Intelligent response method and system for station management operation based on knowledge graph question answering

Abstract

The embodiment of the application provides a station management operation intelligent response method and system based on a knowledge graph question and answer, a three-layer semantic knowledge graph is constructed by performing entity recognition and relationship extraction on an unstructured document to generate a station management knowledge base, a candidate knowledge subgraph is generated by performing permission filtering and graph traversal retrieval based on a query vector and intent classification, and a question and answer response text is output by conversion of an answer generation model, effectively solving the deficiencies of traditional technologies in knowledge graph construction, semantic retrieval and intelligent question and answer, and providing technical support for intelligent response of station management operation.

Description

Technical Field

[0001] This application relates to the field of data processing, specifically to an intelligent response method and system for station management operations based on knowledge graph question answering. Background Technology

[0002] In the field of station management, traditional knowledge management mainly relies on manual review of large amounts of unstructured documents. For example, the power grid operation and maintenance knowledge voice question-and-answer system proposed in CN113836280A (publication date: December 2021), although it achieves intelligent retrieval of business data through knowledge graph technology, its entity recognition model designed for the power industry is difficult to directly apply to station management scenarios. This patent adopts a technical approach based on speech recognition and natural language processing to convert unstructured documents into structured knowledge graphs, but it does not fully consider the nested reference relationships of rules and regulations and the temporal constraints of operational procedures unique to the field of station management, resulting in insufficient semantic integrity.

[0003] In addition, the multi-strategy dialogue method for government services disclosed in CN116628172A (publication date: August 2023) improves the accuracy of question and answer through a multi-stage recall strategy, but its construction method based on a general government knowledge graph is difficult to adapt to the frequently updated temporary regulations and regionally differentiated clauses in station management.

[0004] Existing systems generally have three limitations: First, the accuracy of identifying station administration terminology is insufficient during the knowledge extraction stage, such as misidentifying "interlock release" as a common verb phrase; Second, the semantic association lacks domain-specific modeling, making it impossible to accurately express the conditional triggering relationship between "emergency plan" and "equipment operating procedures"; Third, the question-and-answer response did not take into account the differences in permissions among station staff, resulting in redundant returned information or permissions exceeding the scope.

[0005] These problems mainly stem from the fact that existing technologies have not designed dedicated entity relationship models and access control mechanisms for the field of station administration, making it difficult for the system to balance response speed and answer accuracy in practical applications. Summary of the Invention

[0006] To address the problems in existing technologies, this application provides an intelligent response method and system for station management operations based on knowledge graph question answering. This method effectively solves the shortcomings of traditional technologies in knowledge graph construction, semantic retrieval, and intelligent question answering, providing technical support for intelligent response to station management operations.

[0007] To solve at least one of the above problems, this application provides the following technical solution: Firstly, this application provides an intelligent response method for station management operations based on knowledge graph question answering, including: The system acquires unstructured documents for station administration and performs format conversion and encoding standardization to generate preprocessed text. The preprocessed text is then segmented and lexical labeled to generate a labeled text sequence. The labeled text sequence is input into a pre-trained entity recognition model to perform sequence labeling to generate an entity label sequence. The entity label sequence is then subjected to co-occurrence analysis and dependency resolution based on domain relationship rules to generate a set of entity relationship triples. Based on the entity relationship triple set, a three-layer semantic knowledge graph containing a concept layer, an instance layer, and a relationship layer is constructed, and condition triggering relationships and temporal constraint relationships are established. The three-layer semantic knowledge graph is stored in a graph database to generate a station management knowledge base. Multi-level indexes are established on the node attributes and relationship types of the station management knowledge base to generate a knowledge retrieval index. The system receives a natural language query input from the user and performs vector encoding based on the dialogue context to generate a query vector. The query vector is then input into an intent classification model to perform category determination and generate a query intent label. Based on the query intent label and user role permission parameters, the system performs permission filtering and graph traversal retrieval on the knowledge retrieval index to generate a candidate knowledge subgraph. The candidate knowledge subgraph is then input into an answer generation model to perform natural language conversion to generate a question-and-answer response text, which is then output to the user interaction terminal.

[0008] Furthermore, it also includes: reading unstructured documents for station administration and identifying document format types, performing format parsing to extract text content from the unstructured documents for station administration to generate raw text data, and performing irrelevant content filtering and character encoding unification processing on the raw text data to generate preprocessed text; The preprocessed text is segmented into sentences according to preset sentence boundary rules to generate a sentence sequence. Each sentence in the sentence sequence is then processed by word segmentation and part-of-speech tagging to generate a tagged text sequence.

[0009] Furthermore, it also includes: converting the labeled text sequence into a model input vector and inputting it into a pre-trained entity recognition model; performing sequence labeling inference on the model input vector to generate entity category labels for each word; and performing boundary recognition and entity merging based on the entity category labels to generate an entity labeled sequence. For each entity in the entity annotation sequence, perform intra-sentence co-occurrence statistics to generate an entity co-occurrence matrix. Perform dependency parsing on the entity co-occurrence matrix and the syntactic structure of the annotated text sequence to generate entity dependency pairs. Perform relation type mapping on the entity dependency pairs according to domain relation rules to generate a set of entity relation triples.

[0010] Furthermore, it also includes: reading the entity relationship triple set and performing hierarchical classification on each entity according to the preset entity type classification rules, writing entities classified as concept type into the concept layer and establishing concept nodes, writing entities classified as instance type into the instance layer and establishing instance nodes, and establishing association edges between concept nodes and instance nodes based on the relationship types in the entity relationship triple set to generate a three-layer semantic knowledge graph. Traverse the associated edges in the three-layer semantic knowledge graph and identify relationships with precondition attributes. Mark the relationships with precondition attributes with condition triggering identifiers to generate condition triggering relationships. Mark the relationships with execution order attributes with time constraint identifiers to generate time constraint relationships. Write the condition triggering relationships and the time constraint relationships into the relationship layer of the three-layer semantic knowledge graph.

[0011] Furthermore, it also includes: serializing the concept nodes and instance nodes of the three-layer semantic knowledge graph into graph database node records, serializing the association edges, condition triggering relationships, and temporal constraint relationships of the three-layer semantic knowledge graph into graph database edge records, and writing the graph database node records and graph database edge records into the graph database to generate a station management knowledge base; The system reads the attribute fields of each node in the station management knowledge base and establishes an attribute index according to the preset attribute priority rules. It also reads the relationship types of each edge in the station management knowledge base and establishes a relationship type index. Finally, it organizes the attribute index and the relationship type index in a hierarchical structure to generate a knowledge retrieval index.

