Equipment knowledge data extraction method and system

Abstract

The invention provides an equipment knowledge data extraction method and system. The method is realized through the steps of text data classification and screening, static extraction template determination, static extraction, dynamic extraction template determination, dynamic extraction, equipment knowledge graph construction and the like. According to the method, equipment knowledge data extraction of a whole process is constructed from text data classification and screening, static and dynamic knowledge extraction and an equipment knowledge data knowledge graph, military texts are identified from massive text data with low value density, and high-quality equipment knowledge data are extracted; and support is provided for acquisition, analysis and application of equipment knowledge data in the military field.

Description

Technical Field

[0001] The present invention relates to an equipment knowledge data extraction method and system, belonging to the technical field of knowledge data extraction. Background Art

[0002] With the rapid development of information technology and artificial intelligence technology, open source information data is growing exponentially. How to extract high-quality equipment static functional performance parameters and dynamic combat process knowledge from massive open source data is a hot topic in current research. Knowledge data extraction mainly includes text classification, text mining, language large models, knowledge graphs, entity extraction, event extraction, etc. By extracting high-quality equipment functional performance attribute parameters and dynamic rules from massive open source data, data support can be provided for military intelligence research, equipment design, and equipment application.

[0003] Current existing technologies generally target text relationship mining or entity mining. General knowledge data extraction methods have the following problems when extracting knowledge data in specific equipment fields: First, it is difficult to effectively identify military texts when extracting massive data, and the extraction efficiency is low; second, there is a lack of unified equipment data and knowledge extraction templates, and the extracted knowledge data parameter formats are not unified, making it difficult to support subsequent applications. Summary of the invention

[0004] The purpose of the present invention is to overcome the deficiencies of the prior art and provide a method and system for extracting equipment knowledge data.

[0005] The technical solution of the present invention is: a method for extracting equipment knowledge data, comprising the following steps:

[0006] Step 1: classify and filter the text data from public sources, remove non-military text data, and obtain military text data;

[0007] Step 2, determine a static extraction template. The static extraction template is a set of equipment function performance attribute parameters as elements. Each type of equipment defines the static extraction template of the equipment according to the function performance attribute parameters it is concerned about. The static extraction template at least includes {name, country of origin, function performance attribute parameters};

[0008] Step 3, using the static extraction template determined in step 2 to extract static data of equipment knowledge from the military text data obtained in step 1;

[0009] Step 4: Determine the dynamic extraction template that describes the dynamic rules of the equipment. The dynamic extraction template includes the following relationships:

[0010] The connection relationship is the triggering process: the output of process A can trigger the start of process B;

[0011] The same relationship refers to homogeneous processes: process A and process B are the same process, appearing in different architectures and can replace each other; the inclusion relationship refers to large and small processes): process B is included in the large process A.

[0012] The succession relationship is the continuation of the process: process B can only be executed after process A is executed;

[0013] The calling relationship is the reference process: process A can trigger the start of process B, and process B can also be called by other processes;

[0014] Participation relationship is cross-process: process A includes activity B, but activity B is usually not constructed as a process;

[0015] Step 5: Use the language big model as an extraction tool and combine it with the dynamic extraction template to extract the equipment combat process knowledge from the military text data;

[0016] Step 6: Use the statically extracted equipment knowledge static data and the dynamically extracted equipment combat process knowledge to construct a knowledge equipment knowledge graph.

[0017] A system for extracting equipment knowledge data comprises a memory, a processor and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the steps of the method for extracting equipment knowledge data are implemented.

[0018] The beneficial effects of the present invention compared with the prior art are as follows:

[0019] (1) The present invention constructs a full-process equipment knowledge data extraction from text data classification and screening, static and dynamic knowledge extraction, and equipment knowledge data knowledge graph, realizes the recognition of military texts from massive low-value density text data, and extracts high-quality equipment knowledge data, providing support for the collection, analysis, and application of equipment knowledge data in the military field;

[0020] (2) Compared with the existing artificial intelligence extraction methods, the present invention has high extraction efficiency and standardized equipment data and knowledge generated by extraction, which can effectively support military analysis and application;

[0021] (3) The present invention screens massive text data and preliminarily removes texts that are not related to military themes, thereby effectively improving the efficiency of knowledge data extraction;

[0022] (4) The present invention can form unified standard equipment knowledge data by providing static and dynamic extraction templates. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] Figure 1 It is a diagram of the extraction process of the equipment of the present invention;

[0024] Figure 2 This is a schematic diagram of the general text classification model structure;

[0025] Figure 3 Equipping large language models with dynamic rule extraction based on hints;

[0026] Figure 4 It is a behavior tree of feasible operations on the knowledge graph. DETAILED DESCRIPTION

[0027] The present invention is described in detail below with reference to specific examples and accompanying drawings.

