An accident investigation auxiliary method and system based on knowledge graph enhanced retrieval

Abstract

The application provides a large model accident investigation auxiliary method and system based on knowledge graph enhanced retrieval, and belongs to the technical field of accident investigation and fire accident investigation. The method comprises the following steps: loading and registering a prefabricated basic knowledge graph of accident investigation; analyzing and block processing original accident investigation data to obtain a fragment set; updating the basic knowledge graph according to the fragment set to obtain a unified accident investigation knowledge graph; obtaining user query description information, performing intent analysis and constraint extraction on the user query description information to obtain structured condition data; performing retrieval according to the structured condition data and the unified accident investigation knowledge graph to obtain a unified candidate evidence set; analyzing and reasoning the unified candidate evidence set to obtain explanation result data with evidence reference and graph path; and completing accident investigation assistance according to the explanation result data. The application embodiment can enhance the auxiliary decision-making ability and engineering landing nature in a complex scene.

Description

Technical Field

[0001] This application relates to the field of accident investigation and fire accident investigation technology, and in particular to a large-scale accident investigation auxiliary method and system based on knowledge graph enhanced retrieval. Background Technology

[0002] In related technologies, routine accident investigation requires manual reading of multi-source data. However, manual reading of knowledge from multiple disciplines, such as electrical safety, building structure, material combustion, ventilation and smoke control, and regulations and standards, is prone to omissions, misreadings and forgetting, difficulty in restoring cross-document associations, and difficulty in tracing the basis.

[0003] In summary, the technical problems existing in the relevant technologies need to be improved. Summary of the Invention

[0004] The main objective of this application is to propose a large-scale accident investigation assistance method and system based on knowledge graph-enhanced retrieval, which can improve the intelligence and interpretability of accident investigation.

[0005] To achieve the above objectives, one aspect of this application proposes a large-scale accident investigation assistance method based on knowledge graph-enhanced retrieval, the method comprising the following steps:

[0006] Load and register a pre-built basic knowledge graph of accident investigation; the basic knowledge graph is used to provide a knowledge source for subsequent retrieval and reasoning;

[0007] The original accident investigation data is analyzed and segmented to obtain a set of fragments.

[0008] The basic knowledge graph is updated based on the set of fragments to obtain a unified accident investigation knowledge graph;

[0009] Obtain user query description information, perform intent parsing and constraint extraction on the user query description information to obtain structured condition data;

[0010] A unified candidate evidence set is obtained by searching the structured conditional data and the unified accident investigation knowledge graph.

[0011] The unified candidate evidence set is analyzed and reasoned to obtain interpretation result data with evidence citation and graph path;

[0012] Based on the interpreted data, assist in the accident investigation.

[0013] In some embodiments, the parsing and segmentation of the original accident investigation data to obtain a set of fragments includes:

[0014] Obtain original accident investigation data;

[0015] The original accident investigation data is uniformly coded and standardized to obtain document record data; the document record data includes standardized text and metadata; the metadata includes registration source, time, uploader and version information;

[0016] The document record data is segmented into sentences according to a preset style, and then divided into blocks using a window and overlap strategy to obtain a set of fragments; the set of fragments includes several different fragments.

[0017] In some embodiments, updating the basic knowledge graph based on the fragment set to obtain a unified accident investigation knowledge graph includes:

[0018] Based on preset inspection tags, information is extracted and data is standardized from the fragments in the fragment set to obtain field information;

[0019] Relationship identification is performed based on the field information to obtain relationship identification data;

[0020] Based on the relationship identification data, semantic normalization and alias merging are performed to obtain merged data;

[0021] Based on the merged data, confidence assessment and conflict resolution are performed to obtain assessment record data;

[0022] Based on the assessment record data and the basic knowledge graph, incremental changes are performed and the data is attributed and traced back to its source to obtain a unified accident investigation knowledge graph.

[0023] In some embodiments, the step of performing confidence assessment and conflict resolution based on the merged data to obtain assessment record data includes:

[0024] Based on the merged data, scores are assigned to elements and relationships based on the comprehensive data source type, clarity of expression, and consistency, thus obtaining scoring information;

[0025] The conflicting records in the merged data are adjudicated or marked according to source priority, time relevance, and consistency to obtain conflict-resolved data.

[0026] Based on the scoring information and the conflict resolution data, evaluation record data is obtained.

[0027] In some embodiments, the step of retrieving a unified candidate evidence set based on the structured conditional data and the unified accident investigation knowledge graph includes:

[0028] Based on the inverted index structure, literal matching is performed on the structured condition data, and several results that achieve a preset accuracy are selected to obtain keyword retrieval data;

[0029] The structured conditional data and the fragments of the fragment set are mapped to vectors respectively, and semantic nearest neighbor evidence is recalled based on cosine similarity to obtain semantic retrieval data;

[0030] The elements and constraints parsed from the structured conditional data are mapped to obtain a set of nodes. The relationships between the set of nodes and the unified accident investigation knowledge graph are expanded to obtain graph relationship retrieval data.

[0031] The keyword retrieval data, the semantic retrieval data, and the graph relationship retrieval data are deduplicated and unique source anchors are retained to obtain a unified candidate set;

[0032] Each candidate piece of information in the unified candidate set is labeled with a reason and grouped and archived to obtain a unified candidate evidence set.

[0033] In some embodiments, the analysis and reasoning of the unified candidate evidence set to obtain explanatory result data with evidence citations and graph paths includes:

[0034] A relevance screening model is used to perform pairwise discrimination scoring and reordering screening on the user query description information and the unified candidate evidence set, and evidence within a preset score range is selected to obtain a controlled context.

[0035] Based on the controlled context, a chain-like investigation analysis is performed, and reasoning is carried out step by step according to the preset link to obtain intermediate key points and local conclusions.

[0036] Based on the aforementioned key points and local conclusions, a structured approach is adopted, with each point accompanied by evidence citations and a graphical path, to obtain structured conclusions.

[0037] The structured conclusions were verified to obtain the interpretation results data.

[0038] In some embodiments, the large-model accident investigation assistance method based on knowledge graph enhanced retrieval further includes:

[0039] Provide function switches; the function switches include a search enhancement switch, a graph relationship search switch, a hypothetical answer expansion switch, a reordering and fine screening switch, and a task one-click start / stop and combination switch;

[0040] Provide strategy parameters; the strategy parameters include sampling temperature parameters and maximum number of context lines parameters;

[0041] Provides traceability and version records; the traceability and version records include processing logs, evidence numbers, map versions, and map change details.

[0042] To achieve the above objectives, another aspect of this application proposes a large-scale accident investigation assistance system based on knowledge graph-enhanced retrieval, used to implement the method described above. The system includes:

[0043] The basic knowledge graph loading module is used to load and register pre-made basic knowledge graphs for accident investigation; the basic knowledge graphs are used to provide knowledge sources for subsequent retrieval and reasoning.

[0044] The document access and Chinese parsing module is used to parse and segment the original accident investigation data to obtain a collection of fragments;

[0045] The knowledge graph management module is used to update the basic knowledge graph based on the fragment set to obtain a unified accident investigation knowledge graph;

[0046] The user description parsing module is used to obtain user query description information, perform intent parsing and constraint extraction on the user query description information, and obtain structured condition data.

[0047] The retrieval and indexing module is used to perform retrieval based on the structured conditional data and the unified accident investigation knowledge graph to obtain a unified candidate evidence set;

[0048] The reasoning and generation module is used to analyze and reason about the unified candidate evidence set to obtain interpretation result data with evidence citations and graph paths.

