An entity relationship triple extraction method, system and computer device

Abstract

The application discloses an entity relationship triple extraction method and system and computer equipment, relates to the technical field of knowledge graph construction, and comprises the following steps: acquiring unmanned aerial vehicle unstructured fault knowledge text; inputting the unmanned aerial vehicle unstructured fault knowledge text into a Hor-Ver-Casrel model to extract triples, including: based on the specific position of a head entity, fusing information of horizontal and vertical directions of the head entity, obtaining vector representations of the horizontal and vertical directions of the head entity and concatenating the vector representations to obtain information representation of the head entity; combining vector representations of each character with the final information representation of the head entity, constructing a mapping function of each relationship, predicting a tail entity according to the mapping function of each relationship, and completing triple extraction. The application takes the concatenated vector representations of the horizontal and vertical directions as the final information representation of the head entity, retains semantic information of the head entity, and greatly reduces the influence on subsequent tail entity extraction.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of knowledge graph construction, in particular to an entity relationship triple extraction method and system and a computer device. BACKGROUND

[0002] The traditional unmanned aerial vehicle system fault diagnosis method is to rely on experienced experts or engineering and technical personnel to carry out on-site troubleshooting, which requires a large amount of manpower and high maintenance cost. With the development of information technology, in order to overcome the disadvantages brought by on-site troubleshooting, big data, knowledge graph, artificial intelligence and other methods are applied to the auxiliary diagnosis and intelligent prediction process of various faults, which can effectively promote the rapid and intelligent troubleshooting of unmanned aerial vehicle systems. Among them, the intelligent data processing technology represented by the knowledge graph can be widely applied in product consultation, operation and maintenance services, statistical analysis and other work in various industries by virtue of its fine-grained knowledge management mode, good intelligent interaction process and multi-channel access mode. When applied to the maintenance and support process of unmanned aerial vehicle systems, it has the following advantages: 1) According to user usage data, the fault rate of each device can be counted to realize the visualization of historical data, which is convenient for better maintenance and support work; 2) A large amount of expert knowledge and technical personnel experience can be integrated to build a system fault knowledge base, which can quickly extract the fault phenomenon and fault elimination method of the fault from the field knowledge base when the fault occurs, and provide guidance for the user at the first time. It can also be used for the learning and training of front-line troops and management personnel, and can improve the troubleshooting ability; 3) The correlation between typical faults and flight tasks, aircraft types, maintenance work, flight information, detection results and other data is established to eliminate information gap and time and space obstacles. Through intelligent question and answer and semantic retrieval, various unmanned aerial vehicle system fault information can be quickly provided to effectively support troubleshooting; 4) When the front-end system provides guidance but the fault still cannot be eliminated, artificial service can be cut in, and professional technical personnel will provide online guidance through the back-end service system, playing a part of the role of remote expert diagnosis.

[0003] Entity relation triple extraction is an indispensable step in the process of knowledge graph construction, and entity relation triple extraction refers to extracting entity pairs and determining the relation between entities from unstructured text. The entity relation triple structure is (subject, relation, object), for example (failure, failure location, component unit). Since deep neural networks can well encode semantic information of text, the triple extraction method based on deep neural networks is the most competitive method at present. Most triple extraction methods extract head entities and tail entities, and then treat the relation as a discrete label on the entity pair. This way is not good for the extraction effect of overlapping triplets. In view of the problem of overlapping triplets, the cascade binary tagging framework (A Novel Cascade Binary Tagging Framework for Relational Triple Extraction, Casrel) can alleviate this problem. The basic process of the model is as follows: first, extract the head entity, then model the relation as a mapping function, and map the subject to the object in the sentence, thereby naturally handling the overlapping problem, rather than treating the relation as a discrete label as in previous work.

[0004] In the cascade binary framework, only the sum and average of the start vector and the end vector of the head entity are used as the information of the head entity, which seriously loses the semantic information of the head entity, thereby directly affecting the extraction of the subsequent tail entity. SUMMARY

[0005] The application provides an entity relation triple extraction method and system, and a computer device, and particularly relates to an entity relation triple extraction method and system in the process of constructing a knowledge graph of unmanned aerial vehicle failures, which solves the problem that in the existing cascade binary framework, only the sum and average of the start vector and the end vector of the head entity are used as the information of the head entity, which seriously loses the semantic information of the head entity, thereby directly affecting the extraction of the subsequent tail entity.