[0012] Furthermore, it also includes: receiving a natural language query input by the user and reading the dialogue context record of the current session, performing concatenation processing on the natural language query and the dialogue context record to generate context fusion text, and inputting the context fusion text into a pre-trained semantic encoding model to perform vector mapping to generate a query vector; The query vector is input into a pre-trained intent classification model to perform multi-class probability prediction and generate an intent probability distribution. The intent probability distribution is then filtered by category based on a preset confidence threshold to generate query intent labels.

[0013] Furthermore, it also includes: reading the query intent tag and matching the corresponding search path in the knowledge retrieval index, reading the user role permission parameters and performing access permission determination on the nodes involved in the search path according to the preset permission filtering rules to generate an accessible node set, and performing graph traversal retrieval on the station management knowledge base based on the accessible node set to generate candidate knowledge subgraphs; The node attributes and relational structure of the candidate knowledge subgraph are serialized into a structured knowledge representation. The structured knowledge representation and the query vector are input into a pre-trained answer generation model to perform sequence decoding to generate question-and-answer response text. The question-and-answer response text is encapsulated into a response message and output to the user interaction terminal.

[0014] Secondly, this application provides an intelligent response system for station administration operations based on knowledge graph question answering, including: The entity relation parsing module is used to acquire unstructured documents of station administration and perform format conversion and encoding standardization to generate preprocessed text. The preprocessed text is then segmented and lexical labeled to generate a labeled text sequence. The labeled text sequence is input into a pre-trained entity recognition model to perform sequence labeling to generate an entity labeled sequence. The entity labeled sequence is then subjected to co-occurrence analysis and dependency parsing based on domain relation rules to generate a set of entity relation triples. The knowledge retrieval index module is used to construct a three-layer semantic knowledge graph containing a concept layer, an instance layer, and a relation layer based on the entity relation triple set, and to establish condition triggering relations and temporal constraint relations. The three-layer semantic knowledge graph is stored in a graph database to generate a station administration knowledge base. The node attributes and relation types of the station administration knowledge base are used to establish a multi-level index to generate a knowledge retrieval index. The question-and-answer response module is used to receive natural language queries input by users and perform vector encoding to generate query vectors in combination with the dialogue context. The query vectors are then input into an intent classification model to perform category determination and generate query intent tags. Based on the query intent tags and user role permission parameters, permission filtering and graph traversal retrieval are performed on the knowledge retrieval index to generate candidate knowledge subgraphs. The candidate knowledge subgraphs are then input into an answer generation model to perform natural language conversion to generate question-and-answer response text and output it to the user interaction terminal.

[0015] Thirdly, this application provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps of the intelligent response method for station management operations based on knowledge graph question answering.

[0016] Fourthly, this application provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the intelligent response method for station management operations based on knowledge graph question answering.

[0017] Fifthly, this application provides a computer program product, including a computer program / instruction, which, when executed by a processor, implements the steps of the intelligent response method for station management operations based on knowledge graph question answering.

[0018] As can be seen from the above technical solution, this application provides an intelligent response method and system for station management operations based on knowledge graph question answering. By performing entity recognition and relationship extraction on unstructured documents, a three-layer semantic knowledge graph is constructed to generate a station management knowledge base. Based on query vectors and intent classification, permission filtering and graph traversal retrieval are performed to generate candidate knowledge subgraphs. The question-and-answer response text is then converted and output through an answer generation model. This effectively solves the shortcomings of traditional technologies in knowledge graph construction, semantic retrieval, and intelligent question answering, and provides technical support for intelligent response to station management operations. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below.

[0020] Figure 1 This is a flowchart illustrating the intelligent response method for station management operations based on knowledge graph question answering in the embodiments of this application; Figure 2 This is a flowchart illustrating a specific embodiment of the intelligent response method for station management operations based on knowledge graph question answering in this application. Detailed Implementation

[0021] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are some embodiments of this application, but not all embodiments.

[0022] In view of the problems existing in the prior art, this application provides an intelligent response method and system for station management operations based on knowledge graph question answering. By performing entity recognition and relationship extraction on unstructured documents, a three-layer semantic knowledge graph is constructed to generate a station management knowledge base. Based on query vectors and intent classification, permission filtering and graph traversal retrieval are performed to generate candidate knowledge subgraphs. The question answer response text is then converted and output through an answer generation model. This effectively solves the shortcomings of traditional technologies in knowledge graph construction, semantic retrieval and intelligent question answering, and provides technical support for intelligent response of station management operations.

[0023] To effectively address the shortcomings of traditional technologies in knowledge graph construction, semantic retrieval, and intelligent question answering, and to provide technical support for intelligent response in station administration operations, this application provides an embodiment of an intelligent response method for station administration operations based on knowledge graph question answering. See [link to embodiment]. Figure 1 and Figure 2 The intelligent response method for station management operations based on knowledge graph question answering specifically includes the following: Step S101: Obtain unstructured station management documents and perform format conversion and encoding standardization to generate preprocessed text. Perform sentence segmentation and lexical annotation on the preprocessed text to generate an annotated text sequence. Input the annotated text sequence into a pre-trained entity recognition model to perform sequence annotation to generate an entity annotation sequence. Perform co-occurrence analysis and dependency parsing on the entity annotation sequence according to domain relationship rules to generate a set of entity relationship triples. This embodiment reads unstructured station administration documents from the station administration document storage system as the starting input for the knowledge extraction process. During the reading process, data is located according to document identifier and batch number. If the document format is not supported, a format anomaly flag is registered, and the corresponding document is excluded from subsequent processing.

[0024] In one implementation, the unstructured documents for station management are processed to identify and parse formats to extract text content and generate raw text data. The format identification process determines the document format type, including PDF, Word, and plain text. The format parsing process uses the corresponding document parsing tool to extract the text content. The raw text data is then filtered for irrelevant content and undergoes unified character encoding processing to generate preprocessed text. Filtering includes headers, footers, page number watermarks, and format control characters. The encoding is uniformly converted to UTF-8 standard encoding. The preprocessed text includes a document source identifier (document_id) and a processing timestamp (process_ts).

[0025] Preferably, the preprocessed text is segmented into a sentence sequence according to preset sentence boundary rules, which are determined based on the combination pattern of punctuation marks and line breaks. Each sentence in the sentence sequence is then processed through word segmentation and part-of-speech tagging to generate an annotated text sequence. The word segmentation uses a domain dictionary-enhanced algorithm to improve the accuracy of segmenting website administration terminology. The annotated text sequence includes a text content field, a part-of-speech tag field, and a sentence attribution index field for each word.

[0026] Optionally, the annotated text sequence is converted into a model input vector and input into a pre-trained entity recognition model to perform sequence labeling inference and generate entity category labels for each word. The entity recognition model adopts a BERT-BiLSTM-CRF architecture and is pre-trained on annotated corpus in the station administration domain. Entity categories include the names of rules and regulations, operating procedures, responsible departments, and equipment names. Based on the entity category labels, boundary recognition and entity merging are performed to generate entity annotation sequences. The entity annotation sequences include a text span field, a category label field, and a confidence score field for each entity.

