Knowledge graph construction method surrounding special equipment administrative penalty document

By constructing a knowledge graph of administrative penalty documents for special equipment, using XLNet, BiLSTM and CRF models for entity identification and relationship extraction, the problems of data integration and information sharing in elevator safety supervision are solved, and the intelligent identification and knowledge integration of administrative penalty documents are realized. , improves regulatory efficiency and intelligence.

CN119047551AInactive Publication Date: 2024-11-29CHINA UNIV OF MINING & TECH (BEIJING)
View PDF 3 Cites 0 Cited by

Patent Information

Application Number
CN202410972549.0
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-07-19
Publication Date
2024-11-29
Estimated Expiration
Not applicable · inactive patent

AI Technical Summary

Technical Problem

In the intelligent supervision of elevator safety, how to effectively integrate and manage massive multi-source heterogeneous data, realize information sharing, and improve supervision intelligence, especially the relatively little research on the extraction of information of administrative penalty documents, resulting in the existence of 'knowledge island' in the supervision process .

Method used

A knowledge graph construction method is proposed for special equipment administrative penalty documents. By obtaining the administrative penalty entities to be identified, using the entity recognition model for entity recognition and relationship extraction, combining the similarity algorithm for knowledge fusion, storing it in the Neo4j graph database, and building administrative management Knowledge graph of punishment documents. The method includes the XLNet layer, the BiLSTM layer and the CRF layer, and uses the Transformer-XL mechanism and an improved recurrent neural network for entity recognition and relationship extraction.

Benefits of technology

Effectively identify the objects of punishment, illegal facts, and punishment basis, improve the efficiency of obtaining punishment basis for supervisors, provide decision-making support, reduce "knowledge silos", and improve supervision intelligence.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119047551A_ABST
    Figure CN119047551A_ABST
Patent Text Reader

Abstract

The invention discloses a knowledge graph construction method surrounding a special equipment administrative punishment document. The method comprises the following steps: acquiring an administrative punishment entity to be identified; the to-be-recognized administrative punishment entity is input into an entity recognition model, an entity recognition result is obtained, the entity recognition model is formed through training of a training set, and the training set is an elevator administrative punishment document; processing the entity recognition result, and obtaining an administrative punishment entity relationship by using a rule matching method; and based on the administrative penalty entity relationship, performing knowledge fusion by using a similarity algorithm, and storing the knowledge fusion to a Neo4j graph database to obtain a knowledge graph of the administrative penalty document. According to the method, knowledge extraction based on the administrative penalty documents has important significance on intelligent supervision, meanwhile, the elevator administrative penalty knowledge graph is constructed, administrative penalty digital resources are effectively utilized, the problem of connection of'knowledge islands' in the supervision process can be solved to a certain extent, and a data foundation is laid for intelligent supervision.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention belongs to the technical field of knowledge element identification of administrative penalty documents for special equipment, and in particular, relates to a method for constructing a knowledge graph around administrative penalty documents for special equipment. Background Art

[0002] The new transformation and upgrading from traditional manual supervision to intelligent supervision is not only the deep integration of big data, Internet of Things, blockchain, artificial intelligence and the real economy, but also the inherent requirement for promoting the modernization of market supervision, building a high-standard market system and achieving high-quality development. Against the background of the development of new technologies such as the Internet of Things, big data and artificial intelligence, how to use modern information technology to strengthen the ability and level of elevator intelligent supervision has become a key and difficult point in elevator safety management. In addition, with the improvement of intelligence and information technology, the amount of elevator safety intelligent supervision data has also exploded. So how to integrate and manage these data, realize information sharing, and improve the level of elevator intelligent supervision is still an urgent problem to be solved. In addition, in the face of the growing demand for knowledge retrieval and interaction in interdisciplinary disciplines, how to comprehensively and accurately retrieve the required information from massive and heterogeneous data and recommend it to decision makers in an intelligent way has become a hot research issue.

