain a set of sentences. A domain-fine-tuned entity recognition model is invoked to extract a triple array from the sentence set, which includes a core object, association logic, and domain connotation. The core object is a specific concept or object within the target domain, the association logic is the association type between core objects, and the domain connotation is the target domain knowledge meaning corresponding to the association logic. The core object, association logic, and domain connotation constitute the smallest indivisible semantic unit.

[0048] It should be noted that the preset unit here can be flexibly set according to the type and scale of the original knowledge source. For example, for PDF and Word files, the preset unit can be a page, meaning the file is split into basic data fragments in parallel by page during the Map stage; for web page data, the preset unit can be a web page paragraph; and for database data, the preset unit can be a group of data records. The domain-fine-tuned entity recognition model is based on a general entity recognition model, fine-tuned using specialized corpora in the financial or accounting field. It can accurately identify proprietary concepts and objects within that field. For example, in the financial field, core objects can be wealth management products, loan interest rates, fund shares, etc.; in the accounting field, core objects can be accounting subjects, financial statements, cost accounting items, etc. The association logic is the logical relationship connecting two or more core objects, such as inclusion, interest calculation basis, attribution, accounting standards, etc., while the domain connotation is the explanation of the specific meaning of this association logic within the target domain. For example, if the core objects are time deposits and annual interest rates, and the association logic is the interest calculation basis, the corresponding domain connotation is that time deposits calculate interest according to the agreed annual interest rate, and the interest calculation period is usually consistent with the deposit term. This ternary array-based knowledge scattering method can break down domain knowledge into the smallest semantic units, each of which fully carries a set of independent knowledge logic, avoiding the problems of knowledge redundancy and semantic fragmentation in the traditional whole-segmentation model.

[0049] The method provided by this invention breaks through the limitations of a single knowledge source by clearly defining the diverse types of the original knowledge source and its adaptability to specific target domains such as finance and accounting. It can integrate domain knowledge in different formats and carriers to meet the diverse knowledge processing needs of vertical domains. By leveraging the parallel pipeline parsing mechanism of the Map-Reduce architecture, the processing of the original knowledge source is split into parallel splitting in the Map stage and syntactic segmentation in the Reduce stage, effectively avoiding the efficiency bottleneck caused by serial processing and significantly improving the parsing speed and processing throughput of large-scale knowledge sources. By extracting triple arrays that fit the characteristics of the domain through a domain-fine-tuned entity recognition model and clearly defining them as the smallest indivisible semantic unit, the accuracy, completeness, and independence of knowledge points are ensured. This provides high-quality basic data support for subsequent steps such as dual-view representation and adaptive retrieval, improving the reliability and accuracy of the entire query request processing flow from the source.

[0050] In another possible implementation, the method further includes: adding identification information to each of the knowledge points, the identification information including a source identifier, a version number and a domain classification label, the source identifier being used to record the origin of the original knowledge source, the version number being used to trace the generation and update records of the knowledge points, and the domain classification label being used to distinguish the sub-domain category to which the knowledge points belong.

[0051] It should be understood that the specific form of the source identifier can be flexibly set according to the type of the original knowledge source. For example, for PDF and Word files, the source identifier can be a combination of the file's storage path and name; for web page data, it can be the URL address of the web page; and for database data, it can be a combination of the database name, table name, and unique identifier of the data record. These specific source identifiers allow for quick location of the original source of the knowledge scatterpoint, providing clear evidence for knowledge tracing and verification. The version number uses a combination of generation timestamp and update count. The generation timestamp is accurate to the second and records the specific time when the knowledge scatterpoint was first generated. The update count starts from 0 and increments by 1 each time parameters or content are adjusted for a knowledge scatterpoint. The version number allows for complete traceability of the entire process of a knowledge scatterpoint from generation to each update. Domain classification tags are set based on sub-fields within the financial and accounting domains. For example, in the financial domain, sub-tags can include wealth management products, credit business, securities trading, and insurance terms; in the accounting domain, sub-tags can include accounting, financial statements, cost management, and tax declaration. Each knowledge scatterpoint is assigned one or more sub-domain classification tags according to its core knowledge content, ensuring a high degree of matching between the tags and the knowledge content.

[0052] The method provided by this invention configures multi-dimensional identification information for each knowledge point. The source identifier provides traceable original evidence for the knowledge, allowing users to clearly understand the source background of the knowledge and enhancing the credibility and authority in the knowledge application process. The version number enables refined management of the entire lifecycle of the knowledge points. Whether it is subsequent knowledge verification, update maintenance, or problem troubleshooting, the version number can be used to quickly locate the corresponding knowledge status, improving the efficiency of knowledge management. The domain classification tags accurately subdivide and classify the knowledge in the knowledge base, making the originally complex knowledge system clear and organized. During subsequent retrieval, knowledge points in the target sub-domain can be quickly filtered based on tags, reducing the interference of irrelevant knowledge, thereby improving the efficiency and accuracy of the retrieval process and providing strong support for the efficient advancement of the entire query request processing flow.