[0027] Therefore, an entity co-occurrence matrix is ​​generated by performing intra-sentence co-occurrence statistics on each entity in the entity annotation sequence. The entity co-occurrence matrix records the co-occurrence frequency of entity pairs within the same sentence. Dependency parsing is then performed on the entity co-occurrence matrix and the syntactic structure of the annotated text sequence to generate entity dependency pairs. These entity dependency pairs are then mapped according to domain relation rules to generate a set of entity relation triples. Domain relation rules define matching patterns for reference relations, containment relations, precondition relations, and execution subject relations. The set of entity relation triples includes a head_entity field, a relation_type field, and a tail_entity field, using a combination of the triple sequence number and document identifier as the primary key. This key is used by step S102 at the knowledge graph construction entry point to read nodes and edges for creating the three-layer semantic knowledge graph.

[0028] Step S102: Based on the entity relationship triple set, construct a three-layer semantic knowledge graph containing a concept layer, an instance layer, and a relationship layer, and establish condition triggering relationships and temporal constraint relationships. Store the three-layer semantic knowledge graph in a graph database to generate a station management knowledge base. Build a multi-level index on the node attributes and relationship types of the station management knowledge base to generate a knowledge retrieval index. In this embodiment, entity records and relation records are read from the entity-relation triplet set output in step S101 as the starting input for the knowledge graph construction process. The reading process locates data based on triplet numbers and document identifiers. If an entity category label is missing, a data anomaly flag is registered, and the corresponding triplet is excluded from participating in the graph construction.

[0029] In one implementation, the entity relation triplet set is read, and hierarchical classification is performed on each entity according to a preset entity type classification rule. The hierarchical classification process classifies regulations, operational procedures, and emergency response categories as conceptual entities, and specific regulation names, specific operational steps, and specific responsible departments as instance entities. Entities classified as conceptual are written into the conceptual layer and conceptual nodes are created; entities classified as instance are written into the instance layer and instance nodes are created. Based on the relation types in the entity relation triplet set, association edges are established between conceptual nodes and instance nodes to generate a three-layer semantic knowledge graph. The three-layer semantic knowledge graph includes a conceptual layer node set, an instance layer node set, and an association edge set.

[0030] Preferably, the associated edges in the three-layer semantic knowledge graph are traversed, and relationships with precondition attributes are identified to generate condition-triggered relationships. The identification process filters associated edges containing precondition semantics based on the relationship type field, and labels the relationships with precondition attributes with a condition trigger flag (trigger_flag). Relationships with execution order attributes are labeled with a sequence constraint flag (sequence_flag) to generate sequence constraint relationships, which represent the sequential execution order of operation steps. The condition-triggered relationships and the sequence constraint relationships are written into the relationship layer of the three-layer semantic knowledge graph, which includes a relationship type field, a condition trigger flag field, and a sequence constraint flag field.

[0031] Optionally, the concept nodes and instance nodes of the three-layer semantic knowledge graph are serialized into graph database node records, and the associated edges, conditional triggering relationships, and temporal constraint relationships of the three-layer semantic knowledge graph are serialized into graph database edge records. The graph database node records and graph database edge records are written into a graph database to generate a station management knowledge base. The graph database uses a database system that supports attribute graph models. The station management knowledge base includes a node table, an edge table, and an attribute table, with each table using the node identifier (node_id) or edge identifier (edge_id) as the primary key.

[0032] Therefore, the attribute fields of each node in the station management knowledge base are read and attribute indexes are established according to preset attribute priority rules. The attribute priority rules set the rule name attribute, department name attribute, and operation step attribute as high-priority index fields. The relationship types of each edge in the station management knowledge base are read and relationship type indexes are established, covering reference relationships, containment relationships, conditional trigger relationships, and temporal constraint relationships. The attribute indexes and relationship type indexes are organized hierarchically to generate a knowledge retrieval index. The knowledge retrieval index includes an attribute index layer, a relationship index layer, and a composite index layer, using a combination of index identifier and creation timestamp as the primary key, which is used by subsequent steps at the graph traversal retrieval entry point for query path matching and candidate knowledge subgraph generation.

[0033] Step S103: Receive the natural language query input by the user and perform vector encoding to generate a query vector in combination with the dialogue context. Input the query vector into the intent classification model to perform category determination and generate query intent labels. Based on the query intent labels and user role permission parameters, perform permission filtering and graph traversal retrieval on the knowledge retrieval index to generate candidate knowledge subgraphs. Input the candidate knowledge subgraphs into the answer generation model to perform natural language conversion to generate question-and-answer response text and output it to the user interaction terminal.

[0034] In this embodiment, the knowledge retrieval index and the site management knowledge base output in step S102 are used as the starting input for the intelligent question-and-answer process to read the index structure and knowledge nodes. The reading process locates data according to the index identifier and node identifier. If the index structure is damaged, an index anomaly identifier is registered and the index reconstruction process is triggered.

[0035] In one implementation, a natural language query input by the user is received, and the dialogue context record of the current session is read. The natural language query and the dialogue context record are concatenated to generate context-fused text, with the concatenation process linking historical dialogue rounds to the current query in chronological order. The context-fused text is then input into a pre-trained semantic encoding model to generate a query vector through vector mapping. The semantic encoding model employs a pre-trained Chinese language model architecture. The query vector includes a vector dimension field and a vector value field, with the vector dimension determined based on the hidden layer dimension of the semantic encoding model.

[0036] Preferably, the query vector is input into a pre-trained intent classification model to perform multi-class probability prediction and generate an intent probability distribution. The intent classification model is trained on a question-and-answer corpus in the station administration domain, and the intent categories include regulation queries, operational guidance, equipment malfunctions, and emergency handling. The intent probability distribution is then filtered according to a preset confidence threshold to generate query intent tags. The confidence threshold is determined based on the balance between precision and recall on the validation set; if the value is below the threshold, a low-confidence flag is registered, and a manual review process is triggered. The query intent tag includes an intent category field and a confidence score field.

[0037] Optionally, the query intent tag is read and matched with the corresponding search path in the knowledge retrieval index. User role permission parameters are read, and access permission determination is performed on the nodes involved in the search path according to preset permission filtering rules to generate an accessible node set. The permission filtering rules are determined based on the mapping relationship between user roles and node access levels. Based on the accessible node set, a graph traversal search is performed on the site management knowledge base to generate candidate knowledge subgraphs. The graph traversal search adopts a breadth-first strategy and sets a maximum traversal depth parameter. The candidate knowledge subgraph includes a subgraph node set field, a subgraph edge set field, and a traversal path field.

