CN111597350B - Rail transit event knowledge graph construction method based on deep learning - Google Patents

Rail transit event knowledge graph construction method based on deep learning Download PDF

Info

Publication number
CN111597350B
CN111597350B CN202010365826.3A CN202010365826A CN111597350B CN 111597350 B CN111597350 B CN 111597350B CN 202010365826 A CN202010365826 A CN 202010365826A CN 111597350 B CN111597350 B CN 111597350B
Authority
CN
China
Prior art keywords
event
template
model
events
similarity
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Active
Application number
CN202010365826.3A
Other languages
Chinese (zh)
Other versions
CN111597350A (en
Inventor
黑新宏
彭伟
朱磊
赵钦
王一川
姬文江
姚燕妮
焦瑞
董林靖
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Xian University of Technology
Original Assignee
Xian University of Technology
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Xian University of Technology filed Critical Xian University of Technology
Priority to CN202010365826.3A priority Critical patent/CN111597350B/en
Publication of CN111597350A publication Critical patent/CN111597350A/en
Application granted granted Critical
Publication of CN111597350B publication Critical patent/CN111597350B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
    • G06F16/30—Information retrieval; Database structures therefor; File system structures therefor of unstructured textual data
    • G06F16/36—Creation of semantic tools, e.g. ontology or thesauri
    • G06F16/367—Ontology
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
    • G06F16/30—Information retrieval; Database structures therefor; File system structures therefor of unstructured textual data
    • G06F16/35—Clustering; Classification
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F40/00—Handling natural language data
    • G06F40/20—Natural language analysis
    • G06F40/237—Lexical tools
    • G06F40/247—Thesauruses; Synonyms
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F40/00—Handling natural language data
    • G06F40/20—Natural language analysis
    • G06F40/279—Recognition of textual entities
    • Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
    • Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
    • Y02D—CLIMATE CHANGE MITIGATION TECHNOLOGIES IN INFORMATION AND COMMUNICATION TECHNOLOGIES [ICT], I.E. INFORMATION AND COMMUNICATION TECHNOLOGIES AIMING AT THE REDUCTION OF THEIR OWN ENERGY USE
    • Y02D10/00—Energy efficient computing, e.g. low power processors, power management or thermal management

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • General Engineering & Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Computational Linguistics (AREA)
  • Data Mining & Analysis (AREA)
  • Databases & Information Systems (AREA)
  • Health & Medical Sciences (AREA)
  • Artificial Intelligence (AREA)
  • Audiology, Speech & Language Pathology (AREA)
  • General Health & Medical Sciences (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Animal Behavior & Ethology (AREA)
  • Information Retrieval, Db Structures And Fs Structures Therefor (AREA)

Abstract

The invention discloses a method for constructing a rail transit event knowledge graph based on deep learning; constructing event recognition model training data by adopting a dictionary matching mode and a manual labeling mode; training a canonical event recognition model by adopting a BERT-BiLSTM-CRF algorithm, and automatically extracting a canonical item event from a rail transit design canonical text; adopting a word2vec model, clustering cosine similarity, and logically regressing a two-class model to unify the events output by the event recognition model; adopting a nonowball algorithm to construct training data of an event relation model; and (3) training a relationship identification model by adopting a BERT-BiLSTM-ATTENTION-SOFTMAX algorithm, and automatically extracting the relationship between the events. The informatization of the track traffic construction design engineering is improved, and the workload of constructing the map is reduced.

