CN109902283A - A kind of information output method and device - Google Patents
A kind of information output method and device Download PDFInfo
- Publication number
- CN109902283A CN109902283A CN201810415523.0A CN201810415523A CN109902283A CN 109902283 A CN109902283 A CN 109902283A CN 201810415523 A CN201810415523 A CN 201810415523A CN 109902283 A CN109902283 A CN 109902283A
- Authority
- CN
- China
- Prior art keywords
- text
- term vector
- failure
- semantic
- vector
- 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.)
- Granted
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F40/00—Handling natural language data
- G06F40/20—Natural language analysis
Abstract
This application discloses a kind of information output method and devices, wherein this method comprises: obtaining failure-description text, failure-description text is used for the failure for describing to occur in network；Pass through the semantic semantic vector for generating model and generating failure-description text；The corresponding semantic vector of related text of a plurality of types of target datas is obtained, the target data is for assisting analysis failure Producing reason；Calculate the correlation of the semantic vector and the semantic vector of the related text of every kind of target data of failure-description text；It determines and exports the first data, first data are that the target data or first data of the correlation maximum of the semantic vector of semantic vector and failure-description text in plurality of target data are the target data that semantic vector is greater than preset threshold with the correlation of the semantic vector of failure-description text in plurality of target data.By implementing the present processes, can accurately find out relevant to failure-description text for assisting the data of analyzing failure cause.
Description
Technical field
This application involves field of communication technology more particularly to a kind of information output methods and device.
Background technique
When the network equipment breaks down, it will affect normal communication, bring heavy losses to the work and life of people,
So the reparation in time of network equipment failure is extremely important.Currently, frontline engineer can be from event when the network equipment breaks down
The data that on-site collection is used to assist analyzing failure cause occur for barrier, for example, when one section before and after collection network equipment fault generation
In Key Performance Indicator (KPI), equipment alarm, the supplemental characteristics such as device log.And frontline engineer can be existing to failure
As being described, failure-description text is obtained.Frontline engineer is by the data such as the KPI of collection and failure-description text with failure work
Single form feeds back to O&M department.O&M engineer is according to the failure-description text in fault ticket, by the profession of itself
Knowledge selects the supplemental characteristics such as some KPI, equipment alarm, device log from the data that a line is collected manually.Further
Ground carries out abnormality detection these data chosen and mutually proves, so that analysis is out of order root because of place, to fault network
The reparation of network equipment provides guiding opinion.It is this by manually from supplemental characteristics such as KPI, equipment alarm, device logs
In select the fault detection method that supplemental characteristic relevant to failure-description text check analysis, efficiency low velocity is slow,
It is unable to satisfy increasingly increased network demand.
In the prior art by searching for the text with failure-description text keyword having the same, and according to the text
Associated parameter data progress failure checks analysis.But the high related text that can be used to assist analyzing failure cause of correlation and
It may be there is no identical keyword in failure-description text.Therefore, it cannot accurately be found and event by existing mode
Barrier description text is associated for assisting the data of analyzing failure cause.
Summary of the invention
This application provides a kind of information output method and device, can automatically and accurately find and failure-description
The relevant data for being used to assist analyzing failure cause of text.
In a first aspect, this application provides a kind of information output method, this method comprises: failure-description text is obtained, it should
Failure-description text is used for the failure for describing to occur in network；By it is semantic generate model generate failure-description text it is semantic to
Amount；The corresponding semantic vector of related text of a plurality of types of target datas is obtained, the target data is for assisting analysis
Failure Producing reason；Calculate the semantic vector of the semantic vector of failure-description text and the related text of every kind of target data
Correlation；The first data are determined and export, which is semantic vector and failure-description text in every kind of target data
The target data of the correlation maximum of semantic vector or first data are semantic vector and failure-description in every kind of target data
The correlation of the semantic vector of text is greater than the target data of preset threshold.
The application passes through the semantic vector of the semantic vector of comparison failure-description text and the related text of target data
Correlation can accurately find target data associated with failure-description text.For example, failure-description is that " industry is used
Family online is slow ", entitled " the downlink bandwidth control for the relative critical index for accident analysis that the application analyzes
Packet loss ratio processed ".As can be seen that from it is literal it is both upper it is any can match with associated ingredient, and the application is exactly
Both study, which is excavated, by semantic analysis has arrived domain knowledge as " networking speed and packet loss ratio have relationship ", just realize
Associated analysis.Therefore, it by implementing method described in first aspect, can automatically and accurately find out and event
The relevant data for being used to assist analyzing failure cause of barrier description text.
In a kind of possible embodiment, before obtaining failure-description text, also model can be generated by semanteme and generated
The corresponding semantic vector of the related text of a plurality of types of target datas.
And it also can be reserved for the corresponding semantic vector of related text of plurality of target data；Correspondingly, it obtains a variety of
The specific embodiment of the corresponding semantic vector of the related text of target data are as follows: obtain the plurality of target data of preservation
The corresponding semantic vector of related text.
By implementing the embodiment, the corresponding language of related text of plurality of target data can be pre-generated and saved
Adopted vector can directly use the related text of the plurality of target data saved corresponding after receiving failure-description text
The semantic vector of semantic vector and failure-description text carries out correlation calculations, thus do not have to receive failure-description text it
Afterwards, the corresponding semantic vector of related text of plurality of target data is temporarily generated.As it can be seen that by implementing the embodiment,
Be conducive to the semantic vector of the related text of the semantic vector that failure-description text is rapidly calculated and every kind of target data
Correlation.
In a kind of possible embodiment, the above-mentioned semantic model that generates is that the corresponding word of text is described according to historical failure
Vector matrix training generates, which includes that historical failure describes the corresponding term vector of each word in text, the word
Vector is used to indicate the semanteme of word.
The semantic generation model obtained by implementing embodiment training, can more accurately express the language of text
Justice.
In a kind of possible embodiment, above-mentioned a plurality of types of target datas include Key Performance Indicator, equipment announcement
At least two in police, device log；When above-mentioned target data is Key Performance Indicator, the related text of the target data is
The title of Key Performance Indicator；When above-mentioned target data is equipment alarm, the related text of the target data is equipment alarm
Mark；When above-mentioned target data is device log, the related text of the target data is the contents fragment of device log.
Second aspect, this application provides a kind of semantic training methods for generating model, this method comprises: obtaining training text
This corresponding term vector set, the word in term vector and training text for including in the term vector set correspond, the word to
Measure the semanteme for indicating word；It is to be made of at least one term vector that historical failure, which is described text conversion, according to term vector set
Term vector matrix；Semantic generation model is obtained according to the training of term vector matrix, which generates model for generating text
Semantic vector.
It optionally, can be by the corresponding term vector set of training text after obtaining the corresponding term vector set of training text
It is saved, so as to the subsequent term vector using in term vector set.
As it can be seen that method described in second aspect is gradually to model from the semantic of lexical level to the semanteme of sentence surface
Model is generated to semanteme, this semantic model training mode that generates is to meet the basic principle of language generation.Therefore, pass through reality
The semantic generation model that the training of method described in second aspect obtains is applied, the semanteme of text can be more accurately expressed.
In a kind of possible embodiment, it is by least one that historical failure, which is described text conversion, according to term vector set
The specific embodiment of the term vector matrix of a term vector composition are as follows: text is described to historical failure and carries out word segmentation processing, is obtained
Historical failure describes the corresponding word sequence being made of at least one word of text；Word sequence is obtained from term vector set includes
The corresponding term vector of word；The corresponding term vector of each word for including by word sequence forms term vector matrix.
By implementing the embodiment, historical failure accurately can be described text conversion is by least one term vector group
At term vector matrix.
In a kind of possible embodiment, when there is no the corresponding term vectors of word that word sequence includes in term vector set
When, generate the corresponding term vector of word that random vector includes as word sequence.
By implementing the embodiment, historical failure accurately can be described text conversion is by least one term vector group
At term vector matrix.
In a kind of possible embodiment, the semantic specific embodiment party for generating model is obtained according to the training of term vector matrix
Formula are as follows: obtain historical failure and describe the corresponding faulty equipment type of text；According to term vector matrix and class label training classification
Model, such distinguishing label include the faulty equipment type；Semantic generation model is obtained according to disaggregated model.
The semantic generation model obtained by implementing embodiment training, can more accurately express the language of text
Justice.
In a kind of possible embodiment, according to the specific implementation of term vector matrix and class label train classification models
Mode are as follows: term vector matrix and class label input neural network are iterated training, in each repetitive exercise to input
The parameter of term vector and neural network in the term vector matrix of neural network is adjusted, to generate the disaggregated model.Pass through
The semantic generation model that embodiment training obtains, can more accurately express the semanteme of text.
Optionally, also the term vector in the term vector matrix for using last time repetitive exercise to input can be updated term vector
The corresponding term vector of corresponding words in set.By implementing the embodiment, can be retouched according to the historical failure with domain knowledge
The term vector in corpus of text amendment term vector set is stated, the word of domain knowledge can be expressed by making the term vector in term vector set more
Semantic information.