[0006] The application provides an entity relation triple extraction method, which comprises the following steps:

[0007] Obtaining an unstructured failure knowledge text of an unmanned aerial vehicle;

[0008] The Casrel model is improved to obtain a Hor-Ver-Casrel model; when the information of the head entity is represented, the Hor-Ver-Casrel model fuses the information in the horizontal and vertical directions of the head entity;

[0009] The unstructured failure knowledge text of the unmanned aerial vehicle is input into the Hor-Ver-Casrel model, and the entity relation triple of the failure knowledge text is extracted, comprising:

[0010] inputting the unmanned aerial vehicle unstructured fault knowledge text into a Bert module to obtain a vector representation of each character;

[0011] predicting a start position and an end position of a head entity based on the vector representation of each character to obtain a specific position of the head entity;

[0012] fusing information in horizontal and vertical directions of the head entity based on the specific position of the head entity to obtain vector representations of the head entity in the horizontal and vertical directions and concatenating the vector representations to obtain an information representation of the head entity;

[0013] combining the vector representation of each character and the final information representation of the head entity to construct a mapping function of each relationship corresponding to the head entity, predicting a tail entity according to the mapping function of each relationship, and completing extraction of an entity relationship triple.

[0014] Preferably, the unmanned aerial vehicle unstructured fault knowledge text includes historical fault cases, operation procedures, and troubleshooting manuals.

[0015] Preferably, the Bert module adjusts a weight coefficient matrix by an association degree between characters in the unmanned aerial vehicle unstructured fault knowledge text to output a word vector representation of each character;

[0016] The word vector representation of each character is calculated by the following formula:

[0017]

[0018] In the formula, Q, K, and V are word vector matrices, dk is a nested dimension, Attention is a self-attention mechanism function, Softmax is an activation function, and T is a transpose symbol of a matrix.

[0019] Preferably, the start position and the end position of the head entity are predicted by the following formula:

[0020]

[0021]

[0022] In the formula, F i is the vector representation of each character, is a label of the predicted start position of the head entity of the i-th character, is a label of the predicted end position of the head entity of the i-th character, W start is a weight matrix for predicting the start position of the head entity, b start is a bias for predicting the start position of the head entity, W end is a weight matrix for predicting the end position of the head entity, b end is a bias for predicting the end position of the head entity, and σ is a full connection layer.

[0023] The horizontal and vertical direction vector representation of the head entity is preferably as follows:

[0024]

[0025]

[0026] wherein F a is the word vector of the start position of the head entity, F b is the word vector of the end position of the head entity, average is the vector addition and averaging operation, sum is the vector summation operation, is the horizontal direction vector representation of the head entity, is the vertical direction vector representation of the head entity, W sub is the weight matrix for predicting the vertical direction vector of the head entity, b sub is the bias for predicting the vertical direction vector of the head entity.

[0027] The information representation of the head entity is as follows:

[0028]

[0029] wherein v sub is the information representation of the head entity.

[0030] The tail entity is preferably predicted by the following formula:

[0031]

[0032]

[0033] wherein, is the start position label of the predicted tail entity of the i-th character, is the end position label of the predicted tail entity of the i-th character, is the weight matrix for predicting the start position of the tail entity under the relationship mapping, is the bias for predicting the start position of the tail entity under the relationship mapping, is the weight matrix for predicting the end position of the tail entity under the relationship mapping, is the bias for predicting the end position of the tail entity under the relationship mapping.

[0034] Preferably, before inputting the unmanned aerial vehicle unstructured fault knowledge text into the Hor-Ver-Casrel model, the Hor-Ver-Casrel model needs to be trained by maximum likelihood estimation; the likelihood function of the maximum likelihood estimation includes the likelihood function of the head entity prediction and the likelihood function of the relationship and tail entity prediction.

