A power question and answer multi-object linkage updating method based on an intelligent algorithm

Abstract

The application discloses to the technical field of information system, specifically a power question and answer multi-object linkage updating method based on an intelligent algorithm, first, the original knowledge change in the power knowledge base is converted into a standard event that can participate in the dynamic updating of intelligent question and answer; then, the identified updating event is mapped into a multi-object linkage updating task in the intelligent question and answer link of the power knowledge base; subsequently, the affected knowledge content in the power knowledge base is re-divided into knowledge slices suitable for intelligent question and answer retrieval and evidence positioning; after that, the understanding ability of the intelligent question and answer system to user queries is kept consistent with the latest knowledge state in the power knowledge base; based on this, the application effectively solves the core problems in the power knowledge base intelligent question and answer scene, such as the disconnection between knowledge updating and question and answer ability updating, extensive local revision processing, query understanding lag, retrieval sorting mismatch and unstable answer output.

Description

Technical Field

[0001] This invention relates to the field of information system technology, specifically to a multi-object linkage update method for power-related question-and-answering based on intelligent algorithms. Background Technology

[0002] Currently, power companies continuously accumulate a large number of unstructured documents in their business activities, including regulations, procedures, plans, reports, and technical specifications. These documents typically come in formats such as Word, PDF, and scanned copies, and vary significantly across different business domains in terms of structural hierarchy, terminology, information density, and expression. To enable these documents to serve intelligent question answering, they usually require processing steps such as document parsing, knowledge slicing, semantic retrieval, result sorting, and answer generation to form a searchable, locatable, and answerable power knowledge base. Meanwhile, power grid business rules, terminology, and query expressions are not static but continuously change with policy revisions, process changes, and the introduction of new business scenarios. Therefore, the intelligent question answering system for the power knowledge base not only needs to complete its initial construction but also needs to possess the ability to dynamically update to adapt to changes in knowledge.

[0003] In existing technologies, the closest approach to this invention is a first-class solution for intelligent retrieval and question answering of business documents. This type of solution primarily addresses the problem of users being unable to quickly locate target information from a large number of business documents. It typically involves keyword analysis or semantic understanding of the user's question, performing a search within a document database, and extracting or organizing the answer from candidate document fragments. This type of solution improves document retrieval efficiency and has been widely applied in scenarios such as policy queries and power document retrieval. Its advantage lies in its ability to elevate traditional full-text search to natural language question answering.

[0004] The second type of solution is the knowledge base update or feedback-driven update solution. This type of solution mainly addresses the problems of outdated knowledge base content and high manual maintenance costs. It typically triggers knowledge updates through user questions, user feedback, manual revisions, or changes in answer credibility, and writes the new question-and-answer pairs, answer items, or knowledge entries back into the knowledge base. This type of solution has a relatively straightforward implementation path and is commonly found in scenarios such as FAQ management and customer service knowledge maintenance.

[0005] The third category is dynamic adaptive retrieval-enhanced question-answering solutions or knowledge graph update solutions. These solutions primarily address the problem of frequently changing knowledge and the need for continuous optimization of the question-answering system. They typically involve introducing retrieval enhancement architectures, knowledge representation layers, dynamic learning modules, or knowledge graph update mechanisms to continuously adjust the knowledge structure and question-answering output. This type of solution has a relatively complete overall framework and is suitable for question-answering systems with complex knowledge organization and long-term evolution.

[0006] However, the aforementioned existing technologies still have significant shortcomings in the intelligent question-and-answer scenario of the power knowledge base targeted by this invention. First, existing solutions typically understand knowledge updates as adding document content or supplementing knowledge entries, focusing on "whether it's entered into the database" rather than "whether a complete question-and-answer capability matching the new knowledge is formed." In scenarios where power knowledge documents have heterogeneous formats and complex structures, if only new or revised documents are written into the knowledge base without corresponding adjustments to the knowledge slices and index structure, the system may still use old slices and indexes, making it difficult to accurately recall the updated knowledge. Second, existing solutions mostly rely on static dictionaries, fixed word segmentation rules, or preset intent tags on the query understanding side. Their design logic is to first establish a query preprocessing and classification mechanism and then reuse it long-term. However, in power business scenarios, terminology, abbreviations, and expressions change with institutional and business changes. If the business dictionary, query mapping, and intent recognition configuration cannot be adjusted synchronously with knowledge updates, even though new documents have entered the knowledge base, the query side still cannot correctly understand the user's query, ultimately causing a deviation in the retrieval direction. Third, existing retrieval and ranking capabilities typically employ fixed recall paths and ranking strategies, lacking parameter update mechanisms that adapt to knowledge changes. When new knowledge enters the knowledge base, existing keyword retrieval, semantic retrieval, result fusion, and reordering configurations may no longer be applicable. This can lead to issues such as new knowledge slices not being effectively prioritized and old knowledge fragments still occupying high ranking positions, thus affecting the quality of subsequent answer materials. Fourth, existing technologies in the answer generation stage usually only focus on generating readable answers, lacking business verification and unified effectiveness mechanisms to support knowledge updates. For the power sector, answers must not only be "answerable" but also comply with current business specifications and regulations. If answer templates, rule constraints, and verification mechanisms are not updated synchronously with knowledge changes, even if prior knowledge has been updated, the final output may still use old rule expressions, resulting in unstable or non-compliant answers. Fifth, existing technologies often optimize document processing, query understanding, retrieval sorting, and answer output as independent modules, lacking a unified and interconnected main chain for updating based on knowledge changes. This results in the system only being able to make partial repairs and failing to form a complete transmission link from document updates to answer output updates, leading to problems such as update delays, inconsistent configurations, and fluctuations in question-and-answer results.