[0053] S2. Using the dual-view representation method that combines the ternary array and dense vector, the knowledge scatter points are encoded to form standardized knowledge data and stored in the knowledge base.

[0054] In some embodiments, S2 includes: Semantic vector encoding is performed on the triple arrays corresponding to the knowledge scatter points to generate dense vectors that reflect the core semantic features of the triple arrays; the triple arrays, dense vectors, and the identification information of the knowledge scatter points are structurally integrated to form standardized knowledge data containing knowledge content, semantic features, and management attributes; the standardized knowledge data is written into a vector database in batches, and the vector database supports precise retrieval based on triple array keywords and semantic similarity retrieval based on dense vectors.

[0055] Specifically, semantic vector encoding is implemented using a vector encoding model fine-tuned based on domain corpus. After being trained with professional knowledge data in the financial or accounting field, this model can accurately capture the semantic relationships between the core objects, related logic, and domain connotations in the triple array. The generated dense vector is a numerical vector with fixed dimensions, such as 768 or 1024 dimensions. The value of each dimension corresponds to a specific semantic feature dimension of the triple array, ensuring that the dense vector can comprehensively and accurately reflect the core semantics of the triple array.

[0056] The structured integration process is executed according to a preset data format template. The template clearly defines the storage fields and data types of the three-element arrays, dense vectors, and identification information. For example, the three-element array stores the text content of the core object, the association logic, and the domain connotation in the form of a string, the dense vector stores the values ​​of each dimension in the form of an array, and the identification information is stored in the order of source identifier, version number, and domain classification label. Through this standardized integration method, the standardized knowledge data structure corresponding to different knowledge points is completely unified.

[0057] Meanwhile, batch writing adopts a batch packaging and timed submission method. The batch size and submission interval are set according to the generation speed of knowledge points. For example, a batch is formed every 1,000 standardized knowledge data points, or the generated standardized knowledge data is submitted every 30 seconds. The vector database specifically uses the Elasticsearch vector database. This database uses a dedicated index structure to index the keywords and dense vectors of the triple arrays respectively, ensuring that both retrieval methods can obtain high-efficiency response speed.

[0058] The method provided by this invention generates dense vectors through domain-adaptive semantic vector encoding, enabling the semantic features of scattered knowledge points to be presented in a machine-recognizable numerical form, providing a solid foundation for subsequent semantic-based retrieval operations. Through structured integration, the core content, semantic features, and management information of scattered knowledge points are integrated, eliminating format differences between knowledge data from different sources, making the data in the knowledge base more organized, and reducing the complexity of data management and retrieval. Utilizing a vector database that supports dual-mode retrieval, it achieves dual retrieval capabilities of precise keyword matching and semantic similarity matching. This allows for the rapid location of precise knowledge containing specific keywords and the discovery of related knowledge semantically similar to the query request, fully adapting to different types of query scenarios in the financial or accounting fields, making the retrieval process more flexible and efficient, and further improving the overall effect of query request processing.

[0059] S3. In response to the user's input query request, based on the semantic type, keyword attributes and domain relevance characteristics of the query request, adaptively select a multi-path retrieval strategy, and retrieve the knowledge points corresponding to the query request from the knowledge base.

[0060] In one possible implementation, S3 specifically includes the following steps: Feature extraction is performed on the query request to obtain the semantic type, keyword attributes, and domain relevance features corresponding to the query request. Based on the semantic type, keyword attributes, and domain relevance features, adaptive matching is performed to combine retrieval channels. These retrieval channels include a sparse retrieval channel using the BM25 algorithm, a dense retrieval channel using a vector matching algorithm, and a knowledge graph-based association retrieval channel. Retrieval operations are executed in parallel through the matched retrieval channels to obtain preliminary retrieval results for each channel. Dynamic weighting coefficients are determined based on the fit between the query request and the retrieval channels, and the preliminary retrieval results are dynamically weighted and fused to obtain the knowledge scatter points corresponding to the query request.

[0061] The feature extraction process is achieved by integrating rule matching and semantic analysis models. The semantic type features mainly determine whether the query request is a declarative query, an interrogative query, or an imperative query. For example, "What is the interest rate standard for corporate loans?" is an interrogative query, while "Explain the accounting process for fixed assets" is an imperative query. Keyword attribute features focus on whether the query contains domain-specific terms, vague terms, or rule-based terms. For example, in the financial field, "interbank lending" and "risk reserve" are domain-specific terms, and in the accounting field, "accrual basis" and "bad debt provision" are all domain-specific terms. Domain relevance features are derived by calculating the similarity between query keywords and corpora in the financial and accounting fields. If the similarity is higher than a preset threshold, it is considered to be highly relevant; otherwise, it is considered to be less relevant.

[0062] Each of the three retrieval channels possesses specific advantages. The sparse retrieval channel, employing the BM25 algorithm, excels at precise keyword-based matching, quickly locating knowledge points containing the target keywords. The dense retrieval channel, using a vector matching algorithm, prioritizes semantic similarity, retrieving relevant knowledge points even if the query keywords differ but are semantically similar. The knowledge graph-based association retrieval channel constructs a domain knowledge graph, mining the relationships between keywords to achieve expanded retrieval of related knowledge. Parallel execution of retrieval operations is achieved through a multi-threaded scheduling mechanism. Each retrieval channel is allocated independent thread resources, simultaneously searching the knowledge base to avoid efficiency losses caused by serial execution, ensuring that preliminary retrieval results from multiple channels are obtained in a short time.