The third aspect provides a kind of information output apparatus, which can be performed above-mentioned first aspect or
On the one hand the method in possible embodiment.The function can also be executed corresponding by hardware realization by hardware
Software realization.The hardware or software include one or more units corresponding with above-mentioned function.The unit can be software and/
Or hardware.Based on the same inventive concept, the principle and beneficial effect which solves the problems, such as may refer to above-mentioned
In first aspect or the possible embodiment of first aspect and beneficial effect, overlaps will not be repeated.
Fourth aspect provides a kind of model training apparatus, which can be performed above-mentioned second aspect or
Method in the two possible embodiments of aspect.The function can also be executed corresponding by hardware realization by hardware
Software realization.The hardware or software include one or more units corresponding with above-mentioned function.The unit can be software and/
Or hardware.Based on the same inventive concept, the principle and beneficial effect which solves the problems, such as may refer to above-mentioned
In second aspect or the possible embodiment of second aspect and beneficial effect, overlaps will not be repeated.
5th aspect, provides a kind of information output apparatus, which includes: processor, memory, communication
Interface；Processor, communication interface are connected with memory；Wherein, communication interface can be transceiver.Communication interface for realizing with
Communication between other network elements.Wherein, one or more programs are stored in memory, and processor calling is stored in this and deposits
Program in reservoir is to realize the scheme in above-mentioned first aspect or the possible embodiment of first aspect, the information output apparatus
The embodiment and beneficial effect solved the problems, such as may refer to above-mentioned first aspect or the possible embodiment of first aspect with
And beneficial effect, overlaps will not be repeated.
6th aspect, provides a kind of model training apparatus, which includes: processor, memory, communication
Interface；Processor, communication interface are connected with memory；Wherein, communication interface can be transceiver.Communication interface for realizing with
Communication between other network elements.Wherein, one or more programs are stored in memory, and processor calling is stored in this and deposits
Program in reservoir is to realize the scheme in above-mentioned second aspect or the possible embodiment of second aspect, the model training apparatus
The embodiment and beneficial effect solved the problems, such as may refer to above-mentioned second aspect or the possible embodiment of second aspect with
And beneficial effect, overlaps will not be repeated.
7th aspect, provides a kind of computer program product, when run on a computer, so that computer executes
Above-mentioned first aspect, second aspect, in the possible embodiment of the possible embodiment of first aspect or second aspect
Method.
Eighth aspect provides a kind of chip product of information output apparatus, executes above-mentioned first aspect or first aspect
Any possible embodiment in method.
9th aspect, provides a kind of chip product of model training apparatus, executes above-mentioned second aspect or second aspect
Any possible embodiment in method.
Tenth aspect, has mentioned for a kind of computer readable storage medium, is stored with instruction in computer readable storage medium,
When run on a computer, so that computer executes the method for above-mentioned first aspect or the possible embodiment party of first aspect
Method in formula.
Tenth on the one hand, has mentioned for a kind of computer readable storage medium, is stored with finger in computer readable storage medium
It enables, when run on a computer, so that computer executes the method for above-mentioned second aspect or the possible reality of second aspect
Apply the method in mode.
Detailed description of the invention
Fig. 1 is a kind of flow diagram of information output method provided by the embodiments of the present application；
Fig. 2 is a kind of flow diagram of semantic training method for generating model provided by the embodiments of the present application；
Fig. 3 is the schematic diagram for the neural network that a kind of CBOW algorithm provided by the embodiments of the present application uses；
Fig. 4 is a kind of structural schematic diagram of neural network for train classification models provided by the embodiments of the present application；
Fig. 5 is a kind of structural schematic diagram of information output apparatus provided by the embodiments of the present application；
Fig. 6 is a kind of structural schematic diagram of model training apparatus provided by the embodiments of the present application；
Fig. 7 is the structural schematic diagram of another information output apparatus provided by the embodiments of the present application；
Fig. 8 is the structural schematic diagram of another model training apparatus provided by the embodiments of the present application.
Specific embodiment
The application specific embodiment is described in further detail with reference to the accompanying drawing.
The embodiment of the present application provides a kind of information output method and device, can automatically determine and export and failure-description
The relevant data for being used to assist analyzing failure cause of text.
Information output method and device provided herein are introduced in detail below.
Referring to Figure 1, Fig. 1 is a kind of flow diagram of information output method provided by the embodiments of the present application.Such as Fig. 1 institute
Show, which includes following 101~105 part, in which:
101, information output apparatus obtains failure-description text.
Wherein, failure-description text is the text that phenomenon of the failure is described, i.e., failure-description text is for describing net
The failure occurred in network.For example, failure-description text can be " industry user's online is slow ", " online charging server (online
Charging system, OCS) communicating interrupt " etc..Failure-description text can be other devices and be sent to information output apparatus
's.For example, phenomenon of the failure is described in frontline engineer, failure-description text is obtained, and collection is used to assist to analyze
The data (such as Key Performance Indicator) and failure-description text of failure cause are sent to O&M department in the form of fault ticket
Information output apparatus.
102, information output apparatus passes through the semantic semantic vector for generating model and generating failure-description text.
In a kind of possible embodiment, semanteme generation model, which can be, describes the corresponding word of text according to historical failure
Vector matrix training generates, which includes that historical failure describes the corresponding term vector of each word in text.
Optionally, for details, reference can be made to the training that semanteme described in Fig. 2 generates model for the semantic training method for generating model
Method.That is, semanteme used in information output apparatus, which generates model, can be the language of model training apparatus training in Fig. 2
Justice generates model.The model training apparatus in information output apparatus and Fig. 2 in Fig. 1 can be deployed in same equipment or be deployed in not
Same equipment.When the model training apparatus in the information output apparatus and Fig. 2 in Fig. 1 is deployed in different equipment, model
Training device has trained the semantic transmittable semantic model that generates after generating model to information output apparatus, thus information output dress
Setting can be by the received semantic semantic vector for generating model and generating failure-description text.When in Fig. 1 information output apparatus and
When model training apparatus in Fig. 2 is disposed in the same equipment, information output apparatus can obtain language from model training apparatus
Justice generates model, so that information output apparatus can pass through the semantic semantic vector for generating model and generating failure-description text.
Certainly the semantic model that generates can not also be generated by the training of mode described in Fig. 2, can also be by other means
Training generates, and the embodiment of the present application is without limitation.
In a kind of possible embodiment, information output apparatus generates failure-description text by the semantic model that generates
The specific embodiment of semantic vector are as follows:
Failure-description text conversion is term vector matrix according to term vector set by information output apparatus, then by term vector square
Battle array input is semantic to generate model, to generate the semantic vector of failure-description text, wherein includes multiple words in the term vector set
Vector.Optionally, which can be the model training apparatus in lower Fig. 2 and generates and sends to information output apparatus
's.
Optionally, failure-description text conversion is the specific of term vector matrix according to term vector set by information output apparatus
Embodiment are as follows: information output apparatus carries out word segmentation processing to failure-description text, and it is corresponding by extremely to obtain failure-description text
The word sequence of few word composition；The corresponding term vector of word that the word sequence includes is obtained from term vector set；By the word order
The term vector matrix of the corresponding term vector composition failure-description text of each word that column include.When there is no should in term vector set
When the corresponding term vector of the word that word sequence includes, the corresponding term vector of word that random vector includes as the word sequence is generated.
For example, failure-description text includes 4 words, carries out the word sequence that word segmentation processing obtains to failure-description text
For " industry ", " user ", " online ", " slow ".Information output apparatus finds " industry " corresponding term vector from term vector set
1, " user " corresponding term vector 2, " online " corresponding term vector 3, do not find " slow " corresponding term vector, then generate random vector
Term vector 4 is as " slow " corresponding term vector.Term vector 1~4 is formed the term vector of failure-description text by information output apparatus
Matrix.The term vector Input matrix semanteme is generated to the semantic vector that failure-description text is generated in model again.
103, information output apparatus obtains the corresponding semantic vector of related text of a plurality of types of target datas.
Wherein, the target data is for assisting analysis failure Producing reason.Wherein, the execution of 103 parts and 102 parts
Sequence can first carry out 102 parts and execute 103 parts again, or can first carry out 103 parts and execute 102 parts again in no particular order.
In a kind of possible embodiment, which includes Key Performance Indicator (KPI), equipment
At least two in alarm, device log；When target data is Key Performance Indicator, the related text of the target data is to close
The title of key performance indicator；When target data is equipment alarm, the related text of the target data is the mark of equipment alarm；
When target data is device log, the related text of the target data is the contents fragment of device log.Wherein, each type
Target data be it is multiple.
For example, a plurality of types of target datas include Key Performance Indicator and equipment alarm.Above-mentioned is a plurality of types of
Target data is 100 different Key Performance Indicators and 20 different equipment alarms, and 100 Key Performance Indicators are respectively
1~Key Performance Indicator of Key Performance Indicator 100.20 equipment alarms are respectively equipment alarm 1~20.Information output apparatus obtains
The corresponding semantic vector of the related text of a plurality of types of target datas taken is that 1~key performance of Key Performance Indicator refers to
The corresponding semantic vector of title of mark 100 and the corresponding semantic vector of mark of equipment alarm 1~20.Also
It is to say, information output apparatus can obtain 120 semantic vectors.