[0049] In some embodiments, the system further includes an interaction and parameter control module;

[0050] The interaction and parameter control module is used to provide function switches, strategy parameters, and logging and version records.

[0051] In some embodiments, the system further includes a data storage device;

[0052] The data storage device is used to store basic knowledge graphs, data fragments, inverted and vector indexes, operation logs, and version information.

[0053] To achieve the above objectives, another aspect of this application provides an electronic device, which includes a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the method described above.

[0054] To achieve the above objectives, another aspect of the embodiments of this application proposes a computer-readable storage medium storing a computer program that, when executed by a processor, implements the methods described above.

[0055] To achieve the above objectives, another aspect of the embodiments of this application proposes a computer program product, including a computer program that, when executed by a processor, implements the aforementioned method.

[0056] The embodiments of this application include at least the following beneficial effects: This application provides a method, system, electronic device, storage medium, and program product for assisting large-scale accident investigation based on knowledge graph-enhanced retrieval. The solution includes: loading and registering a pre-built basic knowledge graph for accident investigation; parsing and segmenting the original accident investigation data to obtain a fragment set; updating the basic knowledge graph based on the fragment set to obtain a unified accident investigation knowledge graph; obtaining user query description information, performing intent parsing and constraint extraction on the user query description information to obtain structured condition data; performing retrieval based on the structured condition data and the unified accident investigation knowledge graph to obtain a unified candidate evidence set; analyzing and reasoning on the unified candidate evidence set to obtain interpretation result data with evidence citations and graph paths; and completing the accident investigation assistance based on the interpretation result data. The embodiments of this application can enhance the auxiliary decision-making capability and engineering feasibility in complex scenarios. Attached Figure Description

[0057] Figure 1 This is a flowchart of a large-scale accident investigation auxiliary method based on knowledge graph-enhanced retrieval provided in an embodiment of this application;

[0058] Figure 2 This is a flowchart illustrating the steps of the large-scale accident investigation assistance scheme based on knowledge graph-enhanced retrieval provided in this application embodiment;

[0059] Figure 3 This is a flowchart of user-uploaded data provided in an embodiment of this application;

[0060] Figure 4 This is a flowchart of knowledge graph extraction and generation provided in the embodiments of this application;

[0061] Figure 5 This is a flowchart of the fusion retrieval and evidence aggregation provided in the embodiments of this application;

[0062] Figure 6 This is a flowchart of the fine screening and controlled inference output provided in the embodiments of this application;

[0063] Figure 7 This is a flowchart of the interaction and parameter control provided in the embodiments of this application. Detailed Implementation

[0064] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of this application and are not intended to limit it. In the following description, when referring to the accompanying drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with those of this application; they are merely examples of apparatuses and methods consistent with some aspects of the embodiments of this application as detailed in the appended claims.

[0065] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used herein is for the purpose of describing embodiments of this application only and is not intended to limit this application.

[0066] Before providing a detailed description of the embodiments of this application, some of the nouns and terms involved in the embodiments of this application will be explained first. The nouns and terms involved in the embodiments of this application are subject to the following interpretations.

[0067] 1) Top-K is the method to find the K items with the highest value or the highest score from a dataset.

[0068] This application recognizes that accident investigation is a highly specialized and interdisciplinary task, involving knowledge from multiple fields such as electrical safety, building structure and material combustion, ventilation and smoke control, personnel behavior and evacuation, monitoring and IoT sensing, and legal and regulatory provisions. In practice, it is necessary to comprehensively review multi-source heterogeneous information within a limited time, including investigation records, accident investigation reports, drawings and equipment data, monitoring / sensor logs, historical cases, and standards, and reconstruct the fire scenario, determine the spread path, assign responsibility, and provide compliance opinions under complex, noisy, and even incomplete evidence. This process demands extremely high levels of knowledge coverage, attention to detail, and causal reasoning ability. Relying solely on manual reading and experiential memory can easily lead to information omissions, broken cross-document connections, inconsistent statements, and difficulty in tracing the basis for the investigation. Current information technology methods primarily rely on keyword retrieval, templated forms, or rule engines, which have limited capabilities in handling long, free-text documents and cross-document causal relationships. Traditional full-text retrieval struggles to understand semantic equivalence and differences in aliases and unit / time expressions, making it difficult to establish verifiable relationship links between "ignition source—igniter—combustion-acclimating conditions—spread path—loss consequences—liability elements—applicable clauses." While general-purpose large-scale models that have emerged in recent years possess strong language understanding and generation capabilities, they are prone to "illusions" or inferences based on insufficient information in the absence of controlled context and verifiable evidence, failing to meet the rigid requirements of interpretability, traceability, and compliance in the field of investigation. On the other hand, the construction of knowledge graphs alone, if separated from the retrieval and analysis processes, is also unlikely to play a role in actual question answering and conclusion output. In summary, the industry urgently needs a technical approach that can model cross-source knowledge graphs and deeply integrate them with text retrieval. This approach should explicitly express the elements of the investigation and their causal / attribute / hierarchical relationships, supporting synonym unification and conflict resolution to achieve cross-document evidence association. Furthermore, it should use fusion retrieval (keywords, semantics, and graph relationships) to achieve high-coverage recall of relevant evidence and drive a large model to conduct chain-based investigation analysis and clause verification within a controlled context, outputting interpretable results with citations and graph paths. Simultaneously, it should support incremental updates and parameterized control to ensure the system's sustainable adaptation to knowledge evolution and changes in business scenarios. This is precisely the technological gap and direction for improvement addressed in this application.

[0069] In view of this, this application proposes a large-scale accident investigation assistance method and system based on knowledge graph-enhanced retrieval to address the problems that conventional accident investigations are prone to omissions, misreadings and forgetting, difficulty in restoring cross-document associations, and difficulty in tracing evidence, which require manual reading of multi-source data and involve knowledge from multiple disciplines such as electrical safety, building structure, material combustion, ventilation and smoke control, and regulations and standards. This solution unifies the base knowledge and user-supplemented data into a unified graph model, explicitly encoding investigation elements and their causal, attribute, and hierarchical relationships. It integrates keyword retrieval, semantic retrieval, and graph relationship retrieval around the investigation question, and constructs a controlled context through relevance screening. Based on this, it drives the large-scale model to complete fire cause assessment, spread path analysis, responsibility element identification, and compliance clause verification, outputting interpretable conclusions with evidence citations and graph paths. This significantly reduces omissions and biases while maintaining existing workflows, improving timeliness, consistency, and traceability of conclusions.

[0070] This application relates to the field of accident investigation and fire accident investigation technology, belonging to the interdisciplinary direction of intelligent information processing and emergency safety applications. Specifically, it involves a large-scale accident investigation auxiliary method and system based on knowledge graph-enhanced retrieval. This method is designed for multi-source heterogeneous investigation data and standards, comprehensively utilizing knowledge graph modeling, keyword / semantic / relational fusion retrieval, and relevance screening. Under a controlled context, it drives a large-scale model to complete cause analysis, propagation path analysis, and compliance verification, thereby improving the intelligence and interpretability of accident investigation.

[0071] The large-scale accident investigation assistance method based on knowledge graph augmented retrieval provided in this application can be applied to a terminal, a server, or software running on a terminal or server. In some embodiments, the terminal can be a smartphone, tablet, laptop, or desktop computer, but is not limited thereto; the server can be configured as an independent physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server providing basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, CDN, and big data and artificial intelligence platforms. The server can also be a node server in a blockchain network; the software can be an application that implements the large-scale accident investigation assistance method based on knowledge graph augmented retrieval, but is not limited to the above forms.

