CN109446514A - Construction method, device and the computer equipment of news property identification model - Google Patents
Construction method, device and the computer equipment of news property identification model Download PDFInfo
- Publication number
- CN109446514A CN109446514A CN201811089168.9A CN201811089168A CN109446514A CN 109446514 A CN109446514 A CN 109446514A CN 201811089168 A CN201811089168 A CN 201811089168A CN 109446514 A CN109446514 A CN 109446514A
- Authority
- CN
- China
- Prior art keywords
- model
- label
- chinese character
- news
- random field
- 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.)
- Pending
Links
- 238000010276 construction Methods 0.000 title claims abstract description 30
- 239000013598 vector Substances 0.000 claims abstract description 112
- 238000012549 training Methods 0.000 claims abstract description 104
- 238000013528 artificial neural network Methods 0.000 claims abstract description 52
- 210000005036 nerve Anatomy 0.000 claims abstract description 41
- 238000000605 extraction Methods 0.000 claims abstract description 34
- 238000000034 method Methods 0.000 claims abstract description 18
- 238000012360 testing method Methods 0.000 claims description 26
- 238000004590 computer program Methods 0.000 claims description 25
- 238000003062 neural network model Methods 0.000 claims description 15
- 239000003550 marker Substances 0.000 claims description 9
- 238000011478 gradient descent method Methods 0.000 claims description 8
- 238000012546 transfer Methods 0.000 claims description 8
- 230000000694 effects Effects 0.000 abstract description 4
- 238000013526 transfer learning Methods 0.000 abstract 1
- 238000010586 diagram Methods 0.000 description 6
- 230000008569 process Effects 0.000 description 4
- 210000004218 nerve net Anatomy 0.000 description 3
- 241000208340 Araliaceae Species 0.000 description 2
- 235000005035 Panax pseudoginseng ssp. pseudoginseng Nutrition 0.000 description 2
- 235000003140 Panax quinquefolius Nutrition 0.000 description 2
- 235000008434 ginseng Nutrition 0.000 description 2
- 238000002372 labelling Methods 0.000 description 2
- 230000021615 conjugation Effects 0.000 description 1
- 235000013399 edible fruits Nutrition 0.000 description 1
- 230000002708 enhancing effect Effects 0.000 description 1
- 230000006870 function Effects 0.000 description 1
- 238000001746 injection moulding Methods 0.000 description 1
- 238000010801 machine learning Methods 0.000 description 1
- 239000000463 material Substances 0.000 description 1
- 238000005259 measurement Methods 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 238000012545 processing Methods 0.000 description 1
- 230000011218 segmentation Effects 0.000 description 1
- 239000000243 solution Substances 0.000 description 1
- 230000003068 static effect Effects 0.000 description 1
- 230000001360 synchronised effect Effects 0.000 description 1
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F40/00—Handling natural language data
- G06F40/20—Natural language analysis
- G06F40/279—Recognition of textual entities
- G06F40/289—Phrasal analysis, e.g. finite state techniques or chunking
- G06F40/295—Named entity recognition
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
Abstract
Construction method, device, computer equipment and the storage medium for the news property identification model based on transfer learning that this application involves a kind of.The described method includes: building Named Entity Extraction Model;The neural network parameter of nervus opticus network model in trained part-of-speech tagging model in advance is extracted, and initializes the first nerves network model of Named Entity Extraction Model according to neural network parameter;News corpus training sample is obtained, the first Chinese marks in news corpus training sample have corresponding label;The first word vector is converted by the first Chinese character, and the first word vector is input to first nerves network model, obtains the first eigenvector of Chinese character;Using the corresponding first eigenvector of the first Chinese character and corresponding label, Training is carried out to goal condition random field models, obtains news property identification model.The recognition effect of news property identification model is able to ascend using this method.
Description
Technical field
This application involves machine learning techniques field, more particularly to a kind of news property identification model construction method,
Device, computer equipment and storage medium.
Background technique
Currently, neural network model combination conventional machines learning model --- condition random field (Conditional
Random Field algorithm, CRF) model is to name the current main model of Entity recognition.Mind in the main model
It also can be realized the automatic extraction of feature without artificial Feature Engineering through network model, but the determination of complicated model parameter is past
It is past to need to be trained by a large amount of labeled data.But in newsletter archive corpus, language that Business Name is marked or annotates
Expect that data are in contrast less, be not enough to the model of training complexity, traditional news property is caused to identify to newsletter archive corpus
In Business Name recognition effect it is unsatisfactory.
