Method and device for constructing mapping knowledge domain in high-altitude falling field, and medium

Abstract

The invention discloses a high-altitude falling domain knowledge graph construction method and device and a medium, and the method comprises the steps: collecting and integrating a falling protection equipment standard file and a high-altitude falling accident investigation report, and carrying out the preprocessing of data; concepts, attributes and relationships in the knowledge graph are defined, so that the consistency and expandability of data are ensured; manually extracting entity information according to the falling protection equipment standard file; for a high-altitude falling accident survey report, performing entity extraction based on a self-training domain named entity recognition tool, segmenting a text based on a trigger word, extracting phrases based on unsupervised clustering and a large language model, and performing relation extraction based on rules; and according to FPS design requirements, meeting retrieval requirements, including query through an FFH-ARs entity, query through an FPE-S entity and query through a multi-attribute combination. According to the invention, the characteristics and details of each accident can be displayed from multiple angles, the design of a falling protection system is optimized, and the high-altitude operation safety is improved.

Description

Technical Field

[0001] The present invention belongs to the field of computer technology, and specifically relates to a method, device and medium for constructing a knowledge graph in the field of falling from height. Background Art

[0002] High-altitude work is common in industries such as construction, electricity, and communications. With the increasing risk of falling from heights, effective fall protection systems and equipment are essential to ensure the safety of workers. However, the design and selection of fall protection systems are complex and depend on the actual construction environment and experience. Historical accident data and accident investigation reports provide valuable references for design, but the information is scattered and difficult to integrate and utilize. Knowledge graphs improve information retrieval efficiency by integrating and structuring knowledge, but in the field of fall protection, existing knowledge graphs cannot meet the needs and lack systematic integration of fall protection equipment standards and accident experience. Therefore, developing a high-altitude fall knowledge graph that integrates two heterogeneous files, fall protection equipment standards and accident investigation reports, and integrating relevant standards and actual experience is of great significance for improving the scientificity and practicality of fall protection system design, enhancing the safety awareness and skills of high-altitude workers, and reducing accidents. Summary of the invention

[0003] Purpose of the invention: The present invention proposes a method, device and medium for constructing a knowledge graph in the field of falls from heights, which significantly improves the efficiency and safety of the design of fall protection systems for high-altitude operations.

[0004] Invention content: The method for constructing a knowledge graph in the field of falling from heights described in the present invention comprises the following steps:

[0005] Data collection: Collect and integrate fall protection equipment standard documents and height fall accident investigation reports, and pre-process the data;

[0006] Model layer construction: defines the concepts, attributes, and relationships in the knowledge graph to ensure data consistency and scalability;

[0007] Data layer construction: Automatically extract key information from standard documents and accident investigation reports: Manually extract entity information for standard documents on fall protection equipment; for height fall accident investigation reports, perform entity extraction based on self-training domain named entity recognition tools, segment text based on trigger words, extract phrases based on unsupervised clustering and large language models, and perform relationship extraction based on rules;

[0008] Knowledge graph application layer: Meet retrieval requirements according to FPS design requirements, including querying through FFH-ARs entities, querying through FPE-S entities, and querying through multi-attribute combinations.

[0009] Furthermore, the data preprocessing includes format unification, text segmentation, removal of stop words and cleaning of garbled information.

[0010] Furthermore, the key information includes specific requirements of FPE, actual case experience of accident reporting, high-altitude working scenarios and types of falls from height.

[0011] Furthermore, the entity extraction implementation process based on the self-training domain named entity recognition tool is as follows:

[0012] NER training corpus preparation, prepare high-quality training corpus in the field;

[0013] Construction of entity dictionary in the field of high-altitude falls, using domain-specific entity dictionary to enhance the domain adaptability of the model;

[0014] Automatically identify and label entities in text based on weakly supervised BIO coding automatic annotation;

[0015] Bi-LSTM-CRF-based NER training, combining deep representation capabilities and sequence labeling capabilities;

[0016] Evaluation of NER model in the field of height fall, using precision, recall and F1 value to evaluate model performance;

[0017] The NER model is applied in the field of height falls to extract entities of specific categories and store them in the knowledge graph.