[0007] Therefore, this invention proposes a multi-object linkage update method for power knowledge base question answering based on intelligent algorithms. This method enables the intelligent question answering system for the power knowledge base to not only update the knowledge content itself when faced with new, replaced, or revised documents, but also to synchronously update knowledge slices, index data, business dictionaries, query preprocessing mappings, intent recognition configurations, retrieval sorting configurations, and answer generation and verification configurations. This allows knowledge changes to be transmitted along the complete question answering chain and to take effect in a controlled manner, solving the problem of the disconnect between knowledge updates and intelligent question answering capability updates in existing technologies. Summary of the Invention

[0008] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0009] A multi-object linkage update method for electricity question answering based on intelligent algorithms includes the following steps:

[0010] S1, Power Knowledge Base Update Event Recognition: Transforms the original knowledge changes in the power knowledge base into standard events that can participate in the dynamic update of intelligent question answering;

[0011] S2, Multi-object linkage impact determination for intelligent question answering: Map the update event identified by S1 to the multi-object linkage update task in the intelligent question answering link of the power knowledge base;

[0012] S3, Incremental Reconstruction of Electricity Knowledge Slices for Intelligent Question Answering: The affected knowledge content in the electricity knowledge base is re-sliced ​​into knowledge slices suitable for intelligent question answering retrieval and evidence location, and incremental index replacement is completed;

[0013] S4, Query understanding linkage update for intelligent question answering: to keep the intelligent question answering system's ability to understand user queries consistent with the latest knowledge status in the power knowledge base;

[0014] S5, Verification and Activation of Linked Update Package for Intelligent Question Answering Capabilities: The update results formed by S3 and S4 are encapsulated into a unified linked update package that serves intelligent question answering capabilities, and the entire package takes effect after verification.

[0015] As a preferred embodiment of the intelligent algorithm-based multi-object linkage update method for power-related question answering described in this invention, the specific steps of S1 are as follows:

[0016] S11, Establish basic document records: Generate a unique document number for each document entering the update channel, and write the document name, document type, business domain, version number and status information to form basic document records;

[0017] S12, Perform structural alignment between old and new versions: Perform structural comparison between old and new versions under the same document number to achieve structural alignment; after alignment, a structural record is generated;

[0018] S13, Identify update type: After the structure alignment is completed, determine whether the change is an addition, replacement, partial revision or failure marker;

[0019] S14, Generate difference record: Perform content difference recognition on the aligned corresponding structural units to form a difference record;

[0020] S15, Calculate the basic change measures: calculate the degree of structural change, the degree of terminological change, and the degree of rule change respectively;

[0021] S16, Output standardized update events: Write the document base record, structure record, update event record, and difference record into the update event set.

[0022] As a preferred embodiment of the intelligent algorithm-based multi-object linkage update method for power-related question answering described in this invention, the specific steps of S2 are as follows:

[0023] S21, Construct an object impact mapping table: Pre-establish a mapping table between update type, change nature and object impact relationship;

[0024] S22, Calculate the update impact value: Let the structural change degree be S, the terminology change degree be T, and the rule change degree be R, then the update impact value I is calculated as follows:

[0025] I = 0.35×S + 0.30×T + 0.35×R

[0026] Among them, S, T, and R are output by S1;

[0027] S23, classifying impact levels according to threshold:

[0028] When I < 0.20, the event is marked as a low-impact update;

[0029] When 0.20 ≤ I < 0.45, the event is marked as a low to medium impact update;

[0030] When 0.45 ≤ I < 0.70, the event is marked as a medium-to-high impact update;

[0031] When I ≥ 0.70, the event is marked as a high-impact update;

[0032] S24, Generate object propagation results:

[0033] For low-impact updates, only index correction tasks are generated;

[0034] For updates with low to medium impact, generate slice reconstruction tasks, slice index update tasks, and business dictionary update tasks;

[0035] For medium-to-high impact updates, based on the aforementioned tasks, further query preprocessing rule update tasks and intent recognition configuration update tasks are generated;

[0036] For high-impact updates, further tasks such as search configuration update tasks, reordering configuration update tasks, and answer template and compliance rule update tasks are generated.

[0037] S25, Generate object-level update task records: Generate a task record for each affected object;

[0038] S26, Establish task dependency chain: Link object-level tasks into a directed dependency chain.

[0039] As a preferred embodiment of the intelligent algorithm-based multi-object linkage update method for power-related question answering described in this invention, the specific steps of S3 are as follows:

[0040] S31, Determine the affected area: Based on the change location and title path in the difference record, construct the affected area with the change location as the center;

[0041] S32, Determine whether to retain or reconstruct old slices: For existing old slices, determine their relationship with the affected areas in turn;

[0042] S33, Select slicing strategy:

[0043] For documents containing rules and regulations, slice them according to chapters and clause boundaries;

[0044] For proposals and reports, slice them into thematic paragraphs;

[0045] The scanned document's text is parsed, and sliced ​​using page numbers and paragraph boundaries;

[0046] S34, Perform slice merging or splitting: If the length of a slice is lower than the lower limit of its type and the similarity of its topic with adjacent slices is greater than a preset value, then perform merging; if the length of a slice is higher than the upper limit of its type, or the information density value is greater than a preset value, then perform splitting.

[0047] S35, Generate new slice record: Create a slice record for each reconstructed slice;

[0048] S36, Establish the mapping relationship between old and new slices: For new slices evolved from old slices, establish a mapping record between old and new slices;

[0049] S37, Perform incremental index replacement: Generate new semantic indexes and metadata indexes for the reconstructed slices, and retain the original indexes for the preserved slices; then, based on the mapping relationship between the new and old slices, only replace the index entries corresponding to the affected slices.

[0050] As a preferred embodiment of the intelligent algorithm-based multi-object linkage update method for power-related question answering described in this invention, the specific steps of S4 are as follows:

[0051] S41, Extract candidate terms: Perform word segmentation and phrase recognition on the reconstructed slice text to extract new business terms, abbreviation and full name correspondences, synonym clusters and regular phrases; at the same time, establish a term record for each candidate term;

[0052] S42, Filter term increments:

[0053] If a candidate term appears at a frequency of no less than a preset number of times in the same business domain and its length is within a preset number of Chinese characters, then it will be included in the term increment set.

[0054] If the edit distance between a candidate term and an existing term is less than a preset value, and the context similarity is greater than a preset value, then it is merged into a synonym of the existing term.

[0055] If an abbreviation and its full name appear together a predetermined number of times in the same paragraph or two adjacent paragraphs, a mapping relationship between the abbreviation and the full name is established.