[0003] Knowledge graph technology is a new type of massive knowledge management and service model that has emerged in the context of the big data era. By applying knowledge extraction, knowledge fusion, and knowledge reasoning technology of knowledge graphs, multi-source heterogeneous data in the field of elevator supervision can be integrated and integrated to complete the retrieval and sharing of elevator supervision knowledge. The field of elevator safety supervision contains a large amount of multi-source heterogeneous data, such as relevant laws, administrative regulations, departmental regulations, safety technical specifications, technical standards, judicial precedents, administrative penalty documents, and other data. Among them, the administrative penalty documents record in detail the illegal behavior of the parties, the basis for punishment, the results of the punishment, and other information, which can provide the most intuitive supervision basis for on-site supervisors, thereby improving the supervision efficiency of supervisors.

[0004] Knowledge graphs are mainly divided into general fields and vertical fields according to the scope of knowledge they contain. The characteristics of general field knowledge graphs are large data volume, strong versatility, wide knowledge scope, low authority, etc., while vertical field knowledge graphs have the characteristics of relatively small data volume, high quality, high density, concentrated knowledge, high authority and high knowledge accuracy. At present, the research and application of general knowledge graphs are developing rapidly. There are Satori, Probase, NELL, DBPedia and Freebase abroad, and Baidu Zhixin, Sogou Zhilifang and CN-DBPedia in China. Foreign vertical field knowledge graphs involve GeoNames in the geographical field, DBLife in the academic field and UniProtKb in the biological field. Domestic scholars have also conducted in-depth exploration in the fields of medical care, finance, education, industrial engineering, natural environment, public event governance, safety supervision, smart justice, etc. In the field of safety supervision, scholars have mainly conducted research on coal mine safety supervision, explosion-proof safety supervision, safe driving management, risk vehicle warning, and hazardous chemicals transportation safety supervision. In summary, knowledge graph technology has been widely used in the field of safety supervision and judicial fields. Therefore, the introduction of knowledge graphs can further enrich the theory of elevator safety supervision and expand the perspective from the traditional safety supervision model to the new field of intelligent supervision using big data analysis technology. The knowledge graph based on administrative penalty cases belongs to a type of domain knowledge graph that focuses on the information and relationships of illegal events. It is an important tool for effectively extracting and organizing massive illegal events, which helps to further improve the intelligence of supervision. However, there is relatively little research on the extraction of administrative penalty documents. Therefore, there is an urgent need for a knowledge graph construction method around administrative penalty documents for special equipment. Summary of the invention

[0005] In order to solve the above technical problems, the present invention proposes a knowledge graph construction method around administrative penalty documents for special equipment. The knowledge extraction based on administrative penalty documents is of great significance for intelligent supervision. At the same time, the knowledge graph of elevator administrative penalties is constructed, and the effective use of administrative penalty digital resources can solve the connectivity problem of "knowledge islands" in the supervision process to a certain extent, laying a data foundation for realizing smart supervision.

[0006] To achieve the above objectives, the present invention provides a method for constructing a knowledge graph around administrative penalty documents for special equipment, including:

[0007] Obtain the administrative penalty entity to be identified;

[0008] Input the administrative penalty entity to be identified into an entity recognition model to obtain an entity recognition result, wherein the entity recognition model is trained by a training set, and the training set is an administrative penalty document for elevators;

[0009] Processing the entity recognition results, and obtaining administrative penalty entity relationships based on a rule matching method;

[0010] Based on the entity relationship of administrative penalty, knowledge fusion is performed using a similarity algorithm and stored in the Neo4j graph database to obtain a knowledge graph of administrative penalty documents.

[0011] Optionally, the elevator administrative penalty document includes: penalty authority, penalty object, illegal behavior, illegal facts, penalty basis, penalty result, and penalty date.

[0012] Optionally, the entity recognition model includes: an XLNet layer, a BiLSTM layer and a CRF layer;

[0013] The XLNet layer is used to obtain word vectors;

[0014] The BiLSTM layer is used to perform semantic encoding according to the word vector;

[0015] The CRF layer is used to perform semantic decoding and obtain recognition results.

[0016] Optionally, the Transformer-XL mechanism is used in the XLNet layer to obtain word vectors.