Description

Rail transit event knowledge graph construction method based on deep learning
Technical Field
The invention belongs to an important direction in the field of artificial intelligence, and particularly relates to a track traffic event knowledge graph construction method based on deep learning.
Background
With the rapid development of internet technology, numerous industries are deeply fused with emerging artificial intelligence technologies, and remarkable achievements are achieved. Urban rail transit is used as a standard of urban modernization, and plays an important role in promoting urban economic development. The track traffic construction engineering belongs to complex engineering and has the characteristics of huge scale, long construction period, huge investment and the like. The early design planning stage in the track traffic construction engineering is the basis of the later engineering, and the later construction can be guaranteed only by completing the initial design planning. However, in the design planning stage of the track traffic engineering, the referred design specification standards are various and the information amount of each specification item is huge, and the informatization degree of the whole track traffic construction engineering is low, so that the difficulty of inquiring the specification content in the design planning stage is caused. And there is a very high requirement on the professional ability of the designer in the design stage, making design work very challenging. Therefore, knowledge graph is needed to represent the standard knowledge of the rail transit design, and the informatization of the rail transit construction engineering is promoted.
Most of the current knowledge maps are entity knowledge maps with entities as cores, but entity information is separated from specific contexts, and the one-sided nature of semantic information exists. Compared with the entity, the event can express semantic information more clearly. The event expression is contained in the specification items of the rail transit design standard. The design specification is expressed in the form of an event knowledge graph. Compared with the traditional knowledge graph construction method, most of the knowledge graph construction methods are low in automation degree, time-consuming and labor-consuming, so that the method for constructing the rail transit event knowledge graph based on deep learning is provided, the automation degree is improved, and the workload is reduced.
Disclosure of Invention
The invention aims to provide a rail transit event knowledge graph construction method based on deep learning. The event knowledge graph is used for representing the specification, so that the semantics of the represented content are richer and more accurate. The problems of low automation degree, time consumption and labor consumption in the traditional map construction technology are solved by deep learning.
The method adopts the technical scheme that the training data of the rail transit event recognition model is constructed by adopting an event trigger word dictionary matching mode and a manual labeling mode; training a canonical event recognition model by adopting a BERT-BiLSTM-CRF algorithm, and automatically extracting a canonical item event from a rail transit design canonical text; adopting a word2vec model, clustering cosine similarity, and logically regressing a two-class model to unify the events output by the event recognition model; adopting a nonowball algorithm to construct training data of an event relation model; and (3) training a relationship identification model by adopting a BERT-BiLSTM-ATTENTION-SOFTMAX algorithm, and automatically extracting the relationship between the events to form a rail traffic event knowledge graph. The event knowledge graph construction process comprises the following steps:
and step 1, constructing training data of an event recognition model by adopting an event-triggered word dictionary matching and manual labeling mode for the original text.
And 2, preprocessing a track traffic design specification event extraction training set, dividing texts in the training set by specification items, and marking the parts of speech of the texts.
And 3, training a track traffic design specification event recognition model by using the text processed in the step 2 through a BERT-BiLSTM-CRF algorithm.
And 4, constructing event relation training data by adopting a nonroball algorithm on the original text.
And 5, extracting the track traffic design specification event relation generated in the step 4 to preprocess a training set, and dividing texts in the training set in an event pair mode.
And 6, training a relation recognition model by using the text processed in the step 5 by using a BERT-BiLSTM-ATTENTION-SOFTMAX algorithm.
And 7, preprocessing the rail transit design specification to divide the specification entry.
And 8, inputting the preprocessed track traffic specification text in the step 7 into the event recognition model generated in the step 3, and extracting events in the specification, wherein the events comprise event trigger words and event elements.
And 9, unifying the events identified in the step 8.
And step 10, storing the event identified in the step 9 into an event database.
Step 11, storing the event identified in step 9 into a graph database in the form of a triplet of event element-relationship-event trigger words.
And step 12, taking out the events from the event database generated in the step 10, forming event pairs, inputting the event pairs into the event relation recognition model generated in the step 6, and extracting the relation among the events in the specification.
And 13, storing the event pairs in the step 10 and the event relations extracted in the step 12 into a graph database in the form of a triplet of event trigger words-relation-event trigger words.