In a kind of possible embodiment, information output apparatus can pass through semanteme before receiving failure-description text
Generate the corresponding semantic vector of related text that model generates a plurality of types of target datas.
Optionally, information output apparatus generates the corresponding semantic vector of related text of a plurality of types of target datas
Later, it can be reserved for the corresponding semantic vector of related text of a plurality of types of target datas.Receiving failure-description text
After this, so that it may obtain the corresponding semantic vector of related text for saving the plurality of target data, so as to failure-description
The semantic vector of text carries out correlation calculations.By implementing the embodiment, it can pre-generate and save plurality of target data
The corresponding semantic vector of related text, after receiving failure-description text, can directly with save plurality of target number
According to the corresponding semantic vector of related text and failure-description text semantic vector carry out correlation calculations, to not have to
After receiving failure-description text, the interim corresponding semantic vector of related text for generating plurality of target data.As it can be seen that
By implementing the embodiment, be conducive to rapidly to be calculated the semantic vector of failure-description text and every kind of target data
The correlation of the semantic vector of related text.
In a kind of possible embodiment, information output apparatus passes through the semantic correlation for generating model and generating target data
The principle and information output apparatus of the corresponding semantic vector of text pass through the semantic semanteme for generating model and generating failure-description text
The principle of vector is identical, and this will not be repeated here.
104, the semantic vector of information output apparatus calculating failure-description text and the related text of every kind of target data
The correlation of semantic vector.
For example, there is two kinds of target data, respectively 100 different Key Performance Indicators and 20 are not
Same equipment alarm, 100 Key Performance Indicators are respectively 1~Key Performance Indicator of Key Performance Indicator 100.20 equipment are accused
Alert is respectively equipment alarm 1~20.The semantic vector that information output apparatus calculates failure-description text is key with 100 respectively
The correlation of the semantic vector of the related text of energy index, and the semantic vector of calculating failure-description text are set with 20 respectively
The correlation of the semantic vector of the related text of standby alarm.Therefore, 120 correlations can be obtained.
In a kind of possible embodiment, the angle of vector can be used as the measurement of correlation, failure-description text
The correlation of this semantic vector and the semantic vector of the related text of target data may be expressed as:
Wherein, cos (θ) is the phase of the semantic vector and the semantic vector of the related text of target data of failure-description text
Guan Xing, n are the number of dimensions of the semantic vector of the related text of failure-description text and target data, xiFor failure-description text
The semantic vector of this i-th dimension, yiThe semantic vector of the related text i-th dimension of target data.
105, information output apparatus is determining and exports the first data.
Wherein, the semantic vector and every kind of mesh in every kind of target data that failure-description text is calculated in information output apparatus
After the correlation for marking the semantic vector of the related text of data, information output apparatus is determining and exports the first data, this first
Data are the target data of the semantic vector correlation maximum of semantic vector and failure-description text in this every kind target data, or
First data are that the semantic vector correlation of semantic vector and failure-description text is greater than default threshold in this every kind target data
The target data of value.
For example, the two kinds of target data obtained, respectively 100 different Key Performance Indicators and 20 differences
Equipment alarm, 100 Key Performance Indicators are respectively 1~Key Performance Indicator of Key Performance Indicator 100.20 equipment alarms
Respectively 1~equipment alarm of equipment alarm 20.The semantic vector of failure-description text refers to 1~key performance of Key Performance Indicator
The correlation of the semantic vector of the related text of mark 100 is respectively correlation 1~100.Correlation 1 is maximum correlation,
Then information output apparatus exports Key Performance Indicator 1.The semantic vector of failure-description text and 1~equipment alarm of equipment alarm 20
The correlation of semantic vector of related text be respectively correlation 101~120.Correlation 120 is maximum correlation, then believes
Cease output device output equipment alarm 20.
For another example, correlation 1 and correlation 2 are the correlation greater than preset threshold, then information output apparatus output is key
It can index 1 and Key Performance Indicator 2.Correlation 101 and correlation 102 are the correlation greater than preset threshold, then information exports
Device output equipment alarm 1 and equipment alarm 2.
The semantic vector target data bigger with the semantic vector correlation of failure-description text, illustrate the target data with
Failure-description text is more related, and user may need to check the target data with analyzing failure cause.For example, failure-description text
For " ocs communicating interrupt ", entitled " the ocs communicating interrupt number " of critical index, the semantic vector of the failure-description text
Very big with the semantic vector correlation of the title of critical index, user may need to check the critical index to analyze failure
Reason.As it can be seen that can be found automatically relevant to failure-description text for assisting by implementing method described in Fig. 1
Help the data of analyzing failure cause.
In the prior art by searching for the text with failure-description text keyword having the same, and according to the text
Associated parameter data progress failure checks analysis.But the high related text that can be used to assist analyzing failure cause of correlation and
It may be there is no identical keyword in failure-description text.Therefore, it cannot accurately be found and event by existing mode
Barrier description text is associated for assisting the data of analyzing failure cause.The embodiment of the present application passes through comparison failure-description text
Semantic vector and the semantic vector of the related text of target data correlation, can accurately find and failure-description text
This associated target data.For example, failure-description is " industry user online slow ", the embodiment of the present application analyze with its phase
Entitled " downlink bandwidth controls packet loss ratio " of the critical index for accident analysis closed.As can be seen that from literal
The two do not have it is any can match with associated ingredient, and the application exactly by semantic analysis excavate study arrived " online
Speed and packet loss ratio have relationship " as domain knowledge, both just realize associated analysis.
Therefore, it by implementing method described in Fig. 1, can automatically and accurately find and failure-description text
It is relevant for assisting the data of analyzing failure cause.
Fig. 2 is referred to, Fig. 2 is a kind of process signal of semantic training method for generating model provided by the embodiments of the present application
Figure.As shown in Fig. 2, the training method that the semanteme generates model includes following 201~203 part, in which:
201, model training apparatus obtains the corresponding term vector set of training text.
Wherein, the word in the term vector and training text for including in term vector set corresponds.For example, in training text
It then also include 10000 term vectors in term vector set including 10000 words.The term vector is used to indicate the semanteme of word.It can
Choosing, after obtaining the corresponding term vector set of training text, it can be reserved for the corresponding term vector set of training text, so as to subsequent
Use the term vector in term vector set.
Training text, that is, corpus.In a kind of possible embodiment, training text can be encyclopaedia class text.From encyclopaedia
The term vector that the acquistion of class text middle school is arrived has good general semantics.
In a kind of possible embodiment, model training apparatus first pre-processes training text, by sentence cutting
Word segmentation processing is carried out to every text again afterwards, the training text after being segmented, and pass through word2vec tool or other tools
The corresponding term vector set of training text after obtaining participle.
For example, training text is that " mathematics is one using concepts such as the variation of symbolic language research quantitative structure and spaces
Door subject.I likes mathematics ".Model training apparatus is split as two words to by training text, and respectively " mathematics is to utilize symbol
One Men Xueke of the concepts such as the variation of speech research quantitative structure and space " and " I likes mathematics ".Again to the two sentences point
It carry out not word segmentation processing.Training text after being segmented be " mathematics be changed using symbolic language research quantitative structure and
One Men Xueke of the concepts such as space.I likes mathematics ".Model training apparatus is using word2vec tool to the training text after participle
This progress traverses sentence by sentence, and traversal terminates just to have obtained the corresponding term vector of each word in training text.Model training apparatus will
The term vector set of the corresponding term vector composition of each word in training text is saved.
It is corresponding that model training apparatus can obtain the training text after segmenting by word2vec tool and using CBOW algorithm
Term vector set.The thought of CBOW algorithm is to predict current word by given cliction up and down.The mesh of CBOW algorithm training
When mark is the context for giving some word, so that the maximum probability that the word occurs.After training, each word is obtained in output layer
To a corresponding term vector.Although the idea about modeling of CBOW algorithm is an assorting process, can generate term vector this
Byproduct.
For example, Fig. 3 is the schematic diagram for the neural network that CBOW algorithm uses.As shown in figure 3, the neural network is by three-layered node
Structure is constituted, respectively input layer, mapping layer and output layer.Wherein, output layer includes constructed good Huffman tree.Huffman
A leaf node for tree represents the term vector of a word in training text, the term vector of the corresponding word of each leaf node
It is random initializtion.A weight vectors built in each nonleaf node, the dimension of the vector and the term vector of input layer are identical.
Wherein, input layer is the term vector of n-1 word around some word w (t).N is window size.For example, such as
Fruit n takes 5, the first two and latter two word that n-1 word around word w (t) is word w (t).Before word w (t)
Two and latter two word are respectively w (t-2), w (t-1), w (t+1), w (t+2).It is corresponding, the word of this n-1 word
Vector is denoted as v (w (t-2)), v (w (t-1)), v (w (t+1)), v (w (t+2)).Input layer will arrive this n-1 term vector and be transmitted to
N-1 term vector is added by mapping layer, mapping layer, i.e., by the corresponding addition of each dimension of n-1 term vector.For example, mapping
Layer input is pro (t)=v (w (t-2))+v (w (t-1))+v (w (t+1))+v (w (t+2)).