[0017] Optionally, the BiLSTM layer includes: an improved recurrent neural network;

[0018] The improved recurrent neural network includes: adding an input gate, a forget gate and an output gate on the basis of the recurrent neural network;

[0019] The input gate is used to control the input of variables;

[0020] The forget gate is used to control the state of the cell unit;

[0021] The output gate is used to control the variable output.

[0022] Optionally, processing the recognition result to obtain the administrative penalty entity relationship includes:

[0023] Performing knowledge extraction on the recognition result to obtain a knowledge extraction result;

[0024] The knowledge extraction results are subjected to knowledge fusion to obtain a fusion result, namely, an administrative penalty entity relationship.

[0025] Optionally, the knowledge extraction methods include: entity extraction, relationship extraction and attribute extraction.

[0026] Optionally, performing knowledge fusion on the knowledge extraction result to obtain the fusion result includes:

[0027] Performing cluster analysis on the knowledge extraction results using a similarity algorithm to obtain entity clustering results;

[0028] Based on the entity clustering results, knowledge fusion is performed to obtain the fusion result, that is, the administrative penalty entity relationship.

[0029] Compared with the prior art, the present invention has the following advantages and technical effects:

[0030] The knowledge extraction of administrative penalty documents in the present invention is of great significance to government supervision. It can effectively identify entities such as the objects of punishment, illegal facts, basis for punishment, and results of punishment, and provide supervisors with information such as the basis for punishment, the illegal behaviors of enterprises, and benchmark references for discretion, prompting on-site supervisors to focus on the objects and contents of supervision, thereby improving supervision efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] The drawings constituting a part of the present application are used to provide a further understanding of the present application. The illustrative embodiments and descriptions of the present application are used to explain the present application and do not constitute an improper limitation on the present application. In the drawings:

[0032] Figure 1 It is a flow chart of a method for constructing a knowledge graph around administrative penalty documents for special equipment according to an embodiment of the present invention;

[0033] Figure 2 This is a framework diagram of an ontology model in the field of elevator administrative penalties according to an embodiment of the present invention;

[0034] Figure 3 A diagram of the Transformer-XL cycle mechanism according to an embodiment of the present invention;

[0035] Figure 4 This is a BiLSTM model diagram of an embodiment of the present invention;

[0036] Figure 5 This is a structural diagram of the XLNet-BiLSTM-CRF model of an embodiment of the present invention;

[0037] Figure 6 This is a result analysis diagram based on the XLNet-BiLSTM-CRF model of an embodiment of the present invention;

[0038] Figure 7 A knowledge graph of an elevator administrative penalty case according to an embodiment of the present invention;

[0039] Figure 8 This is a query result diagram based on "penalty object" in an embodiment of the present invention;

[0040] Fig. 9 The query result based on "penalty area" in an embodiment of the present invention;

[0041] Fig.10 This is the query result based on the “penalty result” in an embodiment of the present invention. DETAILED DESCRIPTION

[0042] It should be noted that, in the absence of conflict, the embodiments and features in the embodiments of the present application can be combined with each other. The present application will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0043] It should be noted that the steps shown in the flowcharts of the accompanying drawings can be executed in a computer system such as a set of computer executable instructions, and that, although a logical order is shown in the flowcharts, in some cases, the steps shown or described can be executed in an order different from that shown here.

[0044] Special equipment refers to boilers, pressure vessels (including gas cylinders), pressure pipelines, elevators, lifting machinery, passenger ropeways, large amusement facilities, special motor vehicles within sites (factories), and other special equipment that are subject to this Law as prescribed by laws and administrative regulations, which pose a great risk to the safety of persons and property.

[0045] This invention proposes a knowledge graph construction method around special equipment administrative penalty documents, such as Figure 1 , specifically including:

[0046] Obtain the administrative penalty entity to be identified;

[0047] Input the administrative penalty entity to be identified into the entity recognition model to obtain the entity recognition result. The entity recognition model is trained by the training set, which is the administrative penalty document for elevators;

[0048] Process the entity recognition results and obtain the administrative penalty entity relationship based on the rule matching method;

[0049] Based on the entity relationship of administrative penalty, the similarity algorithm is used to perform knowledge fusion and store it in the Neo4j graph database to obtain the knowledge graph of administrative penalty documents.