In the step 1, an event consists of an event trigger word and an event element; because most event trigger words have fixed vocabulary, a dictionary matching mode is adopted to accelerate manual annotation, and model training data is constructed; the dictionary can be expanded by means of synonym forests.
In step 3, the event recognition model is trained by using a BERT-BiLSTM-CRF algorithm, and the whole model consists of three parts, namely a BERT layer, a BiLSTM layer and a CRF layer. The BERT pre-training model is used for acquiring word vectors containing canonical context feature information, the BiLSTM layer is used for feature extraction, the CRF layer is used for learning constraint conditions of sentences by utilizing sequence information of the whole text, and filtering wrong prediction sequences.
And 4, constructing an event relation recognition model training set by using a semi-supervised swball algorithm. The swball algorithm comprises the following specific steps:
step 4.1, manually marking a small number of event relations to form an event relation table; each event relationship is for an event relationship table.
Step 4.2, matching original sentences containing the events in the event relation table in the original text by using the existing event relation table, and generating a template; the format of the template is in five-tuple form, which is < left >, event 1 type, < middle >, event 2 type, < right >; len is a length that can be set arbitrarily, < left > is a vector representation of len words to the left of event 1, < middle > is a word vector representation between event 1 and event 2, < right > is a vector representation of len words to the right of event; event 1 type is a numeric qualifying event and event 2 type is a numeric qualifying event.
Step 4.3, clustering the generated templates, wherein templates with similarity larger than a threshold value of 0.7 are clustered into one type, generating new templates by using an average method, and adding the new templates into a rule base for storing the templates; . From the steps of4.2 the format of the knowable templates can be written as
Figure BDA0002476714200000041
E 1 ，E 2 Event 1 type and event 2 type, respectively, representing template P,/->
Figure BDA0002476714200000042
Representation E 1 Vector representation of 3 vocabulary lengths on the left, < >>
Figure BDA0002476714200000043
Representation E 1 ，E 2 Vector representation of words between->
Figure BDA0002476714200000044
Representation E 2 Vector representation of three lexical lengths on the right. Similarity calculation between templates, examples are as follows, template 1: />
Figure BDA0002476714200000051
Template 2: />
Figure BDA0002476714200000052
If condition E is satisfied 1 ＝E 1 '&&E 2 ＝E' 2 I.e. to satisfy the template P 1 Event 1 type E of (2) 1 And template P 2 Event 1 type E' 1 Identical and template P 1 Event 2 type E of (2) 2 And template P 2 Event 2 type E 'of (2)' 2 Identical, template P 1 And template P 2 Can be defined by->
Figure BDA0002476714200000053
Calculated, mu 1 μ 2 μ 3 Is the weight, because->
Figure BDA0002476714200000054
Has larger influence on the calculation result of the similarity between templates, and mu can be set 2 >μ 1 >μ 3 The method comprises the steps of carrying out a first treatment on the surface of the If not meeting condition E 1 ＝E 1 '&&E 2 ＝E' 2 Template P 1 And template P 2 The similarity of (2) may be noted as 0.
Step 4.4, firstly, scanning an original text by using the event recognition model trained in the step 3, recognizing event types contained in the text, then, matching the original text by using the template in the rule base generated in the step 4.3, and converting the matched text into a five-tuple form of the template;
step 4.5, calculating the similarity between the new template generated in the step 4.4 and the template in the rule base, discarding the template with the similarity smaller than the threshold value 0.7, and adding the event in the template with the similarity larger than the threshold value 0.7 into the event relation table;
and 4.6, repeatedly executing the steps 4.2-4.5 until the original text processing is finished.
In step 6, the relationship identification model is trained using the BERT-BiLSTM-ATTENTION-SOFTMAX algorithm. The whole model consists of four parts, namely a BERT layer, a BiLSTM layer, an ATTENTION layer and a SOFTMAX layer. The BERT pre-training model is used for acquiring word vectors containing canonical context feature information, the BiLSTM layer is used for feature extraction, the ATTENTION layer is used for calculating ATTENTION probability to highlight importance degree of key words in the text by utilizing sequence information of the whole text, the SOFTMAX layer is used for generating probabilities of various relation categories, and the largest category probability is taken as model prediction category.
In step 9, there is a text referring to the same event in the specification text; to avoid creating a large amount of redundant information in the event database; the event unified processing algorithm is adopted, and the specific steps of the event unified processing algorithm are as follows:
step 9.1, training a word2vec model by utilizing a track traffic original text;
step 9.2, inputting a rail traffic event by using the word2vec model generated in the step 9.1, and generating an event vector;
step 9.3, calculating the similarity between the events by using cosine function values, and gathering the events with the similarity value larger than 0.8 into one class; the cosine function is as follows:
Figure BDA0002476714200000061
step 9.4, generating new events in the step 9.3, combining all the events at will, and calculating the similarity between event pairs;