[0056] S43, Generate query preprocessing rules: Generate query preprocessing rules based on the term increment set;

[0057] S44 defines the rule execution order: after a user query enters, noise cleaning is performed first, followed by abbreviation expansion, then terminology unification and synonym merging, and finally context completion; if the same query hits multiple rules at the same time, they are executed in order of priority.

[0058] S45, Update Intent Configuration: For the newly added question-and-answer scenario, create multiple representative query examples for each intent and form an intent configuration record;

[0059] S46, Execution Intent Classification: For the normalized query, first filter candidate intents according to the triggering rules, then calculate the average similarity and maximum similarity between it and each intent sample set to achieve intent classification;

[0060] S47, Output query understanding update results: Write the term increment set, query preprocessing rule increment set, and intent configuration increment set into the update results of this round.

[0061] As a preferred embodiment of the intelligent algorithm-based multi-object linkage update method for power-related question answering described in this invention, the specific steps of S5 are as follows:

[0062] S51, Construct a linked update package: Encapsulate the document version, slice version, dictionary version, query rule version, intent configuration version, retrieval configuration version, rearrangement configuration version, answer template version, verification status, and effective status involved in this round of updates into a linked update package;

[0063] S52, Update Search Configuration: Configure the weights of keyword search and semantic search channels based on intent type and business scenario;

[0064] S53, Perform multi-channel recall and fusion screening: Use dual-channel recall of candidate slices, normalize the results and calculate the fusion score;

[0065] S54, Perform rearrangement calculation: Calculate the rearrangement score for the candidate slices; then sort them from high to low according to the rearrangement score;

[0066] S55, Match answer templates and generate evidence chains: Match answer templates based on intent type; simultaneously, retain evidence records when generating answers;

[0067] S56, Perform consistency check: Check whether the slices, dictionaries, query rules, retrieval configurations, rearrangement configurations and answer templates in the linked update package all exist, and whether their version identifiers are consistent;

[0068] S57, Perform question-and-answer validation: Use multiple regression test queries for validation;

[0069] S58, perform atomic activation or rollback: If both consistency check and question-and-answer check pass, the entire linked update package is marked as valid, and the slice, dictionary, query rules, intent configuration, retrieval configuration, rearrangement configuration, and answer template are switched at once; if any check fails, the previous valid update package is retained and continues to serve.

[0070] Compared with existing technologies:

[0071] First, by adopting a structured identification mechanism for power knowledge base update events, unified modeling of knowledge changes such as additions, replacements, partial revisions, and failures has been achieved. This solves the problems of scattered sources of knowledge changes, unclear change locations, and difficulty in quantifying the nature of changes in existing technologies, and provides a standardized event entry point for subsequent coordinated updates.

[0072] Second, by adopting a multi-object linkage influence judgment mechanism for intelligent question answering, knowledge changes are propagated to multiple objects such as knowledge slices, indexes, business dictionaries, query preprocessing, intent recognition, retrieval ranking, and answer templates according to their influence intensity. This solves the problem in existing technologies that only update document content without updating other objects in the question answering link, enabling knowledge updates to truly be transformed into question answering capability updates.

[0073] Third, by adopting a difference-aware incremental slice reconstruction and index replacement mechanism, slice reconstruction and index replacement are performed only on the affected areas in local revision scenarios. This solves the problems of high overhead, old slice residue, and mixed use of old and new versions caused by traditional full re-slicing and full reconstruction, thus improving the processing efficiency and structural stability of the power knowledge base in continuous update scenarios.

[0074] Fourth, by adopting a query understanding linkage update mechanism for intelligent question answering, business terms, abbreviations, synonyms, and new scenario intents are updated synchronously. This solves the problem in existing technologies where query understanding relies on static dictionaries and fixed intent configurations, resulting in users' queries still not being accurately understood after new knowledge is introduced. This improves the front-end understanding capability of the intelligent question answering chain.

[0075] Fifth, by adopting a linkage update and unified activation mechanism for retrieval parameters, rearrangement parameters, and answer templates, the overall transmission from knowledge changes to answer generation capabilities is realized. This solves the problems of independent updates of retrieval configuration, rearrangement logic, and answer templates, inconsistent versions, and unstable results after going online in existing technologies, making the intelligent question-and-answer output more in line with current power business rules and regulations.

[0076] Sixth, by adopting a consistency check and atomic activation mechanism for linked update packages, unified version management and overall switching of slices, dictionaries, query rules, intent configurations, retrieval configurations, rearrangement configurations, and answer templates are achieved. This solves the problems of asynchronous updates of various objects, residual old rules, fluctuations in online results, and difficulty in rolling back anomalies in existing technologies, thereby improving the long-term stability and maintainability of the intelligent question-and-answer system for the power knowledge base.

[0077] In summary, this invention effectively solves the core problems in the intelligent question-and-answer scenario of power knowledge base, such as the disconnect between knowledge updates and question-and-answer capability updates, the crude processing of partial revisions, the lag in query understanding, the mismatch between retrieval and sorting, and the unstable output of answers. It provides a new technical solution for the dynamic evolution and intelligent question-and-answer application of power knowledge base. Attached Figure Description

[0078] Figure 1 This is a flowchart illustrating the overall technical solution of the present invention.

[0079] Figure 2 This is a schematic diagram of the system / device structure of the present invention;

[0080] Figure 3 This is a block diagram of the core mechanism principle of the present invention (influence domain propagation + update packet atomic effect). Detailed Implementation

[0081] To make the objectives, technical solutions, and advantages of the present invention clearer, the embodiments of the present invention will be described in further detail below with reference to the accompanying drawings.