[0050] Furthermore, the administrative penalty document for elevators includes: penalty authority, penalty object, illegal behavior, illegal facts, penalty basis, penalty result, and penalty date;

[0051] Specifically, Figure 2 As shown, combined with the text of the elevator administrative penalty case and the identified entity categories in the field of elevator administrative penalty, the present invention defines 6 conceptual relationships to describe the relationship between entities such as the penalty object and the penalty authority, penalty date, penalty basis, and illegal facts.

[0052] Furthermore, the entity recognition model includes: XLNet layer, BiLSTM layer and CRF layer;

[0053] XLNet pre-trained language layer is used to obtain word vectors;

[0054] The BiLSTM layer is used to perform semantic encoding based on word vectors;

[0055] The CRF layer is used for semantic decoding to obtain recognition results.

[0056] Furthermore, the Transformer-XL mechanism is used in the XLNet layer to obtain word vectors.

[0057] Further, the BiLSTM layer includes: an improved recurrent neural network;

[0058] The improved recurrent neural network includes: adding an input gate, a forget gate and an output gate on the basis of the recurrent neural network;

[0059] The input gate is used to control the input of variables;

[0060] The forget gate is used to control the state of the cell unit;

[0061] The output gate is used to control the variable output.

[0062] Specifically, the present invention uses natural language technology and deep learning models such as XLNet, BiLSTM, and CRF to perform named entity recognition on administrative penalty documents, and conducts comparative analysis with pre-trained models such as BERT and ALBERT, aiming to explore a more efficient method for automatic entity recognition in administrative penalty documents, while also providing support for knowledge graph construction, knowledge query, etc.

[0063] Furthermore, the recognition results are processed to obtain the administrative penalty entity relationships including:

[0064] Perform knowledge extraction on the recognition results to obtain knowledge extraction results;

[0065] The knowledge extraction results are fused to obtain the fusion result, i.e., the administrative penalty entity relationship.

[0066] Furthermore, the methods of knowledge extraction include: entity extraction, relationship extraction and attribute extraction.

[0067] Furthermore, the knowledge extraction results are fused to obtain the fusion results including:

[0068] Use similarity algorithm to perform cluster analysis on knowledge extraction results to obtain entity clustering results;

[0069] The entity clustering results are subjected to knowledge fusion to obtain the fusion result, namely the administrative penalty entity relationship.

[0070] This embodiment takes the elevator field as an example:

[0071] Key models:

[0072] (1) XLNet pre-trained language model

[0073] The research on language models has gone through a period of development, from traditional word embedding models Word2Vec and Glove to the currently widely used GPT, ELMo, BERT and XLNet. The XLNet model not only has the advantages of the BERT model but also improves the defects of the BERT model. At the same time, it overcomes the defect of information loss in the Mask mechanism of the Bert model. XLNet is a generalized autoregressive pre-training method proposed by Yang et al. It has made improvements on the basis of previous pre-training models and mainly adopts three mechanisms: Permutation Language Model (PLM), two-stream self-attention mechanism and Transformer-XL structure. PLM is an important method proposed by XLNet to achieve the acquisition of bidirectional semantic information. XLNet is a balanced autoregressive (AR) and autoencoding (AE) language method. It rearranges sentences in the Transformer through the Attentionmask matrix while maintaining the unidirectional mode of the autoregressive model. The role of the two-stream self-attention mechanism is to hope to see not only the position information of xi when predicting xi, but also the position and content information of other words. XLNet uses Transformer-XL to solve the problem of very long sentences. The Transformer architecture used by most existing pre-trained language models does not work well for long text encoding because it is unable to model dependencies that exceed a fixed length. Based on the Transformer structure, Transformer-XL introduces relative position encoding and a loop mechanism. Through the loop mechanism, each segment is individually invested in calculating self-attention, and the output of each layer is stored as a hidden state in the memory storage unit between the segments as an additional input for predicting the next segment, allowing the model to capture longer-distance dependencies. The specific operation is as follows Figure 3 As shown, the dotted area represents the information of the attention layer of the previous segment. The semantic information of the previous segment can be used to predict the next segment. In this way, long-distance dependencies are captured.