[0038] Therefore, the node attributes and relational structure of the candidate knowledge subgraph are serialized into a structured knowledge representation. The serialization process converts the node attributes and relational types into a text sequence format. The structured knowledge representation and the query vector are input into a pre-trained answer generation model to perform sequence decoding and generate question-and-answer response text. The answer generation model employs a sequence-to-sequence architecture based on an attention mechanism. The question-and-answer response text is encapsulated into a response message and output to the user interaction terminal. The response message includes a response text field, a source citation field, and a response timestamp field, using a combination of the session identifier (session_id) and the message sequence number as the primary key. This key is read by the user feedback module at the feedback collection entry point for question-and-answer quality assessment and model optimization.

[0039] As can be seen from the above description, the intelligent response method for station management operations based on knowledge graph question answering provided in this application can generate a station management knowledge base by constructing a three-layer semantic knowledge graph through entity recognition and relationship extraction of unstructured documents, performing permission filtering and graph traversal retrieval based on query vectors and intent classification to generate candidate knowledge subgraphs, and outputting question answer response text through an answer generation model. This effectively solves the shortcomings of traditional technologies in knowledge graph construction, semantic retrieval, and intelligent question answering, and provides technical support for intelligent response to station management operations.

[0040] In one embodiment of the intelligent response method for station management operations based on knowledge graph question answering in this application, it may further include the following: Step S201: Read the unstructured station management document and identify the document format type. Perform format parsing on the unstructured station management document to extract text content and generate raw text data. Perform irrelevant content filtering and character encoding unification processing on the raw text data to generate preprocessed text. Step S202: The preprocessed text is segmented into sentences according to preset sentence boundary rules to generate a sentence sequence. Each sentence in the sentence sequence is then processed by word segmentation and part-of-speech tagging to generate a tagged text sequence.

[0041] This embodiment reads unstructured station administration documents from the station administration document storage system as the starting input for the document preprocessing workflow. During the reading process, data is located according to the document identifier and storage path. If a document file is corrupted or unreadable, a reading error flag is registered, and the corresponding document is excluded from subsequent processing.

[0042] In one implementation, the unstructured documents for station management are processed to generate document format tags based on format type recognition. The format type recognition process determines the format type based on the document file extension and header signature. Document format types include PDF, Word, plain text, and scanned image formats. The document format tag includes a format type field and a parser mapping field. If a format type is not supported, a format not supported flag is registered, and a manual conversion process is triggered.

[0043] Preferably, based on the document format tag, the corresponding format parser is invoked to perform format parsing on the unstructured station management document to extract text content and generate raw text data. PDF format documents are parsed using a text extraction tool, Word format documents are parsed using a Document Object Model (DOM) interface, and scanned image format documents are parsed using an Optical Character Recognition (OCR) engine. The raw text data includes text content fields, paragraph structure fields, and page number mapping fields. If parsing fails, a parsing error flag is registered and the error location information is recorded.

[0044] Optionally, irrelevant content filtering is performed on the original text data to generate filtered text data. The filtering process identifies and removes headers, footers, page number watermarks, table of contents indexes, and format control characters according to preset filtering rules. The filtering rules employ a combination of regular expression patterns and positional features for matching. Character encoding unification processing is then performed on the filtered text data to generate preprocessed text. This unification processing converts the source encoding format to UTF-8 standard encoding after detection. Characters that fail to convert are marked with replacement characters, and their encoding anomalies are recorded. The preprocessed text includes a text content field, a document source identifier (document_id) field, and a processing timestamp (process_ts) field.

[0045] Therefore, the preprocessed text is segmented into sentence sequences according to preset sentence boundary rules. Sentence boundary rules are determined based on the combination pattern of sentence-end punctuation, line breaks, and paragraph markers. Numbered lists and table content in special scenarios are processed using independent sentence segmentation strategies. Each sentence in the sentence sequence undergoes word segmentation and part-of-speech tagging to generate an annotated text sequence. Word segmentation employs a domain dictionary-enhanced algorithm to improve the accuracy of segmenting website administration terminology, and part-of-speech tagging uses a pre-trained part-of-speech tagging model for inference. The annotated text sequence includes a word element text field, a part-of-speech tag field, a sentence attribution index field, and a word element position offset field, using a combination of document identifier and sentence sequence number as the primary key. This key is read at the entity recognition entry point in step S301 for sequence annotation and entity extraction.

[0046] In one embodiment of the intelligent response method for station management operations based on knowledge graph question answering in this application, it may further include the following: Step S301: Convert the labeled text sequence into a model input vector and input it into a pre-trained entity recognition model. Perform sequence labeling inference on the model input vector to generate entity category labels for each word. Based on the entity category labels, perform boundary recognition and entity merging to generate an entity labeling sequence. Step S302: Perform intra-sentence co-occurrence statistics on each entity in the entity annotation sequence to generate an entity co-occurrence matrix. Perform dependency parsing on the entity co-occurrence matrix and the syntactic structure of the annotation text sequence to generate entity dependency pairs. Perform relation type mapping on the entity dependency pairs according to domain relation rules to generate a set of entity relation triples.

[0047] In this embodiment, the text content and part-of-speech tags of each word are read from the annotated text sequence output in step S202 as the starting input for the entity recognition process. During the reading process, data is located according to document identifier and sentence number. When a word text field is empty, a data missing identifier is registered, and the corresponding sentence is excluded from entity recognition.

[0048] In one implementation, the labeled text sequence is converted into a model input vector. The conversion process performs word embedding mapping and positional encoding on each word to generate a word vector representation. The word embedding mapping uses an index lookup from a pre-trained language model's vocabulary; words outside the vocabulary are processed using a sub-word segmentation strategy. The model input vector is then input into a pre-trained entity recognition model to perform sequence labeling inference. The entity recognition model employs a BERT-BiLSTM-CRF architecture, where the BERT layer extracts contextual semantic features, the BiLSTM layer captures sequence dependencies, and the CRF layer constrains the legality of label transitions. The sequence labeling inference outputs entity category labels for each word, including the name of regulations / systems, operating procedures, responsible departments, equipment names, and personnel roles.

[0049] Preferably, entity labeling sequence is generated by performing boundary recognition and entity merging based on the entity category labels. The boundary recognition process identifies the starting position label and internal position label of an entity according to the BIO labeling system, merging consecutive sequences of labels of the same category into a complete entity span. The entity merging process merges adjacent and semantically coherent entities of the same category, with merging rules determined based on entity spacing and contextual semantic similarity. The entity labeling sequence includes an entity text span field, an entity category label field, a confidence score field, and a sentence attribution index field. When the confidence score is lower than a preset threshold, a low-confidence flag is registered for manual review.