[0018] Furthermore, the construction and implementation process of the entity dictionary in the field of falling from height is as follows:

[0019] Extract specialized terms and entities from fall protection equipment standard documents and fall-from-height accident reports;

[0020] Build a domain dictionary containing specific entities and terms to enhance the understanding and adaptability of the NER model to the language in the field of height falls.

[0021] Furthermore, the implementation process of the BIO code automatic annotation based on weak supervision is as follows:

[0022] Automatically match entity words in text using domain-specific dictionaries and heuristic rules;

[0023] Entity words are marked using the BIO encoding format to indicate the start, inside, and outside of the entity.

[0024] Furthermore, the implementation process of extracting phrases based on unsupervised clustering and large language model is as follows:

[0025] The preprocessed text is converted into a numerical feature vector using the TF-IDF algorithm; the feature vector is reduced in dimension using the UMAP algorithm; the optimal number of clusters is determined by the elbow rule and the K-Means clustering method; the optimal number of clusters is 26;

[0026] Based on the determined optimal number of clusters, the large language model Kimi is used to analyze the text content in the accident report and complete the classification of the types of high-altitude fall accidents;

[0027] Extract high-altitude work scenes and fall causes based on the big language model; input high-altitude fall accident reports that have been divided into text modules; guide the big language model Kimi through key text content to extract refined information on high-altitude work scenes and fall causes; adopt a strict manual review and proofreading process to comprehensively review the output results of the Kimi model and compare them with the knowledge and experience of field experts; carefully correct any deviations or errors in the model output to improve the reliability of the data.

[0028] A storage medium described in the present invention stores a computer program on the storage medium, and when the computer program is executed by at least one processor, the steps of the method for constructing a knowledge graph in the field of falling from heights as described above are implemented.

[0029] A device according to the present invention includes a memory and a processor, wherein:

[0030] A memory for storing computer programs that can be run on the processor;

[0031] A processor is used to execute the steps of the method for constructing a knowledge graph in the field of falling from height as described above when running the computer program.

[0032] Beneficial effects: Compared with the prior art, the beneficial effects of the present invention are: the present invention can significantly improve the efficiency and safety of the design of fall protection systems for aerial work; by integrating the technical requirements in national standard documents with the practical experience in historical accident investigation reports, a structured and visual database is provided for designers, so that the lessons and solutions of past accidents can be directly learned and applied during the design process; the present invention can more accurately determine the key issues that need attention; this method of combining theory with practice not only improves the scientificity and systematicness of the design, but also provides a more solid guarantee for the safety of aerial work, brings revolutionary progress to the field of fall protection system design for aerial work, and greatly enhances the reliability and practicality of the design. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1 It is a flow chart of the method for constructing a knowledge graph in the field of falling from height proposed by the present invention;

[0034] Figure 2 It is the ontology mode view of the knowledge graph in the field of falling from height proposed by the present invention;

[0035] Figure 3 It is a flow chart of an entity recognition method based on a self-training named entity recognition tool proposed by the present invention;

[0036] Figure 4 This is a diagram of the Bi-LSTM-CRF-based NER model training architecture proposed in the present invention;

[0037] Figure 5 It is a flow chart of the information extraction method based on unsupervised clustering and large language model proposed by the present invention;

[0038] Figure 6 It is a query result graph of the first embodiment of the application of the knowledge graph in the field of falling from height;

[0039] Figure 7 It is a query result graph of the second embodiment of the application of the knowledge graph in the field of falling from height;

[0040] Figure 8 This is a query result graph of Example 3 of the knowledge graph application in the field of falling from heights. DETAILED DESCRIPTION

[0041] The present invention will be further described below in conjunction with the accompanying drawings.

[0042] like Figure 1 As shown, the present invention provides a method for constructing a fall-from-height knowledge graph (FFH-KG), including data collection, model layer construction, data layer construction and application layer, specifically:

[0043] Step 1: Data collection: Collect various data sources related to FFH accidents and fall protection, such as accident reports and safety standards; and perform data preprocessing.