[0063] It should be noted that the knowledge graph construction method of the association retrieval channel using the knowledge graph association algorithm includes: taking the core objects in the knowledge scatter points of the triple array as entities of the knowledge graph, the association logic as the relationship between entities, and the domain connotation as the relationship attribute to construct a domain knowledge graph.

[0064] The method provided by this invention extracts features from query requests to obtain semantic type, keyword attributes, and domain relevance features. Based on these features, it adaptively matches a combination of retrieval channels, including sparse, dense, and related retrieval channels, achieving precise matching between the retrieval channels and the query request. Parallel execution of retrieval operations through the matched channels improves retrieval efficiency. Dynamic weighting coefficients are determined based on the fit between the query request and the retrieval channels, and the preliminary retrieval results are dynamically weighted and fused. This integrates the retrieval advantages of each channel, effectively filters invalid retrieval results, and improves the accuracy and comprehensiveness of the final retrieval results.

[0065] Furthermore, the step of determining a dynamic weighting coefficient based on the fit between the query request and the retrieval channel, and dynamically weighting and fusing the preliminary retrieval results to obtain the knowledge scatter points corresponding to the query request, includes: pre-constructing a fit score matrix between the query request and the retrieval channel, wherein the fit score matrix records the fit score between the query request and each retrieval channel, wherein the fit score between the query request and the sparse retrieval channel is a first value, the fit score between the query request and the dense retrieval channel is a second value, and the fit score between the query request and the associated retrieval channel is a third value, wherein the second value is less than or equal to the first value and greater than or equal to the third value; extracting the corresponding knowledge scatter points from the fit score matrix. Each retrieval channel's adaptation score is normalized to obtain a dynamic weighting coefficient for each retrieval channel. The sum of the weighting coefficients for all retrieval channels is 1. Preliminary retrieval results for each retrieval channel are obtained, and the dynamic weighting coefficient of the retrieval channel is assigned to each knowledge point in the preliminary retrieval results. A weighted score for each knowledge point is calculated. The weighted score is the product of the semantic matching score and the dynamic weighting coefficient. All knowledge points are sorted from highest to lowest weighted score, and a predetermined number of knowledge points with the highest weighted scores are selected as the knowledge points corresponding to the query request.

[0066] Specifically, the fit score matrix is ​​constructed based on a large amount of historical query data and retrieval performance feedback in the financial and accounting fields. By statistically analyzing the recall accuracy, semantic matching degree and other indicators of each retrieval channel under different query requests, the specific range of fit scores is determined. For example, the first value ranges from 0.7 to 0.9, the second value from 0.5 to 0.7, and the third value from 0.3 to 0.5.

[0067] The normalization process uses a linear normalization method, which divides the fit score of each retrieval channel by the sum of the fit scores of all channels to obtain the corresponding dynamic weighting coefficients. For example, if the fit scores of the three retrieval channels are 0.8, 0.6, and 0.4, respectively, and the sum is 1.8, the corresponding dynamic weighting coefficients after normalization are 0.44, 0.33, and 0.23, respectively, ensuring that the sum of the three is 1.

[0068] Furthermore, the semantic matching score of the knowledge points is calculated by each retrieval channel based on its own algorithm logic. For example, the sparse retrieval channel gives a score between 0 and 1 based on the degree of keyword matching, the dense retrieval channel gives a score between 0 and 1 based on vector similarity, and the related retrieval channel gives a score between 0 and 1 based on the tightness of knowledge graph association. The preset number is set according to the needs of the actual application scenario. For example, in the scenario of querying financial products, the preset number can be set to 50, and in the scenario of querying financial regulations and clauses, the preset number can be set to 30, ensuring that the selected knowledge points can cover the core information without reducing the efficiency of subsequent processing due to an excessive number.

[0069] The method provided by this invention pre-constructs a matching score matrix between query requests and retrieval channels, clarifying the matching score between query requests and each retrieval channel. By reasonably setting the score relationship, a scientific basis is provided for determining the dynamic weighting coefficient. The dynamic weighting coefficient is obtained by normalizing the matching score, ensuring the rationality of weighted fusion. A corresponding dynamic weighting coefficient is assigned to the preliminary retrieval results of each retrieval channel. By calculating the weighted score of knowledge points (the product of semantic matching score and the dynamic weighting coefficient), and sorting and selecting a preset number of knowledge points according to the score, the weight of retrieval results from retrieval channels with high matching scores can be highlighted, while also taking into account the effective information from other channels, further improving the accuracy of retrieval results and ensuring that the selected knowledge points have the optimal matching degree with the query request.

[0070] S4. Perform confidence assessment and optimization on the knowledge scatter points corresponding to the query request, and generate and display the target answer corresponding to the query request based on the optimized knowledge scatter points.