[0072] This application can be used in a wide variety of general-purpose or special-purpose computer system environments or configurations. Examples include: personal computers, server computers, handheld or portable devices, tablet devices, multiprocessor systems, microprocessor-based systems, set-top boxes, programmable consumer electronics devices, network PCs, minicomputers, mainframe computers, and distributed computing environments including any of the above systems or devices. This application can be described in the general context of computer-executable instructions executed by a computer, such as program modules. Generally, program modules include routines, programs, objects, components, data structures, etc., that perform specific tasks or implement specific abstract data types. This application can also be practiced in distributed computing environments where tasks are performed by remote processing devices connected via a communication network. In distributed computing environments, program modules can reside in local and remote computer storage media, including storage devices.

[0073] It should be noted that in all specific embodiments of this application, when processing data related to user identity or characteristics, such as user information, user behavior data, user historical data, and user location information, user permission or consent is obtained first. Furthermore, the collection, use, and processing of this data comply with relevant laws, regulations, and standards. In addition, when embodiments of this application require access to sensitive personal information of users, separate permission or consent from the user is obtained through pop-ups or redirection to confirmation pages. Only after obtaining the user's separate permission or consent is the necessary user-related data required for the proper functioning of these embodiments acquired.

[0074] Figure 1 This is an optional flowchart of a large-model accident investigation assistance method based on knowledge graph augmented retrieval provided in the embodiments of this application. Figure 1 The method may include, but is not limited to, steps S101 to S107.

[0075] Step S101: Load and register the pre-built basic knowledge graph of accident investigation; the basic knowledge graph is used to provide knowledge sources for subsequent retrieval and reasoning;

[0076] Step S102: Analyze and segment the original accident investigation data to obtain a set of fragments;

[0077] Step S103: Update the basic knowledge graph based on the fragment set to obtain a unified accident investigation knowledge graph;

[0078] Step S104: Obtain user query description information, perform intent parsing and constraint extraction on user query description information to obtain structured condition data;

[0079] Step S105: Retrieve based on structured conditional data and unified accident investigation knowledge graph to obtain a unified candidate evidence set;

[0080] Step S106: Analyze and reason about the unified candidate evidence set to obtain interpretation result data with evidence citation and graph path;

[0081] Step S107: Complete the accident investigation assistance based on the interpretation results data.

[0082] Steps S101 to S107, as shown in the embodiments of this application, involve loading a basic knowledge graph, registering the pre-built knowledge graph as a knowledge source for retrieval and reasoning, and recording it as... G 0; User-uploaded data processing, for the original document Standardized encoding and standardization result in standard text. (Depend on and Processing and registering metadata Then, Chinese parsing and sliding window segmentation are performed to output a set of fragments. (The element is a triple) Knowledge graph extraction and generation, for fragments Extracting the set of elements With relation set Through alternative names and synonyms, they are unified. And confidence assessment and conflict resolution Post-construction incremental subgraph and with Merging yields an updated map User description parsing generates structured conditions based on user queries and constraints, denoted as q; fusion retrieval and evidence aggregation simultaneously perform keyword retrieval and semantic retrieval around q, and graph relationship retrieval, deduplicating and fusing the three results to obtain a unified candidate set. Fine-tuning and controlled inference outputs are scored based on relevance. Select a controlled context U and conduct chain-based exploration analysis to output interpretable results with evidence citations and map paths. And formed through standardization and verification This application embodiment also performs interaction and parameter control: receiving the function switch vector b and strategy parameters. Generate runtime snapshot To support reproduction, comparison, and rollback.

[0083] In step S101 of some embodiments, the pre-made accident investigation basic knowledge graph can be constructed by systematically extracting and structuring knowledge from massive amounts of authoritative historical data.

[0084] The generation of a unified accident investigation knowledge graph requires collecting authoritative data from multiple sources uploaded by users, including major accident investigation reports, safety regulations and standards, professional textbooks, and historical investigation records. Then, using natural language processing technology, entities, attributes, and their interrelationships are automatically extracted from these unstructured texts according to a pre-defined investigation element tagging system (such as time, location, equipment, phenomenon, causal relationship, etc.). Next, the extraction results are semantically normalized to unify synonyms and standard terms, and contradictory information from different sources is resolved through confidence assessment and conflict resolution mechanisms. Finally, all standardized knowledge elements are incrementally merged and uniformly stored with the base graph to form a structured knowledge base with complete traceability information that can be directly loaded and used by the system.

[0085] In some embodiments, step S102 may include, but is not limited to, steps S201 to S203:

[0086] Step S201: Obtain the original accident investigation data;

[0087] Step S202: The original accident investigation data is uniformly coded and standardized to obtain document record data; the document record data includes standardized text and metadata; the metadata includes registration source, time, uploader and version information;

[0088] Step S203: Segment the document record data according to a preset style, and divide it into blocks using a window and overlap strategy to obtain a fragment set; the fragment set includes several different fragments.

[0089] In some embodiments, steps S201-S203 involve data reception and metadata registration: each document (original accident investigation data) is recorded. , ),in At least include source, time, uploader, and version; Chinese parsing and segmentation: for After formatting and noise cleanup, sentences are segmented, and fragments are generated using a sliding window strategy based on window length L and overlap rate ρ (the step size is determined by L and ρ). Simultaneously, generate location anchor points for each segment. (Including document number, page number, paragraph and character range), and finally merged into the collection. .

[0090] In some embodiments, step S103 may include, but is not limited to, steps S301 to S305:

[0091] Step S301: Extract information and standardize data from the fragments in the fragment set according to the preset inspection labels to obtain field information;

[0092] Step S302: Relationship identification is performed based on the field information to obtain relationship identification data;

[0093] Step S303: Based on the relationship identification data, perform semantic normalization and alias merging to obtain merged data;

[0094] Step S304: Confidence assessment and conflict resolution are performed based on the merged data to obtain assessment record data;

[0095] Step S305: Based on the assessment record data and the basic knowledge graph, incremental data is generated and attributed to the source to obtain a unified accident investigation knowledge graph.

[0096] In some embodiments, steps S301-S305 involve field extraction: by inspection tag set. from Extracting "tag-value" and standardizing it by time, location, number, and unit, while retaining evidence anchors; Relationship identification: Combining trigger words, sentence structure, and basic syntactic clues, identifying attribute, causal, subordinate, and parallel relationships between elements, allowing merging within the same sentence and across sentences to align with extracted elements; Semantic normalization and alias merging: Performing normalization mapping on {entity, tag, value}. The system categorizes locations, institutions, and equipment hierarchically to reduce fragmentation; confidence assessment and conflict resolution are achieved by considering source weighting, clarity of expression, and cross-evidence consistency. For contradictory records, selective acceptance or coexistence should be marked and a pending verification mark should be set; incremental aggregation and source tracing: the entities and relationships in this batch of standards should be included. and with Merging deduplication and unifying naming results At the same time, the evidence anchors and metadata are attached to the nodes / edges and the version and change logs are recorded.

[0097] In some embodiments, step S304 may include, but is not limited to, steps S341 to S343:

[0098] Step S341: Based on the merged data, score the elements and relationships by comprehensively considering the data source type, clarity of expression, and consistency, and obtain the scoring information;

[0099] Step S342: The conflicting records in the merged data are adjudicated or marked according to source priority, time relevance and consistency to obtain conflict-resolved data;

[0100] Step S343: Based on the scoring information and conflict resolution data, obtain the evaluation record data.

[0101] In steps S341-S343 of some embodiments, a multi-dimensional scoring mechanism is introduced to assign a quantified credibility score to each piece of knowledge, achieving refined management of knowledge quality. Evaluation record data provides a data foundation for system optimization; developers can analyze scoring information and erroneous adjudication cases to optimize the confidence model and adjudication rules. Conflict resolution data (e.g., data marked "to be verified") provides a clear target list for human expert review, making the human-machine collaborative workflow more efficient.