[0074] (2) BiLSTM layer

[0075] Control the input, output and state of the cell units of the variables respectively. It effectively solves the problem of gradient explosion or gradient disappearance in long sequence texts. However, the unidirectional LSTM can only learn forward information but cannot learn backward information. Therefore, Graves et al. proposed BiLSTM composed of forward LSTM and backward LSTM. The basic idea is to perform forward propagation and backward propagation on each word in the sequence respectively, and then connect the results and output them, so as to better capture the bidirectional information of long sequence texts. Bidirectional LSTM applies a forward and backward LSTM network to each training sequence. The two LSTM networks are connected to the same output layer, and the forward and backward states of the sentence are recognized at the same time. LSTM adds three switches on the basis of RNN to control the input and output of the unit, namely forget gate, input gate, and output gate. The structure of the BiLSTM model is as follows Figure 4 shown.

[0076] (3) CRF layer

[0077] The CRF (Conditional Random Field) model has been widely used in sequence labeling problems in recent years and has achieved good results in many research fields. It can effectively solve the problem of unreasonable label prediction. Assume that the sequence length is n and the observation sequence is X{x1,x2,…x n}, the label sequence is y1,y2,…y n}, and its corresponding score function S is shown in formula (1).

[0078]

[0079] Where: A yi,yi+1 For labels from y i Transfer to y i+1 Transfer score; P i,yi Map word i to label y i The non-normalized probability of; S(X,y) is the sum of the scores of each label position. S(X,y) is normalized by the softmax function to obtain the normalized probability. The calculation formula is shown in formula (2).

[0080]

[0081] Where y' is all possible label sequences. Finally, the model is decoded by the Viterbi algorithm to solve the optimal output y * , the calculation formula is shown in formula (3).

[0082] y * =argmax(log P(y|x)) (3)

[0083] Construction of named entity recognition model for administrative penalty documents:

[0084] In the administrative penalty documents for elevators, the penalty authority, the penalty object, the illegal facts and other entity contents often appear continuously as context. By using the XLNet model to extract context features in the text, we can obtain bidirectional context information and mine long-distance text features, which is helpful to improve the recognition effect of administrative penalty documents. Based on this, the naming recognition algorithm of this embodiment is based on the XLNet pre-trained language model, and constructs the XLNet-BiLSTM-CRF model to perform named entity recognition on administrative penalty documents. Figure 5 In the pre-training stage, the XLNet model is embedded to generate word vectors, which are then input into BiLSTM for semantic encoding, and then semantic decoding is performed through CRF to output the label sequence with the highest probability.

[0085] Experiment and result analysis:

[0086] (1) Data source and preprocessing

[0087] This embodiment mainly takes elevator administrative penalty cases as the research object. The data comes from the market supervision administrative penalty website. By searching for elevator-related words, a total of 685 administrative penalty documents on the elevator industry are obtained. The data is preprocessed as follows: format conversion, converting the data text from PDF format to TXT format for subsequent annotation; format normalization, unifying the format of the data text, deleting charts and tables in the file, and removing repeated paragraphs, redundant spaces and blank lines, etc.

[0088] (2) Data Labeling

[0089] The pre-processed elevator administrative penalty cases are manually annotated in the form of BIOE, where "B" represents the position of the first word in the entity to be extracted, "I" represents the middle position of the entity to be extracted, "E" represents the end position of the entity to be extracted, and the remaining non-target words are represented by "O". According to the ontology structure designed in the field of elevator administrative penalties, seven types of entities are defined, including penalty authority, penalty object, penalty date, penalty location, illegal facts, penalty basis, and penalty result, with labels of "AUTHORITY", "TARGET", "DATE", "PLACE", "FACT", "BASISI", and "RESULT". The entity annotation style labels are shown in Table 1, and the data annotation examples are shown in Table 2.