[0071] In one possible implementation, the above S4 includes: The overall confidence score of the knowledge scatter points corresponding to the query request is calculated. The confidence score is based on the semantic matching degree between the knowledge scatter points and the query request, the historical call accuracy of the knowledge scatter points, and the version validity in the identification information. If the overall confidence score is less than a preset confidence threshold, the text expansion model is invoked to expand the query request with synonyms, generating at least two query variants. Based on the query variants, a multi-path retrieval is performed again to obtain supplementary retrieval results. The preliminary retrieval results and the supplementary retrieval results are merged and deduplicated. The deduplicated knowledge scatter points are sorted according to the objective function of minimizing evidence conflict, and the top N knowledge scatter points are selected and output, where N is a positive integer. The top N knowledge scatter points are concatenated with the query request to form a prompt text, which is input into a large language model to generate a target answer, and the target answer is synchronously displayed on the user interface.

[0072] Specifically, the overall confidence score is calculated using a weighted summation method, where the semantic matching degree between knowledge scatter points and query requests accounts for 40%, the historical call accuracy of knowledge scatter points accounts for 35%, and the version validity in the identification information accounts for 25%. After all indicators are normalized to the 0-1 range, the weighted sum is calculated to obtain the overall confidence score.

[0073] Semantic matching score is calculated by the cosine similarity between the dense vector of the knowledge scatter point and the encoded vector of the query request; historical call accuracy is the ratio of the number of times the knowledge scatter point has generated a correct answer in the past to the total number of calls; version validity is determined based on the version number of the knowledge scatter point. If the generation or update time corresponding to the version number is within the preset validity period (e.g., 6 months), it is counted as 1. Otherwise, the score is reduced proportionally according to the time exceeding the validity period.

[0074] The pre-set reliability threshold is 0.3, a value derived from statistical analysis of a large amount of historical query data in the financial and accounting fields, effectively balancing retrieval efficiency and result accuracy. The text expansion model employs a large language model fine-tuned with a domain-specific corpus. Synonym expansion is performed based on a specialized thesaurus for the financial and accounting field. For example, a query request of "What is the interest rate for corporate loans?" can generate query variants such as "What is the interest rate for corporate loans?" or "What is the standard annualized interest rate for company loans?", ensuring that variants retain the core semantics while enriching the retrieval dimensions.

[0075] The merging and deduplication operation compares the contents of the triplet array of knowledge points with the source identifier and version number in the identification information to remove knowledge points that are completely identical or have duplicate core semantics. The evidence conflict minimization objective function calculates the semantic similarity and core information overlap between any two knowledge points. If the degree of semantic conflict between the two is higher than a preset conflict threshold (e.g., 0.6), the knowledge point with higher confidence is retained, ultimately ensuring that the sorted knowledge points are semantically consistent and complementary.

[0076] The value of N is set according to the actual application scenario, with a default value of 10. This ensures the accuracy of the answer while avoiding a decrease in the generation efficiency of the large model due to an excessive number of knowledge points. The concatenated prompt text contains the core requirements of the query request and the key information of the knowledge points. The format is "Answer the question based on the following knowledge: [Knowledge Point Content] Question: [Query Request]". When the generated target answer is displayed synchronously on the user interface, it will also include a summary of the source identifier of the knowledge points, making it convenient for users to trace the information.

[0077] The method provided by this invention comprehensively determines the overall confidence level through multi-dimensional indicators, breaking through the limitations of single-indicator evaluation. It can more comprehensively and objectively reflect the reliability of retrieved knowledge points, providing a scientific basis for subsequent optimization operations. The synonym expansion and secondary retrieval mechanism triggered in low-confidence scenarios effectively makes up for the information gaps that may exist in the initial retrieval, reducing the omission of effective information due to differences in keyword expression or limitations in retrieval strategies. The merged deduplication and evidence conflict minimization sorting process can eliminate redundant information and resolve knowledge conflicts, ensuring that the selected N knowledge points have optimal consistency and completeness. By using a large language model to fuse high-quality knowledge points with query requests, it not only fully leverages the supporting role of domain knowledge points in the accuracy of answers and effectively reduces the risk of large model illusion, but also produces answer content that conforms to users' reading habits through natural language generation capabilities, while attaching information traceability evidence. On the basis of improving the accuracy of answers, it further enhances users' trust in the answers and the user experience.

[0078] As described in S1-S4 above, the method provided by this embodiment of the invention obtains the original knowledge source of the target domain and performs structured parsing to generate knowledge scatter points in the form of triple arrays consisting of core objects, related logic, and domain connotations. Each knowledge scatter point independently carries complete domain knowledge, realizing the smallest semantic unit decomposition of domain knowledge and ensuring the fineness and completeness of knowledge granularity. At the same time, a dual-view representation method combining triple arrays and dense vectors is used to encode and store the knowledge scatter points, which not only retains the structured semantic information of the knowledge but also has the ability to express semantic features in a vectorized manner, providing dual data support for subsequent retrieval. By adaptively selecting a multi-path retrieval strategy based on the semantic type, keyword attributes, and domain relevance features of the query request, the retrieval requirements can be accurately matched, improving the retrieval targeting. After confidence evaluation and optimization processing, the target answer is generated, ensuring the accuracy and reliability of the answer. Overall, it realizes efficient processing of the entire process from knowledge parsing, encoding and storage to retrieval response, greatly improving the accuracy and efficiency of query request processing.