[0035] Preferably, the likelihood function of the head entity prediction is as follows:

[0036]

[0037] wherein θ={W start ,b start ,W end ,b end}, start_s represents a head entity start position identifier, end_s represents a head entity end position identifier, represents that the i-th character is a start position of a head entity, represents that the i-th character is an end position of a head entity, L is a sentence length, I is an indicator function, P is a likelihood function, and s is a head entity identifier, is a probability that the i-th character is a start position of a head entity, is a probability that the i-th character is an end position of a head entity.

[0038] The likelihood function of the relationship and tail entity prediction is as follows:

[0039]

[0040] wherein, represents that the i-th character is a start position of a tail entity, represents that the i-th character is an end position of a head entity, o is a tail entity identifier, start_o represents a tail entity start position identifier, and end_o is a tail entity end position identifier.

[0041] An entity relationship triple extraction system comprises:

[0042] A text acquisition module is configured to acquire an unmanned aerial vehicle unstructured fault knowledge text.

[0043] A model construction module is configured to improve a Casrel model to obtain a Hor-Ver-Casrel model; when the Hor-Ver-Casrel model represents information of a head entity, information of a horizontal direction and a vertical direction of the head entity is fused.

[0044] A triple extraction module is configured to input the unmanned aerial vehicle unstructured fault knowledge text into the Hor-Ver-Casrel model, and extract an entity relationship triple.

[0045] The triple extraction module comprises:

[0046] A character vector representation module is configured to input the unmanned aerial vehicle unstructured fault knowledge text into a Bert module, and acquire a vector representation of each character.

[0047] a head entity position prediction module, configured to predict a start position and an end position of a head entity based on the vector representation of each character, to obtain a specific position of the head entity;

[0048] a head entity information representation module, configured to fuse information in horizontal and vertical directions of the head entity based on the specific position of the head entity, to obtain vector representations in the horizontal and vertical directions of the head entity and concatenate the vector representations, to obtain an information representation of the head entity;

[0049] a tail entity prediction module, configured to combine the vector representation of each character and the final information representation of the head entity, to construct a mapping function of each relation corresponding to the head entity, and to predict a tail entity according to the mapping function of each relation, to complete extraction of an entity relation triple.

[0050] A computer device comprises a memory, a processor, and a computer program stored in the memory and executable on the processor, and the processor implements the above-mentioned entity relation triple extraction method when executing the program.

[0051] Compared with the prior art, the beneficial effects of the present application are:

[0052] The present application obtains the specific position of the head entity, fuses information in the horizontal and vertical directions of the head entity, obtains vector representations in the horizontal and vertical directions of the head entity, and then concatenates the vector representations in the horizontal and vertical directions of the head entity to obtain the final information representation of the head entity. Instead of adding and averaging the start vector and the end vector of the head entity, the present application uses the concatenated vector representations in the horizontal and vertical directions as the final information representation of the head entity, retains the semantic information of the head entity, and greatly reduces the influence on the extraction of the subsequent tail entity. BRIEF DESCRIPTION OF DRAWINGS

[0053] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings needed to be used in the embodiments or the prior art description. Obviously, the drawings in the following description only constitute some embodiments of the present application, and for those skilled in the art, other drawings can also be obtained without creative labor on the basis of these drawings.

[0054] Figure 1 It is a schematic diagram of the existing Casrel model structure;

[0055] Figure 2 It is a schematic diagram of the Hor-Ver-Casrel model structure of the present application;

[0056] Figure 3 It is a schematic diagram of the source classification of the unmanned aerial vehicle fault data of the present application;

[0057] Figure 4 A text fragment of a fault report of a certain type of unmanned aerial vehicle for an embodiment of the application;

[0058] Figure 5 A text data example related to unmanned aerial vehicle maintenance in an embodiment of the application. DETAILED DESCRIPTION

[0059] The technical solutions in the embodiments of the application will be clearly and completely described below with reference to the drawings in the embodiments of the application. Obviously, the described embodiments are only part of the embodiments of the application, rather than all the embodiments of the application. Based on the embodiments in the application, all other embodiments obtained by a person of ordinary skill in the art without creative labor fall within the protection scope of the application.