[0050] Optionally, an entity co-occurrence matrix is ​​generated by performing intra-sentence co-occurrence statistics on each entity in the entity annotation sequence. The co-occurrence statistics process iterates through entity pairs within each sentence and accumulates co-occurrence frequencies. The entity co-occurrence matrix uses entity identifiers as row and column indices and co-occurrence frequencies as matrix element values. Dependency parsing is then performed on the entity co-occurrence matrix and the syntactic structure of the annotated text sequence to generate entity dependency pairs. The dependency parsing process uses a dependency parser to extract syntactic dependency paths between entities. Entity pairs with dependency path lengths exceeding a preset threshold are marked as weakly related. The entity dependency pairs include a head entity field, a tail entity field, and a dependency path field.

[0051] Therefore, the entity dependency pairs are mapped to relation types according to domain relation rules to generate a set of entity relation triples. Domain relation rules define matching patterns for reference relations, containment relations, precondition relations, execution subject relations, and applicable scenario relations. The matching process is triggered by rule combinations based on dependency path features and entity categories. When relation type mapping fails, a mapping failure flag is registered, and the original dependency relations are retained for rule expansion. The set of entity relation triples includes a head_entity field, a relation_type field, a tail_entity field, and a source sentence index field. The triple sequence number and document identifier are used as the primary key, which is read by step S401 at the knowledge graph construction entry point for node creation and edge establishment in the concept layer and instance layer.

[0052] In one embodiment of the intelligent response method for station management operations based on knowledge graph question answering in this application, it may further include the following: Step S401: Read the entity relationship triple set and perform hierarchical classification judgment on each entity according to the preset entity type classification rules. Write the entities that are judged as concept type into the concept layer and establish concept nodes. Write the entities that are judged as instance type into the instance layer and establish instance nodes. Based on the relationship type in the entity relationship triple set, establish association edges between concept nodes and instance nodes to generate a three-layer semantic knowledge graph. Step S402: Traverse the associated edges in the three-layer semantic knowledge graph and identify relationships with precondition attributes. Mark the relationships with precondition attributes with condition triggering identifiers to generate condition triggering relationships. Mark the relationships with execution order attributes with time constraint identifiers to generate time constraint relationships. Write the condition triggering relationships and the time constraint relationships into the relationship layer of the three-layer semantic knowledge graph.

[0053] In this embodiment, entity records and relation records are read from the entity-relation triplet set output in step S302 as the starting input for the knowledge graph construction process. The reading process locates data based on triplet numbers and document identifiers. If an entity field is empty or a relation type field is missing, a data anomaly flag is registered, and the corresponding triplet is excluded from participating in the graph construction.

[0054] In one implementation, the entity relationship triplet set is read, and hierarchical attribution is determined for each entity according to a preset entity type classification rule. The entity type classification rule classifies highly abstract entity categories as conceptual categories and specifically named entity categories as instance categories. Conceptual entities include regulations, operational procedures, emergency response, and equipment systems; instance entities include specific regulation names, specific operational steps, specific responsible departments, and specific equipment models. The hierarchical attribution determination result includes an entity identifier field and a hierarchical attribution field. When the attribution determination is unclear, an attribution pending marker is registered for manual review.

[0055] Preferably, entities determined to be of the concept class are written into the concept layer and concept nodes are created. Concept nodes include a node identifier field, a concept name field, a concept definition field, and a parent concept field. Entities determined to be of the instance class are written into the instance layer and instance nodes are created. Instance nodes include a node identifier field, an instance name field, an attribute set field, and a belonging concept field. Based on the relation types in the entity relation triplet set, association edges are established between concept nodes and instance nodes to generate a three-layer semantic knowledge graph. Association edges include an edge identifier field, a start node field, a stop node field, and a relation type field. The three-layer semantic knowledge graph includes a concept layer node set, an instance layer node set, and a relation layer edge set.

[0056] Optionally, the association edges in the three-layer semantic knowledge graph are traversed and relationships with precondition attributes are identified. The identification process filters association edges containing precondition semantics based on the relationship type field. The determination of the precondition attribute is based on the relationship type: precondition relationship, trigger relationship, or dependency relationship. Relationships with precondition attributes are labeled with a condition trigger flag (trigger_flag) to generate condition trigger relationships. Condition trigger relationships represent the conditional dependency between emergency plans and equipment operating procedures. Relationships with execution order attributes are labeled with a sequence constraint flag (sequence_flag) to generate sequence constraint relationships. Sequence constraint relationships represent the sequential execution order between operation steps. The determination of the execution order attribute is based on the relationship type: successor relationship, preorder relationship, or sequential relationship.

[0057] Therefore, the conditional triggering relationship and the temporal constraint relationship are written into the relation layer of the three-layer semantic knowledge graph. During the writing process, a conditional triggering identifier field and a temporal constraint identifier field are added to the corresponding associated edges in the relation layer edge set. Associated edges without identified special attributes retain their original relation type. The relation layer includes a relation type field, a conditional triggering identifier field, a temporal constraint identifier field, and a constraint parameter field. A combination of the edge identifier and the graph version number is used as the primary key, which is read by step S501 at the knowledge base storage entry point for serialization writing of graph database node records and edge records.

[0058] In one embodiment of the intelligent response method for station management operations based on knowledge graph question answering in this application, it may further include the following: Step S501: Serialize the concept nodes and instance nodes of the three-layer semantic knowledge graph into graph database node records, serialize the association edges, condition triggering relationships, and temporal constraint relationships of the three-layer semantic knowledge graph into graph database edge records, and write the graph database node records and graph database edge records into the graph database to generate the station management knowledge base; Step S502: Read the attribute fields of each node in the station management knowledge base and establish an attribute index according to the preset attribute priority rules; read the relationship types of each edge in the station management knowledge base and establish a relationship type index; organize the attribute index and the relationship type index in a hierarchical structure to generate a knowledge retrieval index.

[0059] In this embodiment, the concept layer node set, instance layer node set, and relation layer edge set are read from the three-layer semantic knowledge graph output in step S402 as the starting input for the knowledge base storage process. The reading process locates data based on node and edge identifiers. If a node's data structure is incomplete or an edge's associated node is missing, a data integrity anomaly is registered, and a data repair process is triggered.

[0060] In one implementation, the concept nodes and instance nodes of the three-layer semantic knowledge graph are serialized into graph database node records. The serialization process converts the concept name field, concept definition field, and parent concept field of the concept node into graph database node attribute format, and converts the instance name field, attribute set field, and belonging concept field of the instance node into graph database node attribute format. Concept nodes and instance nodes are labeled with node type tags to distinguish their hierarchical affiliation. The graph database node record includes a node identifier field, a node type field, an attribute key-value pair set field, and a creation timestamp field.