and 9.5, inputting the similarity between the event pairs and the event into a trained logistic regression classification model, and judging the similarity of the events. The logistic regression mathematical model is as follows:
Figure BDA0002476714200000062
step 9.6, according to the classification result of step 9.5, discarding one event if the events are similar, and storing both events if the events are dissimilar.
The beneficial effects of the invention are as follows:
aiming at the problems of complex engineering information, defects of the traditional knowledge graph and large graph construction workload in the track traffic construction design stage, the invention provides a method for constructing a track traffic event knowledge graph based on deep learning. Constructing training data of a rail transit event recognition model by adopting an event trigger word dictionary matching mode and a manual labeling mode; training a canonical event recognition model by adopting a BERT-BiLSTM-CRF algorithm, and automatically extracting a canonical item event from a rail transit design canonical text; adopting a word2vec model, clustering cosine similarity, and logically regressing a two-class model to unify the events output by the event recognition model; adopting a nonowball algorithm to construct training data of an event relation model; and (3) training a relationship identification model by adopting a BERT-BiLSTM-ATTENTION-SOFTMAX algorithm, and automatically extracting the relationship between the events to form a rail traffic event knowledge graph. The informatization of the track traffic construction design engineering is improved, and the workload of constructing the map is reduced.
Drawings
FIG. 1 is a general flow chart of a method for constructing a rail transit event knowledge graph based on deep learning;
FIG. 2 is a process of constructing an event training data set by adopting dictionary matching and manual labeling in the track traffic event knowledge graph construction method based on deep learning;
FIG. 3 is a process of constructing a canonical event recognition model based on BERT-BiLSTM-CRF algorithm by using the deep learning-based rail transit event knowledge graph construction method;
FIG. 4 is a process of event unification for events output by an event recognition model by adopting a word2vec model, cosine similarity clustering and logistic regression two-class model in the track traffic event knowledge graph construction method based on deep learning;
FIG. 5 is a process of constructing training data of an event relationship model by adopting a swball algorithm according to the deep learning-based rail transit event knowledge graph construction method;
FIG. 6 is a process of constructing a relationship recognition model based on BERT-BiLSTM-ATTENTION-SOFTMAX algorithm in the method for constructing a rail transit event knowledge graph based on deep learning.
Detailed Description
The invention will be described in detail below with reference to the drawings and the detailed description.
Referring to fig. 1, the method for constructing the rail transit event knowledge graph based on deep learning specifically comprises the following steps:
step 1, as shown in fig. 2, training data of an event recognition model is constructed by adopting an event trigger word dictionary matching and manual labeling mode on an original text. The pseudo code of the labeling training set algorithm is as follows:
Figure BDA0002476714200000081
and 2, preprocessing a track traffic design specification event extraction training set, dividing texts in the training set by specification items, and marking the parts of speech of the texts.
And 3, as shown in fig. 3, training a track traffic design specification event recognition model by using the BERT-BiLSTM-CRF algorithm on the text processed in the step 2. The pseudo code for constructing the event recognition model is as follows:
Figure BDA0002476714200000091
and 4, as shown in fig. 5, constructing an event relation recognition model training set by adopting a semi-supervised swball algorithm on the original text. The swball algorithm comprises the following specific steps:
step 4.1, manually marking a small number of event relations to form an event relation table; each event relationship is for an event relationship table.
Step 4.2, matching original sentences containing the events in the event relation table in the original text by using the existing event relation table, and generating a template; the format of the template is in five-tuple form, which is < left >, event 1 type, < middle >, event 2 type, < right >; len is a length that can be set arbitrarily, < left > is a vector representation of len words to the left of event 1, < middle > is a word vector representation between event 1 and event 2, < right > is a vector representation of len words to the right of event; event 1 type is a numeric qualifying event and event 2 type is a numeric qualifying event.
Step 4.3, clustering the generated templates, wherein templates with similarity larger than a threshold value of 0.7 are clustered into one type, generating new templates by using an average method, and adding the new templates into a rule base for storing the templates; . The format of the template known from step 4.2 can be written as
Figure BDA0002476714200000101
E 1 ，E 2 Event 1 type and event 2 type, respectively, representing template P,/->
Figure BDA0002476714200000102
Representation E 1 Vector representation of 3 vocabulary lengths on the left, < >>
Figure BDA0002476714200000103
Representation E 1 ，E 2 Vector representation of words between->
Figure BDA0002476714200000104
Representation E 2 Vector representation of three lexical lengths on the right. Similarity calculation between templates, examples are as follows, template 1: />