[0090] Table 1

[0091] Entity Class Entity begins Entity Content Entity tail Punishment authority B-AUTHORITY I-AUTHORITY E-AUTHORITY Punishment B-TARGET I-TARGET E-TARGET Date of penalty B-DATE I-DATE E-DATE Place of punishment B-PLACE I-PLACE E-PLACE Illegal Facts B-FACT I-FACT E-FACT Basis of punishment B-BASIS I-BASIS E-BASIS Penalty Result B-RESULT I-RESULT E-RESULT

[0092] Table 2

[0093] sequence Annotation sequence Annotation sequence Annotation sequence Annotation machine O 4 O city I-AUTHORITY Penalty O Structure O 3 O field I-AUTHORITY Decision O generation O 6 O Supervisor I-AUTHORITY Certainly O code O 2 O Supervision I-AUTHORITY Book O : O 8 O Tube I-AUTHORITY 2 O superior B-AUTHORITY reason I-AUTHORITY l O ocean I-AUTHORITY Bureau E-AUTHORITY 3 O rainbow I-AUTHORITY OK O l O mouth I-AUTHORITY Politics O 9 O district I-AUTHORITY Department O

[0094] (3) Test method and parameter setting

[0095] The experimental model parameters are set as follows: The XLNet-base model version selected in this embodiment is chinese-xlnet-base, which has 12 layers, 768 dimensions of hidden layers, 12 attention heads, and 128 hidden layers of the BiLSTM model. To avoid overfitting, dropout is set to 0.1. The evaluation indicators of experimental performance include precision, recall, and F1 value. The labeled data is divided into training set, test set, and validation set in a ratio of 8:1:1, and input into the model for training.

[0096] (4) Evaluation indicators

[0097] This embodiment uses precision (P), recall (R) and F1-score, which are commonly used in natural language processing, as measurement indicators to test the effect of each model in named entity recognition

[38] . The larger the F1-score, the higher the quality of the model. The calculation process is shown in formulas (4)-(6).

[0098]

[0099] Among them, TP represents the positive examples with correct output in the model operation, FP represents the positive examples with incorrect output in the model operation, and FN represents the negative examples with incorrect output in the model.

[0100] (5) Analysis of test results

[0101] In order to verify the effectiveness of the XLNet-BiLSTM-CRF administrative penalty document naming entity method proposed in this embodiment, this embodiment conducts comparative experiments on the commonly used named entity recognition models BiLSTM-CRF, BERT-BiLSTM-CRF, ALBERT-BiLSTM-CRF, and XLNet-BiLSTM-CRF models. The experimental results are shown in Table 3.

[0102] Table 3

[0103] Model Accuracy Recall F1 value BiLSTM-CRF 0.6415 0.7138 0.7757 BERT-BiLSTM-CRF 0.8260 0.8621 0.8437 ALBERT-BiLSTM-CRF 0.8590 0.8529 0.8560 XLNet-BiLSTM-CRF 0.9467 0.9449 0.9458

[0104] Similarly, for the 7 types of penalty entities in the field of elevator administrative penalties, the evaluation indicators of each type of penalty entity are measured by P, R and F1, and the optimal XLNet-BiLSTM-CRF model is selected for display. The results are as follows: Figure 6 shown.

[0105] (5) Demonstration of named entity recognition application

[0106] In order to analyze the automatic recognition effect of the model of this embodiment in administrative penalty documents, an administrative penalty document is randomly selected and the entity recognition results are displayed. The specific recognition results are shown in Table 7.

[0107] Table 7

[0108]

[0109] Relation Extraction

[0110] Relation extraction refers to extracting specific relationships between entities from unstructured text data sets. Relation extraction can be done using template-based, supervised learning-based, and weakly supervised learning-based methods. Since this embodiment is to extract entity relationships in a limited field, and the relationships between entities are relatively clear, it is a better choice to use a relationship extraction method based on rule matching. This embodiment designs a rule extraction process based on trigger words based on the characteristics of the elevator administrative penalty text. First, the relationship set base relationship dictionary definition table that needs to be extracted is defined, as shown in Table 5.