Summary of the invention
Based on this, it is necessary to be imitated for traditional news property identification to the identification of the Business Name in newsletter archive corpus
The unsatisfactory technical problem of fruit provides construction method, device, computer equipment and the storage of a kind of news property identification model
Medium.
A kind of construction method of news property identification model, which comprises
Named Entity Extraction Model is constructed, the Named Entity Extraction Model includes first nerves network model and target
Conditional random field models;
The neural network parameter of nervus opticus network model in trained part-of-speech tagging model in advance is extracted, and according to institute
It states neural network parameter and initializes the first nerves network model;
News corpus training sample is obtained, the first Chinese marks in the news corpus training sample have corresponding
Label;
The first word vector is converted by first Chinese character, and the first word vector is input to first mind
Through network model, the first eigenvector of the Chinese character is obtained;
Using the corresponding first eigenvector of first Chinese character and corresponding label, to the goal condition with
Airport model carries out Training, obtains news property identification model.
It is described in one of the embodiments, to utilize the corresponding first eigenvector of first Chinese character and correspondence
Label, Training, the step of obtaining news property identification model, packet are carried out to the goal condition random field models
It includes:
The first eigenvector is input to goal condition random field models, obtains the news corpus training sample
Prediction label sequence;
The default sequence label of the news corpus training sample is determined according to the corresponding label of first Chinese character,
And calculate the cross entropy between the predictive marker sequence and the default sequence label;
The parameter of the goal condition random field layer is adjusted, so that the intersection entropy minimization.
It is described in one of the embodiments, to extract nervus opticus network model in trained part-of-speech tagging model in advance
Neural network parameter the step of before, comprising:
Part-of-speech tagging model is constructed, the part-of-speech tagging model includes nervus opticus network model and source condition random field
Model;
Part-of-speech tagging training sample is obtained, wherein the second Chinese marks in the part-of-speech tagging training sample are corresponding
Part of speech label;
The second word vector is converted by second Chinese character, and the second word vector is input to nervus opticus net
Network model obtains the second feature vector of second Chinese character;
The second feature vector is input in the source conditional random field models, second Chinese character is obtained
Predict part of speech label;
According to the prediction part of speech label of second Chinese character and corresponding part of speech label, by reverse transfer and
Gradient descent method is adjusted the parameter of nervus opticus network model and source conditional random field models.
In one of the embodiments, the first nerves network model include forward recursive neural network hidden layer and
Backward recursive neural network hidden layer;
It is described that the first word vector is input to the first nerves network model, obtain the first of the Chinese character
The step of feature vector, comprising:
The first word vector is input in the forward recursive neural network hidden layer, to hidden state sequence before obtaining
Column;
The first word vector is input in the backward recursive neural network hidden layer, to hidden state sequence after acquisition
Column;
Merge the hidden status switch of the forward direction and the backward hidden status switch generates the of first Chinese character
One feature vector.
In one of the embodiments, after described the step of obtaining news property identification model, further includes:
Obtain news corpus test sample, the corresponding mark of third Chinese marks in the news corpus test sample
Label;
The word vector of the third Chinese character is input in news property identification model, the news corpus is obtained and surveys
The prediction label sequence of sample sheet;
It is calculated according to the label of the prediction label sequence of the news corpus test sample and the third Chinese character
The error rate of Business Name recognition result;
If the error rate is greater than preset threshold, to the goal condition random field mould in the news property identification model
Type carries out parameter adjustment.
A kind of construction device of news property identification model, described device include:
Model construction module, for constructing Named Entity Extraction Model, the Named Entity Extraction Model includes the first mind
Through network model and goal condition random field models;
Neural network parameter obtains module, for extracting nervus opticus network mould in preparatory trained part-of-speech tagging model
The neural network parameter of type, and the first nerves network model is initialized according to the neural network parameter;
Training sample obtains module, for obtaining news corpus training sample, in the news corpus training sample the
One Chinese marks have corresponding label;
Feature vector obtains module, for converting the first word vector for first Chinese character, and by described first
Word vector is input to the first nerves network model, obtains the first eigenvector of the Chinese character;
Model training module, for utilizing the corresponding first eigenvector of first Chinese character and corresponding mark
Label carry out Training to the goal condition random field models, obtain news property identification model.