Figure BDA0002476714200000105
Template 2:
Figure BDA0002476714200000106
if condition E is satisfied 1 ＝E 1 '&&E 2 ＝E' 2 I.e. to satisfy the template P 1 Event 1 type E of (2) 1 And template P 2 Event 1 type E' 1 Identical and template P 1 Event 2 type E of (2) 2 And template P 2 Event 2 type E 'of (2)' 2 Identical, template P 1 And template P 2 Can be defined by->
Figure BDA0002476714200000107
Calculated, mu 1 μ 2 μ 3 Is the weight, because->
Figure BDA0002476714200000108
Has larger influence on the calculation result of the similarity between templates, and mu can be set 2 >μ 1 >μ 3 The method comprises the steps of carrying out a first treatment on the surface of the If not meeting condition E 1 ＝E 1 '&&E 2 ＝E' 2 Template P 1 And template P 2 The similarity of (2) may be noted as 0.
Step 4.4, firstly, scanning an original text by using the event recognition model trained in the step 3, recognizing event types contained in the text, then, matching the original text by using the template in the rule base generated in the step 4.3, and converting the matched text into a five-tuple form of the template;
step 4.5, calculating the similarity between the new template generated in the step 4.4 and the template in the rule base, discarding the template with the similarity smaller than the threshold value 0.7, and adding the event in the template with the similarity larger than the threshold value 0.7 into the event relation table;
and 4.6, repeatedly executing the steps 4.2-4.5 until the original text processing is finished.
And 5, preprocessing the track traffic design specification event relation extraction training set generated in the step 4, and dividing the text in an event pair mode.
And 6, training a relation recognition model by using the text processed in the step 5 by using a BERT-BiLSTM-ATTENTION-SOFTMAX algorithm. The pseudo code for constructing the event relationship identification model is as follows, as shown in FIG. 6:
Figure BDA0002476714200000111
and 7, preprocessing the rail transit design specification to divide the specification entry.
And 8, inputting the preprocessed track traffic specification text in the step 7 into the event recognition model generated in the step 3, and extracting events in the specification, wherein the events comprise event trigger words and event elements.
And 9, unifying the events identified in the step 8 as shown in fig. 4. Text referring to the same event exists in the specification text; to avoid creating a large amount of redundant information in the event database; the event unified processing algorithm is adopted, and the specific steps of the event unified processing algorithm are as follows:
step 9.1, training a word2vec model by utilizing a track traffic original text;
step 9.2, inputting a rail traffic event by using the word2vec model generated in the step 9.1, and generating an event vector;
step 9.3, calculating the similarity between the events by using cosine function values, and gathering the events with the similarity value larger than 0.8 into one class; the cosine function is as follows:
Figure BDA0002476714200000121
step 9.4, generating new events in the step 9.3, combining all the events at will, and calculating the similarity between event pairs;
and 9.5, inputting the similarity between the event pairs and the event into a trained logistic regression classification model, and judging the similarity of the events. The logistic regression mathematical model is as follows:
Figure BDA0002476714200000122
step 9.6, according to the classification result of step 9.5, discarding one event if the events are similar, and storing both events if the events are dissimilar.
And step 10, storing the event identified in the step 9 into an event database.
Step 11, storing the event identified in step 9 into a graph database in the form of a triplet of event element-relationship-event trigger words. For example, "track center bed surface as emergency evacuation channel" stores the map database with < track center bed surface, subject, as > and < emergency evacuation channel, object, as >.
And step 12, taking out the events from the event database generated in the step 10, forming event pairs, inputting the event pairs into the event relation recognition model generated in the step 6, and extracting the relation among the events in the specification.
And 13, storing the event pairs in the step 10 and the event relations extracted in the step 12 into a graph database in the form of a triplet of event trigger words-relation-event trigger words. For example, "a track center bed surface is used as an emergency evacuation channel" and "a train end vehicle should be provided with a special end door and a get-off facility is configured with an event relationship of < as a condition relationship, and > is stored in a map database.
The invention adopts an event trigger word dictionary matching mode and a manual labeling mode to construct the training data of the rail transit event recognition model; training a canonical event recognition model by adopting a BERT-BiLSTM-CRF algorithm, and automatically extracting a canonical item event from a rail transit design canonical text; adopting a word2vec model, clustering cosine similarity, and logically regressing a two-class model to unify the events output by the event recognition model; adopting a nonowball algorithm to construct training data of an event relation model; and (3) training a relationship identification model by adopting a BERT-BiLSTM-ATTENTION-SOFTMAX algorithm, and automatically extracting the relationship between the events to form a rail traffic event knowledge graph. The informatization of the track traffic construction design engineering is improved, and the workload of constructing the map is reduced.