[0082] This invention provides a multi-object linkage update method for power-related question answering based on intelligent algorithms, the process of which is shown in the attached figure. Figure 1 As shown, firstly, changes such as document additions, replacements, partial revisions, and invalidations in the power knowledge base are identified, forming standard events for intelligent question-and-answer updates. Secondly, based on the update events, the impact of these changes on multiple key objects in the intelligent question-and-answer chain of the power knowledge base is determined, and a linked update task is generated. Subsequently, incremental slice reconstruction and index replacement with difference awareness are performed on the affected knowledge content, enabling the updated power knowledge to re-enter the searchable state of intelligent question-and-answer. On this basis, terminology increments are further extracted, and query preprocessing rules and intent recognition configurations are updated in a linked manner, so that user queries can be correctly understood according to the new power knowledge caliber. Finally, search parameters, rearrangement parameters, and answer templates are adjusted synchronously, and the results of this round of updates are packaged into a unified update package, which takes effect after consistency verification and question-and-answer verification. Through the above technical links, this invention realizes the continuous maintenance of the intelligent question-and-answer capability of the power knowledge base under the condition of knowledge change, enabling newly added or revised power knowledge to be promptly transformed into new question-and-answer understanding capabilities, search capabilities, and answer generation capabilities. To achieve the above-mentioned objectives, this invention adopts the following technical solution:

[0083] S1, Power Knowledge Base Update Event Recognition: Transforms the original knowledge changes in the power knowledge base into standard events that can participate in the dynamic update of intelligent question answering;

[0084] The inputs for this step include the document to be processed in the power knowledge base, the content of the old and new versions of the document, the basic attributes of the document, and the update signal triggered by an external source. The basic attributes of the document include at least the document number, document name, document type, business domain, source path, version number, effective time, expiration time, and current status.

[0085] Establish a basic document record: Generate a unique document number for each document entering the update channel, and write the document name, document type, business domain, version number and status information to form a basic document record; this record serves as the master record for all subsequent steps to avoid confusion between different versions of the same document in subsequent processing;

[0086] Perform structural alignment between old and new versions: Perform a structural comparison of old and new versions under the same document number to achieve structural alignment: For documents with explicit heading levels, perform first-level alignment based on headings, chapter numbers, clause numbers, and paragraph boundaries; for scanned documents converted to text, first restore page numbers, paragraph boundaries, and heading features before alignment; after alignment, a structural record is generated; the structural record must include at least the structural unit number, parent structure number, heading path, page number start and end positions, character start and end positions, and unit type; the unit type must include at least chapter, clause, paragraph, table, and attachment description;

[0087] Identify update type: After structural alignment is completed, determine whether the change is an addition, replacement, partial revision, or invalidation: If there is no valid historical version with the same document number in the system, the change is marked as an addition; if the old version is completely removed from the valid state and fully taken over by the new version, it is marked as a replacement; if only some structural units are changed, it is marked as a partial revision; if the document is obsolete and there is no replacement version, it is marked as invalidation.

[0088] Generate difference records: Perform content difference identification on the aligned corresponding structural units to form difference records; the difference record includes at least the difference number, event number, changed structural unit number, title path, content summary before change, content summary after change, change location range, change nature, and change intensity; the change nature is preferably divided into three categories: structural change, terminology change, and rule definition change;

[0089] Calculate the basic change metrics: calculate the degree of structural change, the degree of terminology change, and the degree of rule change separately; where, the degree of structural change is the proportion of the number of changed structural units to the total number of original structural units; the degree of terminology change is the proportion of the number of newly added or replaced terms to the total number of original terms in the changed area; the degree of rule change is the proportion of the number of changed rule sentences to the total number of original rule sentences; if the proportion of changes in chapter level, clause boundary, or paragraph division exceeds 15%, it is judged as a significant structural change; if the proportion of newly added terms or abbreviations exceeds 20%, it is judged as a significant terminology change; if the proportion of changed sentences in applicable conditions, restrictive descriptions, or concluding statements exceeds 25% of the rule sentences, it is judged as a significant change in rule scope;

[0090] Output standardized update events: Write the document base record, structure record, update event record, and difference record into the update event set as input for subsequent impact domain propagation determination; if the old version is missing, the structure restoration fails, or the difference identification conflict is found in this step, the event is marked as an abnormal event and will not enter the subsequent chain.

[0091] Through the above processing, the output of S1 is no longer a simple "document update," but rather standard event data for the dynamic updating of the intelligent question-and-answer system for the power knowledge base. This output is directly used as the input of S2 to determine which question-and-answer objects will be affected by this update and to what extent.

[0092] S2, Multi-object linkage impact determination for intelligent question answering: Map the update event identified by S1 to the multi-object linkage update task in the intelligent question answering link of the power knowledge base;

[0093] The inputs for this step are update event records, difference records, and structure records; the outputs are the impact domain determination results and a set of object-level update tasks. The objects here include at least: knowledge slice objects, slice index objects, business dictionary objects, query preprocessing objects, intent recognition configuration objects, retrieval configuration objects, rearrangement configuration objects, and answer template and compliance rule objects.

[0094] Construct an object impact mapping table: Pre-establish a mapping table between update type, change nature, and object impact relationship; for example, structural changes first affect slice and index objects; terminology changes first affect business dictionary, query preprocessing, and intent recognition objects; rule definition changes first affect answer template, retrieval configuration, and reordering configuration objects; this mapping table provides the basic rules for subsequent automatic propagation.

[0095] Calculating the update impact value: Let the degree of structural change be S, the degree of terminological change be T, and the degree of rule change be R, then the update impact value I is calculated as follows:

[0096] I = 0.35×S + 0.30×T + 0.35×R

[0097] Among them, S, T, and R are output by S1; this value is used to measure the overall impact of this update on the intelligent question-answering capability of the power knowledge base;

[0098] Impact levels are categorized by threshold:

[0099] When I < 0.20, the event is marked as a low-impact update;

[0100] When 0.20 ≤ I < 0.45, the event is marked as a low to medium impact update;

[0101] When 0.45 ≤ I < 0.70, the event is marked as a medium-to-high impact update;

[0102] When I ≥ 0.70, the event is marked as a high-impact update;

[0103] Meanwhile, if the rule change degree R > 0.40, it will be unconditionally upgraded to at least a medium-high impact update; if the term change degree T > 0.30, it will be upgraded to at least a query understanding update level in the medium-low impact update category.

[0104] Propagation results of generated objects:

[0105] For low-impact updates, only index correction tasks are generated;

[0106] For updates with low to medium impact, generate slice reconstruction tasks, slice index update tasks, and business dictionary update tasks;

[0107] For medium-to-high impact updates, based on the aforementioned tasks, further query preprocessing rule update tasks and intent recognition configuration update tasks are generated;

[0108] For high-impact updates, further tasks such as search configuration update tasks, reordering configuration update tasks, and answer template and compliance rule update tasks are generated.