The model training module in one of the embodiments, for the first eigenvector to be input to target
Conditional random field models obtain the prediction label sequence of the news corpus training sample;According to first Chinese character pair
The label answered determines the default sequence label of the news corpus training sample, and calculate the predictive marker sequence with it is described pre-
If the cross entropy between sequence label;The parameter of the goal condition random field layer is adjusted, so that the intersection entropy minimization.
The construction device of news property identification model further includes that originating task obtains module in one of the embodiments,;
The originating task obtains module, and for constructing part-of-speech tagging model, the part-of-speech tagging model includes nervus opticus
Network model and source conditional random field models;Part-of-speech tagging training sample is obtained, wherein in the part-of-speech tagging training sample
The corresponding part of speech label of the second Chinese marks;The second word vector is converted by second Chinese character, and will be described
Second word vector is input to nervus opticus network model, obtains the second feature vector of second Chinese character;By described
Two feature vectors are input in the source conditional random field models, obtain the prediction part of speech label of second Chinese character;Root
According to the prediction part of speech label and corresponding part of speech label of second Chinese character, pass through reverse transfer and gradient descent method
The parameter of nervus opticus network model and source conditional random field models is adjusted.
A kind of computer equipment, including memory and processor, the memory are stored with computer program, the processing
Device performs the steps of when executing the computer program
Named Entity Extraction Model is constructed, the Named Entity Extraction Model includes first nerves network model and target
Conditional random field models;
The neural network parameter of nervus opticus network model in trained part-of-speech tagging model in advance is extracted, and according to institute
It states neural network parameter and initializes the first nerves network model;
News corpus training sample is obtained, the first Chinese marks in the news corpus training sample have corresponding
Label;
The first word vector is converted by first Chinese character, and the first word vector is input to first mind
Through network model, the first eigenvector of the Chinese character is obtained;
Using the corresponding first eigenvector of first Chinese character and corresponding label, to the goal condition with
Airport model carries out Training, obtains news property identification model.
A kind of computer readable storage medium, is stored thereon with computer program, and the computer program is held by processor
It is performed the steps of when row
Named Entity Extraction Model is constructed, the Named Entity Extraction Model includes first nerves network model and target
Conditional random field models;
The neural network parameter of nervus opticus network model in trained part-of-speech tagging model in advance is extracted, and according to institute
It states neural network parameter and initializes the first nerves network model;
News corpus training sample is obtained, the first Chinese marks in the news corpus training sample have corresponding
Label;
The first word vector is converted by first Chinese character, and the first word vector is input to first mind
Through network model, the first eigenvector of the Chinese character is obtained;
Using the corresponding first eigenvector of first Chinese character and corresponding label, to the goal condition with
Airport model carries out Training, obtains news property identification model.
Construction method, device, computer equipment and the storage medium of above-mentioned news property identification model utilize preparatory training
Good part-of-speech tagging model moves to name Entity recognition mould as originating task, using neural network model as reusable feature
In type, neural network model output is the text information feature for inputting the Chinese character of corpus, so that name Entity recognition mould
Type only needs this layer of training objective condition random field, less in the corpus data for having had mark or annotation of Business Name
In the case of, mention the Named Entity Extraction Model for training to the accuracy rate of the Business Name identification in newsletter archive corpus
Height promotes the recognition effect of news property identification model.
Detailed description of the invention
Fig. 1 is the schematic diagram of internal structure of computer equipment in one embodiment of the invention;
Fig. 2 is the flow chart of the construction method of news property identification model in one embodiment of the invention;
Fig. 3 is the structural block diagram of the construction device of news property identification model in one embodiment of the invention;
Fig. 4 is the structural block diagram of the construction device of news property identification model in another embodiment of the present invention;
Fig. 5 is the structural block diagram of the construction device of news property identification model in another embodiment of the invention.
Specific embodiment
It is with reference to the accompanying drawings and embodiments, right in order to which the objects, technical solutions and advantages of the application are more clearly understood
The application is further elaborated.It should be appreciated that specific embodiment described herein is only used to explain the application, not
For limiting the application.