[0061] Preferably, the associated edges, conditional triggering relationships, and temporal constraint relationships of the three-layer semantic knowledge graph are serialized into graph database edge records. The serialization process converts the start node field, end node field, and relationship type field of the associated edges into graph database edge attribute format, and appends the conditional triggering flag (trigger_flag) and temporal constraint flag (sequence_flag) as edge attributes. The graph database edge record includes an edge identifier field, a start node identifier field, an end node identifier field, a relationship type field, a conditional triggering flag field, a temporal constraint flag field, and a creation timestamp field.

[0062] Optionally, the graph database node records and graph database edge records are written into the graph database to generate a station management knowledge base. The writing process uses a batch import method to improve writing efficiency, and the graph database uses a database system that supports attribute graph models. Before writing, the uniqueness of node identifiers and edge identifiers is checked. If identifiers conflict, an update overwrite or version append is performed according to preset conflict handling rules. The station management knowledge base includes a node table, an edge table, and an attribute table. If a writing failure occurs, a writing exception identifier is registered and the failure record is recorded for retry processing.

[0063] Therefore, the attribute fields of each node in the station management knowledge base are read and attribute indexes are established according to preset attribute priority rules. The attribute priority rules set the rule name, department name, equipment name, and operation step attributes as high-priority index fields. These high-priority attributes are established using a combination of full-text indexing and exact matching indexing. The relationship types of each edge in the station management knowledge base are read and a relationship type index is established, covering reference relationships, containment relationships, conditional trigger relationships, temporal constraint relationships, and execution subject relationships. The attribute indexes and relationship type indexes are organized hierarchically to generate a knowledge retrieval index. This knowledge retrieval index includes an attribute index layer, a relationship index layer, and a composite index layer, using a combination of index identifier and creation timestamp as the primary key. This key is read by step S601 at the graph traversal retrieval entry point for query path matching and candidate knowledge subgraph generation.

[0064] In one embodiment of the intelligent response method for station management operations based on knowledge graph question answering in this application, it may further include the following: Step S601: Receive the natural language query input by the user and read the dialogue context record of the current session. Perform concatenation processing on the natural language query and the dialogue context record to generate context fusion text. Input the context fusion text into a pre-trained semantic encoding model to perform vector mapping to generate a query vector. Step S602: Input the query vector into the pre-trained intent classification model to perform multi-class probability prediction to generate an intent probability distribution, and perform category filtering on the intent probability distribution according to the preset confidence threshold condition to generate query intent labels.

[0065] This embodiment receives natural language queries input by the user from the user interaction terminal and reads the current session's dialogue context record from the session management module as the starting input for the query comprehension process. The reading process locates data based on the session identifier (session_id) and the dialogue round number. When the dialogue context record is empty, a single-turn dialogue mode is used. When the natural language query is empty or contains only invalid characters, an input error flag is registered and an input prompt message is returned.

[0066] In one implementation, the natural language query and the dialogue context record are concatenated to generate context-fused text. The concatenation process sequentially joins user queries and system responses from previous dialogue rounds with the current query. The number of historical rounds is truncated based on a preset context window length parameter, which is determined by the maximum input length limit of the semantic coding model. Separators are inserted between each dialogue round to distinguish round boundaries. The context-fused text includes a historical dialogue text field, a current query text field, and a round separator field.

[0067] Preferably, the context-fused text is input into a pre-trained semantic encoding model, and vector mapping is performed to generate a query vector. The semantic encoding model adopts a pre-trained Chinese language model architecture and undergoes domain-specific fine-tuning on question-and-answer corpora in the station administration domain. The vector mapping process performs word segmentation and word embedding mapping on the context-fused text, then inputs it into the semantic encoding model, extracting the pooled representation of the last hidden state as the query vector. The query vector includes a vector identifier field, a vector dimension field, and a vector value field. The vector dimension is determined based on the hidden layer dimension of the semantic encoding model. If encoding fails, an encoding error flag is registered, and a default vector is used as a replacement.

[0068] Optionally, the query vector is input into a pre-trained intent classification model to perform multi-class probability prediction and generate an intent probability distribution. The intent classification model employs a multi-layer fully connected network architecture and is trained on intent-annotated corpora in the station administration domain. Intent categories include rule queries, operation guidance, equipment malfunctions, emergency handling, and permission inquiries. The multi-class probability prediction process performs forward propagation on the query vector and outputs the probability value of each intent category using a softmax function. The intent probability distribution includes an intent category field and a corresponding probability value field.

[0069] Therefore, the intent probability distribution is subjected to category filtering based on a preset confidence threshold to generate query intent tags. The category filtering process selects the intent category with the highest probability value as a candidate tag and determines whether this probability value meets the confidence threshold condition. The confidence threshold is determined based on the balance between precision and recall on the validation set. When the probability value is lower than the confidence threshold, a low-confidence flag (low_confidence_flag) is registered, triggering multi-intent candidate processing or manual review. The query intent tag includes an intent category field, a confidence score field, and a low-confidence flag field, using the combination of the query flag (query_id) and the session flag as the primary key. This key is read by step S701 at the permission filtering and graph traversal retrieval entry point for retrieval path matching and candidate knowledge subgraph generation.

[0070] In one embodiment of the intelligent response method for station management operations based on knowledge graph question answering in this application, it may further include the following: Step S701: Read the query intent tag and match the corresponding search path in the knowledge retrieval index; read the user role permission parameters and perform access permission determination on the nodes involved in the search path according to the preset permission filtering rules to generate an accessible node set; perform graph traversal retrieval on the station management knowledge base based on the accessible node set to generate candidate knowledge subgraphs. Step S702: Serialize the node attributes and relation structure of the candidate knowledge subgraph into a structured knowledge representation, input the structured knowledge representation and the query vector into a pre-trained answer generation model to perform sequence decoding to generate question-and-answer response text, encapsulate the question-and-answer response text into a response message and output it to the user interaction terminal.

[0071] In this embodiment, the query intent tag output in step S602 and the knowledge retrieval index output in step S502 are used as the starting input for the knowledge retrieval process to read the intent category and index structure. The reading process locates data according to the query identifier query_id and the index identifier. When the low confidence identifier of the query intent tag is true, the multi-intent candidate retrieval mode is triggered. When the index structure is damaged, the index anomaly identifier is registered and the index reconstruction process is triggered.

[0072] In one implementation, the query intent tag is read and matched against the corresponding retrieval path in the knowledge retrieval index. The retrieval path matching process locates the corresponding index entry in the combined index layer of the knowledge retrieval index based on the intent category field. Each intent category is pre-configured with associated node type ranges and relationship type ranges. For example, regulatory query intents match regulatory system nodes with reference relationship paths, operational guidance intents match operational step nodes with time-series constraint relationship paths, and emergency handling intents match emergency plan nodes with condition trigger relationship paths. The retrieval path includes a starting node type field, a target node type field, and a set of allowed relationship types field.