[0060] Reference Figure 1 The specific idea of the existing Casrel model is as follows:

[0061] The input sentence is input into the pre-trained Bert module, the key part of the encoding is the Transformer structure, and the weight coefficient matrix is adjusted according to the correlation degree between characters in the same sentence to output the word vector representation of each character.

[0062] After obtaining the vector representation of each character, the starting position and the ending position of the subject entity are predicted based on the vector representation of each character.

[0063] Based on the predicted head and tail of the subject entity, the representation vector of each subject entity is obtained by using the sum and average method.

[0064] Using the vector representation of each character and the representation vector of the predicted subject entity, a mapping function of each relationship is constructed by traversing the relationship dictionary, which is used to predict the object entity.

[0065] Finally, the entity relationship triple is output.

[0066] Therefore, in the existing Casrel model, only the sum and average of the starting vector and the ending vector of the head entity are used as the information of the head entity, which seriously loses the semantic information of the head entity.

[0067] The application provides an entity relationship triple extraction method, and particularly relates to an entity relationship triple extraction method in a process of constructing an unmanned aerial vehicle fault knowledge graph, which comprises the following steps:

[0068] Step 1: Obtain the unstructured fault knowledge text of the unmanned aerial vehicle.

[0069] With the rapid development of unmanned aerial vehicles, the fault data generated in the field of unmanned aerial vehicles is growing, and the data types are various, mainly including historical flight fault data of grassroots units, various fault articles accumulated at the same period, troubleshooting manuals, professional fault articles, and equipment maintenance support teaching materials and fault analysis and research reports. These fault data can be divided into two kinds according to different storage methods, namely structured data and semi / non-structured data. Figure 3 As shown in the figure, the structured data refers to the data logically expressed and realized by a two-dimensional table structure, and this kind of data is mainly stored and managed by a relational database. The semi / non-structured data refers to the data in the form of documents such as text, image and video, such as the related text materials of a large number of unmanned aerial vehicle troubleshooting manuals and fault analysis research reports, which belong to non-structured data.

[0070] However, there are still a large amount of valuable non-structured data to be mined and utilized in the whole life cycle of military unmanned aerial vehicles, and the extraction difficulty is relatively complex. Therefore, the present application focuses on the knowledge extraction of non-structured data. As shown in the figure, Figure 4 the text information highlighted and displayed in the non-structured text needs to be extracted, including fault mode, fault unit and signal parameter. As can be seen from the figure, Figure 4 the above-mentioned fault entities have certain correlation, and technical means needs to be taken to extract the relationship contained in the non-structured fault knowledge text, which is the difficulty of knowledge extraction. For example, in the sentence "the reason for unsuccessful engine starting is fuel nozzle blockage", the relationship between "fault mode-engine unsuccessful starting" and "fault reason-fuel nozzle blockage" needs to be extracted. This is a relatively complex process for a computer to recognize the relationship between the two as a person does through professional knowledge learning and maintenance practice.

[0071] Second step: based on the Casrel model, the information of the head entity is represented by fusing the horizontal and vertical information, and the Hor-Ver-Casrel model is obtained.

[0072] In order to alleviate the problem that in the existing cascading binary framework Casrel model, only the sum and average of the start vector and the end vector of the head entity are used as the information of the head entity, which seriously loses the semantic information of the head entity, a horizontal and vertical information fusion method is designed to more richly represent the vector of the head entity, and the method is called a cascading binary framework based on horizontal and vertical information fusion, as shown in the figure, and is recorded as Hor-Ver-Casrel model. Figure 2

[0073] ​Step 3: input the unmanned aerial vehicle unstructured fault knowledge text into the Hor-Ver-Casrel model to extract triples, including the following steps:

[0074] S1: input the unmanned aerial vehicle unstructured fault knowledge text into the Bert module to obtain the vector representation of each character.

[0075] The Bert module adjusts the weight coefficient matrix through the correlation degree between characters in the unmanned aerial vehicle unstructured fault knowledge text to output the word vector representation of each character. The correlation degree between characters in the unmanned aerial vehicle unstructured fault knowledge text is calculated by the following formula:

[0076]

[0077] In the formula, Q, K, and V are word vector matrices, dk is the nesting dimension, Attention is the self-attention mechanism function, Softmax is the activation function, and T is the transpose symbol of the matrix. Finally, the vector representation of each character is denoted as F i , and the vector representation of the entire sentence is denoted as h N .