Claims (2)

1. The track traffic event knowledge graph construction method based on deep learning is characterized in that an event trigger word dictionary matching mode and a manual labeling mode are adopted to construct track traffic event recognition model training data; training a canonical event recognition model by adopting a BERT-BiLSTM-CRF algorithm, and automatically extracting a canonical item event from a rail transit design canonical text; adopting a word2vec model, clustering cosine similarity, and logically regressing a two-class model to unify the events output by the event recognition model; adopting a nonowball algorithm to construct training data of an event relation model; the BERT-BiLSTM-ATTENTION-SOFTMAX algorithm is adopted to train a relationship identification model, and relationships among events are automatically extracted to form a rail traffic event knowledge graph;
the method specifically comprises the following steps:
step 1, training data of an event recognition model is built by adopting an event trigger word dictionary matching and manual labeling mode for an original text, wherein in the step 1, an event consists of an event trigger word and an event element; because most event trigger words have fixed vocabulary, a dictionary matching mode is adopted to accelerate manual annotation, and model training data is constructed; the expansion of the dictionary may be by means of a synonym forest;
step 2, preprocessing a track traffic design specification event extraction training set, dividing texts in the training set by specification items, and marking the parts of speech of the texts;
step 3, training a track traffic design specification event recognition model by using the text processed in the step 2 through a BERT-BiLSTM-CRF algorithm; in the step 3, a BERT-BiLSTM-CRF algorithm is used for training an event recognition model, and the whole model consists of three parts, namely a BERT layer, a BiLSTM layer and a CRF layer; the BERT pre-training model is used for acquiring word vectors containing canonical context feature information, the BiLSTM layer is used for feature extraction, the CRF layer is used for learning constraint conditions of sentences by utilizing sequence information of the whole text, and filtering false prediction sequences;
and 4, constructing event relation training data by adopting a swball algorithm on the original text, wherein in the step 4, an event relation recognition model training set is constructed by utilizing a semi-supervision swball algorithm, and the specific steps of the swball algorithm are as follows:
step 4.1, manually marking a small number of event relations to form an event relation table; each event relationship is for an event relationship table;
step 4.2, matching original sentences containing the events in the event relation table in the original text by using the existing event relation table, and generating a template; the format of the template is in five-tuple form, which is < left >, event 1 type, < middle >, event 2 type, < right >; len is a length that can be set arbitrarily, < left > is a vector representation of len words to the left of event 1, < middle > is a word vector representation between event 1 and event 2, < right > is a vector representation of len words to the right of event; event 1 type is a numeric qualifying event, and event 2 type is a numeric qualifying event;
step 4.3, clustering the generated templates, wherein templates with similarity larger than a threshold value of 0.7 are clustered into one type, generating new templates by using an average method, and adding the new templates into a rule base for storing the templates; the format of the template known from step 4.2 can be written as
Figure FDA0004087374820000021
E 1 ，E 2 Event 1 type and event 2 type, respectively, representing template P,/->
Figure FDA0004087374820000022
Representation E 1 Vector representation of 3 vocabulary lengths on the left, < >>
Figure FDA0004087374820000023
Representation E 1 ，E 2 Vector representation of words between->
Figure FDA0004087374820000024
Representation E 2 Vector representation of three word lengths on the right, similarity calculation between templates, template 1: />
Figure FDA0004087374820000025
Template 2: />
Figure FDA0004087374820000026
If condition E is satisfied 1 ＝E 1 '&&E 2 ＝E' 2 I.e. to satisfy the template P 1 Event 1 type E of (2) 1 And template P 2 Event 1 type E' 1 Identical and template P 1 Event 2 type E of (2) 2 And template P 2 Event 2 type E 'of (2)' 2 Identical, template P 1 And template P 2 Can be defined by the similarity of
Figure FDA0004087374820000031
Calculated, mu 1 μ 2 μ 3 Is the weight, because->
Figure FDA0004087374820000032