[0044] These data serve as the basic content of KG, and the relevant documents come from government platforms and websites. Text preprocessing includes file format unification, text sentence-level segmentation, and manual deletion of garbled information.

[0045] Step 2: Schema layer construction: According to the characteristics of the collected documents, preliminary organization and text module extraction are performed. At the schema layer, it is necessary to define the concept and relationship model of the KG, namely the ontology. This includes determining entity classes, attributes, and relationships between entities. The core concepts and relationships are determined based on the characteristics of the two document types. The design of the schema layer should effectively integrate expertise from heterogeneous sources: Fall protection equipment standard (FPE-S) documents and Fall-from-height accident reports (FFH-ARs), support information extraction and retrieval, and accommodate future expansion.

[0046] The main body of FFH-KG is as follows Figure 2 As shown in the figure, the ontology is derived from the respective characteristics of the standard documents of fall protection equipment and historical accident investigation reports. It mainly includes two parts: "fall protection system" and "accident" (blue nodes). The key elements of the "fall protection system" include the corresponding "fall protection equipment", the "components" contained in the equipment, the "test equipment" of the equipment involved, the "test methods" that the test equipment needs to adopt, the "performance" of the fall protection system, and the "subcategories" of its further subdivision (green nodes), while "mark", "English name", "definition", "file source", "file number", "equipment requirements", "test steps", "performance requirements", and "technical requirements" are attributes sorted out from the documents (yellow nodes) to further fully explain the entity. The "accident" part sorts out the various content modules such as "accident process", "accident solution", "direct cause of the accident", "indirect cause of the accident", "accident summary", basic information of the accident "province", "year", accident consequences "number of deaths", "number of injuries", and supplements the "file source" attribute of the accident. At the same time, we add knowledge entities extracted from LLM, such as "type of high-altitude fall", "cause of fall", and "high-altitude work scene" (orange nodes). Among them, "cause of fall" is used as the attribute of "high-altitude work scene" (yellow node), and "building elements", "components", "fall protection equipment", and "fall prevention system" entities extracted using the self-training NER tool. The two parts are connected through the "components", "fall protection equipment", and "fall prevention system" entities.

[0047] Step 3: Data layer construction: In the data layer, knowledge instances, including specific entities and relations are extracted. The collected data is mapped to the entities and relations defined in the schema layer. The data layer construction consists of two parts: structuring of FPE-S documents and structuring of FFH-ARs. This process involves named entity recognition (NER), unsupervised clustering, LLM (Large Language Model)-based classification, and information refinement. In the data layer, FFH-KG is constructed using a hybrid natural language processing approach, including expert-based manual extraction, entity recognition based on self-training NER tools, text segmentation based on trigger words, information refinement based on unsupervised clustering and large language models, and rule-based relationship extraction.

[0048] Figure 3 The process of completing entity extraction by a self-training named entity recognition tool in an embodiment of the present invention includes:

[0049] S1: Preparation of NER training corpus: The first step in training a specific named entity recognition (NER) tool for the field of height falls is to prepare high-quality training corpus in the field. The quality, diversity and coverage of the training corpus directly affect the generalization ability and accuracy of the model. The present invention uses the collected 941 FFH-ARs to perform text preprocessing on these files to obtain basic training corpus. These texts are used as the basic corpus for training field NER tools.

[0050] S2: Construction of entity dictionary in the field of falling from height: Using domain-specific entity dictionaries, the NER model can better understand and adapt to the language usage and terminology in a specific field, thereby improving recognition accuracy. The entity dictionary in the field of falling from height contains entities and terminology specific to the field of falling from height, which can enhance the domain adaptability of the model and is also an important tool for completing automatic annotation in the BIO encoding format based on weak supervision. The present invention uses entities manually sorted from FPE-S and entities automatically extracted from Baidu UIE tools to construct an entity dictionary in the field of falling from height, with a total of 1,681 words. The entity dictionary in this field contains entity classes such as "fall prevention system", "fall protection equipment", "components", and "building elements". Among them, the "building element" class is extracted by defining trigger words using Baidu UIE tools, and the defined trigger words include "building elements", "components", "equipment", etc.