[0078] S2: based on the vector representation of each character, the start and end positions of the head entity are predicted to obtain the specific position of the head entity:

[0079]

[0080]

[0081] In the formula, F i is the vector representation of each character, is the label of the predicted start position of the head entity for the i-th character, is the label of the predicted end position of the head entity for the i-th character, W start is the weight matrix for predicting the start position of the head entity, b start is the bias for predicting the start position of the head entity, W end is the weight matrix for predicting the end position of the head entity, b end is the bias for predicting the end position of the head entity, and σ is the full connection layer.

[0082] S3: based on the specific position of the head entity, the horizontal and vertical directions of the head entity are fused to obtain the vector representation of the horizontal and vertical directions and concatenate them to obtain the information representation of the head entity.

[0083] Based on the predicted head entity in the previous step, assume that it is the 8th and 9th characters in Figure 2 , through the information fusion of the horizontal and vertical directions, the vector representations of the horizontal and vertical directions are obtained respectively and Finally, the head entity final information representation v is obtained by cascading sub :

[0084]

[0085]

[0086]

[0087] where average is an averaging operation, sum is a vector sum operation, is a horizontal vector representation of the head entity, is a vertical vector representation of the head entity, W sub is a weight matrix for predicting the vertical vector of the head entity, b sub is a bias for predicting the vertical vector of the head entity.

[0088] S4: Combine the vector representation of each character with the head entity final information representation to construct a mapping function for each relationship, and predict the tail entity according to the mapping function of each relationship to complete the extraction of the entity relationship triplets.

[0089] The tail entity is predicted by the following formula:

[0090]

[0091]

[0092] where, is the start position label of the predicted tail entity of the i-th character, is the end position label of the predicted tail entity of the i-th character, is a weight matrix for predicting the start position of the tail entity under the relationship mapping, is a bias for predicting the start position of the tail entity under the relationship mapping, is a weight matrix for predicting the end position of the tail entity under the relationship mapping, is a bias for predicting the end position of the tail entity under the relationship mapping.

[0093] Before inputting the unmanned aerial vehicle unstructured fault knowledge text into the Hor-Ver-Casrel model, the Hor-Ver-Casrel model needs to be trained by maximum likelihood estimation. The likelihood function of the maximum likelihood estimation includes the likelihood function logP θ (s|X) of predicting the head entity and the likelihood function

[0094]

[0095] wherein,

[0096]

[0097]

[0098] where L denotes a sentence length, is the label of the start and end positions of the head entity of the i-th character, is the label of the start and end positions of the tail entity of the i-th character, θ = {W start ,b start ,W end ,b end},

[0099] By further calculating for the likelihood loss, we can get:

[0100]

[0101]

[0102] Thus, the following algorithm is used to solve:

[0103]

[0104]

[0105]

[0106] Separately calculate the gradient:

[0107]

[0108] Thus, we can get, wherein a is a learning rate.

[0109] Similarly, we can get, wherein a is a learning rate.

[0110] Set the gradient ascent method for the learning stage of the model to update the iterative parameters w j,0 and w j,o , and train the model by maximizing the above likelihood function. Finally, a trained model is obtained.

[0111] Based on the same concept, this invention also provides an entity relation triple extraction system, including a text acquisition module, a model building module, and a triple extraction module. The text acquisition module is used to acquire unstructured fault knowledge text of UAVs. The model building module is used to improve the Casrel model to obtain the Hor-Ver-Casrel model; the Hor-Ver-Casrel model integrates the horizontal and vertical information of the head entity when representing the information of the head entity. The triple extraction module is used to input the unstructured fault knowledge text of the UAV into the Hor-Ver-Casrel model and extract the triples.

[0112] The triplet extraction module includes a character vector representation module, a head entity position prediction module, a head entity information representation module, and a tail entity prediction module.