[0028] The present invention realizes the extraction of equipment attribute data and combat rules through the ERNIE model, the universal information extraction framework (UIE) and the language large model, and realizes the extraction of equipment data and rules with higher value density from the original mixed knowledge data with low value density.

[0029] The present invention provides an equipment knowledge data extraction method, comprising the following steps:

[0030] Step 1: classify and screen text data from public sources, eliminate non-military text data, and obtain military text data.

[0031] In this step, the ERNIE model is used to classify and filter text data, and military text data is extracted from massive open source data. The similarity score is calculated based on the text category label and open source data to predict the category to which the sentence belongs.

[0032] Step 2: determine a static extraction template. The static extraction template is a set of equipment function performance attribute parameters as elements. Each type of equipment defines the static extraction template of the equipment according to the function performance attribute parameters it is concerned with.

[0033] Furthermore, the static extraction template at least includes {name, country of origin, function performance attribute parameters}, and the function performance attribute parameters in the template can be set as needed.

[0034] Further, in this step, a static extraction template is designed according to the functional performance attribute parameters of the equipment. For example, the functional performance attribute parameters of missiles include combat targets, external dimensions, detection parameters, missile range, killing parameters, target discovery, flight performance parameters, guidance parameters, combat time, combat mission, typical strike target, maintainability, and supportability. The static extraction template of the missile is: {name, country of origin, functional performance attribute parameters}, and the template element functional performance attribute parameters can be selected from the above categories. Based on this principle, the static extraction template of combat aircraft is: {name, country of origin, infrared characteristics, radar characteristics, optical characteristics, external dimensions, maneuverability, take-off weight, ceiling, flight speed, range, mounted weapons, combat mission, typical confrontation target}; the static extraction template of combat ships (non-aircraft carriers, destroyers, cruisers) is: {name, country of origin, external dimensions, displacement, range, speed, detection performance, missile type, gun type}, etc.

[0035] Step 3: Extract static data of equipment knowledge.

[0036] In this step, a general information extraction model is used to extract static functional performance of equipment from the military text data obtained in step 1 using the equipment static extraction template determined in step 2, and military text data in the format of {name, country of origin, functional performance attribute parameters}{original text} is obtained.

[0037] Furthermore, the general information extraction model adopts the unified framework model of general information extraction (UIE). The preferred general information extraction model uses the military relationship extraction dataset of the PaddleNLP resource platform to train the model, which specifically involves three steps: data set division, model fine-tuning, and model evaluation:

[0038] (1) Dataset division: Negative samples are constructed by setting non-military samples (negative-ratio) as the training set data. In the experiment, no non-military samples are set for 0-shot and 1-shot, and it is set to 3 in other experiments. The sum of the number of positive and negative samples is regarded as the specific number of training data used.

[0039] (2) Model fine-tuning: Model fine-tuning is to calculate the binary cross-entropy loss (BCELoss) based on the divided training set data using the model prediction results and the data annotation truth value. Specifically, the model predicts the starting position L of the extracted content respectively. start and end position L end The probability of each position in the input text is calculated as follows:

[0040]

[0041]

[0042]

[0043] The model parameters are updated by gradient descent, and the calculation process can be expressed as

[0044]

[0045] where ystart and The predicted and target values ​​for the starting position, yend and The predicted value and target value of the end position, θ represents the model parameters, η represents the learning rate, Represents the gradient of the loss function with respect to the model parameters.

[0046] (3) After the model is fine-tuned, in order to measure the predictive ability of the model, the F1 value is calculated on the test set to evaluate the predictive ability of the model.

[0047] Step 4: Determine the dynamic extraction template that describes the dynamic rules of the equipment.

[0048] Define dynamic extraction templates that describe equipment dynamic rules as input for equipment dynamic knowledge extraction; the defined equipment dynamic rule standard description templates are as follows:

[0049] Connection relationship (triggering process): The output of process A can trigger the start of process B; for example, the command behavior of a fighter jet can trigger the start of the strike behavior.