The vector pro (t) that adduction obtains is input to the root node of Huffman tree by projection layer.Pro (t) is being inputted into root section
After point, can calculate root node to each leaf node probability, the training process of model be desirable to obtain by root node to
Up to the maximum probability of the corresponding leaf node of w (t), since in the training text of magnanimity, identical context environmental can be multiple
Occur, so can constantly correct each weight vectors during traversing training training text, reaches such effect.To training
After all words traversal in text is completed, the corresponding term vector of each leaf node of Huffman tree is just each of training text
The corresponding term vector of word.Here " all words in training text " include dittograph in training text.
Wherein, when passing through an intermediate node every time from root node to the corresponding leaf node of word w (t), be the equal of into
It has gone one time two to classify, classifier can return classifier using softmax.Wherein, the class probability of every subseries are as follows:
Wherein, θiIndicate i-th of weight vectors.Pro (t) is the sum of the term vector of context of w (t), and e is that nature is normal
Number.
If containing L intermediate node by the path that root node traverses the corresponding leaf node of word w (t), these nodes
On parameter composition parameter vector be [θ1, θ2, θ3..., θL], then the probability of root node to the corresponding leaf node of word w (t) is
The product of the probability of two classification, i.e. probability of the root node to the corresponding leaf node of word w (t) every time are as follows:
Wherein, P (w (t) | context (w (t))) is probability of the root node to the corresponding leaf node of word w (t),Indicate that i is incremented by P (context (w (t)), θ one by one from 1 to Li) even multiply quadrature.From root node
To other leaf nodes probability calculation method similarly, this will not be repeated here.
202, model training apparatus according to term vector set by historical failure describe text conversion be from least one word to
Measure the term vector matrix of composition.
Specifically, it is term vector matrix that a large amount of historical failure can be described text conversion by model training apparatus.Model instruction
Practice device and semantic generation model is obtained according to the training of a large amount of term vector matrix.For example, having historical failure to describe text 1~go through
Historical failure, can be described 1~historical failure of text and describe text 100 to be respectively converted into term vector by history failure-description text 100
Matrix is to get to 100 term vector matrixes.Model training apparatus obtains semantic generation according to this 100 term vector matrix training
Model.
In a kind of possible embodiment, historical failure is described text according to term vector set and turned by model training apparatus
The specific embodiment for being changed to the term vector matrix being made of at least one term vector can be with are as follows: model training apparatus to history therefore
Barrier description text carries out word segmentation processing, obtains historical failure and describes the corresponding word sequence being made of at least one word of text；From
The corresponding term vector of word that the word sequence includes is obtained in term vector set；The corresponding word of each word for including by the word sequence to
Amount forms the term vector matrix that the historical failure describes text.When being not present in term vector set, the word that the word sequence includes is corresponding
Term vector when, produce the random vector corresponding term vector of word that includes as the word sequence.As it can be seen that by implementing the implementation
Mode, historical failure accurately can be described text conversion is the term vector matrix being made of at least one term vector.
For example, it includes 4 words that historical failure, which describes text 1, describes the progress of text 1 word segmentation processing to historical failure and obtains
To word order be classified as " industry ", " user ", " online ", " slow ".Model training apparatus finds " industry " from term vector set
Corresponding term vector 1, " user " corresponding term vector 2, " online " corresponding term vector 3, do not find " slow " corresponding term vector, then give birth to
At random vector term vector 4 as " slow " corresponding term vector.Term vector 1~4 is formed the historical failure by model training apparatus
The term vector matrix 1 of text 1 is described.Other historical failures describe the principle and history that text 2~100 is converted to term vector matrix
The principle that failure-description text 1 is converted to term vector matrix is identical, and this will not be repeated here.
203, model training apparatus obtains semantic generation model according to the training of term vector matrix.
Specifically, model training apparatus is after obtaining term vector matrix, can by term vector Input matrix neural network into
Row training, to obtain semantic generation model.The semanteme generates the semantic vector that model is used to generate text.The semantic vector is used for
Indicate the semanteme of text.
As it can be seen that method described in Fig. 2 is to obtain language from the semantic semantic gradually modeling to sentence surface of lexical level
Justice generates model, and this semantic model training mode that generates is to meet the basic principle of language generation.Therefore, by implementing to scheme
The semantic generation model that the training of method described in 2 obtains, can more accurately express the semanteme of text.
In a kind of possible embodiment, model training apparatus obtains semantic generation model according to the training of term vector matrix
Specific embodiment are as follows: model training apparatus obtain historical failure the corresponding faulty equipment type of text is described；Model training
For device according to term vector matrix and class label train classification models, such distinguishing label includes the faulty equipment type；Model instruction
Practice device and semantic generation model is obtained according to disaggregated model.The semantic generation model obtained by implementing embodiment training,
The semanteme of text can more accurately be expressed.
For example, historical failure, which describes the corresponding faulty equipment type of text, can be router, wireline equipment or wirelessly set
It is standby etc..For example, historical failure describes the failure that the failure of text description generates for router, then it is corresponding to describe text for historical failure
Faulty equipment type be router.Frontline engineer collects the corresponding faulty equipment type of each failure-description text, so
By failure-description text, the corresponding faulty equipment type of failure-description text and it is used to assist the data of analyzing failure cause to add afterwards
It adds in work order, and work order is sent to O&M terminal and carries out failure reason analysis.Therefore, model training apparatus can be from work order
It obtains historical failure and describes the corresponding faulty equipment type of text.
Wherein, the disaggregated model that training obtains is the mould for generating the corresponding faulty equipment type of failure-description text
Type.For example, by the corresponding term vector Input matrix disaggregated model of failure-description text 1, the exportable failure-description of the disaggregated model
The corresponding faulty equipment type of text 1.
In a kind of possible embodiment, model training apparatus is according to term vector matrix and class label training classification mould
The specific embodiment of type are as follows: term vector matrix and class label input neural network are iterated training, in each iteration
The parameter of term vector and neural network in the term vector matrix of input neural network is adjusted when training, to generate classification
Model.By implementing the embodiment, the disaggregated model that training can be enable to obtain accurately carries out failure-description text
Classification.
Optionally, also the term vector in term vector matrix adjusted can be used to update term vector set for model training apparatus
The corresponding term vector of middle corresponding words.By implementing the optional mode, can be described according to the historical failure with domain knowledge
Corpus of text corrects the term vector in term vector set, makes the term vector in term vector set more and can express the word in failure field
Semantic information.
For example, Fig. 4 is a kind of structural schematic diagram of neural network for train classification models.As shown in figure 4, should
Neural network includes convolutional layer, pond layer and full articulamentum.The term vector matrix 1 that historical failure describes text 1 includes term vector
{ w1, w2, w3, w4, w5, w6 }.The dimension of each term vector is 128 dimensions.Model training apparatus obtains term vector matrix 1
Later, term vector matrix 1 is inputted into neural network.As shown in figure 4, there are two convolution kernels for the tool in neural network.Certainly in reality
Can also be there are two above convolution kernel in the application of border, the embodiment of the present application is illustrated with two convolution kernels.The left side
Convolution kernel 1 includes that term vector carries out convolution two-by-two to term vector matrix 1.For example, w1 and w2, which carries out convolution, obtains C1, w2 and w3
Carry out convolution obtain C2, w3 and w4 carry out convolution obtain C3, w4 and w5 progress convolution obtain C4, w5 and w6 progress convolution obtain
C5.The convolution kernel 2 on the right includes that term vector carries out three or three convolution to term vector matrix 1.It is obtained for example, w1, w2 and w3 carry out convolution
Convolution is carried out to C6, w2, w3 and w4 and obtains C7, and w3, w4 and w5 carry out convolution and obtain C8, and w4, w5 and w6 carry out convolution and obtain
C9.Convolution also carried out to the term vector of other quantity in practical application, the embodiment of the present application with three or three convolution of convolution sum two-by-two into
Row illustrates.
As it can be seen that convolution kernel 1 produces characteristic pattern (feature map) C=[C1, C2 ..., C5], convolution kernel 2 is generated
One characteristic pattern C=[C6, C7, C8, C9].After model training apparatus obtains the characteristic pattern of each convolution karyogenesis, for every
A characteristic pattern, by the operation of maximum pondization choose the maximum value in each dimension as current convolution karyogenesis text feature to
Amount.Model training apparatus splices all Text eigenvectors, obtain final historical failure describe text 1 it is semantic to
Amount.I.e. as shown in figure 4, model training apparatus chooses maximum value from first dimension of C1~C5, from the 2nd of C1~C5 the
Maximum value is chosen in a dimension, chooses maximum value from the 3rd dimension of C1~C5, and so on, until from C1~C5
The 128th dimension in choose arrive maximum value.The maximum value of 128 dimensions of selection is formed convolution kernel by model training apparatus
1 corresponding Text eigenvector 1.Similarly, model training apparatus also obtains the corresponding Text eigenvector 2 of convolution kernel 2.Model instruction
Practice device to splice Text eigenvector 1 and Text eigenvector 2, obtains the language that final historical failure describes text 1
Adopted vector.