[0113] The character vector representation module is used to input unstructured fault knowledge text of the UAV into the Bert module and obtain the vector representation of each character.

[0114] The head entity position prediction module is used to predict the start and end positions of the head entity based on the vector representation of each character, so as to obtain the specific position of the head entity.

[0115] The head entity information representation module is used to fuse the horizontal and vertical information of the head entity based on its specific location, obtain the horizontal and vertical vector representations of the head entity, and concatenate them to obtain the information representation of the head entity.

[0116] The tail entity prediction module combines the vector representation of each character with the final information representation of the head entity to construct a mapping function for each relation corresponding to the head entity. Based on the mapping function of each relation, the tail entity is predicted, thus completing the extraction of entity relation triples.

[0117] Example

[0118] The experiment used 100 text articles about drone maintenance. The sample text data is as follows: Figure 5 As shown, the data format marked by the Daguan platform is: {"text":"II. Fault Analysis: The engine speed of the Attack-1 UAV is mainly determined by factors such as the fuel supply system, intake boosting system, turbocharger control system, ignition subsystem, and engine load control.","spo_list":[{"predicate":"Composition","object_type":"Model","subject_type":"Component Unit","object":"Attack-1 UAV Engine","subject":"Ignition Subsystem"}]}. Based on this data format, the Hor-Ver-Casrel model proposed in this invention is trained using training samples.

[0119] The final model reaches an accuracy of 87% and a recall of 70% on the test set. Through the verification of the experiment, it is found that the effect of the Hor-Ver-Casrel model meets the requirements in the technical index list, that is, the extraction accuracy is not less than 80%, and the recall is not less than 70%.

[0120] The application further provides a computer device, comprising a memory, a processor, and a computer program stored in the memory and capable of running on the processor, and the processor implements the above-mentioned entity relationship triple extraction method when executing the program.

[0121] Although preferred embodiments of the application have been described, those skilled in the art will be able to make additional changes and modifications to these embodiments once they have been given the basic inventive concept. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications falling within the scope of the application.

[0122] Obviously, those skilled in the art can make various modifications and variations to the application without departing from the spirit and scope of the application. Thus, if these modifications and variations of the application fall within the scope of the claims of the application and equivalent technologies thereof, the application also intends to include these modifications and variations.

Claims

1. A method for extracting entity relation triples, characterized in that, Includes the following steps: Acquire unstructured fault knowledge text of drones; The Casrel model is improved to obtain the Hor-Ver-Casrel model; the Hor-Ver-Casrel model integrates the horizontal and vertical information of the head entity when representing the information of the head entity. The Hor-Ver-Casrel model is trained using maximum likelihood estimation; the likelihood function of the maximum likelihood estimation includes the likelihood function for predicting the head entity and the likelihood function for predicting the relation and tail entity. The likelihood function for the head entity prediction is shown below: In the formula, , The weight matrix is ​​used to predict the starting position of the head entity. To predict the deviation of the head entity's starting position, The weight matrix is ​​used to predict the ending position of the head entity. To predict the deviation of the head entity's ending position, , Indicates the end position marker of the header entity. Indicates the first i The first character is the starting position of the header entity. } indicates the first i This character marks the end of the header entity. L Sentence length For indicator functions, Let be the likelihood function. s For the head entity identifier, For the first The probability that each character represents the starting position of the header entity. For the first Each character represents the probability of the head entity ending position; The likelihood functions for predicting relations and tail entities are shown below: In the formula, , This is the weight matrix for predicting the starting position of the tail entity under relation mapping. The deviation in predicting the starting position of the tail entity under relation mapping. This is the weight matrix for predicting the ending position of the tail entity under relation mapping. This represents the deviation in predicting the ending position of the tail entity under relation mapping. W sub The weight matrix is ​​used to predict the vertical vector of the head entity. To predict the deviation of the vertical vector of the head entity, Indicates the first i The last character is the starting position of the tail entity. } indicates the first i This character marks the end of the header entity. o For the tail entity identifier, , This serves as the marker for the end position of the tail entity; Unstructured drone fault knowledge text is input into the Hor-Ver-Casrel model, and entity relation triples are extracted from the fault knowledge text, including: Input the unstructured fault knowledge text of the UAV into the Bert module to obtain the vector representation of each character; The start and end positions of the head entity are predicted based on the vector representation of each character, thus obtaining the specific position of the head entity; Based on the specific location of the head entity, the horizontal and vertical information of the head entity is fused to obtain the vector representation of the head entity in the horizontal and vertical directions. The vector representations of the head entity in the horizontal and vertical directions are then concatenated to obtain the information representation of the head entity. The vector representations of the head entity in the horizontal and vertical directions are shown below: In the formula, The word vector at the starting position of the head entity. The word vector at the end position of the head entity. For the operation of summing and averaging vectors, sum For vector summation operations, This is the vector representation of the head entity in the horizontal direction. This is the vector representation of the head entity in the vertical direction. W sub The weight matrix is ​​used to predict the vertical vector of the head entity. To predict the deviation of the vertical direction vector of the head entity; The information of the head entity is represented as follows: In the formula, Information representation for the head entity; By combining the vector representation of each character with the information representation of the head entity, a mapping function for each relation corresponding to the head entity is constructed. The tail entity is predicted based on the mapping function of each relation, thus completing the extraction of entity relation triples.