[0050] Same relationship (homogeneous process): Process A and Process B are the same process, appearing in different architectures and can replace each other; for example, different missile launchers of a certain equipment launch missiles.

[0051] Inclusion relationship (large and small processes): Process B is included in the large process A.

[0052] Succession relationship (continuation process): Process B can only be executed after process A is executed.

[0053] Calling relationship (reference process): Process A can trigger the start of process B, and process B can also be called by other processes; for example, the reconnaissance behavior of space-based reconnaissance equipment can trigger the fine reconnaissance behavior of ground-to-air jamming equipment.

[0054] Participation relationship (cross-process): Process A includes activity B, but activity B is usually not constructed as a process.

[0055] Step 5: Use the language big model as an extraction tool and combine it with the dynamic extraction template to extract the equipment combat process knowledge from the military text data obtained in step 4.

[0056] In this step, the equipment combat process knowledge is extracted by using a large language model combined with the equipment dynamic rules described by a dynamic extraction template, and statically extracted {name, country of origin, function performance attribute parameters} and dynamically extracted equipment combat process knowledge are obtained.

[0057] Furthermore, in this step, the language model realizes the extraction of equipment combat process knowledge through prompt words. Preferably, the prompt words of the language model are determined according to the dynamic extraction template, as follows:

[0058] Connection relationship (trigger process): There is a connection relationship between different tasks executed in sequence by the same equipment. The connection relationship of all tasks in the document is mined, and the equipment combat process knowledge is output in the format of [equipment; task A->task B->task C->…].

[0059] Same relationship (homogeneous process): Different equipments perform the same task, and these equipments have the same relationship. Mining the same relationship of all equipments in the document, outputting the equipment combat process knowledge in the format of [task; equipment A, equipment B, equipment C, ...].

[0060] Inclusion relationship (large and small processes): If the task performed by equipment A is a sub-link of the task performed by equipment B, then there is an inclusion relationship between equipment A and equipment B. Mining the inclusion relationship of all equipment in the document, outputting equipment combat process knowledge in the format of [equipment A->inclusion->equipment B].

[0061] Succession relationship (continuation process): If the task performed by equipment A is a continuation of the task performed by equipment B, then there is a succession relationship between equipment A and equipment B. Mining the succession relationship of all equipment in the document, outputting equipment combat process knowledge in the format of [equipment A-> succession-> equipment B].

[0062] Calling relationship (reference process): Equipment A calls equipment B for support to achieve the specified purpose, and equipment B can also be called by other equipment. Mining the calling relationship of all equipment in the document, outputting equipment combat process knowledge in the format of [equipment A->call->equipment B].

[0063] Participation relationship (cross-process): If equipment A receives support from equipment B when performing a mission, there is a participation relationship between these equipment A and equipment B. Mining the participation relationship of all equipment in the document, outputting equipment combat process knowledge in the format of [equipment A->participation->equipment B].

[0064] Step 6: Construct an equipment knowledge graph based on triple knowledge representation for the statically extracted {name, country of origin, function performance attribute parameters} and the dynamically extracted equipment combat process knowledge.

[0065] Furthermore, in this step, the extracted equipment static attribute data and equipment combat process knowledge are visualized and managed through graph de-redundancy, graph update and graph visualization.

[0066] Knowledge graph redundancy removal is mainly calculated through node semantic similarity, which is specifically calculated as follows:

[0067] Assume that the name of node (A) is "X-YYY equipment", after word segmentation, the word sequence (w A =[X, -, YYY, equipment]). The word vector is represented as (v A =[v X , v - , v YYY , v 装备 ]), where (v X ) is the vector representation of the vocabulary (X).

[0068] Assume that the two equipment vectors are (v A ) and (v B ), calculate the cosine similarity cos_sim(v A , v B ), when (cos_sim(v A , v B )>0.9), node (A) and node (B) are considered redundant nodes and are merged.

[0069] Furthermore, in step 6, the graph update mainly involves adding, deleting, modifying and checking nodes and edges in the equipment knowledge data graph.

[0070] The present invention realizes the extraction of high-quality equipment static functional performance parameters and dynamic combat process knowledge from massive mixed text data through artificial intelligence methods (ERNIE model, universal information extraction framework model (UIE) and language large model).