[0073] Preferably, user role permission parameters are read, and access permission determination is performed on the nodes involved in the search path according to preset permission filtering rules to generate an accessible node set. User role permission parameters are extracted from user session information; user roles include ordinary station staff, duty station managers, equipment maintenance personnel, and security administrators. Permission filtering rules are determined based on the mapping relationship between user roles and node access levels; node access levels include public, internal, and confidential levels. The access permission determination process traverses the nodes involved in the search path and compares user role permissions with node access levels; nodes with insufficient permissions are marked as inaccessible and excluded from the candidate range. The accessible node set includes a node identifier field, a node type field, and an access permission identifier field.

[0074] Optionally, a graph traversal retrieval is performed on the station management knowledge base based on the accessible node set to generate candidate knowledge subgraphs. The graph traversal retrieval process starts from the starting node in the accessible node set and performs a breadth-first traversal based on the set of allowed relation types for the retrieval path. The traversal depth is limited by a preset maximum traversal depth parameter, which is determined based on query complexity and response time requirements. The traversal process records the access path and collects the nodes and edges on the path. The candidate knowledge subgraph includes a subgraph node set field, a subgraph edge set field, and a traversal path set field.

[0075] Therefore, the node attributes and relational structures of the candidate knowledge subgraph are serialized into a structured knowledge representation. The serialization process converts the attribute key-value pairs of subgraph nodes and the relational types of subgraph edges into a text sequence format. Node attributes and relational structures are organized according to the traversal path order to maintain semantic coherence. The structured knowledge representation and the query vector are input into a pre-trained answer generation model to perform sequence decoding and generate question-and-answer response text. The answer generation model adopts a sequence-to-sequence architecture based on an attention mechanism and is trained on a question-and-answer corpus in the station administration domain. The question-and-answer response text is encapsulated into a response message and output to the user interaction terminal. The response message includes response text fields, a source citation field, a response timestamp field, and a session identifier field, using the combination of the message identifier (message_id) and the session identifier as the primary key. This is read by the user feedback module at the feedback collection entry point for question-and-answer quality assessment and model iteration optimization.

[0076] To effectively address the shortcomings of traditional technologies in knowledge graph construction, semantic retrieval, and intelligent question answering, and to provide technical support for intelligent response to station management operations, this application provides an embodiment of an intelligent response system for station management operations based on knowledge graph question answering, which implements all or part of the aforementioned intelligent response method for station management operations based on knowledge graph question answering. The intelligent response system for station management operations based on knowledge graph question answering specifically includes the following components: The entity relation parsing module is used to acquire unstructured documents of station administration and perform format conversion and encoding standardization to generate preprocessed text. The preprocessed text is then segmented and lexical labeled to generate a labeled text sequence. The labeled text sequence is input into a pre-trained entity recognition model to perform sequence labeling to generate an entity labeled sequence. The entity labeled sequence is then subjected to co-occurrence analysis and dependency parsing based on domain relation rules to generate a set of entity relation triples. The knowledge retrieval index module is used to construct a three-layer semantic knowledge graph containing a concept layer, an instance layer, and a relation layer based on the entity relation triple set, and to establish condition triggering relations and temporal constraint relations. The three-layer semantic knowledge graph is stored in a graph database to generate a station administration knowledge base. The node attributes and relation types of the station administration knowledge base are used to establish a multi-level index to generate a knowledge retrieval index. The question-and-answer response module is used to receive natural language queries input by users and perform vector encoding to generate query vectors in combination with the dialogue context. The query vectors are then input into an intent classification model to perform category determination and generate query intent tags. Based on the query intent tags and user role permission parameters, permission filtering and graph traversal retrieval are performed on the knowledge retrieval index to generate candidate knowledge subgraphs. The candidate knowledge subgraphs are then input into an answer generation model to perform natural language conversion to generate question-and-answer response text and output it to the user interaction terminal.

[0077] As can be seen from the above description, the intelligent response system for station management operations based on knowledge graph question answering provided in this application can generate a station management knowledge base by constructing a three-layer semantic knowledge graph through entity recognition and relationship extraction of unstructured documents, performing permission filtering and graph traversal retrieval based on query vectors and intent classification to generate candidate knowledge subgraphs, and outputting question answer response text through an answer generation model. This effectively solves the shortcomings of traditional technologies in knowledge graph construction, semantic retrieval, and intelligent question answering, and provides technical support for intelligent response to station management operations.

[0078] This invention also provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the intelligent response method for station management operations based on knowledge graph question answering.

[0079] This invention also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described intelligent response method for station management operations based on knowledge graph question answering.

[0080] This invention also provides a computer program product, which includes a computer program that, when executed by a processor, implements the above-described intelligent response method for station management operations based on knowledge graph question answering.

[0081] In this embodiment of the invention, a three-layer semantic knowledge graph is constructed by performing entity recognition and relationship extraction on unstructured documents to generate a station management knowledge base. Based on query vectors and intent classification, permission filtering and graph traversal retrieval are performed to generate candidate knowledge subgraphs. The question-and-answer response text is then converted and output through an answer generation model. This effectively solves the shortcomings of traditional technologies in knowledge graph construction, semantic retrieval, and intelligent question answering, and provides technical support for intelligent response to station management operations.

[0082] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for intelligent response to station management operations based on knowledge graph question answering, characterized in that, The method includes: The system acquires unstructured documents for station administration and performs format conversion and encoding standardization to generate preprocessed text. The preprocessed text is then segmented and lexical labeled to generate a labeled text sequence. The labeled text sequence is input into a pre-trained entity recognition model to perform sequence labeling to generate an entity label sequence. The entity label sequence is then subjected to co-occurrence analysis and dependency resolution based on domain relationship rules to generate a set of entity relationship triples. Based on the entity relationship triple set, a three-layer semantic knowledge graph containing a concept layer, an instance layer, and a relationship layer is constructed, and condition triggering relationships and temporal constraint relationships are established. The three-layer semantic knowledge graph is stored in a graph database to generate a station management knowledge base. Multi-level indexes are established on the node attributes and relationship types of the station management knowledge base to generate a knowledge retrieval index. The system receives a natural language query input from the user and performs vector encoding based on the dialogue context to generate a query vector. The query vector is then input into an intent classification model to perform category determination and generate a query intent label. Based on the query intent label and user role permission parameters, the system performs permission filtering and graph traversal retrieval on the knowledge retrieval index to generate a candidate knowledge subgraph. The candidate knowledge subgraph is then input into an answer generation model to perform natural language conversion to generate a question-and-answer response text, which is then output to the user interaction terminal.