The semantic vector that obtained historical failure describes text 1 is inputted full articulamentum by model training apparatus, and by history
The corresponding faulty equipment type (such as router) of failure-description text 1 is used as class label, inputs full articulamentum.Model training dress
It sets and is analyzed in the semantic vector that full articulamentum describes text 1 to historical failure, it is most general that analysis obtains faulty equipment type
Rate is interchanger.Due to the faulty equipment class for the maximum probability that the semantic vector that historical failure describes text 1 is analyzed
It is not identical that type (i.e. interchanger) and historical failure describe the corresponding class label (i.e. router) of text 1, therefore model training fills
Set the faulty equipment type for the maximum probability that record is analyzed by the semantic vector for describing text 1 to historical failure not
Correctly.Similarly, historical failure is described the term vector Input matrix neural network of text 2 by model training apparatus according to the above process
It is trained, obtains the semantic vector that historical failure describes text 2, and it is right in full articulamentum input historical failure to describe text 2
The faulty equipment type (such as interchanger) answered is used as class label.Model training apparatus describes the semanteme of text 2 to historical failure
Vector is analyzed, and it is firewall that analysis, which obtains faulty equipment type maximum probability,.Therefore, model training apparatus record by pair
The faulty equipment type for the maximum probability that the semantic vector that historical failure describes text 2 is analyzed is incorrect.Assuming that tool
There are 100 historical failures to describe text, remaining 98 historical failure describes text similarly, describes text according to above-mentioned historical failure
This 1 mode inputs the training that neural network carries out disaggregated model.The first round is completed describing text 1~100 to historical failure
After training, it is assumed that the failure for the maximum probability that the corresponding semantic vector of text 1~50 is analyzed is described according to historical failure
Device type is incorrect, and model training apparatus describes the corresponding word of text 1~50 to the parameter and historical failure of neural network
Term vector in vector matrix is adjusted.It is again right with the parameter of neologisms vector matrix and neural network after adjustment finishes
Historical failure describes text 1~100 and is trained, until describing the corresponding semantic vector of text 1~100 according to historical failure
The faulty equipment type for analyzing obtained maximum probability matches with tag along sort, just generates disaggregated model, i.e., is instructed by iteration
Practice neural network next life ingredient class model.
Last model training apparatus uses the term vector more neologisms in the term vector matrix of last wheel repetitive exercise input
The corresponding term vector of corresponding words in vector set.For example, it is that " networking speed is slow " last wheel changes that historical failure, which describes text 1,
It describes 1 equivalent vector matrix of text to historical failure before generation training to be adjusted, by " online " corresponding term vector tune
Whole is term vector 1, then corresponding using " online " in 1 substitute vector set of term vector after the completion of last time repetitive exercise
Term vector.Historical failure describes to describe text to historical failure before text 2 is " OCS communicating interrupt " last wheel repetitive exercise
This 2 equivalent vector matrix is adjusted, and " interruption " corresponding term vector is adjusted to term vector 2, then last wheel iteration
After the completion of training, " interruption " corresponding term vector in 2 substitute vector set of term vector is used.Other historical failures description text
Originally similarly, this will not be repeated here.
In a kind of possible embodiment, model training apparatus obtains the specific of semantic generation model according to disaggregated model
Embodiment are as follows: part more than articulamentum complete in disaggregated model is generated model by model training apparatus.Pass through reality
It applies the semantic of embodiment generation and generates model, can accurately generate the semantic vector of text.
The embodiment of the present invention can carry out the division of functional module according to above method example to equipment, for example, can be right
The each functional module of each function division is answered, two or more functions can also be integrated in a module.It is above-mentioned
Integrated module both can take the form of hardware realization, can also be realized in the form of software function module.It needs to illustrate
, be to the division of module in the embodiment of the present invention it is schematical, only a kind of logical function partition in actual implementation may be used
To there is other division mode.
Fig. 5 is referred to, Fig. 5 is that the present invention implements a kind of information output apparatus provided.The information output apparatus includes: to obtain
Modulus block 501, generation module 502, computing module 503 and output module 504.Wherein:
Module 501 is obtained, for obtaining failure-description text；Generation module 502, for being generated by the semantic model that generates
The semantic vector of failure-description text, the failure-description text are used for the failure for describing to occur in network；Module 501 is obtained, also
The corresponding semantic vector of related text for obtaining a plurality of types of target datas, the target data is for assisting analysis
Failure Producing reason；Computing module 503 is related to every kind of target data for calculating the semantic vector of failure-description text
The correlation of the semantic vector of text；Output module 504, for determining and exporting the first data, the first data are every kind of target
The target data or the first data of the correlation maximum of the semantic vector of semantic vector and failure-description text are every kind in data
Semantic vector is greater than the target data of preset threshold with the correlation of the semantic vector of failure-description text in target data.
In a kind of possible embodiment, generation module 502 is also used to obtaining the acquisition failure-description text of module 501
Before this, pass through the semantic corresponding semantic vector of related text for generating model and generating plurality of target data.
In a kind of possible embodiment, it is that the corresponding term vector of text is described according to historical failure that semanteme, which generates model,
Matrix training generates, and term vector matrix includes that historical failure describes the corresponding term vector of each word in text, which uses
In the semanteme for indicating word.
In a kind of possible embodiment, a plurality of types of target datas include Key Performance Indicator, equipment alarm,
At least two in device log；When target data is Key Performance Indicator, the related text of target data is key performance
The title of index；When target data is equipment alarm, the related text of target data is the mark of equipment alarm；Work as number of targets
When according to for device log, the related text of target data is the contents fragment of device log.
Fig. 6 is referred to, Fig. 6 is that the present invention implements a kind of model training apparatus provided.The model training apparatus includes obtaining
Modulus block 601, conversion module 602 and training module 603, in which:
Module 601 is obtained, for obtaining the corresponding term vector set of training text, the term vector for including in term vector set
It is corresponded with the word in training text；Conversion module 602, for historical failure to be described text conversion according to term vector set
For the term vector matrix being made of at least one term vector；Training module 603 is also used to obtain language according to the training of term vector matrix
Justice generates model, and semanteme generates the semantic vector that model is used to generate text.
In a kind of possible embodiment, conversion module 602 is specifically used for: describing text to historical failure and segments
Processing, obtains historical failure and describes the corresponding word sequence being made of at least one word of text；Word is obtained from term vector set
The corresponding term vector of the word that sequence includes；The corresponding term vector of each word for including by word sequence forms term vector matrix.
In a kind of possible embodiment, conversion module 602 also particularly useful for: when in term vector set be not present word order
When the corresponding term vector of the word that column include, the corresponding term vector of word that random vector includes as word sequence is generated.
In a kind of possible embodiment, training module 603 obtains semantic generation model according to the training of term vector matrix
Mode specifically: obtain historical failure the corresponding faulty equipment type of text is described；According to term vector matrix and class label
Train classification models, such distinguishing label include the faulty equipment type；Semantic generation model is obtained according to disaggregated model.
In a kind of possible embodiment, training module 603 is according to term vector matrix and class label training classification mould
The mode of type specifically: term vector matrix and class label input neural network are iterated training, in each repetitive exercise
When in the term vector matrix of input neural network term vector and the parameter of neural network be adjusted, to generate classification mould
Type.
Fig. 7 is referred to, Fig. 7 is a kind of structural schematic diagram of information output apparatus disclosed in the embodiment of the present application.Such as Fig. 7 institute
Show, which includes processor 701, memory 702 and communication interface 703.Wherein, processor 701, storage
Device 702 is connected with communication interface 703.
Wherein, processor 701 can be central processing unit (central processing unit, CPU), general procedure
Device, coprocessor, digital signal processor (digital signal processor, DSP), specific integrated circuit
(application-specific integrated circuit, ASIC), field programmable gate array (field
Programmable gate array, FPGA) or other programmable logic device, transistor logic, hardware component or
Person's any combination thereof.The processor 701 is also possible to realize the combination of computing function, such as includes one or more microprocessors
Combination, DSP and the combination of microprocessor etc..
Wherein, communication interface 703 is for realizing the communication between other network elements.
Wherein, processor 701 calls the program code stored in memory 702, believes in executable above method embodiment
Cease step performed by output device.
Fig. 8 is referred to, Fig. 8 is a kind of structural schematic diagram of model training apparatus disclosed in the embodiment of the present application.Such as Fig. 8 institute
Show, which includes processor 801, memory 802 and communication interface 803.Wherein, processor 801, storage
Device 802 is connected with communication interface 803.
Wherein, processor 801 can be central processing unit (central processing unit, CPU), general procedure
Device, coprocessor, digital signal processor (digital signal processor, DSP), specific integrated circuit
(application-specific integrated circuit, ASIC), field programmable gate array (field
Programmable gate array, FPGA) or other programmable logic device, transistor logic, hardware component or
Person's any combination thereof.The processor 801 is also possible to realize the combination of computing function, such as includes one or more microprocessors
Combination, DSP and the combination of microprocessor etc..
Wherein, communication interface 803 is for realizing the communication between other network elements.
Wherein, processor 801 calls the program code stored in memory 802, and mould in above method embodiment can be performed
Step performed by type training device.