Fig. 1 is the schematic diagram of internal structure of computer equipment in one embodiment.The computer equipment can be server,
As shown in Figure 1, the computer equipment includes processor, memory, network interface and the database connected by system bus.Its
In, the processor of the computer equipment is for providing calculating and control ability.The memory of the computer equipment includes non-volatile
Property storage medium, built-in storage.The non-volatile memory medium is stored with operating system, computer program and database.This is interior
Memory provides environment for the operation of operating system and computer program in non-volatile memory medium.The computer equipment
Database is for storing the data such as news corpus training sample, neural network parameter.The network interface of the computer equipment is used for
It is communicated with external terminal by network connection.To realize a kind of news property identification when the computer program is executed by processor
Model method.
It will be understood by those skilled in the art that structure shown in Fig. 1, only part relevant to application scheme is tied
The block diagram of structure does not constitute the restriction for the computer equipment being applied thereon to application scheme, specific computer equipment
It may include perhaps combining certain components or with different component layouts than more or fewer components as shown in the figure.
In one embodiment, as shown in Fig. 2, a kind of construction method of news property identification model is provided, with the party
Method is applied to be illustrated for the server in Fig. 1, comprising the following steps:
Step S210: building Named Entity Extraction Model, Named Entity Extraction Model include first nerves network model with
And goal condition random field models.
In this step, server construction Named Entity Extraction Model, Named Entity Extraction Model is neural network model knot
It is built-up to close conditional random field models.
Step S220: the neural network ginseng of nervus opticus network model in trained part-of-speech tagging model in advance is extracted
Number, and first nerves network model is initialized according to neural network parameter.
In this step, part-of-speech tagging model is the building of neural network model conjugation condition random field models, in advance in Chinese
It is trained under language environment;Server utilizes the nerve net of part-of-speech tagging model using the part-of-speech tagging model as originating task
The neural network parameter of network model initializes the neural network model of Named Entity Extraction Model.
Step S230: obtaining news corpus training sample, and the first Chinese marks in news corpus training sample have
Corresponding label.
In this step, server, which obtains Chinese marks, has the news corpus of label as training sample, wherein news
The corresponding Chinese character of Business Name is come out by label for labelling as name entity in corpus training sample;Specifically, can
The label of the Chinese character in news corpus training sample to be made an addition to the rear end of each Chinese character, scratched with underscore;
Wherein, label for labelling rule can use BIOES mode, and B label, i.e. Begin indicate beginning character;I label, i.e.,
Intermediate indicates intermediate character;E label, i.e. End indicate ending character;S label, i.e. Single indicate single word
Symbol;O label, i.e. Other indicate other characters, for marking unrelated character, i.e. Business Name pair in news corpus training sample
The Chinese character segmentation markers answered have B label, I label and E label, and in news corpus training sample except Business Name with
Outer Chinese character is labeled as O label.
Step S240: the first word vector is converted by the first Chinese character, and the first word vector is input to first nerves
Network model obtains the first eigenvector of Chinese character.
In this step, the Chinese character in news corpus training sample is converted word vector by server, and by word vector
It is input in the neural network model of Named Entity Extraction Model, to obtain each Chinese character in news corpus training sample
Feature vector;Specifically, server can use general word2vec model for first in instruction news corpus training sample
Middle text is converted into corresponding word vector.
Step S250: the corresponding first eigenvector of the first Chinese character and corresponding label are utilized, to goal condition
Random field models carry out Training, obtain news property identification model.
In this step, the corresponding first eigenvector of the first Chinese character of server by utilizing and corresponding label, to life
The conditional random field models of name entity recognition model carry out Training, and the Named Entity Extraction Model after training is as final
News property identification model, wherein only need the parameter to the conditional random field models of Named Entity Extraction Model to adjust
Whole, the neural network model of Named Entity Extraction Model does not need re -training.
In the construction method of above-mentioned news property identification model, appoint using preparatory trained part-of-speech tagging model as source
Business exports neural network model moving in Named Entity Extraction Model as reusable feature, neural network model
It is the text information feature for inputting the Chinese character of corpus, so that Named Entity Extraction Model only needs training objective condition random
This layer of field makes to train the name come in the case where the corpus data for having had mark or annotation of Business Name is less
Entity recognition model improves the accuracy rate of the Business Name identification in newsletter archive corpus, promotes news property identification model
Recognition effect.
In one embodiment, using the corresponding first eigenvector of the first Chinese character and corresponding label, to mesh
Mark the step of conditional random field models carry out Training, obtain news property identification model, comprising: by first eigenvector
Goal condition random field models are input to, the prediction label sequence of news corpus training sample is obtained;According to the first Chinese character
Corresponding label determines the default sequence label of news corpus training sample, and calculates predictive marker sequence and default sequence label
Between cross entropy;The parameter of goal condition random field layer is adjusted, so as to intersect entropy minimization.