[0071] Furthermore, the present invention also provides an equipment knowledge data extraction system, including a memory, a processor, and a computer program stored in the memory and executable on the processor, and the processor implements the steps of the above-mentioned equipment knowledge data extraction method when executing the computer program.

[0072] The present invention provides a method for extracting equipment knowledge data. Figure 1 The steps shown are as follows:

[0073] Step 1: Classify and filter text data from public sources and eliminate non-military text data.

[0074] In this embodiment, in step 1, the text data from public sources is classified and screened, non-military text data is eliminated, and general text classification is achieved by introducing the ERNIE model. The specific algorithm flow is as follows:

[0075]

[0076]

[0077] Take the two sentences "The F-16 fighter is a third-generation fighter independently developed by the United States" and "Children's toys are hot on the market, click the link to buy now" as examples. Figure 2 As shown in the figure, the category labels of the text to be classified are set as "military", "entertainment" and "advertisement", and the output is whether the input sentence is military content. The general text classification is realized by the ERNIE model to judge the sentence. The output results of the above example are: "True" and "False" respectively. Through the above operation, the noise and redundant information in the original text data crawled from the web page are cleaned up, and unified and standardized text data related to the topic and rich in available information is obtained as the object of subsequent information extraction.

[0078] Step 2: Define the equipment static extraction template as the input for equipment static data extraction.

[0079] The static extraction template of equipment can be defined according to the needs, that is, according to the equipment characteristics that the user is concerned about, the equipment attributes to be extracted can be flexibly defined as the input of the general information extraction model. For example, taking the air defense missile as an example, its attribute parameter template can be defined as: {name, country of origin, combat target, missile length, missile diameter, detection range, far limit, near limit, high limit, low limit, single shot kill probability, target detection probability, cruising speed, maximum flight speed, maximum maneuver overload, guidance method, system response time, multi-target capability}, etc.

[0080] Step 3: Use the general information extraction model combined with the equipment static extraction template to extract the static functional performance of the equipment.

[0081] In this step, the general information extraction model adopts the unified framework model of general information extraction (UIE), and uses the military relationship extraction dataset of the PaddleNLP resource platform to train the model. Specifically, it involves three steps: data set division, model fine-tuning, and model evaluation:

[0082] (1) Dataset division: Negative samples are constructed by setting non-military samples (negative_ratio) as the training set data. In the experiment, no non-military samples are set for 0-shot and 1-shot, and it is set to 3 in other experiments. The sum of the number of positive and negative samples is regarded as the specific number of training data used.

[0083] (2) Model fine-tuning: Model fine-tuning is to calculate the binary cross-entropy loss (BCELoss) based on the divided training set data using the model prediction results and the data annotation truth value. Specifically, the model predicts the starting position of the extracted content. With end position The loss at the input text position is calculated as

[0084]

[0085]

[0086]

[0087] The model parameters are updated by gradient descent, and the calculation process can be expressed as

[0088]

[0089] where y start and The predicted and target values ​​for the starting position, y end and The predicted value and target value of the end position, θ represents the model parameters, η represents the learning rate, Represents the gradient of the loss function with respect to the model parameters.

[0090] (3) After the model is fine-tuned, in order to measure the predictive ability of the model, the F1 value is calculated on the test set to evaluate the predictive ability of the model.

[0091] Based on the above general modeling ideas, the military relationship extraction dataset of the PaddleNLP resource platform is used to train a general information extraction model, so that it can effectively learn on multiple tasks, making the model have strong general information extraction capabilities. The combined input is sent to the model, and the required information extraction can be completed according to the extraction method learned in the training phase.

[0092] In terms of specific model training, the military relationship extraction dataset based on the PaddleNLP resource platform contains the annotation of four types of information: "weapon name", "country of origin", "type" and "research and development unit" of weapons described in 121 military texts, totaling 484 data. Different shot values ​​were set respectively, and after training the model, the model F1 value was tested under 0-shot, 1-shot, 7-shot and 16-shot respectively. The results are shown in the following table:

[0093] 0-shot 1-shot 7-shot 16-shot F1 value 0.6174 0.6617 0.7667 0.8193

[0094] Based on the UIE model, the result of extracting equipment static attributes ['country of origin', 'weapon name', 'mission'] from the unstructured text "MQ-9 UAV is a U.S. Chinese air force drone. It can perform reconnaissance, surveillance and target strike missions" is: ['Country of origin: United States', 'Weapon name: MQ-9 UAV', 'Mission: reconnaissance, surveillance'].