Based on the same inventive concept, the principle and the application method that each equipment provided in the embodiment of the present application solves the problems, such as
Embodiment is similar, therefore the implementation of each equipment may refer to the implementation of method, for succinct description, repeats no more herein.
In the above-described embodiments, it all emphasizes particularly on different fields to the description of each embodiment, there is no the portion being described in detail in some embodiment
Point, reference can be made to the related descriptions of other embodiments.
Finally, it should be noted that the above various embodiments is only to illustrate the technical solution of the application, rather than its limitations；To the greatest extent
Pipe is described in detail the application referring to foregoing embodiments, those skilled in the art should understand that: its according to
So be possible to modify the technical solutions described in the foregoing embodiments, or to some or all of the technical features into
Row equivalent replacement；And these are modified or replaceed, each embodiment technology of the application that it does not separate the essence of the corresponding technical solution
The range of scheme.
Claims (20)
1. a kind of information output method, which is characterized in that the described method includes:
Failure-description text is obtained, the failure-description text is used for the failure for describing to occur in network；
Pass through the semantic semantic vector for generating model and generating the failure-description text；
The corresponding semantic vector of related text of a plurality of types of target datas is obtained, the target data is divided for assisting
Analyse the failure Producing reason；
Calculate the phase of the semantic vector and the semantic vector of the related text of target data described in every kind of the failure-description text
Guan Xing；
The first data are determined and export, first data are semantic vector and the failure-description in every kind of target data
The target data of the correlation maximum of the semantic vector of text or first data be in every kind of target data semanteme to
Measure the target data for being greater than preset threshold with the correlation of the semantic vector of the failure-description text.
2. the method according to claim 1, wherein the method is also before the acquisition failure-description text
Include:
Pass through the semantic corresponding semantic vector of related text for generating model and generating a plurality of types of target datas.
3. method according to claim 1 or 2, which is characterized in that the semantic model that generates is retouched according to historical failure
State what text corresponding term vector matrix training generated, the term vector matrix include the historical failure describe it is each in text
The corresponding term vector of word, the term vector are used to indicate the semanteme of word.
4. method according to any one of claims 1 to 3, which is characterized in that a plurality of types of target packets
Include Key Performance Indicator, equipment alarm, at least two in device log；When the target data is the Key Performance Indicator
When, the related text of the target data is the title of the Key Performance Indicator；When the target data is equipment announcement
When alert, the related text of the target data is the mark of the equipment alarm；When the target data is the device log
When, the related text of the target data is the contents fragment of the device log.
5. a kind of semantic training method for generating model, which is characterized in that the described method includes:
The corresponding term vector set of training text is obtained, in the term vector and the training text for including in the term vector set
Word correspond, the term vector is used to indicate the semanteme of word；
It is the term vector square being made of at least one term vector that historical failure, which is described text conversion, according to the term vector set
Battle array；
Semantic generation model, the semantic semanteme for generating model and being used to generate text are obtained according to term vector matrix training
Vector.
6. according to the method described in claim 5, it is characterized in that, described describe historical failure according to the term vector set
Text conversion is the term vector matrix being made of at least one term vector, comprising:
Text is described to historical failure and carries out word segmentation processing, obtaining the historical failure, to describe text corresponding by least one word
The word sequence of composition；
The corresponding term vector of word that the word sequence includes is obtained from the term vector set；
The corresponding term vector of each word for including by the word sequence forms term vector matrix.
7. according to the method described in claim 6, it is characterized in that, the method also includes:
When the corresponding term vector of the word sequence word that includes is not present in the term vector set, random vector conduct is generated
The corresponding term vector of the word that the word sequence includes.
8. according to method described in claim 5~7 any one, which is characterized in that described to be instructed according to the term vector matrix
Get semantic generation model, comprising:
It obtains the historical failure and describes the corresponding faulty equipment type of text；
According to the term vector matrix and class label train classification models, the class label includes the faulty equipment class
Type；
Semantic generation model is obtained according to the disaggregated model.
9. according to the method described in claim 8, it is characterized in that, described according to the term vector matrix and the class label
Train classification models, comprising:
The term vector matrix and class label input neural network are iterated training, in each repetitive exercise pair
The parameter of the term vector and the neural network that input in the term vector matrix of the neural network is adjusted, described in generating
Disaggregated model.
10. a kind of information output apparatus, which is characterized in that the information output apparatus includes:
Module is obtained, for obtaining failure-description text, the failure-description text is used for the failure for describing to occur in network；
Generation module, for passing through the semantic semantic vector for generating model and generating the failure-description text；
The acquisition module is also used to obtain the corresponding semantic vector of related text of a plurality of types of target datas, institute
State target data for assist the analysis failure Producing reason；
Computing module, for calculating the semantic vector of the failure-description text and the related text of every kind of target data
The correlation of semantic vector；
Output module, for determining and exporting the first data, first data are semantic vector in every kind of target data
It is every kind of mesh with the target data of the correlation maximum of the semantic vector of the failure-description text or first data
Mark the target data that semantic vector in data is greater than preset threshold with the correlation of the semantic vector of the failure-description text.
11. device according to claim 10, which is characterized in that
The generation module is also used to before the acquisition module obtains failure-description text, passes through the semantic generation mould
The corresponding semantic vector of related text of type generation plurality of target data.
12. device described in 0 or 11 according to claim 1, which is characterized in that the semantic model that generates is according to historical failure
Describe what the corresponding term vector matrix training of text generated, the term vector matrix includes that the historical failure describes in text respectively
The corresponding term vector of a word, the term vector are used to indicate the semanteme of word.
13. device described in 0~12 any one according to claim 1, which is characterized in that a plurality of types of target datas
Including at least two in Key Performance Indicator, equipment alarm, device log；When the target data is that the key performance refers to
When mark, the related text of the target data is the title of the Key Performance Indicator；When the target data is the equipment
When alarm, the related text of the target data is the mark of the equipment alarm；When the target data is the equipment day
When will, the related text of the target data is the contents fragment of the device log.
14. a kind of model training apparatus, which is characterized in that the model training apparatus includes:
Obtain module, for obtaining the corresponding term vector set of training text, the term vector for including in the term vector set with
Word in the training text corresponds；
Conversion module is by least one term vector group for historical failure to be described text conversion according to the term vector set
At term vector matrix；
The training module is also used to obtain semantic generation model, the semantic generation mould according to term vector matrix training
Type is used to generate the semantic vector of text.
15. device according to claim 14, which is characterized in that the conversion module is specifically used for:
Text is described to historical failure and carries out word segmentation processing, obtaining the historical failure, to describe text corresponding by least one word
The word sequence of composition；
The corresponding term vector of word that the word sequence includes is obtained from the term vector set；
The corresponding term vector of each word for including by the word sequence forms term vector matrix.
16. device according to claim 15, which is characterized in that the conversion module also particularly useful for:
When the corresponding term vector of the word sequence word that includes is not present in the term vector set, random vector conduct is generated
The corresponding term vector of the word that the word sequence includes.
17. device described in 4~16 any one according to claim 1, which is characterized in that the training module is according to institute's predicate
Vector matrix training obtains the semantic mode for generating model specifically:
It obtains the historical failure and describes the corresponding faulty equipment type of text；
According to the term vector matrix and the class label train classification models, the class label includes the faulty equipment
Type；
Semantic generation model is obtained according to the disaggregated model.
18. device according to claim 17, which is characterized in that the training module is according to the term vector matrix and institute
State the mode of class label train classification models specifically:
The term vector matrix and class label input neural network are iterated training, in each repetitive exercise pair
The parameter of the term vector and the neural network that input in the term vector matrix of the neural network is adjusted, described in generating
Disaggregated model.
19. a kind of computer program product, which is characterized in that when run on a computer, so that computer executes above-mentioned 1
Method described in any one of~9.
20. a kind of computer readable storage medium, which is characterized in that instruction is stored in computer readable storage medium, when it
When running on computers so that computer execute above-mentioned any one of 1~9 described in method.