2. The entity relation triple extraction method as described in claim 1, characterized in that, The unstructured fault knowledge text for UAVs includes historical fault cases, operating procedures, and troubleshooting manuals.

3. The entity relation triple extraction method as described in claim 1, characterized in that, The Bert module adjusts the weight coefficient matrix based on the correlation between characters in the unstructured fault knowledge text of the UAV to output a vector representation of each character; The vector representation of each character is calculated using the following formula: In the formula, For word vector matrices, Nested dimensions Attention For the self-attention mechanism function, Softmax For activation function, T This is the transpose symbol for a matrix.

4. The entity relation triple extraction method as described in claim 1, characterized in that, The start and end positions of the head entity are predicted using the following formula: In the formula, F i A vector representation for each character. For the first i The label of the starting position of the predicted header entity for each character. For the first i The label at the end position of the predicted header entity for each character. The weight matrix is ​​used to predict the starting position of the head entity. To predict the deviation of the head entity's starting position, The weight matrix is ​​used to predict the ending position of the head entity. To predict the deviation of the head entity's ending position, It is a fully connected layer.

5. The entity relation triple extraction method as described in claim 1, characterized in that, The tail entity is predicted using the following formula: In the formula, For the first i The start position label of the predicted tail entity for each character. For the first i The end position label of the predicted tail entity for each character. This is the weight matrix for predicting the starting position of the tail entity under relation mapping. The deviation in predicting the starting position of the tail entity under relation mapping. This is the weight matrix for predicting the ending position of the tail entity under relation mapping. This represents the deviation in predicting the ending position of the tail entity under relation mapping.

6. An extraction system for the entity relation triple extraction method according to any one of claims 1-5, characterized in that, include: The text acquisition module is used to acquire unstructured fault knowledge text of UAVs; The model building module is used to improve the Casrel model to obtain the Hor-Ver-Casrel model; the Hor-Ver-Casrel model integrates the horizontal and vertical information of the head entity when representing the information of the head entity. The triple extraction module is used to input unstructured fault knowledge text of UAVs into the Hor-Ver-Casrel model and extract entity relation triples; The triplet extraction module includes: The character vector representation module is used to input the unstructured fault knowledge text of the UAV into the Bert module and obtain the vector representation of each character; The head entity position prediction module is used to predict the start and end positions of the head entity based on the vector representation of each character, so as to obtain the specific position of the head entity. The head entity information representation module is used to fuse the horizontal and vertical information of the head entity based on its specific location, obtain the horizontal and vertical vector representations of the head entity, and concatenate them to obtain the information representation of the head entity. The tail entity prediction module combines the vector representation of each character with the final information representation of the head entity to construct a mapping function for each relation corresponding to the head entity. Based on the mapping function of each relation, the tail entity is predicted, thus completing the extraction of entity relation triples.

7. A computer device, characterized in that, The system includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the entity relation triple extraction method according to any one of claims 1-5.

Images