[0111] Table 5

[0112] Serial number Relationship Type Label Relationship Dictionary 1 The punishment authority is Authority_is 'Administrative penalty','Decision','Administrative penalty decision' 2 The penalty date is Date_is 'According to', 'in', 'to', 'according to' 3 The area is Area_is 'Residential address', 'residence', 'address' 4 The illegal fact is Fact_is 'Party','Existence','Facts','Above' 5 The penalty is based on Basis_is 'Based on', 'according to', 'violated', 'referenced' 6 The penalty result is Results_is 'Subject', 'Order', 'Total', 'The penalties are as follows', 'Give'

[0113] In the process of relation extraction, first input the data set that needs to be subjected to the relation extraction task, and use ";,?,!,." as delimiters to segment the sentences in the input data set to obtain sentences. Secondly, find the trigger word of the sentence according to Table 5, retain the entity pair with the closest distance, remove other content, and judge the relation type according to the relation type feature word to which the trigger word belongs. Judge the extracted relation type. If the relation type is correct, the relation extraction task of the sentence is completed, and go to the next sentence, and repeat the above steps; if the relation type is wrong, re-extract the relation until it is correct, and then the relation extraction of the sentence is completed, and go to the next sentence. By traversing the database, the relationship extraction results performed according to the above process are shown in Table 6.

[0114] Table 6

[0115]

[0116] Knowledge Fusion

[0117] Since there are differences in the written expressions in administrative penalty documents, there are often multiple ways of expressing the same concept, such as "order the party to immediately stop the illegal behavior", "order to stop the illegal behavior", "order the party to stop the illegal behavior", etc. In order to further improve the quality of the knowledge graph, this embodiment performs entity clustering analysis on the entities identified by the model, and uses a similarity algorithm to align and analyze the results. The results are shown in Table 4. According to the results of entity clustering, manual intervention is used to perform entity fusion, and the entities and relationships are regenerated after fusion.

[0118] Table 4

[0119]

[0120] Knowledge storage and query

[0121] Through named entity recognition, relationship extraction and knowledge fusion, the "entity-relationship-entity" triple data form is formed, and the extracted data is converted into csv format. On this basis, Python's Py2neo module is used to use Node, Relationship and other methods to complete the creation of entities and relationships, and the merge method is used to complete the import of nodes and edges. A total of 1508 entities and 2682 relationships are imported. Figure 7 The overall visualization effect of the elevator administrative penalty field is shown.

[0122] The entity of the penalty authority is queried through the Cypher language of the Neo4j graph database. According to the Neo4j query statement "match(n{name'Guizhou Elevator Co., Ltd.'})--(b)return n, b", the following can be obtained: Figure 8 result.

[0123] Depend on Figure 8 From the results, we can see that when searching for the punished object "Guizhou Elevator Co., Ltd.", the relevant punishment agency, the illegal facts of the enterprise, the punishment results borne by the enterprise, etc. will appear. According to the display of the knowledge graph, it can be clearly seen that "Guizhou Elevator Co., Ltd." was punished by the Xingyi Municipal Market Supervision Administration on December 28, 2020. The main illegal facts were failure to report to the local special equipment safety supervision and management department and failure to report to the local special equipment safety supervision and management department. The punishment result was an order to stop the illegal behavior and a fine of RMB 10,000. This can provide the on-site supervisory personnel with a basis for punishment when they impose penalties, thereby improving the efficiency of on-site supervision. In addition, by inquiring about the companies punished in various regions, the number of punished companies in each region and the names of related companies can be determined, providing supervisory personnel with a list of key supervision targets. The results are as follows Fig. 9 shown.

[0124] In addition, by searching for "fine of 10,000 yuan", you can identify the names of related companies and understand the illegal facts of related companies, such as Fig.10 As shown, it provides a basis for on-site supervisory personnel to exercise discretion.