Priority Applications (2)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810415523.0A CN109902283B (en) | 2018-05-03 | 2018-05-03 | Information output method and device |
| PCT/CN2019/084814 WO2019210820A1 (en) | 2018-05-03 | 2019-04-28 | Information output method and apparatus |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810415523.0A CN109902283B (en) | 2018-05-03 | 2018-05-03 | Information output method and device |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN109902283A true CN109902283A (en) | 2019-06-18 |
| CN109902283B CN109902283B (en) | 2023-06-06 |
Family
ID=66943185
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201810415523.0A Active CN109902283B (en) | 2018-05-03 | 2018-05-03 | Information output method and device |
Country Status (2)
| Country | Link |
|---|---|
| CN (1) | CN109902283B (en) |
| WO (1) | WO2019210820A1 (en) |
Cited By (14)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110245233A (en) * | 2019-06-19 | 2019-09-17 | 北京航空航天大学 | A kind of fault determination method and device |
| CN110378486A (en) * | 2019-07-15 | 2019-10-25 | 出门问问信息科技有限公司 | Internet startup disk method, apparatus, electronic equipment and storage medium |
| CN110569330A (en) * | 2019-07-18 | 2019-12-13 | 华瑞新智科技（北京）有限公司 | text labeling system, device, equipment and medium based on intelligent word selection |
| CN111046674A (en) * | 2019-12-20 | 2020-04-21 | 科大讯飞股份有限公司 | Semantic understanding method and device, electronic equipment and storage medium |
| CN111429155A (en) * | 2020-03-25 | 2020-07-17 | 中国银行股份有限公司 | Bank card dispute processing method and device |
| CN111651601A (en) * | 2020-06-02 | 2020-09-11 | 全球能源互联网研究院有限公司 | Training method and classification method for fault classification model of power information system |
| CN111858725A (en) * | 2020-04-30 | 2020-10-30 | 北京嘀嘀无限科技发展有限公司 | Event attribute determination method and system |
| CN112183994A (en) * | 2020-09-23 | 2021-01-05 | 南方电网数字电网研究院有限公司 | Method and device for evaluating equipment state, computer equipment and storage medium |
| CN112383421A (en) * | 2020-11-03 | 2021-02-19 | 中国联合网络通信集团有限公司 | Fault positioning method and device |
| CN112667805A (en) * | 2019-09-30 | 2021-04-16 | 北京沃东天骏信息技术有限公司 | Work order category determination method, device, equipment and medium |
| CN112925668A (en) * | 2021-02-25 | 2021-06-08 | 北京百度网讯科技有限公司 | Server health evaluation method, device, equipment and storage medium |
| CN113657022A (en) * | 2021-07-15 | 2021-11-16 | 华为技术有限公司 | Chip fault identification method and related equipment |
| CN113722494A (en) * | 2021-09-10 | 2021-11-30 | 中国航空工业集团公司西安飞行自动控制研究所 | Equipment fault positioning method based on natural language understanding |
| CN113657022B (en) * | 2021-07-15 | 2024-05-14 | 华为技术有限公司 | Chip fault recognition method and related equipment |
Families Citing this family (27)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110909550B (en) * | 2019-11-13 | 2023-11-03 | 北京环境特性研究所 | Text processing method, text processing device, electronic equipment and readable storage medium |
| CN111078822A (en) * | 2019-11-29 | 2020-04-28 | 北京百卓网络技术有限公司 | Reader information extraction method and system based on Chinese novel text |
| CN112988921A (en) * | 2019-12-13 | 2021-06-18 | 北京四维图新科技股份有限公司 | Method and device for identifying map information change |
| CN111124925B (en) * | 2019-12-25 | 2024-04-05 | 斑马网络技术有限公司 | Scene extraction method, device, equipment and storage medium based on big data |
| CN111291564B (en) * | 2020-03-03 | 2023-10-31 | 腾讯科技（深圳）有限公司 | Model training method, device and storage medium for word vector acquisition |
| CN113495949A (en) * | 2020-03-18 | 2021-10-12 | 北京沃东天骏信息技术有限公司 | Method, system, computer system and medium for text recognition |
| CN111274366A (en) * | 2020-03-25 | 2020-06-12 | 联想(北京)有限公司 | Search recommendation method and device, equipment and storage medium |
| CN111767721A (en) * | 2020-03-26 | 2020-10-13 | 北京沃东天骏信息技术有限公司 | Information processing method, device and equipment |
| CN112749553B (en) * | 2020-06-05 | 2023-07-25 | 腾讯科技（深圳）有限公司 | Text information processing method and device for video file and server |
| CN113779975B (en) * | 2020-06-10 | 2024-03-01 | 北京猎户星空科技有限公司 | Semantic recognition method, device, equipment and medium |
| CN113822016B (en) * | 2020-06-19 | 2024-03-22 | 阿里巴巴集团控股有限公司 | Text data processing method and device, electronic equipment and readable storage medium |
| CN112069833B (en) * | 2020-09-01 | 2024-04-30 | 北京声智科技有限公司 | Log analysis method, log analysis device and electronic equipment |
| CN112507720A (en) * | 2020-11-12 | 2021-03-16 | 西安交通大学 | Graph convolution network root identification method based on causal semantic relation transfer |
| CN112463378B (en) * | 2020-11-27 | 2023-12-22 | 北京浪潮数据技术有限公司 | Server asset scanning method, system, electronic equipment and storage medium |
| CN112529104A (en) * | 2020-12-23 | 2021-03-19 | 东软睿驰汽车技术（沈阳）有限公司 | Vehicle fault prediction model generation method, fault prediction method and device |
| CN112711947B (en) * | 2021-01-09 | 2023-08-22 | 国网湖北省电力有限公司电力科学研究院 | Text vectorization-based fault power failure emergency repair handling reference method |
| CN112818008A (en) * | 2021-01-21 | 2021-05-18 | 中广核工程有限公司 | Intelligent diagnosis method, system, medium and electronic equipment for nuclear power debugging faults |
| CN113610112B (en) * | 2021-07-09 | 2024-04-16 | 中国商用飞机有限责任公司上海飞机设计研究院 | Auxiliary decision-making method for aircraft assembly quality defects |
| CN113591477B (en) * | 2021-08-10 | 2023-09-15 | 平安银行股份有限公司 | Fault positioning method, device, equipment and storage medium based on associated data |
| CN114036293B (en) * | 2021-11-03 | 2023-06-06 | 腾讯科技（深圳）有限公司 | Data processing method and device and electronic equipment |
| CN113961708B (en) * | 2021-11-10 | 2024-04-23 | 北京邮电大学 | Power equipment fault tracing method based on multi-level graph convolutional network |
| CN114265930A (en) * | 2021-11-19 | 2022-04-01 | 国电南京自动化股份有限公司 | Low-voltage user fault report repairing and processing method based on event extraction |
| CN114218402A (en) * | 2021-12-17 | 2022-03-22 | 迈创企业管理服务股份有限公司 | Method for recommending computer hardware fault replacement part |
| CN115687031A (en) * | 2022-11-15 | 2023-02-03 | 北京优特捷信息技术有限公司 | Method, device, equipment and medium for generating alarm description text |
| CN115994217B (en) * | 2022-11-29 | 2024-01-23 | 南京审计大学 | Financial report fraud detection method and system |
| CN116502058B (en) * | 2023-06-28 | 2023-09-26 | 长园深瑞能源技术有限公司 | AI fault detection analysis method and system applied to charging pile system and cloud platform |
| CN116738323B (en) * | 2023-08-08 | 2023-10-27 | 北京全路通信信号研究设计院集团有限公司 | Fault diagnosis method, device, equipment and medium for railway signal equipment |
Citations (18)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| JP2003173286A (en) * | 2001-12-05 | 2003-06-20 | Nippon Telegr & Teleph Corp <Ntt> | Method of acquiring sematic information on distributed network |
| CN101795210A (en) * | 2010-01-11 | 2010-08-04 | 浪潮通信信息系统有限公司 | Method for processing communication network failure |
| KR101078751B1 (en) * | 2011-02-23 | 2011-11-02 | 한국과학기술정보연구원 | Method and apparatus for detecting association errors of lexical semantic networks |
| CN102650960A (en) * | 2012-03-31 | 2012-08-29 | 奇智软件（北京）有限公司 | Method and device for eliminating faults of terminal equipment |
| US20120232905A1 (en) * | 2011-03-10 | 2012-09-13 | GM Global Technology Operations LLC | Methodology to improve failure prediction accuracy by fusing textual data with reliability model |
| US20130110496A1 (en) * | 2011-10-28 | 2013-05-02 | Sap Ag | Calculating Term Similarity Using A Meta-Model Semantic Network |
| US20130339787A1 (en) * | 2012-06-15 | 2013-12-19 | International Business Machines Coporation | Systematic failure remediation |
| CN104361026A (en) * | 2014-10-22 | 2015-02-18 | 北京航空航天大学 | Failure knowledge storage and push method for FMEA (failure mode and effects analysis) process |
| CN106326346A (en) * | 2016-08-06 | 2017-01-11 | 上海高欣计算机系统有限公司 | Text classification method and terminal device |
| CN106815252A (en) * | 2015-12-01 | 2017-06-09 | 阿里巴巴集团控股有限公司 | A kind of searching method and equipment |
| CN106941423A (en) * | 2017-04-13 | 2017-07-11 | 腾讯科技（深圳）有限公司 | Failure cause localization method and device |
| CN107171819A (en) * | 2016-03-07 | 2017-09-15 | 北京华为数字技术有限公司 | A kind of network fault diagnosis method and device |
| CN107248927A (en) * | 2017-05-02 | 2017-10-13 | 华为技术有限公司 | Generation method, Fault Locating Method and the device of fault location model |
| CN107291699A (en) * | 2017-07-04 | 2017-10-24 | 湖南星汉数智科技有限公司 | A kind of sentence semantic similarity computational methods |
| CN107291693A (en) * | 2017-06-15 | 2017-10-24 | 广州赫炎大数据科技有限公司 | A kind of semantic computation method for improving term vector model |