[0102] In some embodiments, step S105 may include, but is not limited to, steps S501 to S505:

[0103] Step S501: Based on the inverted index, perform literal matching on the structured conditional data, select several results that achieve the preset accuracy, and obtain keyword retrieval data;

[0104] Step S502: Map the structured conditional data and the fragments in the fragment set into vectors respectively, and recall semantic nearest neighbor evidence based on cosine similarity to obtain semantic retrieval data;

[0105] Step S503: Map the parsed elements and constraints in the structured condition data to obtain a node set. Expand the relationships based on the node set and the unified accident investigation knowledge graph to obtain graph relationship retrieval data.

[0106] Step S504: Deduplicate the keyword retrieval data, semantic retrieval data, and graph relationship retrieval data and retain the unique source anchor points to obtain a unified candidate set;

[0107] Step S505: Mark the reasons for each candidate information in the unified candidate set and group and archive it to obtain the unified candidate evidence set.

[0108] In some embodiments, steps S501-S505 involve performing keyword retrieval: obtaining literal matching of q and fragments based on the inverted index structure. Select several high-scoring results to form Semantic retrieval: Mapping q and the fragment to vectors (from...) and (Obtained) and based on cosine similarity Several results were formed before the recall. Graph relation retrieval: Maps the elements and constraints resolved in q to a set of nodes. In the map The above uses the relationship expansion radius h to search for relevant paths and transcribes them into readable fragments. ;Result merging and deduplication: For , and Perform homogeneous and approximate deduplication while retaining the unique source anchor point, and simultaneously apply the normalized channel score and fusion weights { Calculate the fusion score and obtain the sorted candidate set. Reason labeling and hierarchical organization: Add reason labels such as "keyword hit, semantic neighbor, graph path" to candidates, and group and archive them by theme or element for subsequent fine screening.

[0109] In some embodiments, step S106 may include, but is not limited to, steps S601 to S604:

[0110] Step S601: Use a relevance screening model to perform pairwise discrimination scoring and reordering screening on user query description information and unified candidate evidence set, select evidence within a preset score range, and obtain controlled context;

[0111] Step S602: Perform chain-based investigation analysis based on the controlled context, and reason step by step according to the preset link to obtain intermediate key points and local conclusions;

[0112] Step S603: Based on the intermediate key points and local conclusions, perform structured processing, attach evidence citations and graph paths for each point, and obtain structured conclusions;

[0113] Step S604: Verify the structured conclusions to obtain the interpretation result data.

[0114] In some embodiments, steps S601-S604 involve performing a reordering fine screening: based on The length budget B selects a controlled context U, taking into account the diversity of sources and content and reducing redundancy during the selection process; chain-based investigation analysis: under the constraints of U, intermediate points and comprehensive judgments are generated according to the chain of "cause of fire → spread path → responsibility elements → compliance verification", only citing evidence in U and providing explanations and branches when there are conflicts; interpretable result generation: outputting structured results. (including textual conclusions) Key Points List Timeline / Process Diagram With atlas view (and attached evidence citations and graph paths for each item); result standardization and verification: for Formatting is required to obtain The system checks the chronology, dimensions, location / numbering consistency, and legal references, generating a checklist that needs to be reviewed. .

[0115] In some embodiments, the large-scale accident investigation assistance method based on knowledge graph augmented retrieval further includes:

[0116] Step S108 provides function switches; the function switches include search enhancement switch, graph relationship search switch, hypothetical answer expansion switch, reordering and fine screening switch, and task one-click start / stop and combination switch;

[0117] Step S109: Provide strategy parameters; the strategy parameters include sampling temperature parameters and maximum number of context lines parameters;

[0118] Step S10A provides traceability and version records; traceability and version records include processing logs, evidence numbers, map versions, and map change details.

[0119] In some embodiments, steps S108-S10A involve interaction and parameter control, including function switches: receiving a switch vector b (containing the start / stop states of retrieval enhancement, graph relation retrieval, hypothetical answer expansion, and reordering screening) and starting / stopping the corresponding modules accordingly; and strategy parameters: receiving parameter pairs. (Including sampling temperature τ and maximum number of context lines) ), and accordingly limit the amount and diversity of evidence entering U; Traceability and Versioning: Generate runtime snapshots (At least include b) Current atlas version Cited evidence and its anchor points The timestamp (t) is written to the processing log and map change record for result reproduction, difference comparison and rollback.

[0120] The relation expansion radius h and fusion weight of the graph relation retrieval in this application embodiment are { } It can be configured according to the task scenario or performance goals, and compared with the parameters mentioned above. Collaborative design aims to strike a balance between recall coverage, discrimination accuracy, and interpretability.

[0121] The solutions of this application embodiment will be described in detail and explained below with reference to specific application examples:

[0122] As an optional implementation, a method and system for assisting large-scale accident investigation based on knowledge graph-enhanced retrieval includes the following steps:

[0123] S1. Loading the basic knowledge graph: The system loads and registers the pre-built basic knowledge graph of accident investigation as a knowledge source for retrieval and reasoning. Its generation method is the same as S3. This step only loads and registers the graph.

[0124] S2. Users can choose to upload supplementary data. When users provide supplementary data, the system receives this data (as the original accident investigation data) and completes the parsing, cleaning and segmentation to generate processable segments for subsequent map input.

[0125] S3. Knowledge graph extraction and generation: Extract "tag-value" elements from the fragments and identify attributes, causal relationships, subordinate relationships and parallel relationships. Complete synonym unification and conflict resolution and merge with the base to form a unified accident investigation knowledge graph.

[0126] S4. The user describes known information. The user provides known facts, clues and constraints in natural language. The system completes the analysis of elements and intents and generates structured retrieval and reasoning conditions (structured condition data).

[0127] S5. Fusion retrieval and evidence aggregation: Keyword retrieval, semantic retrieval and graph relationship retrieval are performed in parallel around the conditions analyzed in S4. The results are merged and deduplicated to obtain a unified candidate evidence set, and the source and reason are marked.

[0128] S6. Fine screening and controlled reasoning output: The system reorders the relevance of candidate evidence, selects the top few to form a controlled context, and conducts chain analysis according to the investigation link, outputting interpretable conclusions, map views and key point lists (interpretation result data).

[0129] S7. Interaction and parameter control: The system provides switches for search enhancement, graph search, hypothetical answer expansion, reordering, etc., as well as parameters such as sampling temperature and number of context entries, and records processing logs and graph versions.

[0130] Furthermore, the process of user uploading data (S2) includes:

[0131] S21. Data Receiving and Metadata Registration: Supports Chinese PDF, DOCX, TXT and scanned documents, etc., with automatic unified encoding and the ability to unify simplified and traditional Chinese characters; registers metadata such as source, time, uploader and version.

[0132] S22. Chinese parsing and segmentation: Intelligent sentence segmentation and line merging based on Chinese punctuation and common heading styles; window and overlap strategies are used to divide into segments, and anchor information such as document name, page number or paragraph position is retained for each segment to facilitate source tracing.

[0133] Furthermore, the knowledge graph extraction and generation (S3) process includes:

[0134] S31. Field Extraction: Extract “tag-value” from the fragment according to the preset inspection tags, and standardize the time, location, number, quantity and unit, etc.; each result retains the document name, page number or paragraph and start and end positions as evidence anchors.

[0135] S32. Relationship Recognition: Combining trigger words, sentence structure, and basic syntactic clues, it identifies attribute, causal, subordinate, and parallel relationships; it supports merging within the same sentence and across sentences, and automatically aligns entities with extracted elements.