Has larger influence on the calculation result of the similarity between templates, and mu can be set 2 >μ 1 >μ 3 The method comprises the steps of carrying out a first treatment on the surface of the If not meeting condition E 1 ＝E 1 '&&E 2 ＝E' 2 Template P 1 And template P 2 The similarity of (2) may be noted as 0; />
Step 4.4, firstly, scanning an original text by using the event recognition model trained in the step 3, recognizing event types contained in the text, then, matching the original text by using the template in the rule base generated in the step 4.3, and converting the matched text into a five-tuple form of the template;
step 4.5, calculating the similarity between the new template generated in the step 4.4 and the template in the rule base, discarding the template with the similarity smaller than the threshold value 0.7, and adding the event in the template with the similarity larger than the threshold value 0.7 into the event relation table;
step 4.6, repeatedly executing the steps 4.2-4.5 until the original text processing is finished;
step 5, extracting a training set from the track traffic design specification event relation generated in the step 4 for preprocessing, and dividing texts in the training set in an event pair mode;
step 6, training a relation recognition model by using the BERT-BiLSTM-ATTENTION-SOFTMAX algorithm to the text processed in the step 5, wherein in the step 6, the relation recognition model is trained by using the BERT-BiLSTM-ATTENTION-SOFTMAX algorithm; the whole model consists of four parts, namely a BERT layer, a BiLSTM layer, an ATTENTION layer and a SOFTMAX layer; the BERT pre-training model is used for acquiring word vectors containing canonical context feature information, the BiLSTM layer is used for feature extraction, the ATTENTION layer is used for calculating ATTENTION probability to highlight importance degree of key words in the text by utilizing sequence information of the whole text, the SOFTMAX layer is used for generating probabilities of various relation categories, and the largest category probability is taken as a model prediction category;
step 7, preprocessing the rail transit design specification, and dividing the rail transit design specification by specification items;
step 8, inputting the track traffic specification text preprocessed in the step 7 into the event recognition model generated in the step 3, and extracting an event in the specification, wherein the event comprises an event trigger word and an event element;
step 9, unifying the events identified in the step 8;
step 10, storing the event identified in the step 9 into an event database;
step 11, storing the event identified in step 9 into a graph database in the form of a triplet of event element-relationship-event trigger word;
step 12, taking out the events from the event database generated in the step 10, forming event pairs, inputting the event pairs into the event relation recognition model generated in the step 6, and extracting the relation among the events in the specification;
and 13, storing the event pairs in the step 10 and the event relations extracted in the step 12 into a graph database in the form of a triplet of event trigger words-relation-event trigger words.
2. The method for constructing a knowledge graph of a rail transit event based on deep learning according to claim 1, wherein in the step 9, text referring to the same event exists in the canonical text; to avoid creating a large amount of redundant information in the event database; the event unified processing algorithm is adopted, and the specific steps of the event unified processing algorithm are as follows:
step 9.1, training a word2vec model by utilizing a track traffic original text;
step 9.2, inputting a rail traffic event by using the word2vec model generated in the step 9.1, and generating an event vector;
step 9.3, calculating the similarity between the events by using cosine function values, and gathering the events with the similarity value larger than 0.8 into one class; the cosine function is as follows:
Figure FDA0004087374820000051
step 9.4, generating new events in the step 9.3, combining all the events at will, and calculating the similarity between event pairs;
step 9.5, inputting the similarity between the event pairs and the event into a trained logistic regression classification model, and judging the similarity of the events; the logistic regression mathematical model is as follows:
Figure FDA0004087374820000052
step 9.6, according to the classification result of step 9.5, discarding one event if the events are similar, and storing both events if the events are dissimilar.
CN202010365826.3A 2020-04-30 2020-04-30 Rail transit event knowledge graph construction method based on deep learning Active CN111597350B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202010365826.3A CN111597350B (en) 2020-04-30 2020-04-30 Rail transit event knowledge graph construction method based on deep learning