| CN107340766A (en) * | 2017-07-10 | 2017-11-10 | 浙江大学 | Power scheduling alarm signal text based on similarity sorts out and method for diagnosing faults |
| CN107391727A (en) * | 2017-08-01 | 2017-11-24 | 北京航空航天大学 | The method for digging and device of equipment fault sequence pattern |
| CN107704563A (en) * | 2017-09-29 | 2018-02-16 | 广州多益网络股份有限公司 | A kind of question sentence recommends method and system |
Family Cites Families (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN103617157B (en) * | 2013-12-10 | 2016-08-17 | 东北师范大学 | Based on semantic Text similarity computing method |
| CN103744905B (en) * | 2013-12-25 | 2018-03-30 | 新浪网技术(中国)有限公司 | Method for judging rubbish mail and device |
-
2018
- 2018-05-03 CN CN201810415523.0A patent/CN109902283B/en active Active
-
2019
- 2019-04-28 WO PCT/CN2019/084814 patent/WO2019210820A1/en active Application Filing
Patent Citations (18)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| JP2003173286A (en) * | 2001-12-05 | 2003-06-20 | Nippon Telegr & Teleph Corp <Ntt> | Method of acquiring sematic information on distributed network |
| CN101795210A (en) * | 2010-01-11 | 2010-08-04 | 浪潮通信信息系统有限公司 | Method for processing communication network failure |
| KR101078751B1 (en) * | 2011-02-23 | 2011-11-02 | 한국과학기술정보연구원 | Method and apparatus for detecting association errors of lexical semantic networks |
| US20120232905A1 (en) * | 2011-03-10 | 2012-09-13 | GM Global Technology Operations LLC | Methodology to improve failure prediction accuracy by fusing textual data with reliability model |
| US20130110496A1 (en) * | 2011-10-28 | 2013-05-02 | Sap Ag | Calculating Term Similarity Using A Meta-Model Semantic Network |
| CN102650960A (en) * | 2012-03-31 | 2012-08-29 | 奇智软件（北京）有限公司 | Method and device for eliminating faults of terminal equipment |
| US20130339787A1 (en) * | 2012-06-15 | 2013-12-19 | International Business Machines Coporation | Systematic failure remediation |
| CN104361026A (en) * | 2014-10-22 | 2015-02-18 | 北京航空航天大学 | Failure knowledge storage and push method for FMEA (failure mode and effects analysis) process |
| CN106815252A (en) * | 2015-12-01 | 2017-06-09 | 阿里巴巴集团控股有限公司 | A kind of searching method and equipment |
| CN107171819A (en) * | 2016-03-07 | 2017-09-15 | 北京华为数字技术有限公司 | A kind of network fault diagnosis method and device |
| CN106326346A (en) * | 2016-08-06 | 2017-01-11 | 上海高欣计算机系统有限公司 | Text classification method and terminal device |
| CN106941423A (en) * | 2017-04-13 | 2017-07-11 | 腾讯科技（深圳）有限公司 | Failure cause localization method and device |
| CN107248927A (en) * | 2017-05-02 | 2017-10-13 | 华为技术有限公司 | Generation method, Fault Locating Method and the device of fault location model |
| CN107291693A (en) * | 2017-06-15 | 2017-10-24 | 广州赫炎大数据科技有限公司 | A kind of semantic computation method for improving term vector model |
| CN107291699A (en) * | 2017-07-04 | 2017-10-24 | 湖南星汉数智科技有限公司 | A kind of sentence semantic similarity computational methods |
| CN107340766A (en) * | 2017-07-10 | 2017-11-10 | 浙江大学 | Power scheduling alarm signal text based on similarity sorts out and method for diagnosing faults |
| CN107391727A (en) * | 2017-08-01 | 2017-11-24 | 北京航空航天大学 | The method for digging and device of equipment fault sequence pattern |
| CN107704563A (en) * | 2017-09-29 | 2018-02-16 | 广州多益网络股份有限公司 | A kind of question sentence recommends method and system |
Cited By (18)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110245233A (en) * | 2019-06-19 | 2019-09-17 | 北京航空航天大学 | A kind of fault determination method and device |
| CN110378486A (en) * | 2019-07-15 | 2019-10-25 | 出门问问信息科技有限公司 | Internet startup disk method, apparatus, electronic equipment and storage medium |
| CN110569330A (en) * | 2019-07-18 | 2019-12-13 | 华瑞新智科技（北京）有限公司 | text labeling system, device, equipment and medium based on intelligent word selection |
| CN112667805A (en) * | 2019-09-30 | 2021-04-16 | 北京沃东天骏信息技术有限公司 | Work order category determination method, device, equipment and medium |
| CN112667805B (en) * | 2019-09-30 | 2024-04-09 | 北京沃东天骏信息技术有限公司 | Work order category determining method, device, equipment and medium |
| CN111046674A (en) * | 2019-12-20 | 2020-04-21 | 科大讯飞股份有限公司 | Semantic understanding method and device, electronic equipment and storage medium |
| CN111429155A (en) * | 2020-03-25 | 2020-07-17 | 中国银行股份有限公司 | Bank card dispute processing method and device |
| CN111858725A (en) * | 2020-04-30 | 2020-10-30 | 北京嘀嘀无限科技发展有限公司 | Event attribute determination method and system |
| CN111651601B (en) * | 2020-06-02 | 2023-04-18 | 全球能源互联网研究院有限公司 | Training method and classification method for fault classification model of power information system |
| CN111651601A (en) * | 2020-06-02 | 2020-09-11 | 全球能源互联网研究院有限公司 | Training method and classification method for fault classification model of power information system |
| CN112183994A (en) * | 2020-09-23 | 2021-01-05 | 南方电网数字电网研究院有限公司 | Method and device for evaluating equipment state, computer equipment and storage medium |
| CN112383421A (en) * | 2020-11-03 | 2021-02-19 | 中国联合网络通信集团有限公司 | Fault positioning method and device |
| CN112383421B (en) * | 2020-11-03 | 2023-03-24 | 中国联合网络通信集团有限公司 | Fault positioning method and device |
| CN112925668B (en) * | 2021-02-25 | 2024-04-05 | 北京百度网讯科技有限公司 | Method, device, equipment and storage medium for evaluating server health |
| CN112925668A (en) * | 2021-02-25 | 2021-06-08 | 北京百度网讯科技有限公司 | Server health evaluation method, device, equipment and storage medium |
| CN113657022A (en) * | 2021-07-15 | 2021-11-16 | 华为技术有限公司 | Chip fault identification method and related equipment |
| CN113657022B (en) * | 2021-07-15 | 2024-05-14 | 华为技术有限公司 | Chip fault recognition method and related equipment |
| CN113722494A (en) * | 2021-09-10 | 2021-11-30 | 中国航空工业集团公司西安飞行自动控制研究所 | Equipment fault positioning method based on natural language understanding |
Also Published As
| Publication number | Publication date |
|---|---|
| CN109902283B (en) | 2023-06-06 |
| WO2019210820A1 (en) | 2019-11-07 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109902283A (en) | A kind of information output method and device | |
| Li et al. | The emerging graph neural networks for intelligent fault diagnostics and prognostics: A guideline and a benchmark study | |
| Arora | Comparative analysis of classification algorithms on different datasets using WEKA | |
| JP2021182430A (en) | Method for training semantic analysis model, device, electronic apparatus, and storage medium | |
| CN111858123A (en) | Fault root cause analysis method and device based on directed graph network | |
| CN109614495B (en) | Related company mining method combining knowledge graph and text information | |
| CN110516757A (en) | A kind of transformer fault detection method and relevant apparatus | |
| CN109471793A (en) | A kind of webpage automatic test defect positioning method based on deep learning | |
| CN113312447A (en) | Semi-supervised log anomaly detection method based on probability label estimation | |
| CN112217674A (en) | Alarm root cause identification method based on causal network mining and graph attention network | |
| Matas | Comparing Network Centrality Measures as Tools for Identifying Key Concepts in Complex Networks: A Case of Wikipedia. | |
| CN111709244A (en) | Deep learning method for identifying causal relationship of contradictory dispute events | |
| CN106874963A (en) | A kind of Fault Diagnosis Method for Distribution Networks and system based on big data technology | |
| Khoshraftar et al. | Dynamic graph embedding via lstm history tracking | |
| Pavlenko et al. | Criterion of cyber-physical systems sustainability | |
| CN115328782A (en) | Semi-supervised software defect prediction method based on graph representation learning and knowledge distillation | |
| CN113268370A (en) | Root cause alarm analysis method, system, equipment and storage medium | |
| CN112015890B (en) | Method and device for generating movie script abstract | |
| Qi et al. | Solving asymmetric decision problems with influence diagrams | |
| Fu et al. | Nonlinear dynamic measurement method of software reliability based on data mining | |
| CN113378178A (en) | Deep learning-based graph confidence learning software vulnerability detection method | |
| CN110908919B (en) | Response test system based on artificial intelligence and application thereof | |
| CN112507720A (en) | Graph convolution network root identification method based on causal semantic relation transfer | |
| Pokorádi | Methodolody of Advanced Graph Model-based Vehicle Systems' Analysis | |
| CN116578336A (en) | Software clone detection method based on plagiarism-detector countermeasure |
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 |