[0079] In one possible implementation, the method further includes: obtaining user feedback information on the target answer, the feedback information being generated through user interaction and including a binary evaluation label for the accuracy of the target answer; based on the binary evaluation label, locating and generating all knowledge scatter points on which the target answer depends, and calling a preset referee model to re-estimate the confidence level of the knowledge scatter points, the preset referee model being used to perform secondary calibration of the confidence level score based on the historical call frequency, matching accuracy, and domain relevance of the knowledge scatter points; if the binary evaluation label is a positive evaluation, increasing the confidence level score of the corresponding knowledge scatter point and recording the number of valid calls; if it is a negative evaluation, decreasing the confidence level score of the corresponding knowledge scatter point and marking it as pending review; based on the confidence level re-estimated results of the knowledge scatter points, correcting the vector parameters of the low-confidence knowledge scatter points to obtain corrected knowledge scatter points, and structurally organizing the knowledge content corresponding to the high-confidence knowledge scatter points to form new knowledge scatter points; writing the corrected knowledge scatter points and the new knowledge scatter points into a knowledge base, and synchronously updating the version number and associated knowledge scatter point mapping relationships.

[0080] Specifically, the user interaction operation includes two feedback buttons, "Accurate" and "Inaccurate", set in the user interaction interface. When the user clicks the corresponding button, a binary evaluation label is generated, and the system will automatically record the correlation between the feedback label and the corresponding target answer.

[0081] When locating the knowledge scatter points on which the target answer depends, the system log is queried to find the call association records between the target answer and the knowledge scatter points, so as to accurately locate all the knowledge scatter points involved in the generation of the answer and ensure that no object is missed in the confidence re-estimation.

[0082] The pre-defined referee model is trained based on the knowledge characteristics and historical interaction data in the financial field. Historical call frequency accounts for 30% of the weight, matching accuracy accounts for 40%, and domain relevance accounts for 30%. The confidence score of the knowledge points is re-estimated through weighted calculation. The adjustment range of the confidence score follows fixed rules: a positive evaluation increases the confidence score by 0.1-0.2, while a negative evaluation decreases it by 0.2-0.3. Only the call scenarios corresponding to positive evaluations are recorded for valid calls, ensuring that the statistical results truly reflect the practical value of the knowledge points.

[0083] The criterion for low-confidence knowledge scatter points is a confidence score below 0.5. Their vector parameters are adjusted using a domain-adaptive fine-tuning algorithm, optimizing parameters based on the vector characteristics of similar high-confidence knowledge scatter points. The criterion for high-confidence knowledge scatter points is a confidence score above 0.8. When structuring their knowledge content, new knowledge relationships are extracted to form new knowledge scatter points according to a triple array format of core object, related logic, and domain connotation. When updating the version number, the update count is incremented by 1. Updating the mapping relationships of related knowledge scatter points adjusts the association links between new knowledge scatter points and existing related knowledge scatter points, ensuring the coherence of the knowledge system in the knowledge base.

[0084] The method provided by this invention collects user feedback through simple and intuitive binary evaluation labels, lowering the operational threshold for user feedback and enabling the rapid accumulation of a large amount of real and effective optimization evidence, providing a fresh user demand orientation for the iteration of the knowledge base. By using a pre-set referee model to integrate multi-dimensional indicators for confidence re-estimation of knowledge scatter points, it avoids the one-sidedness of single-dimensional evaluation, making the confidence calibration results more scientific and reliable. By adopting differentiated processing strategies for knowledge scatter points corresponding to different evaluation labels, it can accurately identify high-quality knowledge resources and areas for improvement, achieving refined management of the knowledge base. Based on the confidence re-estimation results, it performs parameter correction and addition work on knowledge scatter points, and simultaneously updates the version and relationships of the knowledge base, constructing a closed-loop mechanism of feedback-evaluation-optimization-update. This allows the knowledge base to continuously absorb high-quality knowledge and correct inferior information, constantly improving knowledge quality and adaptability, thereby providing more accurate and reliable knowledge support for subsequent query request processing and driving continuous performance optimization of the entire query request processing system.

[0085] To facilitate understanding of this solution, the following detailed explanation of the method provided in this embodiment of the invention will be provided with reference to a specific example.