[0109] Generate object-level update task records: Generate a task record for each affected object; the task record should include at least the task number, event number, object type, impact level, preceding task, subsequent task, target version number, and task status; preferably, slice and index update tasks have the highest priority, query understanding update tasks are the next highest, retrieval and rearrangement update tasks are the next lowest, and answer configuration update tasks have the lowest priority; this ensures that the data for the next step comes from the effective output of the previous step;

[0110] Establish a task dependency chain: Link object-level tasks into a directed dependency chain: After slice reconstruction and index update are completed, business dictionary and query preprocessing updates can be executed; after query rules and intent configuration updates are completed, retrieval and rearrangement parameter updates can be executed; after retrieval and rearrangement updates are completed, answer templates and unified validity verification can be executed; this forms the basis for the execution order of subsequent steps.

[0111] The output of S2 is the set of object-level update tasks and the result of the influence domain determination. This output will flow into subsequent steps respectively, with slicing and indexing related tasks going into S3, query understanding related tasks going into S4 based on the results of S3, and retrieval, reordering, and answer configuration related tasks going into S5 based on the results of S4.

[0112] S3, Incremental Reconstruction of Electricity Knowledge Slices for Intelligent Question Answering: The affected knowledge content in the electricity knowledge base is re-sliced ​​into knowledge slices suitable for intelligent question answering retrieval and evidence location, and incremental index replacement is completed;

[0113] The inputs for this step are the update tasks, change locations, structure records, and document content related to slices and indexes in S2; the outputs are the reconstructed slice set, the retained slice set, and the mapping relationship between the old and new slices.

[0114] Determine the affected area: Based on the change location and title path in the difference record, construct the affected area with the change location as the center; preferably, the affected area includes the directly changed text block, the upward extension of the first-level title path, and the downward extension of the two adjacent semantic segments; the reason for doing this is that local revisions may change the context and interpretation boundaries of the clause, rather than just affecting a single sentence;

[0115] Determining whether to retain or reconstruct old slices: For existing old slices, determine their relationship with the affected areas in sequence:

[0116] If the overlap between the old slice and the character range of the affected area is less than 10%, and the title path remains unchanged, then the slice is determined to be retained.

[0117] If the overlap ratio between the old slice and the character range of the affected area is greater than or equal to 10%, it is determined to be a reconstructed slice;

[0118] If, although they do not directly overlap, the title path changes, or the semantic similarity with the changed area is greater than 0.78, it is determined to be an associated reconstructed slice.

[0119] Among them, semantic similarity is preferably calculated using the cosine similarity of sliced ​​text vectors;

[0120] Choose a slicing strategy:

[0121] For documents containing rules and regulations, it is preferable to slice them by chapter and clause boundaries, with the length of a single slice controlled between 280 and 420 Chinese characters;

[0122] For proposals and reports, it is preferable to slice them by topic paragraphs, with the length of a single slice controlled between 350 and 600 Chinese characters;

[0123] When parsing text from scanned documents, it is preferable to combine page numbers and paragraph boundaries into slices, with the length of a single slice controlled between 220 and 400 Chinese characters.

[0124] This categorization strategy makes the slices more suitable for subsequent intelligent question-answering retrieval;

[0125] Perform slice merging or splitting: If a slice's length is below the lower limit of its type and its topic similarity with adjacent slices is greater than the preset value (0.82), then merge it; if a slice's length is above the upper limit of its type, or its information density value is greater than the preset value (0.68), then split it.

[0126] The information density value D is calculated as follows:

[0127] D = 0.40 × Terminology density + 0.35 × Rule sentence ratio + 0.25 × Indicator field ratio;

[0128] This calculation ensures that the slices retain semantic integrity without being too long and affecting retrieval accuracy.

[0129] Generate new slice records: Create a slice record for each restructured slice; the slice record should include at least: slice number, document number, version number, slice order, slice text, title path, page number range, character position range, business domain label, source chapter, and event number; for retained slices, use the original slice number; for restructured slices, generate a new slice number;

[0130] Establish a mapping relationship between old and new slices: For new slices evolved from old slices, establish a mapping record between old and new slices; the mapping record should include at least the old slice number, the new slice number, the mapping type, and the associated event number; the mapping type preferably includes four categories: inheritance, splitting, merging, and replacement. This mapping relationship is used for subsequent evidence backtracking and index replacement;

[0131] Perform incremental index replacement: generate new semantic and metadata indexes for reconstructed slices, and retain the original indexes for preserved slices; then, based on the mapping relationship between the new and old slices, only replace the index entries corresponding to the affected slices, without rebuilding the entire knowledge base index; this can significantly reduce the reconstruction cost in local revision scenarios.

[0132] The output of this step includes the reconstructed slice set, the retained slice set, the mapping relationship between the old and new slices, and the incremental index results. These outputs will serve as the direct basis for term extraction, query rule reconstruction, and intent configuration updates in S4, and as the knowledge source for retrieval and answer generation in S5.

[0133] S4, Query understanding linkage update for intelligent question answering: to keep the intelligent question answering system's ability to understand user queries consistent with the latest knowledge status in the power knowledge base;

[0134] The inputs to this step are the reconstructed slice set, the retained slice set, and the mapping relationship between the old and new slices output from S3, as well as the update tasks related to the business dictionary, query preprocessing rules, and intent recognition configuration in S2; the outputs are the term increment set, the query preprocessing rule increment set, and the intent configuration increment set.

[0135] Extracting candidate terms: Perform word segmentation and phrase recognition on the reconstructed slice text to extract new business terms, abbreviation and full name correspondences, synonym clusters and regular phrases; at the same time, establish a term record for each candidate term, which includes at least the term text, standard term, term type, source slice number, business domain label, frequency of occurrence and contextual examples;

[0136] Filter term increments:

[0137] If a candidate term appears at least 3 times in the same business domain and its length is between the preset number of Chinese characters (2 to 12 Chinese characters), then it will be included in the term increment set.

[0138] If the edit distance between a candidate term and an existing term is less than the preset value (0.25) and the context similarity is greater than the preset value (0.80), then it is merged into a synonym of the existing term.