[0136] S33. Semantic unification and alias merging: unify the expressions of synonyms, near-synonyms and aliases into standardized names and retain mappings, and handle the differences between full-width and half-width characters, numbers and abbreviations; merge locations, institutions and equipment by level to reduce node fragmentation.

[0137] S34. Confidence assessment and conflict resolution: Scoring is based on the source type, clarity of expression, and consistency. When conflicting records are encountered, they are adjudicated according to the source priority, time relevance, and consistency. One can be accepted, and both can be marked or marked for verification. The supporting evidence is recorded simultaneously.

[0138] S35. Incremental Merging and Source Tracing: Merge the elements and relationships of this batch with the base map, remove duplicates, and unify the naming to generate an updated unified map; record the version and change log, and continuously associate evidence and sources on nodes and edges to achieve traceability.

[0139] Furthermore, the process of integrating retrieval and evidence aggregation (S5) includes:

[0140] S51 Keyword retrieval: Based on inverted index and Chinese word segmentation, perform literal matching on user descriptions, prioritize the retrieval of high-precision fragments, and record the hit words, hit locations, and source anchors.

[0141] S52. Semantic Retrieval: A semantic representation model is used to encode user descriptions and fragments into a unified vector space, and semantic nearest neighbor evidence is recalled based on similarity. The fragment vectors have been indexed in S1 or S2. During retrieval, only the query is encoded online, and the similarity score is retained for fusion and fine screening.

[0142] S53. Knowledge Graph Relationship Retrieval: Map the elements and constraints parsed in the description to knowledge graph nodes, expand the relationships according to the adjacency or limited step size path, extract structured evidence and convert it into readable fragments, and save the hit nodes, edges and paths as interpretable evidence.

[0143] S54. Results Merging and Deduplication: Deduplication is performed on the results from the keyword, semantic, and graph sources to remove duplicates based on similarity and approximation. The concentrated proportion of single sources is broken up, and identical or highly similar segments are merged to form a unified candidate set while retaining the unique source anchor.

[0144] S55. Reason Labeling and Hierarchical Organization: Add reason labels to each candidate, such as keyword hit, semantic proximity, or graph path, and give a basic score; group and archive by theme, element, or timeline for use in subsequent S6.

[0145] Furthermore, the process of fine screening and controlled inference output (S6) includes:

[0146] S61, Re-ranking and Fine Screening: The relevance-based fine screening model is used to perform pairwise discrimination and scoring on "user description - candidate fragments". High-scoring evidence is selected by combining threshold, Top-K or length budget, prioritizing the diversity of sources and element coverage, outputting controlled context and retaining source anchors.

[0147] S62, Chain-based investigation analysis: Reasoning step by step within a controlled context according to a predetermined chain, such as the cause of the fire, the spread path, the responsible factors, and the handling or rectification suggestions. Each step only uses controlled evidence to generate intermediate key points and partial conclusions.

[0148] S63. Interpretable Result Generation: Produces structured conclusions, including textual conclusions, a list of key points, a timeline or process diagram, and a knowledge graph path snapshot. It attaches a list of evidence and reference fragments for each key conclusion and provides the metadata required for export and archiving, such as version number, timestamp, and evidence number.

[0149] S64. Result Standardization and Verification: Without altering the substance of the conclusion, perform format standardization and consistency checks on the output, control the number of words or tokens and apply templates, unify time, units, numbering and terminology, mark obvious contradictory items with "requires verification" and generate a brief list.

[0150] Furthermore, the interaction and parameter control (S7) process includes:

[0151] S71, Function Switches: Provides switches for search enhancement, graph relationship search, hypothetical answer expansion and reordering fine screening, and supports one-click start / stop and combination by task.

[0152] S72. Strategy Parameters: The currently adjustable parameters are sampling temperature and maximum number of context entries. Other parameters, such as block size, fusion weight, graph expansion radius, and Top-K or length budget, are built-in defaults and can be enabled in future versions.

[0153] S73. Traceability and Versioning: Record processing logs, evidence numbers, map versions and change details, as well as parameter snapshots during retrieval and generation.

[0154] As an optional implementation method, refer to Figure 2 This application provides a method for assisting in the investigation of large-scale accidents based on knowledge graph-enhanced retrieval, comprising the following steps:

[0155] S1. Loading the Basic Knowledge Graph: When the system starts, it loads and registers a pre-built accident investigation (such as fire investigation) foundation knowledge graph (basic knowledge graph) as an authoritative knowledge source for retrieval and reasoning. This graph explicitly expresses the investigation elements and their causal, attribute, and hierarchical relationships. It only performs loading and availability verification, without re-extracting or rewriting; at the same time, it establishes necessary indexes to support subsequent relationship retrieval and interpretable display.

[0156] S2. User-uploaded data: When the user provides supplementary report data (original accident investigation data), the system performs Chinese-friendly parsing, cleaning, and segmentation, generating processable segments with source anchors for image integration. If no data is uploaded, this step is automatically skipped, and the process proceeds directly from S1 to S4.

[0157] S3. Knowledge Graph Extraction and Generation: For the knowledge graph generated from the fragments produced in S2, the S1 base graph (basic knowledge graph) is generated using the same method. "Label-value" is extracted according to the investigation labels, identifying attributes, causal relationships, subordinate relationships, and parallel relationships. Synonym unification and alias merging are performed, and confidence assessment and conflict resolution are conducted for contradictory records. The processed elements and relationships are merged with the incremental base graph for deduplication and unified naming, forming a traceable, unified accident investigation knowledge graph.

[0158] S4. User describes known information: Users describe known facts, clues and constraints in natural language (user query description information); the system completes element identification, intent parsing and constraint extraction, and maps key information to a graph or standardized fields to form a structured retrieval request and reasoning starting point (structured conditional data), providing targets and boundaries for subsequent fusion retrieval and controlled reasoning.

[0159] S5. Fusion Retrieval and Evidence Aggregation: Based on the structured conditional data from S4, a three-way retrieval is performed: keyword retrieval for high-precision literal matching, semantic retrieval using a semantic representation model to recall semantic neighbors in the vector space, and graph relationship retrieval obtaining structured evidence through neighbor and path expansion. The results from the three approaches are merged and deduplicated to form a unified candidate evidence set, accompanied by its source and brief explanation, ensuring comprehensive and traceable evidence.

[0160] S6. Fine Screening and Controlled Inference Output: A relevance fine screening model is used to perform pairwise discrimination scoring on "user description - candidate evidence" and select highly relevant evidence to form a controlled context. Under the constraints of this context, chain analysis is performed according to the investigation link to cover the cause of fire and the spread path. Finally, interpretable conclusions with citations and map paths are output, and timelines, key point lists and necessary visualization views are generated simultaneously.

[0161] S7. Interaction and Parameter Control: Provides function switches for search enhancement, graph relationship retrieval, hypothetical answer expansion and reordering screening, as well as strategy parameters such as sampling temperature and maximum number of context entries; records processing logs, evidence numbers and graph versions.

[0162] Furthermore, user-uploaded data is triggered when the user provides supplementary information. The system opens the data access channel, performs unified encoding and formatting processing on Chinese documents and scanned copies, and cleans up noise such as headers and footers; then, it intelligently segments sentences based on Chinese punctuation and title styles, and divides them into standardized segments using a "window + overlap" strategy. Simultaneously, it binds each segment with anchor points such as document name, page number / paragraph, and character range for subsequent traceability and auditing. See also Figure 3 Specifically, it includes:

[0163] S21. Data Receipt and Metadata Registration. The system uniformly encodes and standardizes the uploaded documents (original accident investigation data), forming standard text and registering metadata, represented as follows:

[0164]

[0165]

[0166] in For the first k Original document (original accident investigation data); For unified coding; Includes unified simplified and traditional Chinese characters, full-width / half-width characters, and common unit standards; To standardize the text; The metadata for registration includes source (registration source), time (time), uploader (uploader), and version (version information).