[0086] In one example, query request processing system 100 processes a user's query request for "What is the standard annual interest rate for corporate working capital loans?", and the specific process is as follows: The query request processing system 100 first acquires raw knowledge sources containing financial domain knowledge, including loan business management regulations in PDF format, interest rate pricing guidelines in Word format, and loan policy data in the database. Employing a parallel pipeline parsing mechanism with a Map-Reduce architecture, the Map phase splits the raw knowledge sources into basic data fragments by "page" or "record group." The Reduce phase performs syntactic segmentation to obtain a set of 156 sentences. A financial domain-tuned entity recognition model is then called to extract 89 triplet knowledge points, such as "Core object: corporate working capital loans, annual interest rate; Association logic: pricing basis; Domain connotation: determining the spread based on market interest rates combined with credit rating." Each triplet is the smallest indivisible semantic unit. The query request processing system 100 adds identification information to each knowledge point. The source identifier records the original file and page number or database information, the version number uses the "generation timestamp-update count" format, and the domain classification tags are assigned according to subdivisions such as "corporate loans - interest rate pricing." Subsequently, the triple array is semantically vectorized to generate a 768-dimensional dense vector. The triple array, dense vector, and identification information are then structurally integrated into standardized knowledge data and batch-written into the Elasticsearch vector database, which supports keyword-based precise retrieval and semantic similarity retrieval. After a user inputs a query request, the query request processing system 100 extracts its semantic type, keyword attributes, and domain relevance features to determine that the query request is a professional category query request. Then, adaptive matching sparse retrieval, dense retrieval, and relevance retrieval channels are executed in parallel, yielding 18, 15, and 12 preliminary search results, respectively. Based on a preset fit score matrix, the dynamic weighting coefficients of the three channels are obtained after normalization: 0.44, 0.33, and 0.23. The weighted scores of each knowledge point are calculated and sorted, and the top 50 knowledge points are selected as the corresponding query results. The query request processing system 100 calculates the overall confidence of these knowledge points to be 0.75, which is higher than the preset confidence threshold of 0.3. After sorting according to the objective function of minimizing evidence conflict, the top 10 knowledge points are selected. The query request is concatenated with the prompt text, which is then input into a large language model to generate the target answer and display it on the user interface. The interface includes a summary of the knowledge scatter point source identifiers.

[0087] Furthermore, when a user clicks "Accurate," positive feedback is generated. The query processing system then identifies 10 knowledge points corresponding to the location of 100, calls a pre-defined referee model to re-estimate the confidence level, improves the confidence score, and records the number of valid calls. These high-confidence knowledge points are then structured to form two new knowledge points. These new and corrected knowledge points are written into the knowledge base, with version numbers and associated mappings updated synchronously, completing the closed-loop reinforcement of the knowledge base.

[0088] The foregoing mainly describes the solutions of the embodiments of the present invention from a methodological perspective. It is understood that, in order to achieve the above-mentioned functions, the query request processing system 100 includes at least one of the hardware structures and software modules corresponding to the execution of each function. Those skilled in the art should readily recognize that, in conjunction with the units and algorithm steps of the various examples described in the embodiments disclosed herein, the embodiments of the present invention can be implemented in hardware or a combination of hardware and computer software. Whether a function is executed in hardware or by computer software driving hardware depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of the embodiments of the present invention.

[0089] In this embodiment of the invention, the query request processing system 100 can be divided into functional units according to the above method example. For example, the query request processing system 100 can be divided into functional units corresponding to various functions, or two or more functions can be integrated into one processing unit. The integrated unit can be implemented in hardware or as a software functional unit. It should be noted that the unit division in this embodiment of the invention is illustrative and only represents one logical functional division; other division methods may be used in actual implementation.

[0090] For example, Figure 3 The diagram illustrates a hardware structure of a query request processing system according to an embodiment of the present invention. The query request processing system 100 includes: a knowledge scatter plot generation module 110, used to acquire an original knowledge source containing target domain knowledge, perform structured parsing on the original knowledge source, and generate knowledge scatter plots of the smallest semantic units; the knowledge scatter plots are triple arrays including core objects, association logic, and domain connotations, each triple array independently carrying a complete set of domain knowledge; The knowledge scatter plot writing module 120 is used to encode the knowledge scatter plots using a dual-view representation method combining ternary arrays and dense vectors, forming standardized knowledge data and storing it in the knowledge base; the knowledge scatter plot retrieval module 130 is used to respond to user-input query requests, adaptively select a multi-path retrieval strategy based on the semantic type, keyword attributes, and domain relevance characteristics of the query request, and retrieve the knowledge scatter plots corresponding to the query request from the knowledge base; the target answer generation module 140 is used to perform confidence evaluation and optimization processing on the knowledge scatter plots corresponding to the query request, and generate and display the target answer corresponding to the query request based on the optimized knowledge scatter plots.

[0091] It should be understood that a detailed description of the above-mentioned optional methods can be found in the foregoing method embodiments, and will not be repeated here. Furthermore, explanations of any of the query request processing systems 100 provided above, as well as descriptions of their beneficial effects, can be found in the corresponding method embodiments described above, and will not be repeated here.

[0092] This invention also provides a computer-readable storage medium storing at least one computer instruction, which is loaded and executed by a processor to implement the methods of the various embodiments described above. Explanations of the relevant content and descriptions of the beneficial effects of any of the computer-readable storage media provided above can be found in the corresponding embodiments described above, and will not be repeated here.

[0093] This invention also provides a chip. This chip integrates a control circuit for implementing the functions of the query request processing system 100 described above, and one or more ports. Optionally, the functions supported by this chip are as described above, and will not be repeated here.