[0051] S3: Automatic annotation based on weakly supervised BIO coding: When training domain-specific named entity recognition (NER) tools, the training corpus needs to contain a large amount of entity annotation data. The NER model recognizes and understands the key information in the text by learning these annotated data, and then accurately locates and classifies different entity types. This requires the completion of the entity data annotation work of the training corpus. Manual data annotation work is time-consuming and laborious, and the automatic annotation method based on external dictionaries can effectively improve the efficiency of data annotation. The present invention automatically annotates the text corpus by writing a Python program, using a weakly supervised (i.e., external domain dictionary supervision) learning method and the BIO (Begin, Inside, Outside) encoding format. Using a domain-specific dictionary and heuristic rules for falling from heights, entities in the text are automatically identified and marked, thereby avoiding a large amount of manual annotation work. The heuristic rule adopts the principle of maximum character matching, which can ensure the boundary information of complete terms in the field. This step is crucial in converting unstructured text data into structured data suitable for processing by machine learning algorithms.

[0052] S4: NER training based on Bi-LSTM-CRF: The training of the named entity recognition (NER) model is based on the Bidirectional Long Short-Term Memory Network and Conditional Random Field (Bi-LSTM-CRF) architecture. Bi-LSTM is used to capture contextual information in the text, and CRF is used to constrain the tag sequence. This architecture combines the deep representation capabilities of the Bidirectional Long Short-Term Memory Network and the sequence labeling expertise of the Conditional Random Field to improve its ability to identify and classify entities unique to the field of high-altitude falls. This step divides the labeled data from the previous step into training, validation, and test sets. Use the training set to train the model to allow the model to learn text features; use the validation set to automatically adjust the model parameters; and finally use the test set to evaluate the entity recognition effect of the model. The entire model is trained in a Python environment, calling the Tensorflow deep learning framework.

[0053] The overall structure of the high-altitude fall entity recognition model trained by Bi-LSTM-CRF is as follows: Figure 4 As shown in the figure, the training process of the model is as follows: input natural language sequence; pass the pre-trained word embedding vector as character feature into the model; use Bi-LSTM to encode the feature sequence; then use CRF to predict the tag sequence; the final output of the model is the predicted label corresponding to each character. The pre-trained word embedding vector is trained on the Chinese Wikipedia corpus based on the gensim version of word2vec.

[0054] S5: Evaluation of NER model in the field of high-altitude fall: Precision, recall and F1 value are three evaluation indicators used to evaluate the performance of the high-altitude fall NER model. The model tends to be stable after the 20th iteration, and the evaluation results of the model on each evaluation indicator are: precision: 97.48%; recall: 98.10%; F1 Score: 0.9779.

[0055] S6: Application of NER model in the field of falling from height: The present invention uses a trained domain-specific NER model to perform named entity recognition on the training corpus to extract entities of categories such as "fall prevention system", "fall protection equipment", "components", and "building elements". It was analyzed that there are 5 text modules related to fall protection equipment, namely "accident summary", "accident process", "direct cause of the accident", "indirect cause of the accident", and "accident solution", and 131 FFH-ARs (fall from height accident reports), i.e. 655 text modules, were successfully extracted. These text paragraphs contain both detailed descriptions of various aspects of the high-altitude falling accident and entity information such as "fall prevention system", "fall protection equipment", "components", and "building elements". Entities were extracted from these text modules using the trained domain NER tool, and a total of 2597 entities were obtained. After deduplication, a total of 694 unique entities were extracted, including 24 new entities. The entities extracted from FFH-ARs will be stored in the knowledge graph later.

[0056] Figure 5 The process of information extraction based on unsupervised clustering and large language model in an embodiment of the present invention includes:

[0057] L1: Determination of the number of types of falls from height based on machine learning: The present invention adopts an unsupervised clustering method to determine the number of categories of fall-from-height (TOFFH) accidents. First, 941 FFH accident reports (FFH-ARs) were preprocessed, including word segmentation and removal of stop words. Then, the TF-IDF algorithm was used to convert the text into a numerical feature vector, and the UMAP algorithm was used for dimensionality reduction. The optimal number of clusters was determined to be 8 through the elbow rule and K-Means clustering method.