[0167] S22. Chinese parsing and segmentation. After removing noise such as headers and footers and segmenting sentences, a sliding window strategy is used to... Slice the data to obtain fragments and their anchor points:

[0168]

[0169] ;

[0170] ;

[0171] in, L Window length (in characters) The overlap rate, S Step size; The character range of the fragment in the text; `page` is the anchor point for the segment; `page` is the page number of the segment; `para` is the paragraph position of the segment.

[0172] This step outputs the standardized set:

[0173] ;

[0174] As input for subsequent S3 knowledge graph extraction and generation.

[0175] Furthermore, the knowledge graph extraction and generation process takes the standardized fragment set produced by S2 as input, and sequentially completes element extraction, relation identification, normalization and merging, confidence determination, and incremental graph input according to preset inspection tags and relationship patterns. See also Figure 4 The details are as follows:

[0176] S31, Field Extraction. For each fragment... According to the label list Extract the "tag-value" information (such as fire time / location, combustibles, damage, electrical and structural elements, etc.) and standardize the format to obtain the element set:

[0177] ;

[0178] in, For inspection of labels, Standardized values ​​(time / location / number / unit are consistent). Anchor points for evidence (including document name, page / segment, and character range).

[0179] S32. Relationship Identification. Combining trigger words, sentence structure, and basic syntactic clues, identify attribute, causal, subordinate, and parallel relationships between elements to form a relationship set.

[0180] ;

[0181] in, For a set of relation types, For entities mapped from elements, Indicates the relation type, It serves as the anchor point for supporting sentences; clues within the same sentence and across sentences can be grouped into one relationship.

[0182] S33. Semantic unification and alias merging. This involves unifying entity names and values ​​using synonyms / aliases and merging their hierarchical levels.

[0183] ;

[0184] in, To merge data, To achieve a unified mapping (handling aliases, full / half-width characters, numeric notation, and abbreviations; merging locations / organizations / equipment by level), thereby reducing node fragmentation and improving consistency.

[0185] S34. Confidence Assessment and Conflict Resolution. Calculate confidence levels for elements and relationships, and adjudicate or label conflicting records:

[0186] ;

[0187] in, To evaluate the recorded data, x The object to be evaluated; , and These are weighting coefficients, used to balance the relative importance of the three factors—source weight, expressive clarity, and cross-evidence consistency—in the final confidence score. Source weight (determined by metadata such as upload time / record type). To express clarity (e.g., whether it is quantified, whether it is numbered). For consistency across evidence, The function is a normalization function; when the confidence difference between candidates is less than the threshold. The records are simultaneously marked and marked with "Pending Verification".

[0188] S35. Incremental Merging and Source Tracing. The standardized entities and relations obtained in this batch are combined into an incremental subgraph. , and base map After merging, deduplication, and unifying naming, we obtain the updated unified graph:

[0189] ;

[0190] At the same time, the evidence anchor points and metadata are attached to the corresponding nodes / edges, and the version number and change log are recorded to achieve full traceability.

[0191] Furthermore, the fusion retrieval and evidence aggregation, using the structured conditions obtained from S4 parsing as input, simultaneously conduct keyword retrieval, semantic retrieval, and graph relationship retrieval around the user's question. At the merging point, deduplication and organization are performed to form a candidate evidence set that can be used for subsequent fine-tuning. See also Figure 5 The details are as follows:

[0192] S51. Keyword Search. Query based on inverted index. q With document fragments d Perform literal matching and calculate literal similarity:

[0193] ;

[0194] Before selecting scores fragments as a set It also records the hit words and source anchors.

[0195] in, This represents a commonly used scoring function based on term frequency / inverse document frequency and length normalization. This represents the recall count for the keyword channel.

[0196] S52. Semantic Retrieval. A semantic representation model is used to map text to a vector space, denoted as embedding. Similarity:

[0197] ;

[0198] Before recalling similarity on a pre-built vector index A collection of fragments And retain the similarity score.

[0199] S53. Knowledge Graph Relationship Retrieval. Mapping the elements and constraints parsed from the query to a set of knowledge graph nodes. In the map Press the upper limit of step size Expand to obtain the path set:

[0200] ;

[0201] The structured evidence involved in the path is transcribed into readable fragments to form a set. Its path description.

[0202] in, For the mapping of elements to nodes, Expand the radius (step size) of the relationship.

[0203] S54. Result Merging and Deduplication. Merge and combine the three results:

[0204] ;

[0205] in This indicates that duplicates from the same source / approximate source are removed while retaining the unique source anchor point; simultaneously, a fusion score is calculated for the retained items:

[0206] ;

[0207] Obtain the sorted unified candidate set . Normalize the scores for each channel. For weight fusion.

[0208] S55, Reason Labeling and Hierarchical Organization. For Each candidate is accompanied by a tag with a summary of evidence:

[0209] ;

[0210] Grouped by theme or element bucket:

[0211]

[0212] The output "candidate fragment + source anchor + reason label (and optional grouping)" serves as the input for S6 reordering and controlled reasoning.

[0213] Furthermore, the refined screening and controlled inference outputs use the unified candidate evidence set obtained from S5 as input. First, a controlled context is constructed through relevance-based refined screening. Then, under the constraints of this context, analysis and generation are carried out according to the investigation link, resulting in interpretable results with citations and graph paths. See also Figure 6 The details are as follows.

[0214] S61. Re-ranking and Refinement Screening. A relevance-based refinement screening model is used to score the "user description - candidate fragment" pairs, and the relevance is denoted as:

[0215] ;

[0216] In length budget Select controlled context :

[0217]

[0218] in, For user description, For candidate set, For the segment length, For similarity, This is a redundancy reduction coefficient. The objective balances relevance and source / content diversity, outputting controlled contexts sorted in descending order of score. And retain the source anchor and reason label.

[0219] S62, Chain-based exploration analysis. In a controlled context. Within, reasoning is conducted step-by-step according to the established chain, forming intermediate key points and interim conclusions:

[0220] ;

[0221] Each step can only be referenced. The evidence is presented, and when conflicts arise, branches and explanations are provided (corresponding evidence numbers and graph paths are marked). Ultimately, a comprehensive problem-oriented judgment is obtained. .

[0222] S63. Interpretable Result Generation. Generate structured output based on a template:

[0223]

[0224] Each key conclusion is accompanied by a list of evidence and cited excerpts, and a snapshot of the graph path and metadata (version, timestamp, evidence number) are exported simultaneously for easy review and archiving.

[0225] S64. Result Normalization and Verification. Perform format and consistency processing on the output:

[0226] ;

[0227] in Standardize time / unit / numbering and terminology. The time sequence, unit range, location / numbering consistency, and legal clause references were checked. Mark the items that "require review" on the list. If necessary, return the problematic points to S61 / S5 for partial recalculation.

[0228] Furthermore, interaction and parameter control are used to configure strategies before execution and to solidify traces and versions after execution, supporting reproduction, comparison, and rollback. See also Figure 7 The details are as follows:

[0229] S71, Function Switch. System receives switch vector:

[0230] ;

[0231] This will start or stop the corresponding process module accordingly. Influences whether or not a fusion search is performed. Control map relationship retrieval. Controlling the pre-hypothetical expansion of semantic retrieval Control reordering and fine-tuning. Enable the set denoted as... .