[0094] Those skilled in the art will understand that the program for implementing all or part of the steps of the above embodiments, which can be executed by a program instructing related hardware, can be stored in a computer-readable storage medium. The storage medium mentioned above can be a read-only memory, a random access memory, etc. The processing unit or processor mentioned above can be a central processing unit, a general-purpose processor, an application-specific integrated circuit (ASIC), a microprocessor (DSP), a field-programmable gate array (FPGA), or other programmable logic devices, transistor logic devices, hardware components, or any combination thereof.

[0095] This invention also provides a computer program product containing instructions that, when executed on a computer, cause the computer to perform any of the methods described in the above embodiments. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the flow or function according to the embodiments of this invention is generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, computer instructions can be transmitted from one website, computer, server, or data center to another via wired (e.g., coaxial cable, fiber optic, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium accessible to a computer or a data storage device such as a server or data center that integrates one or more available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium (e.g., SSD), etc.

[0096] It should be noted that the devices for storing computer instructions or computer programs provided in the embodiments of the present invention, such as, but not limited to, the aforementioned memory, computer-readable storage medium, and communication chip, are all non-transitory. Those skilled in the art should recognize that the functions described in the embodiments of the present invention in one or more of the above examples can be implemented using hardware, software, firmware, or any combination thereof. When implemented using software, these functions can be stored in a computer-readable storage medium or transmitted as one or more instructions or code on a computer-readable storage medium. Computer-readable storage media include computer storage media and communication media, wherein communication media include any medium that facilitates the transmission of computer programs from one place to another. Storage media can be any available medium accessible to general-purpose or special-purpose computers.

[0097] Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of the present invention.

Claims

1. A query request processing method based on a scattered neural network, characterized in that, include: Obtain the original knowledge source containing knowledge of the target domain, perform structured parsing on the original knowledge source, and generate knowledge scatter points of the smallest semantic unit; The knowledge scatter points are triple arrays that include core objects, related logic, and domain connotations. Each triple array independently carries a complete set of domain knowledge. The knowledge scatter points are encoded using a dual-view representation method that combines the ternary array and dense vectors to form standardized knowledge data and store it in a knowledge base. In response to a user's query request, based on the semantic type, keyword attributes, and domain relevance features of the query request, an adaptive multi-path retrieval strategy is selected, and knowledge points corresponding to the query request are retrieved from the knowledge base. The confidence level of the knowledge scatter points corresponding to the query request is evaluated and optimized. Based on the optimized knowledge scatter points, the target answer corresponding to the query request is generated and displayed.

2. The query request processing method based on a scattered neural network according to claim 1, characterized in that, The original knowledge source includes at least one of PDF files, Word files, web page data, and database data; the target domain includes the financial or accounting domain; the process of acquiring the original knowledge source containing knowledge of the target domain, performing structured parsing on the original knowledge source, and generating knowledge scatter points of the smallest semantic units includes: The original knowledge source is processed hierarchically using a parallel pipeline parsing mechanism based on the Map-Reduce architecture. In the Map phase, the data is split into basic data fragments according to a preset unit. In the Reduce phase, the basic data fragments are syntactically segmented to obtain a set of sentences. The entity recognition model with domain fine-tuning is invoked to extract a triple array from the sentence set, which includes a core object, association logic, and domain connotation. The core object is a specific concept or object in the target domain, the association logic is the association type between core objects, and the domain connotation is the target domain knowledge meaning corresponding to the association logic. The core object, association logic, and domain connotation constitute the smallest indivisible semantic unit.

3. The query request processing method based on a scattered neural network according to claim 2, characterized in that, The method further includes: Add identification information to each knowledge point. The identification information includes a source identifier, a version number, and a domain classification label. The source identifier is used to record the origin of the original knowledge source, the version number is used to trace the generation and update records of the knowledge point, and the domain classification label is used to distinguish the sub-domain category to which the knowledge point belongs.

4. The query request processing method based on a scattered neural network according to claim 3, characterized in that, The method of using a dual-view representation combining ternary arrays and dense vectors to encode the knowledge scatter points, forming standardized knowledge data and storing it in a knowledge base includes: Semantic vector encoding is performed on the triple arrays corresponding to the knowledge scatter points to generate dense vectors that reflect the core semantic features of the triple arrays; The identification information of the ternary array, dense vector and knowledge scatter points are structurally integrated to form standardized knowledge data containing knowledge content, semantic features and management attributes; The standardized knowledge data is written into a vector database in batches. The vector database supports precise retrieval based on triple array keywords and semantic similarity retrieval based on dense vectors.

5. The query request processing method based on a scattered neural network according to claim 3, characterized in that, The process of adaptively selecting a multi-path retrieval strategy based on the semantic type, keyword attributes, and domain relevance features of the query request, and retrieving knowledge points corresponding to the query request from the knowledge base, includes: Feature extraction is performed on the query request to obtain the semantic type, keyword attributes, and domain relevance features corresponding to the query request; The retrieval channels are adaptively matched based on the semantic type, keyword attributes, and domain relevance features. The retrieval channels include a sparse retrieval channel using the BM25 algorithm, a dense retrieval channel using the vector matching algorithm, and a relevance retrieval channel based on a knowledge graph. The search operation is performed in parallel through the matched search channels to obtain the preliminary search results for each channel; Based on the compatibility between the query request and the retrieval channel, a dynamic weighting coefficient is determined, and the preliminary retrieval results are dynamically weighted and fused to obtain the knowledge scatter points corresponding to the query request.