[0058] L2: Classification of high-altitude fall accidents based on large language models: On the basis of determining the number of high-altitude fall categories, the present invention uses the large language model Kimi to conduct in-depth analysis of the accident report to complete the 8-category classification of high-altitude fall accident types. Kimi supports long text processing and can understand multiple file formats. Based on Kimi's classification results, by reading the information of text modules such as "accident occurrence" and "accident cause", combined with manual review and modification, it is further subdivided into 26 TOFFH categories. Specifically, the five categories of "falling from edge operations", "falling from rooftop operations", "falling related to scaffolding", "falling from building beams, ladders or other structural steels" and "falling related to mechanical equipment" are subdivided into the second level.

[0059] "Falling while working on an edge" can be subdivided into four categories: "falling from the edge of an edge", "falling from an edge opening", "falling from the stairwell" and "falling while working in an elevator shaft"; "rooftop operation accidents" can be subdivided into three categories: "falling from the edge of the roof", "falling from a roof opening" and "falling due to stepping on light roofing materials"; "scaffolding-related falls" can be subdivided into three categories: "falling from the scaffolding due to missing or instability", "falling due to scaffolding collapse" and "falling during scaffolding dismantling or erection"; "falling from building beams, ladders or other structural steel" can be subdivided into three categories: "falling from building beams", "falling from ladders" and "falling from other structural steel"; "mechanical equipment-related falls" can be subdivided into ten categories: "construction elevator operation accidents", "crane operation accidents", "crane operation accidents", "tower crane operation accidents", "single-pole crane operation accidents", "material hoist operation accidents", "hoisting machine operation accidents", "flat car operation accidents", "crane basket operation accidents" and "other mechanical equipment operation accidents".

[0060] Including the remaining three of the original eight categories, “falling during climbing operations,” “falling during mobile platform operations,” and “falling during exterior wall operations,” there are a total of 26 types of falls from heights.

[0061] These five types of data were sorted, input, and answered using the big model. The big language model Kimi further subdivided the 131 accident investigation reports into 26 types of falls from heights. This step not only improved the accuracy of classification, but also provided a basis for the subsequent refinement of high-altitude work scenarios.

[0062] L3: Extraction of high-altitude work scenes and causes of falls based on large language models: The present invention further uses the Kimi model to refine high-altitude work scenes and causes of falls. Input the FFH-ARs that have been divided into text modules, and guide the model to extract refined information through key text content. To ensure the accuracy of the refined results, the present invention adopts a strict manual review and proofreading process, comprehensively reviews the output results of the Kimi model, and compares the knowledge and experience of experts in the field. When any deviations or errors are found in the model output, detailed corrections are performed to improve the reliability of the data.

[0063] Step 4: The application layer is where the KG interacts with the end user. It meets various retrieval requirements according to the FPS design requirements, including querying by FFH-ARs entities, querying by FPE-S entities, and querying by multi-attribute combinations. The query is completed through Cypher query statements. The retrieval results in Neo4j return relevant instances, detailed information, and various requirements from standard documents.

[0064] The knowledge graph is constructed in Neo4j, using nodes and links to complete the structured display of knowledge, organize domain knowledge in a certain logical form, and support query and retrieval using the graph database query language Cypher.

[0065] Example 1: Query relevant instance experience through entities from FFH-ARs, such as high-altitude work scenes, etc.