[0095] Step 4: Define the equipment dynamic extraction template as the input for equipment dynamic knowledge extraction.

[0096] In this embodiment, in step 4, based on the definition of the static extraction template of the equipment, the large model prompt words are defined around the six types of dynamic extraction templates of equipment, namely, connection relationship (trigger process), identical relationship (homogeneous process), inclusion relationship (large and small processes), succession relationship (continuation process), call relationship (reference process), and participation relationship (cross process), as the input for the language large model to extract the dynamic rule knowledge of equipment.

[0097] Step 5: Use the language model as an extraction tool and combine it with the equipment dynamic extraction template to extract the equipment combat process knowledge.

[0098] In this embodiment, Figure 3 In step 5 shown, the specific algorithm flow of equipment combat process knowledge extraction based on the language large model is as follows:

[0099]

[0100]

[0101] The split(·) function is used to split the text data into a length acceptable to the language model.

[0102] Take the text as an example: "Six X-1s of the 310th Independent Carrier-Based Fighter Aviation Regiment deployed at the Yelizovo Air Force Base were deployed in anti-submarine warfare. At 02:25 in the morning, the sonar buoy discovered an unidentified underwater target. The X-1 fired a Y-1 torpedo at the target. At the same time, the sonar signature was identified as the Z-1 submarine. The torpedo accurately hit the target."

[0103] Extract the cohesion relationship in the text, and define the prompt word of the cohesion relationship as follows: different tasks performed in sequence by the same equipment have a cohesion relationship. Mine the cohesion relationship of all tasks in the document and output it in the format of [equipment; task A->task B->task C->…]. Based on the language model and prompt words, the extraction result can be "[X-1; establish anti-submarine zone->discover enemy submarine->sink enemy submarine]".

[0104] Step 6: Construct an equipment knowledge graph based on triple knowledge representation, and realize the visualization and management of equipment combat knowledge data through graph de-redundancy, graph update and graph visualization.

[0105] In this embodiment, Figure 4 In step 6 shown, the equipment knowledge graph constructed based on triple knowledge representation realizes graph redundancy removal, graph update, and graph visualization to realize the visualization and management of equipment combat knowledge data.

[0106] The redundancy removal of knowledge graph is mainly achieved by converting equipment into word vectors and calculating language similarity, and merging equipment nodes with high similarity to achieve redundancy removal.

[0107] Graph updates include adding and deleting graph nodes, adding and deleting equipment links (combat process), and modifying, adding and deleting equipment static function performance attribute parameters.

[0108] Graph visualization is built through the vis.js library, which provides functions such as adding, deleting, editing, and searching nodes and edges. It aims to explore how to improve the efficiency and effectiveness of data analysis through graph visualization technology.

[0109] Parts of the present invention that are not described in detail are well known to those skilled in the art.

Claims

1. A method for extracting equipment knowledge data, characterized in that: The following steps are included: Step 1: classify and filter the text data from public sources, remove non-military text data, and obtain military text data; Step 2, determine a static extraction template. The static extraction template is a set of equipment function performance attribute parameters as elements. Each type of equipment defines the static extraction template of the equipment according to the function performance attribute parameters it is concerned about. The static extraction template at least includes {name, country of origin, function performance attribute parameters}; Step 3, using the static extraction template determined in step 2 to extract static data of equipment knowledge from the military text data obtained in step 1; Step 4: Determine the dynamic extraction template that describes the dynamic rules of the equipment. The dynamic extraction template includes the following relationships: The connection relationship is the triggering process: the output of process A can trigger the start of process B; The same relationship refers to homogeneous processes: process A and process B are the same process, appearing in different architectures and can replace each other; the inclusion relationship refers to large and small processes): process B is included in the large process A. The succession relationship is the continuation of the process: process B can only be executed after process A is executed; The calling relationship is the reference process: process A can trigger the start of process B, and process B can also be called by other processes; Participation relationship is cross-process: process A includes activity B, but activity B is usually not constructed as a process; Step 5: Use the language big model as an extraction tool and combine it with the dynamic extraction template to extract the equipment combat process knowledge from the military text data; Step 6: Use the statically extracted equipment knowledge static data and the dynamically extracted equipment combat process knowledge to construct a knowledge equipment knowledge graph.