2. The intelligent response method for station management operations based on knowledge graph question answering according to claim 1, characterized in that, The process of acquiring unstructured station management documents, performing format conversion and encoding standardization to generate preprocessed text, and then performing sentence segmentation and lexical tagging on the preprocessed text to generate a tagged text sequence includes: Read unstructured documents from the station administration system and identify the document format type. Perform format parsing on the unstructured documents from the station administration system to extract text content and generate raw text data. Perform irrelevant content filtering and character encoding unification processing on the raw text data to generate preprocessed text. The preprocessed text is segmented into sentences according to preset sentence boundary rules to generate a sentence sequence. Each sentence in the sentence sequence is then processed by word segmentation and part-of-speech tagging to generate a tagged text sequence.

3. The intelligent response method for station management operations based on knowledge graph question answering according to claim 1, characterized in that, The step of inputting the annotated text sequence into a pre-trained entity recognition model to perform sequence annotation and generate an entity annotation sequence, and then performing co-occurrence analysis and dependency resolution on the entity annotation sequence according to domain relation rules to generate a set of entity relation triples, includes: The labeled text sequence is converted into a model input vector and input into a pre-trained entity recognition model. Sequence labeling inference is performed on the model input vector to generate entity category labels for each word. Based on the entity category labels, boundary recognition and entity merging are performed to generate an entity labeled sequence. For each entity in the entity annotation sequence, perform intra-sentence co-occurrence statistics to generate an entity co-occurrence matrix. Perform dependency parsing on the entity co-occurrence matrix and the syntactic structure of the annotated text sequence to generate entity dependency pairs. Perform relation type mapping on the entity dependency pairs according to domain relation rules to generate a set of entity relation triples.

4. The intelligent response method for station management operations based on knowledge graph question answering according to claim 1, characterized in that, The construction of a three-layer semantic knowledge graph, comprising a concept layer, an instance layer, and a relation layer, based on the entity relation triple set, and the establishment of condition-triggered relations and temporal constraint relations, includes: The entity relationship triple set is read and the hierarchical classification of each entity is determined according to the preset entity type classification rules. Entities determined to be of the concept class are written into the concept layer and concept nodes are established. Entities determined to be of the instance class are written into the instance layer and instance nodes are established. Based on the relationship type in the entity relationship triple set, an association edge is established between the concept node and the instance node to generate a three-layer semantic knowledge graph. Traverse the associated edges in the three-layer semantic knowledge graph and identify relationships with precondition attributes. Mark the relationships with precondition attributes with condition triggering identifiers to generate condition triggering relationships. Mark the relationships with execution order attributes with time constraint identifiers to generate time constraint relationships. Write the condition triggering relationships and the time constraint relationships into the relationship layer of the three-layer semantic knowledge graph.

5. The intelligent response method for station management operations based on knowledge graph question answering according to claim 1, characterized in that, The step of storing the three-layer semantic knowledge graph in a graph database to generate a station administration knowledge base, and establishing a multi-level index for the node attributes and relationship types of the station administration knowledge base to generate a knowledge retrieval index, includes: The concept nodes and instance nodes of the three-layer semantic knowledge graph are serialized into graph database node records. The associated edges, condition triggering relationships, and temporal constraint relationships of the three-layer semantic knowledge graph are serialized into graph database edge records. The graph database node records and graph database edge records are written into the graph database to generate a station management knowledge base. The system reads the attribute fields of each node in the station management knowledge base and establishes an attribute index according to the preset attribute priority rules. It also reads the relationship types of each edge in the station management knowledge base and establishes a relationship type index. Finally, it organizes the attribute index and the relationship type index in a hierarchical structure to generate a knowledge retrieval index.

6. The intelligent response method for station management operations based on knowledge graph question answering according to claim 1, characterized in that, The process of receiving a natural language query input from a user and performing vector encoding based on the dialogue context to generate a query vector, then inputting the query vector into an intent classification model to perform category determination and generate a query intent label, includes: The system receives a natural language query input from the user and reads the dialogue context record of the current session. It then performs concatenation processing on the natural language query and the dialogue context record to generate context fusion text. Finally, it inputs the context fusion text into a pre-trained semantic encoding model to perform vector mapping to generate a query vector. The query vector is input into a pre-trained intent classification model to perform multi-class probability prediction and generate an intent probability distribution. The intent probability distribution is then filtered by category based on a preset confidence threshold to generate query intent labels.

7. The intelligent response method for station management operations based on knowledge graph question answering according to claim 1, characterized in that, The process of performing permission filtering and graph traversal retrieval on the knowledge retrieval index based on the query intent tag and user role permission parameters to generate candidate knowledge subgraphs, inputting the candidate knowledge subgraphs into an answer generation model to perform natural language conversion to generate question-and-answer response text, and outputting it to the user interaction terminal includes: Read the query intent tag and match the corresponding search path in the knowledge retrieval index; read the user role permission parameters and perform access permission determination on the nodes involved in the search path according to the preset permission filtering rules to generate an accessible node set; and perform graph traversal retrieval on the station management knowledge base based on the accessible node set to generate candidate knowledge subgraphs. The node attributes and relational structure of the candidate knowledge subgraph are serialized into a structured knowledge representation. The structured knowledge representation and the query vector are input into a pre-trained answer generation model to perform sequence decoding to generate question-and-answer response text. The question-and-answer response text is encapsulated into a response message and output to the user interaction terminal.

8. A station administration operation intelligent response system based on knowledge graph question answering, characterized in that, The system includes: The entity relation parsing module is used to acquire unstructured documents of station administration and perform format conversion and encoding standardization to generate preprocessed text. The preprocessed text is then segmented and lexical labeled to generate a labeled text sequence. The labeled text sequence is input into a pre-trained entity recognition model to perform sequence labeling to generate an entity labeled sequence. The entity labeled sequence is then subjected to co-occurrence analysis and dependency parsing based on domain relation rules to generate a set of entity relation triples. The knowledge retrieval index module is used to construct a three-layer semantic knowledge graph containing a concept layer, an instance layer, and a relation layer based on the entity relation triple set, and to establish condition triggering relations and temporal constraint relations. The three-layer semantic knowledge graph is stored in a graph database to generate a station administration knowledge base. The node attributes and relation types of the station administration knowledge base are used to establish a multi-level index to generate a knowledge retrieval index. The question-and-answer response module is used to receive natural language queries input by users and perform vector encoding to generate query vectors in combination with the dialogue context. The query vectors are then input into an intent classification model to perform category determination and generate query intent tags. Based on the query intent tags and user role permission parameters, permission filtering and graph traversal retrieval are performed on the knowledge retrieval index to generate candidate knowledge subgraphs. The candidate knowledge subgraphs are then input into an answer generation model to perform natural language conversion to generate question-and-answer response text and output it to the user interaction terminal.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the steps of the intelligent response method for station management operations based on knowledge graph question answering as described in any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the computer program implements the steps of the intelligent response method for station management operations based on knowledge graph question answering as described in any one of claims 1 to 7.

Images