[0066] Designers can use FFH-KG to retrieve historical accident cases for specific aerial work scenarios, analyze the relationship between the cause of the accident and the fall protection system, as well as the fall protection equipment and precautions involved. For example, for the rooftop color steel tile laying operation, designers can search for accidents in similar scenarios, search for related accidents and their specific information for the aerial work scenario "color steel tile laying" through the following query statement, and return the aerial work scenarios, fall protection equipment, and connected nodes involved in the accident:

[0067] MATCH(node1:`High-altitude work scene`)-[*..3]-(node2:`Fall protection equipment`)

[0068] WHERE node1.name CONTAINS "Colored Steel Tile Laying"

[0069] RETURN node1,node2

[0070] MATCH(node1:`High-altitude operation scenario`)-[*..3]-(node2:`Fall protection equipment`): This indicates matching two class nodes (node1 and node2) in the graph, where the label of node1 is "High-altitude operation scenario" and the label of node2 is "Fall protection equipment". -[*..3]- means that there can be any relationship with 0 to 3 layers (including directly connected or indirectly connected through other nodes) between these two nodes.

[0071] WHERE node1.name CONTAINS "Color steel tile laying": This condition filters out all node1 nodes whose name attribute contains the string "Color steel tile laying".

[0072] RETURN node1,node2: The query result returns the node1 and node2 nodes that meet the conditions.

[0073] The Match query form is used to retrieve all nodes and relationships in the graph database. The Return query form is used to return a subset of all variables or constrained nodes and relationships. The Where query expression is used to include nodes or relationships with a certain attribute value. The Contains query form is used to retrieve all relevant content containing a certain keyword in the graph database. The variable node1 represents the "High-altitude operation scenario" class node with the name "Color steel tile laying". node2 represents all "Fall protection equipment" class nodes involved.

[0074] More information can be obtained for the nodes in the query result through the node expansion function, such as Figure 6 As shown, entity nodes of different colors represent different attributes and text modules of the accident, including the location, time, number of casualties, fall type, building elements, etc. of the accident, as well as text modules such as the accident process, cause, and solution. This information is presented in a logical structure and intuitive way, facilitating designers to quickly understand the accident situation, learn from experience, and improve the design idea.

[0075] Example 2: Query relevant instance experiences through the fall protection system or fall protection equipment.

[0076] Designers can query accident reports related to specific fall protection systems or equipment through FFH-KG, expand to view the accident details, and the query statement is as follows:

[0077] MATCH(a:`Fall prevention system`{name:"Safety belt"})-[r]-(b),

[0078] (a:`Fall prevention system`{name:"Safety belt"})-[r1]-(node)-[r2]-(c:`Accident`)

[0079] RETURN a,r,b,c

[0080] This query language is used to query information related to the "safety belt" of the fall protection system, and also includes related information from standard documents and accident reports. Figure 7 The visualization results after clicking the node, combined with the search results, allow designers to think about and answer questions about the actual use of FPE, how accidents occur, and the solutions and reference standards to avoid similar accidents.

[0081] Embodiment 3: Match relevant instance experience through customized multi-condition combinations such as high-altitude fall type and high-altitude work scene.

[0082] Designers can use FFH-KG to search for relevant keywords in different types of entity nodes and customize multiple condition combinations to match relevant instance experiences. For example, the query statement used to query rooftop fall accidents is as follows:

[0083]

[0084] This query language is used to query roof fall accidents, where the aerial work scenario is installation work on the "roof". Through Cypher query, find all paths starting from a specific high-altitude fall type node and connecting to a specific high-altitude work scenario node, where the name of the high-altitude fall type node contains the keyword "roof" and the name of the aerial work scenario node contains the keyword "installation". Figure 8 The query results that meet the parameter conditions are returned, including high-altitude fall accidents and their attribute information related to roof installation work. By filtering with specific conditions, designers can quickly retrieve relevant accident instances, gain an in-depth understanding of the causes of falls in specific high-altitude work scenarios, and design fall protection systems in a targeted manner.

[0085] The present invention also provides a storage medium having a computer program stored thereon, and when the computer program is executed by at least one processor, the steps of the method for constructing a knowledge graph in the field of falling from height as described above are implemented.

[0086] A device described in the present invention includes a memory and a processor, wherein: the memory is used to store a computer program that can be run on the processor; the processor is used to execute the steps of the method for constructing a knowledge graph in the field of falling from heights as described above when running the computer program.