[0232] S72, Strategy Parameters. The system receives the following parameter pairs:

[0233] ;

[0234] in The sampling temperature is used to adjust the generation of diversity and determinism. This is the maximum number of context entries (limiting the maximum amount of evidence allowed to enter a controlled context). Controlled contexts are defined by...

[0235]

[0236] Confirmed, among which As a candidate set for S5, This is the relevance score.

[0237] S73, Traceability and Versioning. System generates runtime snapshots.

[0238]

[0239] And write it into the processing log and map change record, among which This is the current version number of the knowledge graph. For the set of cited evidence and its anchor points, t This is a timestamp. Snapshots are used for result reproduction, difference comparison, and version rollback; audit summaries can be exported in batches if necessary.

[0240] This application also provides a large-scale accident investigation auxiliary system based on knowledge graph enhanced retrieval to implement the aforementioned method. The system includes:

[0241] The basic knowledge graph loading module is used to load and register pre-built basic knowledge graphs for accident investigation; the basic knowledge graphs are used to provide knowledge sources for subsequent retrieval and reasoning.

[0242] The document access and Chinese parsing module is used to parse and segment the original accident investigation data to obtain a collection of fragments;

[0243] The knowledge graph management module is used to update the basic knowledge graph based on the fragment set to obtain a unified accident investigation knowledge graph;

[0244] The user description parsing module is used to obtain user query description information, perform intent parsing and constraint extraction on user query description information, and obtain structured conditional data.

[0245] The retrieval and indexing module is used to perform retrieval based on structured conditional data and a unified accident investigation knowledge graph to obtain a unified candidate evidence set;

[0246] The reasoning and generation module is used to analyze and reason about the unified candidate evidence set to obtain explanatory result data with evidence citations and graph paths.

[0247] The large-model accident investigation auxiliary system based on knowledge graph enhanced retrieval provided in this application also includes an interaction and parameter control module;

[0248] The interaction and parameter control module is used to provide function switches, policy parameters, and logging and version records.

[0249] The large-model accident investigation auxiliary system based on knowledge graph enhanced retrieval provided in this application also includes a data storage device;

[0250] The data storage device is used to store basic knowledge graphs, data fragments, inverted and vector indexes, operation logs, and version information.

[0251] The system in this embodiment includes a document access and Chinese parsing module, a knowledge graph management module, a retrieval and indexing module, a relevance screening module, a reasoning and generation module, an interaction and parameter control module, and a data storage device; under the control of a computing processor, the above modules work together to complete loading. Process and divide into blocks Incremental image generation Fusion retrieval Fine screening construction U and output and will and Information such as archives is used to support reproduction and auditing.

[0252] It is understood that the content of the above method embodiments is applicable to this system embodiment. The specific functions implemented in this system embodiment are the same as those in the above method embodiments, and the beneficial effects achieved are also the same as those achieved in the above method embodiments.

[0253] This application also provides an electronic device, which includes a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the above-described method. This electronic device can be any smart terminal, including tablet computers, desktop computers, etc.

[0254] It is understood that the content of the above method embodiments is applicable to this device embodiment. The specific functions implemented by this device embodiment are the same as those of the above method embodiments, and the beneficial effects achieved are also the same as those achieved by the above method embodiments.

[0255] This application also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described method.

[0256] It is understood that the content of the above method embodiments is applicable to this storage medium embodiment. The specific functions implemented in this storage medium embodiment are the same as those in the above method embodiments, and the beneficial effects achieved are also the same as those achieved in the above method embodiments.

[0257] This application also provides a computer program product, including a computer program that, when executed by a processor, implements the above-described method.

[0258] It is understood that the content of the above method embodiments is applicable to the embodiments of this program product. The specific functions implemented by the embodiments of this program product are the same as those of the above method embodiments, and the beneficial effects achieved are also the same as those achieved by the above method embodiments.

[0259] Memory, as a non-transitory computer-readable storage medium, can be used to store non-transitory software programs and non-transitory computer-executable programs. Furthermore, memory may include high-speed random access memory, and may also include non-transitory memory, such as at least one disk storage device, flash memory device, or other non-transitory solid-state storage device. In some embodiments, memory may optionally include memory remotely located relative to the processor, and these remote memories can be connected to the processor via a network. Examples of such networks include, but are not limited to, the Internet, intranets, local area networks, mobile communication networks, and combinations thereof.

[0260] This application provides a method, system, electronic device, storage medium, and program product for assisting large-scale accident investigation based on knowledge graph-enhanced retrieval. The method includes: loading a pre-built basic knowledge graph for accident investigation as an authoritative knowledge source; receiving and parsing user-supplemented information as needed, cleaning and segmenting it according to Chinese rules; extracting elements and identifying relationships from fragments, completing synonym unification and conflict resolution, and incrementally incorporating them into the knowledge graph; simultaneously performing keyword retrieval, semantic retrieval, and graph relationship retrieval based on the user's natural language description, and merging and deduplicating the results; reordering candidate evidence based on relevance, constructing a controlled context, and driving the large model to complete a chain analysis of the cause of fire and its spread path; outputting an interpretable conclusion and key point list with supporting evidence citations and graph paths, and archiving parameters and version information for verification. This application's solution can achieve high coverage recall and traceable reasoning of multi-source evidence, significantly improving the accuracy, timeliness, and consistency of accident investigation, and is suitable for professional scenarios such as fire accident investigation and on-site inspection. It reduces the risk of omissions and missing connections in manual retrieval by using graph-based modeling of existing interdisciplinary professional knowledge to explicitly express the elements of the investigation and their causal, attribute, and hierarchical relationships. Based on the mechanisms of Chinese data parsing and segmentation, element extraction and synonym unification, conflict resolution and source tracing, it unifies the records from multiple sources and ensures the traceability of evidence. It adopts a fusion strategy of keyword retrieval, semantic retrieval and graph relationship retrieval to comprehensively recall text and structured evidence related to the problem. It introduces relevance screening to construct a controlled context, suppressing irrelevant content and reasoning bias while ensuring the verifiability of cited basis. It implements chain-like investigation analysis within the controlled context, which can cover the cause of fire, the path of fire spread, the identification of responsible elements and disposal / rectification suggestions, forming an interpretable closed loop of "evidence-reasoning-conclusion". Combined with result standardization and consistency verification, it improves the professionalism and stability of the output. Through parameterized switching and incremental aggregation, it continuously absorbs new cases, new standards and on-site data, so as to dynamically optimize the knowledge graph and retrieval effect. At the same time, it retains processing logs, versions and evidence anchors to facilitate reproduction, auditing and rollback. The system of this application consists of a computing processor, a document access and Chinese parsing module, a knowledge graph management module, a retrieval and re-ranking module, a reasoning and report generation module, an interaction and parameter control module, and a data storage device. The data storage device is used to store the base knowledge graph, user data fragments, inverted and vector indexes, operation logs, and version information. The computing processor, in collaboration with the modules, completes the following: loading the base graph; optionally receiving and parsing user data and incrementally importing it into the graph; performing a three-way fusion retrieval and re-ranking based on user descriptions using keywords, semantics, and the knowledge graph, constructing a controlled context; driving the large model to complete analyses such as fire cause, spread path, responsibility factors, and compliance verification, outputting an interpretable report with evidence citations and graph paths; simultaneously supporting on / off switches for retrieval enhancement, graph retrieval, hypothetical expansion, and re-ranking, as well as parameter settings such as sampling temperature and the number of context entries, and archiving the results and version information for easy reproduction and auditing.In summary, this application significantly improves the information coverage, analysis accuracy, processing efficiency, and traceability of conclusions in accident investigation, and enhances the ability to support decision-making and the feasibility of engineering implementation in complex scenarios.