6. The query request processing method based on a scattered neural network according to claim 5, characterized in that, The process of determining dynamic weighting coefficients based on the compatibility between the query request and the retrieval channel, and dynamically weighting and fusing the preliminary retrieval results to obtain the knowledge scatter points corresponding to the query request, includes: A matching score matrix for query requests and retrieval channels is pre-constructed. The matching score matrix records the matching score between query requests and each retrieval channel. The matching score between query requests and sparse retrieval channels is the first value, the matching score between query requests and dense retrieval channels is the second value, and the matching score between query requests and associated retrieval channels is the third value. The second value is less than or equal to the first value and greater than or equal to the third value. Extract the corresponding search channel fit scores from the fit score matrix, normalize the fit scores to obtain the dynamic weighting coefficients for each search channel, and the dynamic weighting coefficients satisfy that the sum of the weighting coefficients of all search channels is 1. Obtain the preliminary search results corresponding to each search channel, assign the dynamic weighting coefficient of the search channel to the knowledge points in each preliminary search result, and calculate the weighted score of each knowledge point; wherein, the weighted score is the product of the semantic matching score of the knowledge point and the dynamic weighting coefficient; All knowledge points are sorted in descending order of weighted score, and a predetermined number of knowledge points with the highest weighted scores are selected as the knowledge points corresponding to the query request.

7. The query request processing method based on a scattered neural network according to claim 6, characterized in that, The step of performing confidence evaluation and optimization on the knowledge scatter points corresponding to the query request, and generating and displaying the target answer corresponding to the query request based on the optimized knowledge scatter points, includes: Calculate the overall confidence of the knowledge scatter points corresponding to the query request. The confidence is based on the semantic matching degree between the knowledge scatter points and the query request, the historical call accuracy of the knowledge scatter points, and the version validity in the identification information. If the overall confidence level is less than the preset confidence threshold, the text expansion model is invoked to expand the query request with synonyms, generating at least two query variants. Based on the query variants, multi-path retrieval is performed again to obtain supplementary retrieval results. The preliminary search results and the supplementary search results are merged and deduplicated. The deduplicated knowledge points are sorted according to the objective function of minimizing evidence conflict. The top N knowledge points are selected and output, where N is a positive integer. The top N knowledge points are concatenated with the query request to form a prompt text, which is then input into a large language model to generate the target answer. The target answer is then displayed synchronously on the user interface.

8. The query request processing method based on a scattered neural network according to claim 1, characterized in that, The method further includes: Obtain user feedback information on the target answer, the feedback information being generated through user interaction, including a binary evaluation label on the accuracy of the target answer; Based on the binary evaluation labels, all knowledge points on which the target answer depends are located and generated. A preset referee model is called to re-estimate the confidence of the knowledge points. The preset referee model is used to perform secondary calibration of the confidence score based on the historical call frequency, matching accuracy and domain relevance of the knowledge points. If the binary evaluation label is a positive evaluation, the confidence score of the corresponding knowledge point is increased and the number of valid calls is recorded; if it is a negative evaluation, the confidence score of the corresponding knowledge point is decreased and it is marked as pending review. Based on the confidence re-estimation results of the knowledge scatter points, the vector parameters of the low-confidence knowledge scatter points are corrected to obtain the corrected knowledge scatter points. The knowledge content corresponding to the high-confidence knowledge scatter points is structured and sorted to form new knowledge scatter points. Write the revised and newly added knowledge points into the knowledge base, and update the version number and associated knowledge point mapping relationships synchronously.

9. A query request processing system based on a scattered neural network, characterized in that, include: The knowledge scattering generation module is used to obtain the original knowledge source containing target domain knowledge, perform structured parsing on the original knowledge source, and generate knowledge scattering points of the smallest semantic unit; the knowledge scattering points are triple arrays including core objects, association logic and domain connotation, and each triple array independently carries a complete set of domain knowledge; The knowledge scatter plot writing module is used to encode the knowledge scatter plots using a dual-view representation method that combines the triple array and dense vector, forming standardized knowledge data and storing it in the knowledge base; The knowledge scatter point retrieval module is used to respond to the query request input by the user, and adaptively select a multi-path retrieval strategy based on the semantic type, keyword attributes and domain relevance characteristics of the query request, and retrieve the knowledge scatter points corresponding to the query request from the knowledge base; The target answer generation module is used to perform confidence assessment and optimization processing on the knowledge scatter points corresponding to the query request, and generate and display the target answer corresponding to the query request based on the optimized knowledge scatter points.

10. An electronic device, characterized in that, include: processor; Memory used to store the processor's executable instructions; The processor is configured to execute the instructions to implement the query request processing method based on a scatter neural network as described in any one of claims 1-8.

Images