[0139] If an abbreviation and its full name appear together a predetermined number of times (at least 2 times) in the same paragraph or two adjacent paragraphs, then a mapping relationship between the abbreviation and the full name is established.

[0140] The above screening process ensures that newly added terms have a certain degree of stability, rather than being occasional textual noise.

[0141] Generate query preprocessing rules: Generate query preprocessing rules based on the term increment set. The rules include at least noise removal rules, abbreviation expansion rules, terminology unification rules, colloquialism unification rules, and context completion rules. Each rule includes at least the source expression, target expression, rule type, applicable business domain, and priority.

[0142] The rule execution order is defined as follows: after a user query enters, noise cleaning is performed first, followed by abbreviation expansion, then terminology unification and synonym merging, and finally context completion. If the same query matches multiple rules, they are executed in order of priority, with the following order of preference: abbreviation expansion is the highest, followed by terminology unification, then colloquialism unification, and finally context completion. This ensures that technical terms and abbreviations are expanded correctly first, and then subjected to lightweight normalization.

[0143] Update intent configuration: For new question-and-answer scenarios, create multiple (5 to 10) representative query examples for each intent and form an intent configuration record; the intent configuration record should include at least the intent tag, typical example set, scenario, trigger rule and confidence condition;

[0144] Execution Intent Classification: For the normalized query, candidate intents are first filtered according to the triggering rules, and then the average similarity and maximum similarity with each intent sample set are calculated to classify the intent. If the average similarity is greater than 0.76 and the maximum similarity is greater than 0.82, it is classified into the corresponding intent. If multiple intents meet the conditions at the same time, the intent with the highest maximum similarity is selected. If none of them meet the conditions, it falls back to the default search intent. This mechanism enables new question-answering scenarios introduced by new knowledge to quickly enter the intelligent question-answering link.

[0145] Output query understanding update results: Write the term increment set, query preprocessing rule increment set, and intent configuration increment set into the update results of this round, as a prerequisite for retrieval parameter update, reordering calculation, and answer template selection in S5; if a term has strong ambiguity, a rule conflict cannot be resolved, or a new intent has insufficient distinguishability, the corresponding configuration will not be included in the effective update package of this round.

[0146] Through the above processing, S4 enables changes in the knowledge content of the power knowledge base to be directly transmitted to the intelligent question-answering system's new understanding of user queries.

[0147] S5, Verification and activation of the linked update package for intelligent question answering capabilities: The update results formed by S3 and S4 are encapsulated into a unified linked update package that serves intelligent question answering capabilities, and the entire package becomes effective after verification;

[0148] The inputs to this step are the slice and index results output by S3, and the term increments, query rule increments, and intent configuration increments output by S4; the output is a valid update package that can be officially put into use.

[0149] Construct a linked update package: Encapsulate the document version, slice version, dictionary version, query rule version, intent configuration version, retrieval configuration version, rearrangement configuration version, answer template version, verification status, and effective status involved in this round of updates into a linked update package; this update package is a complete upgrade result of the intelligent question answering capability of the power knowledge base;

[0150] Update search configuration: Configure the weights of keyword search and semantic search based on intent type and business scenario; for clause location questions, set the keyword search weight to 0.65 and the semantic search weight to 0.35; for explanation and summary questions, set the keyword search weight to 0.40 and the semantic search weight to 0.60; for comprehensive question-and-answer questions, set both to 0.50.

[0151] Perform multi-channel recall and initial screening: Candidate slices are recalled using dual channels, and the results are normalized before calculating the fusion score. If a slice is hit in both channels, the scores are summed according to the channel weights. If the fusion score is below 0.42, it is removed during the initial screening stage. A set of candidate slices is obtained after the initial screening.

[0152] Perform rearrangement calculation: Calculate the rearrangement score for the candidate slices, using the following formula:

[0153] P = 0.45 × semantic similarity + 0.20 × title path matching + 0.20 × business domain consistency + 0.15 × version freshness;

[0154] Then, sort the slices by their reordering scores from highest to lowest, and select the top 5 slices as answer material; this ensures that the slices that best match the current query and the current knowledge version are prioritized for answer generation.

[0155] Match answer templates and generate evidence chains: Match answer templates according to intent type, preferably including three types: clause citation type, inductive explanation type, and step summary type; at the same time, when generating answers, in addition to the answer text, evidence records are also retained; the evidence records include at least the evidence slice number, source document number, title path, page number range, and character position range to support the original text tracing of intelligent question answering results;

[0156] Perform consistency verification: Check whether all slices, dictionaries, query rules, retrieval configurations, rearrangement configurations, and answer templates in the linked update package exist and whether their version identifiers are consistent; if any key object is missing or the version is inconsistent, the update package is not allowed to enter the effective state;

[0157] Perform Q&A validation: It is preferable to use multiple (20) regression test queries for validation, requiring that the proportion of the target slice entering the top 5 is not less than 85%, the proportion of answers referencing the current valid version content reaches 100%, and the old version clause residual rate is less than 5%; the update package is only ready to go live when the Q&A validation passes.

[0158] Perform atomic activation or rollback: If both consistency and question-and-answer checks pass, the entire linked update package will be marked as valid, and the slices, dictionary, query rules, intent configuration, retrieval configuration, rearrangement configuration, and answer template will be switched at once; if any check fails, the previous valid update package will continue to serve; if more than 3 out of 10 consecutive online queries after activation contain references to the old version or the key slice is not hit, the update will automatically roll back to the previous valid update package.

[0159] Through the above processing, S5 truly transforms the results of the preceding steps into an effective online state of the intelligent question-and-answer capability of the power knowledge base, rather than a fragmented local replacement.

[0160] In summary, the present invention includes, but is not limited to, the following embodiments:

[0161] To verify the feasibility, effectiveness, and technical advantages of the multi-object linkage dynamic update method for intelligent question answering in a power knowledge base, this implementation case is based on anonymized real business data from a power company's knowledge base construction scenario. The selected data comes from knowledge documents such as policy documents, regulations, project plans, technical specifications, and acceptance materials, covering various formats including Word, editable PDF, formatted PDF, and scanned PDF. Newly added policy documents and revised regulations are also included as dynamic update samples. The implementation objective is to verify the practical effects of this invention in areas such as update event identification, multi-object impact propagation, incremental reconstruction of knowledge slices, query-understanding linkage updates, and unified activation of intelligent question answering capabilities.