[0261] The embodiments described in this application are for the purpose of more clearly illustrating the technical solutions of the embodiments of this application, and do not constitute a limitation on the technical solutions provided by the embodiments of this application. As those skilled in the art will know, with the evolution of technology and the emergence of new application scenarios, the technical solutions provided by the embodiments of this application are also applicable to similar technical problems.

[0262] Those skilled in the art will understand that the technical solutions shown in the figures do not constitute a limitation on the embodiments of this application, and may include more or fewer steps than shown, or combine certain steps, or different steps.

[0263] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs.

[0264] Those skilled in the art will understand that all or some of the steps in the methods disclosed above, as well as the functional modules / units in the systems and devices, can be implemented as software, firmware, hardware, or suitable combinations thereof.

[0265] The terms “first,” “second,” “third,” “fourth,” etc. (if present) in the specification and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this application described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms “comprising” and “having,” and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0266] It should be understood that in this application, "at least one (item)" means one or more, and "more than" means two or more. "And / or" is used to describe the relationship between related objects, indicating that three relationships can exist. For example, "A and / or B" can represent three cases: only A exists, only B exists, and both A and B exist simultaneously, where A and B can be singular or plural. The character " / " generally indicates that the preceding and following related objects are in an "or" relationship. "At least one (item) of the following" or similar expressions refer to any combination of these items, including any combination of single or plural items. For example, at least one (item) of a, b, or c can represent: a, b, c, "a and b", "a and c", "b and c", or "a and b and c", where a, b, and c can be single or multiple.

[0267] In the several embodiments provided in this application, it should be understood that the disclosed apparatus and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of the units described above is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or units may be electrical, mechanical, or other forms.

[0268] The units described above as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0269] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0270] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes multiple instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this application. The aforementioned storage medium includes various media capable of storing programs, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0271] The preferred embodiments of the present application have been described above with reference to the accompanying drawings, but this does not limit the scope of the claims of the present application. Any modifications, equivalent substitutions, and improvements made by those skilled in the art without departing from the scope and substance of the embodiments of the present application shall be within the scope of the claims of the present application.

Claims

1. A large model accident investigation auxiliary method based on knowledge graph enhanced retrieval, characterized in that, The method comprises the following steps: loading and registering a prefabricated accident investigation basic knowledge graph; the basic knowledge graph is used to provide a knowledge source for subsequent retrieval and reasoning; analyzing and block processing original accident investigation material data to obtain a fragment set; updating the basic knowledge graph according to the fragment set to obtain a unified accident investigation knowledge graph; obtaining user query description information, performing intent analysis and constraint extraction on the user query description information to obtain structured condition data; retrieving according to the structured condition data and the unified accident investigation knowledge graph to obtain a unified candidate evidence set; analyzing and reasoning the unified candidate evidence set to obtain explanation result data with evidence reference and graph path; completing accident investigation assistance according to the explanation result data.

2. The method of claim 1, wherein, The analyzing and block processing of the original accident investigation material data to obtain the fragment set comprises: obtaining original accident investigation material data; uniformly encoding and text normalizing the original accident investigation material data to obtain document record data; the document record data comprises normalized text and metadata; the metadata comprises registration source, time, uploader and version information; performing sentence breaking on the document record data according to a preset style, adopting window and overlapping strategy for block processing to obtain a fragment set; the fragment set comprises a plurality of different fragments.

3. The method of claim 1, wherein, The updating of the basic knowledge graph according to the fragment set to obtain the unified accident investigation knowledge graph comprises: performing information extraction and data standardization processing on the fragments in the fragment set according to a preset investigation label to obtain field information; performing relationship recognition according to the field information to obtain relationship recognition data; performing semantic normalization and alias merging according to the relationship recognition data to obtain merged data; performing confidence evaluation and conflict resolution according to the merged data to obtain evaluation record data; performing increment and homology and tracing according to the evaluation record data and the basic knowledge graph to obtain the unified accident investigation knowledge graph.

4. The method of claim 3, wherein, The confidence evaluation and conflict resolution according to the merged data to obtain the evaluation record data comprises: according to the merged data, taking data source type, expression clarity and consistency as elements and relationship scoring to obtain scoring information; performing arbitration or marking on contradictory records of the merged data according to source priority, time recency and consistency to obtain conflict resolution data; obtaining evaluation record data according to the scoring information and the conflict resolution data.

5. The method of claim 1, wherein, The retrieving according to the structured condition data and the unified accident investigation knowledge graph to obtain the unified candidate evidence set comprises: performing literal matching on the structured condition data based on an inverted structure, selecting a plurality of results reaching a preset accuracy to obtain keyword retrieval data; mapping the structured condition data and the fragments of the fragment set into vectors respectively and recalling semantic neighbor evidence according to cosine similarity to obtain semantic retrieval data; Map the parsed elements in the structured condition data to constraints to obtain a node set, perform relation expansion according to the node set and the unified accident investigation knowledge graph to obtain graph relation retrieval data; De-duplicate the keyword retrieval data, the semantic retrieval data and the graph relation retrieval data and retain unique source anchors to obtain a unified candidate set; Reasonably annotate and group archive each piece of candidate information in the unified candidate set to obtain a unified candidate evidence set.

6. The method of claim 1, wherein, The unified candidate evidence set is analyzed and reasoned to obtain explanation result data with evidence references and graph paths, including: The user query description information and the unified candidate evidence set are pair-wise judged and scored and re-ordered by a correlation precision screening model, and evidence in a preset score range is selected to obtain a controlled context; Chain investigation analysis is performed according to the controlled context, and pre-set link step-by-step reasoning is performed to obtain intermediate points and local conclusions; The intermediate points and local conclusions are structured to sequentially attach evidence references and graph paths to obtain structured conclusions; The structured conclusions are checked to obtain explanation result data.

7. The method of claim 1, wherein, The method further includes: Providing function switches; the function switches include retrieval enhancement switches, graph relation retrieval switches, hypothetical answer expansion switches, re-ordering precision screening switches and task one-key start-stop and combination switches; Providing strategy parameters; the strategy parameters include sampling temperature parameters and maximum context number parameters; Providing traces and version records; the traces and version records include processing logs, evidence numbers, graph versions and graph change details.

8. A large model accident investigation auxiliary system based on knowledge graph enhanced retrieval, for implementing the method of any one of claims 1 to 7, characterized in that, The system includes: A basic knowledge graph loading module for loading and registering a prefabricated basic knowledge graph of accident investigation; the basic knowledge graph is used to provide a knowledge source for subsequent retrieval and reasoning; A document access and Chinese analysis module for analyzing and block processing original accident investigation material data to obtain a fragment set; A knowledge graph management module for updating the basic knowledge graph according to the fragment set to obtain a unified accident investigation knowledge graph; A user description analysis module for obtaining user query description information, performing intent analysis and constraint extraction on the user query description information to obtain structured condition data; A retrieval and indexing module for retrieving according to the structured condition data and the unified accident investigation knowledge graph to obtain a unified candidate evidence set; A reasoning and generation module for analyzing and reasoning the unified candidate evidence set to obtain explanation result data with evidence references and graph paths.

9. The system of claim 8, wherein, The system further includes an interaction and parameter control module; The interaction and parameter control module is used to provide function switches, strategy parameters and traces and version records.

10. The system of claim 8, wherein, The system further includes a data storage; The data storage is used to save basic knowledge graphs, material fragments, inverted and vector indexes, running logs and version information.

Images