[0087] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principle of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. A method for constructing a knowledge graph in the field of falling from heights, characterized in that: The implementation process is as follows: Data collection: Collect and integrate fall protection equipment standard documents and height fall accident investigation reports, and pre-process the data; Model layer construction: defines the concepts, attributes, and relationships in the knowledge graph to ensure data consistency and scalability; Data layer construction: Automatically extract key information from standard documents and accident investigation reports: Manually extract entity information for standard documents on fall protection equipment; for height fall accident investigation reports, perform entity extraction based on self-training domain named entity recognition tools, segment text based on trigger words, extract phrases based on unsupervised clustering and large language models, and perform relationship extraction based on rules; Knowledge graph application layer: Meet retrieval requirements according to FPS design requirements, including querying through FFH-ARs entities, querying through FPE-S entities, and querying through multi-attribute combinations.

2. The method for constructing a knowledge graph in the field of falling from height according to claim 1, characterized in that: The data preprocessing includes format unification, text segmentation, removal of stop words and cleaning of garbled information.

3. The method for constructing a knowledge graph in the field of falling from height according to claim 1, characterized in that: The key information includes the specific requirements of FPE, actual case experience of accident reporting, high-altitude working scenarios and types of falls from height.

4. The method for constructing a knowledge graph in the field of falling from height according to claim 1, characterized in that: The entity extraction implementation process based on the self-training domain named entity recognition tool is as follows: NER training corpus preparation, prepare high-quality training corpus in the field; Construction of entity dictionary in the field of high-altitude falls, using domain-specific entity dictionary to enhance the domain adaptability of the model; Automatically identify and label entities in text based on weakly supervised BIO coding automatic annotation; Bi-LSTM-CRF-based NER training, combining deep representation capabilities and sequence labeling capabilities; Evaluation of NER model in the field of height fall, using precision, recall and F1 value to evaluate model performance; The NER model is applied in the field of height falls to extract entities of specific categories and store them in the knowledge graph.

5. The method for constructing a knowledge graph in the field of falling from height according to claim 4, characterized in that: The implementation process of constructing the entity dictionary in the field of falling from height is as follows: Extract specialized terms and entities from fall protection equipment standard documents and fall-from-height accident reports; Build a domain dictionary containing specific entities and terms to enhance the understanding and adaptability of the NER model to the language in the field of height falls.

6. The method for constructing a knowledge graph in the field of falling from height according to claim 4, characterized in that: The implementation process of the BIO coding automatic annotation based on weak supervision is as follows: Automatically match entity words in text using domain-specific dictionaries and heuristic rules; Entity words are marked using the BIO encoding format to indicate the start, inside, and outside of the entity.

7. The method for constructing a knowledge graph in the field of falling from height according to claim 1, characterized in that: The implementation process of extracting phrases based on unsupervised clustering and large language model is as follows: The preprocessed text is converted into a numerical feature vector using the TF-IDF algorithm; the feature vector is reduced in dimension using the UMAP algorithm; the optimal number of clusters is determined using the elbow rule and K-Means clustering method; Based on the determined optimal number of clusters, the large language model Kimi is used to analyze the text content in the accident report and complete the classification of the types of high-altitude fall accidents; Extract high-altitude work scenes and fall causes based on the big language model; input high-altitude fall accident reports that have been divided into text modules; guide the big language model Kimi through key text content to extract refined information on high-altitude work scenes and fall causes; adopt a strict manual review and proofreading process to comprehensively review the output results of the Kimi model and compare them with the knowledge and experience of field experts; carefully correct any deviations or errors in the model output to improve the reliability of the data.

8. The method for constructing a knowledge graph in the field of falling from height according to claim 7, characterized in that: The optimal number of clusters is 26.

9. A storage medium, characterized in that: The storage medium stores a computer program, which, when executed by at least one processor, implements the steps of a method for constructing a knowledge graph in the field of falls from heights as described in any one of claims 1 to 8.

10. A device, characterized in that: comprising a memory and a processor, wherein: A memory for storing computer programs that can be run on the processor; A processor, used to execute the steps of a method for constructing a knowledge graph in the field of falling from height as described in any one of claims 1 to 8 when running the computer program.

Images