[0125] This embodiment takes the elevator field as an example, introduces the XLNet pre-trained model into the BILSTM-CRF model, and constructs a named entity recognition model for administrative penalty documents based on XLNet-BILSTM-CRF. It also proves that the XLNet pre-trained model is superior to the BERT and ALBERT models in the recognition process of administrative penalty documents. This method can effectively solve the problem of long entities in administrative penalty texts. At the same time, a trigger word rule extraction process is designed to extract the relationship in the administrative penalty document, and the identified entities are aligned using a similarity algorithm. The knowledge graph of elevator administrative penalty documents is constructed, and the knowledge base of administrative penalty documents is formed. The query and retrieval are performed on it, and the accuracy of the database is further verified. In future research, the corpus scale can be further increased, and the effect of the named entity recognition model can be improved. At the same time, it is necessary to further apply the intelligent retrieval and similar case recommendation of the administrative penalty knowledge graph database, so as to give full play to the role of the knowledge graph database and further promote the process of smart supervision.

[0126] The above are only preferred specific implementations of the present application, but the protection scope of the present application is not limited thereto. Any changes or substitutions that can be easily thought of by a person skilled in the art within the technical scope disclosed in the present application should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be based on the protection scope of the claims.

Claims

1. A method for constructing a knowledge graph around administrative penalty documents for special equipment, characterized in that: include: Obtain the administrative penalty entity to be identified; Input the administrative penalty entity to be identified into an entity recognition model to obtain an entity recognition result, wherein the entity recognition model is trained by a training set, and the training set is an administrative penalty document for elevators; Processing the entity recognition results, and obtaining the administrative penalty entity relationship using a rule matching method; Based on the entity relationship of administrative penalty, knowledge fusion is performed using a similarity algorithm and stored in the Neo4j graph database to obtain a knowledge graph of administrative penalty documents.

2. According to claim 1, a knowledge graph construction method around administrative penalty documents for special equipment is characterized in that: The elevator administrative penalty document includes: penalty authority, penalty object, illegal behavior, illegal facts, penalty basis, penalty result, and penalty date.

3. According to claim 1, a knowledge graph construction method around administrative penalty documents for special equipment is characterized in that: The entity recognition model includes: an XLNet layer, a BiLSTM layer and a CRF layer; The XLNet layer is used to obtain word vectors; The BiLSTM layer is used to perform semantic encoding according to the word vector; The CRF layer is used to perform semantic decoding and obtain recognition results.

4. A method for constructing a knowledge graph around administrative penalty documents for special equipment according to claim 3, characterized in that: The Transformer-XL mechanism is used in the XLNet layer to obtain word vectors.

5. The method for constructing a knowledge graph around administrative penalty documents for special equipment according to claim 3 is characterized in that: The BiLSTM layer includes: an improved recurrent neural network; The improved recurrent neural network includes: adding an input gate, a forget gate and an output gate on the basis of the recurrent neural network; The input gate is used to control the input of variables; The forget gate is used to control the state of the cell unit; The output gate is used to control the variable output.

6. The method for constructing a knowledge graph around administrative penalty documents for special equipment according to claim 1 is characterized in that: Processing the identification results to obtain the administrative penalty entity relationship includes: Performing knowledge extraction on the recognition result to obtain a knowledge extraction result; The knowledge extraction results are subjected to knowledge fusion to obtain a fusion result, namely, an administrative penalty entity relationship.

7. A method for constructing a knowledge graph around administrative penalty documents for special equipment according to claim 6, characterized in that: The methods of knowledge extraction include: entity extraction, relationship extraction and attribute extraction.

8. The method for constructing a knowledge graph around administrative penalty documents for special equipment according to claim 6 is characterized in that: Performing knowledge fusion on the knowledge extraction results to obtain the fusion results includes: Performing cluster analysis on the knowledge extraction results using a similarity algorithm to obtain entity clustering results; Based on the entity clustering results, knowledge fusion is performed to obtain the fusion result, that is, the administrative penalty entity relationship.

Citation Information

Patent Citations

  • Legal field text analysis and recognition method and system, storage medium and terminal

    CN113011185A

  • Hazardous chemical substance transport vehicle supervision method and system based on knowledge graph, and storage bank

    CN116304096A

  • Tax field-oriented knowledge map construction method and system

    WO2021196520A1