2. The equipment knowledge data extraction method according to claim 1, characterized in that: In step 5, the equipment combat process knowledge is extracted by using the language big model in combination with the equipment dynamic rules described by the dynamic extraction template to obtain the dynamically extracted equipment combat process knowledge. The language big model realizes the extraction of equipment combat process knowledge through prompt words. The prompt words and dynamic extraction of the language big model include: The connection relationship triggers the process: different tasks executed in sequence by the same equipment have a connection relationship. The connection relationship of all tasks in the document is mined, and the equipment combat process knowledge is output in the format of [equipment; task A->task B->task C->…]; The same relationship means homogeneous process: different equipments perform the same task, so these equipments have the same relationship. The same relationship of all equipments in the document is mined, and the equipment combat process knowledge is output in the format of [task; equipment A, equipment B, equipment C, ...]; Inclusion relationship refers to large and small processes: if the task performed by equipment A is a sub-link of the task performed by equipment B, then there is an inclusion relationship between equipment A and equipment B. The inclusion relationship of all equipment in the document is mined, and the equipment combat process knowledge is output in the format of [equipment A->inclusion->equipment B]; The succession relationship is the continuation process: if the task performed by equipment A is the continuation of the task performed by equipment B, then there is a succession relationship between equipment A and equipment B. The succession relationship of all equipment in the document is mined, and the equipment combat process knowledge is output in the format of [equipment A->succession->equipment B]; The calling relationship is the reference process: Equipment A calls Equipment B for support to achieve the specified purpose, and Equipment B can also be called by other equipment. The calling relationship of all equipment in the document is mined, and the equipment combat process knowledge is output in the format of [Equipment A->Call->Equipment B]; Participation relationship refers to cross-process: if equipment A receives support from equipment B when performing a mission, then there is a participation relationship between equipment A and equipment B. The participation relationship of all equipment in the document is mined, and the equipment combat process knowledge is output in the format of [equipment A->participation->equipment B].

3. The equipment knowledge data extraction method according to claim 2, characterized in that: In the step 3, a general information extraction model is used to extract static functional performance of the military text data obtained in step 1 using the equipment static extraction template determined in step 2, and the military text data in the format of {name, country of origin, functional performance attribute parameters} {original text} is obtained.

4. The equipment knowledge data extraction method according to claim 3 is characterized in that: The general information extraction model in step 3 adopts the unified framework model of general information extraction (UIE).

5. The equipment knowledge data extraction method according to claim 4, characterized in that: The general information extraction model in step 3 is trained using the military relationship extraction dataset of the PaddleNLP resource platform. The following steps are included: (1) Dataset division: Negative samples are constructed by setting non-military samples as training set data. In the experiments, no non-military samples are set in 0-shot and 1-shot. In the other experiments, it is set to 3. The sum of the number of positive and negative samples is regarded as the specific number of training data used. (2) Model fine-tuning: Based on the divided training set data, the binary cross entropy loss is calculated using the model prediction results and the data annotation truth values; (3) After the model is fine-tuned, in order to measure the predictive ability of the model, the F1 value is calculated on the test set to evaluate the predictive ability of the model.

6. The equipment knowledge data extraction method according to claim 5, characterized in that: In step 3, the model predicts the starting position L of the extracted content respectively start and end position L end The probability of each position in the input text is calculated as follows: The model parameters are updated by gradient descent, and the calculation process can be expressed as where y start and The predicted and target values ​​for the starting position, y end and The predicted value and target value of the end position, θ represents the model parameters, η represents the learning rate, Represents the gradient of the loss function with respect to the model parameters.

7. The equipment knowledge data extraction method according to claim 3, characterized in that: In step 6, visualization and management of extracted equipment static attribute data and equipment combat process knowledge are achieved through graph de-redundancy, graph update and graph visualization.

8. The equipment knowledge data extraction method according to claim 7, characterized in that: In step 6, the knowledge graph redundancy removal is calculated by the semantic similarity of the nodes; The graph is updated to add, delete, modify and query nodes and edges in the equipment knowledge data graph.

9. The equipment knowledge data extraction method according to claim 3, characterized in that: In step 1, the text data classification and screening adopts the ERNIE model, which takes the category label of the text to be classified and the open source data as input, and calculates the similarity score to predict the category to which the sentence belongs.

10. An equipment knowledge data extraction system, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the steps of any equipment knowledge data extraction method according to claim 1 when executing the computer program.

Images