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202010365826.3A CN111597350B (en) 2020-04-30 2020-04-30 Rail transit event knowledge graph construction method based on deep learning

Publications (2)

Publication Number Publication Date
CN111597350A CN111597350A (en) 2020-08-28
CN111597350B true CN111597350B (en) 2023-06-02

Family

ID=72186939

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202010365826.3A Active CN111597350B (en) 2020-04-30 2020-04-30 Rail transit event knowledge graph construction method based on deep learning

Country Status (1)

Country Link
CN (1) CN111597350B (en)

Families Citing this family (9)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112131401B (en) * 2020-09-14 2024-02-13 腾讯科技（深圳）有限公司 Concept knowledge graph construction method and device
CN112733874B (en) * 2020-10-23 2023-04-07 招商局重庆交通科研设计院有限公司 Suspicious vehicle discrimination method based on knowledge graph reasoning
CN112418696B (en) * 2020-11-27 2024-06-18 北京工业大学 Construction method and device of urban traffic dynamic knowledge graph
CN112463989A (en) * 2020-12-11 2021-03-09 交控科技股份有限公司 Knowledge graph-based information acquisition method and system
CN112800762A (en) * 2021-01-25 2021-05-14 上海犀语科技有限公司 Element content extraction method for processing text with format style
CN113268591B (en) * 2021-04-17 2022-11-01 中国人民解放军战略支援部队信息工程大学 Air target intention evidence judging method and system based on affair atlas
CN113535979A (en) * 2021-07-14 2021-10-22 中国地质大学（北京） Method and system for constructing knowledge graph in mineral field
CN113546426B (en) * 2021-07-21 2023-08-22 西安理工大学 Security policy generation method for data access event in game service
CN115269931B (en) * 2022-09-28 2022-11-29 深圳技术大学 Rail transit station data map system based on service drive and construction method thereof

Family Cites Families (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN107729309B (en) * 2016-08-11 2022-11-08 中兴通讯股份有限公司 Deep learning-based Chinese semantic analysis method and device
CN107908671B (en) * 2017-10-25 2022-02-01 南京擎盾信息科技有限公司 Knowledge graph construction method and system based on legal data
CN110633409B (en) * 2018-06-20 2023-06-09 上海财经大学 Automobile news event extraction method integrating rules and deep learning

Also Published As

Publication number Publication date
CN111597350A (en) 2020-08-28

Similar Documents

Publication Publication Date Title
CN111597350B (en) Rail transit event knowledge graph construction method based on deep learning
CN107315737B (en) Semantic logic processing method and system
CN109783818B (en) Enterprise industry classification method
CN111931506B (en) Entity relationship extraction method based on graph information enhancement
CN110275936B (en) Similar legal case retrieval method based on self-coding neural network
CN111209401A (en) System and method for classifying and processing sentiment polarity of online public opinion text information
CN107818141B (en) Biomedical event extraction method integrated with structured element recognition
CN113191148B (en) Rail transit entity identification method based on semi-supervised learning and clustering
CN112906397B (en) Short text entity disambiguation method
CN111832293B (en) Entity and relation joint extraction method based on head entity prediction
CN110717045A (en) Letter element automatic extraction method based on letter overview
CN111274817A (en) Intelligent software cost measurement method based on natural language processing technology
CN112084336A (en) Entity extraction and event classification method and device for expressway emergency
CN115292461B (en) Man-machine interaction learning method and system based on voice recognition
CN110991185A (en) Method and device for extracting attributes of entities in article
CN112364132A (en) Similarity calculation model and system based on dependency syntax and method for building system
CN113934909A (en) Financial event extraction method based on pre-training language and deep learning model
CN111597349B (en) Rail transit standard entity relation automatic completion method based on artificial intelligence
CN114239574A (en) Miner violation knowledge extraction method based on entity and relationship joint learning
CN114491024A (en) Small sample-based specific field multi-label text classification method
CN113312922A (en) Improved chapter-level triple information extraction method
CN113553400A (en) Construction method and device of enterprise knowledge graph entity link model
CN116010553A (en) Viewpoint retrieval system based on two-way coding and accurate matching signals
CN115618883A (en) Business semantic recognition method and device
CN116522165B (en) Public opinion text matching system and method based on twin structure

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination
GR01 Patent grant
GR01 Patent grant