[0162] This embodiment selects 100 anonymized business documents as the initial knowledge base sample, including 45 policy documents, 25 procedure documents, 20 scheme materials, and 10 technical specifications; 10 newly added policy documents and 10 partially revised procedure documents are selected as dynamic update verification samples. The implementation environment uses Python 3.9, document parsing tools include python-docx, PyPDF2, and pdfplumber, scanned document recognition uses PP-OCRv4, semantic vector retrieval uses an Elasticsearch vector database, and metadata storage uses MySQL.

[0163] Step S1: Update the power knowledge base event identification.

[0164] First, the initial sample documents and incremental sample documents, along with their source system identifier, business domain identifier, upload time, and original file path, were input into the system. The system assigned a unique document number to each document and recorded its name, type, version number, effective date, and status, forming a basic document record. Subsequently, the system aligned the new and revised documents to their original structure, restoring heading levels, chapter numbers, clause numbers, and paragraph boundaries, and generating a structure record. For revised documents, the system further identified the location of changes, the summary of content before and after the changes, and the nature of the changes, forming update event records and difference records. Implementation results show that all 20 dynamically updated samples could be identified as new, partial revision, or replacement events, forming unified update event data.

[0165] Step S2: Determine the impact of multi-object linkage.

[0166] After acquiring update events, the system calculates the degree of structural change, terminology change, and rule change, and classifies the impact level according to the update impact value formula. For newly added institutional documents, the system determines them to be high-impact updates, generating tasks such as slice reconstruction, index update, business dictionary update, query rule update, intent configuration update, retrieval and rearrangement parameter update, and answer template update. For procedural documents that only revise a few clauses, the system prioritizes generating slice reconstruction, index update, and partial query understanding update tasks, without triggering full object reconstruction. Implementation results show that the system can differentiate the update scope according to different change types, avoiding processing all updates as full tasks.

[0167] Step S3: Incremental reconstruction of knowledge slices.

[0168] For affected documents, the system constructs the affected area based on the changed location and determines whether to retain or reconstruct the old slices. For policy documents, slices are prioritized by chapter and clause boundaries; for explanatory paragraphs of regulations, slices are prioritized by topic paragraphs; and for scanned text, slices are combined with page numbers and paragraph positions. Old slices with low overlap and unchanged title paths are retained, while slices that overlap with or are strongly related to the affected area are reconstructed, and a mapping relationship between the old and new slices is established. Implementation results show that in 10 partially revised policy documents, only slices related to the changed area were reconstructed, while most unchanged slices were directly retained, significantly reducing the scope of index replacement.

[0169] Step S4: Query and understand the linked updates.

[0170] The system extracts new business terms, abbreviation-full name relationships, and synonyms from the reconstructed slices and updates the business dictionary. Based on this, it automatically generates query preprocessing rules such as abbreviation expansion, terminology standardization, and colloquialism. For new business topics appearing in newly added policy documents, the system further expands intent configurations based on a small number of representative query examples. Implementation results show that the system can quickly update the dictionary and query rule layers for new terms and expressions appearing in incremental samples, ensuring that the new knowledge can be accurately matched by user queries.

[0171] Step S5: Verify and activate the linked update package.

[0172] The system encapsulates slices, dictionaries, query rules, intent configurations, retrieval configurations, reordering configurations, and answer templates into a unified update package, and performs consistency checks and question-and-answer checks on each. Question-and-answer checks use 20 anonymized typical queries for verification, focusing on whether the target slice ranks in the top 5, whether the answer references the latest version's clauses, and whether there are any remnants of the old version. If the check passes, the system switches the entire update package to the valid version; if the check fails, the old version continues to be used. Implementation results show that after new regulations and revised procedures are added to the knowledge base, the system can update its intelligent question-and-answer capabilities without rebuilding all historical knowledge, while maintaining the stability and traceability of answer output.

[0173] As can be seen from the above implementation examples, the present invention can complete multi-object linkage dynamic updates along the knowledge slicing, query understanding, retrieval sorting and answer output links when power knowledge base documents are added, replaced and partially revised, so that knowledge changes can be transformed into new intelligent question answering capabilities in a timely manner, verifying the feasibility and technical effect of the present invention in the intelligent question answering scenario of power knowledge base.

[0174] Although the present invention has been described above with reference to embodiments, various modifications can be made and components can be replaced with equivalents without departing from the scope of the invention. In particular, as long as there is no structural conflict, the features in the disclosed embodiments can be combined with each other in any manner. The lack of an exhaustive description of these combinations in this specification is merely for the sake of brevity and resource conservation. Therefore, the present invention is not limited to the specific embodiments disclosed herein, but includes all technical solutions falling within the scope of the claims.

Claims

1. A multi-object linkage update method for electricity question answering based on intelligent algorithms, characterized in that, Includes the following steps: S1, Power Knowledge Base Update Event Recognition: Transforms the original knowledge changes in the power knowledge base into standard events that can participate in the dynamic update of intelligent question answering; S2, Multi-object linkage impact determination for intelligent question answering: Map the update event identified by S1 to the multi-object linkage update task in the intelligent question answering link of the power knowledge base; S3, Incremental Reconstruction of Electricity Knowledge Slices for Intelligent Question Answering: The affected knowledge content in the electricity knowledge base is re-sliced ​​into knowledge slices suitable for intelligent question answering retrieval and evidence location, and incremental index replacement is completed; S4, Query understanding linkage update for intelligent question answering: to keep the intelligent question answering system's ability to understand user queries consistent with the latest knowledge status in the power knowledge base; S5, Verification and Activation of Linked Update Package for Intelligent Question Answering Capabilities: The update results formed by S3 and S4 are encapsulated into a unified linked update package that serves intelligent question answering capabilities, and the entire package takes effect after verification.

2. The method for multi-object linkage update of power-related question answering based on intelligent algorithms according to claim 1, characterized in that, The specific steps of S1 are as follows: S11, Establish basic document records: Generate a unique document number for each document entering the update channel, and write the document name, document type, business domain, version number and status information to form basic document records; S12, Perform structural alignment between old and new versions: Perform structural comparison between old and new versions under the same document number to achieve structural alignment; Once the alignment is complete, a structure record is created; S13, Identify update type: After the structure alignment is completed, determine whether the change is an addition, replacement, partial revision or failure marker; S14, Generate difference record: Perform content difference recognition on the aligned corresponding structural units to form a difference record; S15, Calculate the basic change measures: calculate the degree of structural change, the degree of terminological change, and the degree of rule change respectively; S16, Output standardized update events: Write the document base record, structure record, update event record, and difference record into the update event set.

3. The method for multi-object linkage update of power-related question answering based on intelligent algorithms according to claim 1, characterized in that, The specific steps of S2 are as follows: S21, Construct an object impact mapping table: Pre-establish a mapping table between update type, change nature and object impact relationship; S22, Calculate the update impact value: Let the structural change degree be S, the terminology change degree be T, and the rule change degree be R, then the update impact value I is calculated as follows: I = 0.35×S + 0.30×T + 0.35×R Among them, S, T, and R are output by S1; S23, Impact levels are classified according to threshold: When I < 0.20, the event is marked as a low-impact update; When 0.20 ≤ I < 0.45, the event is marked as a low to medium impact update; When 0.45 ≤ I < 0.70, the event is marked as a medium-to-high impact update; When I ≥ 0.70, the event is marked as a high-impact update; S24, Generate object propagation results: For low-impact updates, only index correction tasks are generated; For updates with low to medium impact, generate slice reconstruction tasks, slice index update tasks, and business dictionary update tasks; For medium-to-high impact updates, based on the aforementioned tasks, further query preprocessing rule update tasks and intent recognition configuration update tasks are generated; For high-impact updates, further tasks such as search configuration update tasks, reordering configuration update tasks, and answer template and compliance rule update tasks are generated. S25, Generate object-level update task records: Generate a task record for each affected object; S26, Establish task dependency chain: Link object-level tasks into a directed dependency chain.

4. The method for multi-object linkage update of power-related question answering based on intelligent algorithms according to claim 1, characterized in that, The specific steps of S3 are as follows: S31, Determine the affected area: Based on the change location and title path in the difference record, construct the affected area with the change location as the center; S32, Determine whether to retain or reconstruct old slices: For existing old slices, determine their relationship with the affected areas in turn; S33, Select slicing strategy: For documents containing rules and regulations, slice them according to chapters and clause boundaries; For proposals and reports, slice them into thematic paragraphs; The scanned document's text is parsed, and sliced ​​using page numbers and paragraph boundaries; S34, Perform slice merging or splitting: If the length of a slice is lower than the lower limit of its type and the similarity of its topic with adjacent slices is greater than a preset value, then perform merging; if the length of a slice is higher than the upper limit of its type, or the information density value is greater than a preset value, then perform splitting. S35, Generate new slice record: Create a slice record for each reconstructed slice; S36, Establish the mapping relationship between old and new slices: For new slices evolved from old slices, establish a mapping record between old and new slices; S37, Perform incremental index replacement: Generate new semantic indexes and metadata indexes for the reconstructed slices, and retain the original indexes for the preserved slices; then, based on the mapping relationship between the new and old slices, only replace the index entries corresponding to the affected slices.

5. The method for multi-object linkage update of power-related question answering based on intelligent algorithms according to claim 1, characterized in that, The specific steps of S4 are as follows: S41, Extract term candidates: Perform word segmentation and phrase recognition on the reconstructed slice text to extract new business terms, abbreviations and full names correspondences, synonym clusters and regular phrases; Simultaneously, a term record is created for each candidate term; S42, Filter term increments: If a candidate term appears at a frequency of no less than a preset number of times in the same business domain and its length is within a preset number of Chinese characters, then it will be included in the term increment set. If the edit distance between a candidate term and an existing term is less than a preset value, and the context similarity is greater than a preset value, then it is merged into a synonym of the existing term. If an abbreviation and its full name appear together a predetermined number of times in the same paragraph or two adjacent paragraphs, a mapping relationship between the abbreviation and the full name is established. S43, Generate query preprocessing rules: Generate query preprocessing rules based on the term increment set; S44 defines the rule execution order: after a user query enters, noise cleaning is performed first, followed by abbreviation expansion, then terminology unification and synonym merging, and finally context completion; if the same query hits multiple rules at the same time, they are executed in order of priority. S45, Update Intent Configuration: For the newly added question-and-answer scenario, create multiple representative query examples for each intent and form an intent configuration record; S46, Execution Intent Classification: For the normalized query, first filter candidate intents according to the triggering rules, then calculate the average similarity and maximum similarity between it and each intent sample set to achieve intent classification; S47, Output query understanding update results: Write the term increment set, query preprocessing rule increment set, and intent configuration increment set into the update results of this round.

6. The method for multi-object linkage update of power-related question answering based on intelligent algorithms according to claim 1, characterized in that, The specific steps of S5 are as follows: S51, Construct a linked update package: Encapsulate the document version, slice version, dictionary version, query rule version, intent configuration version, retrieval configuration version, rearrangement configuration version, answer template version, verification status, and effective status involved in this round of updates into a linked update package; S52, Update Search Configuration: Configure the weights of keyword search and semantic search channels based on intent type and business scenario; S53, Perform multi-channel recall and fusion screening: Use dual-channel recall of candidate slices, normalize the results and calculate the fusion score; S54, Perform rearrangement calculation: Calculate the rearrangement score for the candidate slices; then sort them from high to low according to the rearrangement score; S55, Match answer templates and generate evidence chains: Match answer templates based on intent type; simultaneously, retain evidence records when generating answers; S56, Perform consistency check: Check whether the slices, dictionaries, query rules, retrieval configurations, rearrangement configurations and answer templates in the linked update package all exist and whether their version identifiers are consistent; S57, Perform question-and-answer validation: Use multiple regression test queries for validation; S58, perform atomic activation or rollback: If both consistency check and question-and-answer check pass, the entire linked update package is marked as valid, and the slice, dictionary, query rules, intent configuration, retrieval configuration, rearrangement configuration, and answer template are switched at once; if any check fails, the previous valid update package is retained and continues to serve.

Images