In the present embodiment, the feature vector of the middle character in news corpus training sample is input to goal condition by server
It will be that default sequence label is defeated as the true value of goal condition random field models in news corpus training sample in random field models
Out, by being minimised as target with the cross entropy between predictive marker sequence and default sequence label, adjustment updates goal condition
The parameter of random field layer, the predictive marker sequence for improving the output of goal condition random field models are consistent with true sequence label
Property, the accuracy rate of the Business Name identification in newsletter archive corpus is improved to improve Named Entity Extraction Model, in addition, with
Cross entropy effectively avoids the problem that learning rate reduces when gradient declines as loss function.
In one embodiment, after the step of obtaining news property identification model, further includes: obtain news corpus test
Sample, the corresponding label of third Chinese marks in news corpus test sample;The word vector of third Chinese character is defeated
Enter into news property identification model, obtains the prediction label sequence of news corpus test sample;According to news corpus test specimens
The error rate of label Computer Corp. title recognition result of prediction label sequence originally and third Chinese character;If error rate is big
In preset threshold, then parameter adjustment is carried out to the goal condition random field models in news property identification model.
The above-mentioned test process for the goal condition random field models in news property identification model, by obtaining news language
The character vector of Chinese character is input in the first nerves network of news property identification model in material test sample, obtains news
The feature vector of Chinese character in corpus test sample, it is then that the feature vector of Chinese character in news corpus test sample is defeated
Enter into the goal condition random field models of news property identification model, obtains the prediction label sequence of news corpus test sample
Column;Prediction label sequence and the sequence label of Chinese character in news corpus test sample are compared, news property is calculated
The error rate of Business Name recognition result in identification model, if error rate is greater than preset threshold, i.e., when threshold value is not achieved in accuracy rate,
Again parameter adjustment is carried out to the goal condition random field models in news property identification model, to guarantee the accurate of sequence label
Property, thus the accuracy of guarantee company's title identification.
In one embodiment, first nerves network model includes forward recursive neural network hidden layer and backward recursive
Neural network hidden layer;First word vector is input to first nerves network model, obtains the first eigenvector of Chinese character
The step of, comprising: the first word vector is input in forward recursive neural network hidden layer, to hidden status switch before obtaining;It will
First word vector is input in backward recursive neural network hidden layer, to hidden status switch after acquisition;To hidden state sequence before merging
Column and backward hidden status switch generate the first eigenvector of the first Chinese character.
In the present embodiment, the word vector of the Chinese character in news corpus training sample is input to name entity by server
In the neural network model of identification model, by forward recursive neural network hidden layer, according to the previous word of current word vector
The hidden state vector of vector calculates the hidden status switch of forward direction of current word vector;And pass through backward recursive neural network hidden layer,
The backward hidden status switch of current word vector is calculated according to the hidden state vector of the latter word vector of current word vector, then will
The hidden status switch of forward direction is cascaded with backward hidden status switch, obtains the feature vector of Chinese character, wherein feature vector packet
The dependence for containing Chinese character Yu front and back Chinese character can in the subsequent progress Entity recognition to newsletter archive corpus
It provides more to language, semantic relevant feature, effectively reduces identification Business Name task to the labeled data in professional domain
Dependence.
In one embodiment, the nerve net of nervus opticus network model in trained part-of-speech tagging model in advance is extracted
Before the step of network parameter, comprising: building part-of-speech tagging model, part-of-speech tagging model includes nervus opticus network model and source
Conditional random field models;Part-of-speech tagging training sample is obtained, wherein the second Chinese marks in part-of-speech tagging training sample
Corresponding part of speech label;The second word vector is converted by the second Chinese character, and the second word vector is input to nervus opticus net
Network model obtains the second feature vector of the second Chinese character;Second feature vector is input in the conditional random field models of source,
Obtain the prediction part of speech label of the second Chinese character;According to the prediction part of speech label of the second Chinese character and corresponding part of speech mark
Label carry out the parameter of nervus opticus network model and source conditional random field models by reverse transfer and gradient descent method
Adjustment.
The present embodiment is the training process of part-of-speech tagging model, server construction by neural network model and source condition with
The part-of-speech tagging model that airport model is constituted, and obtain the training sample of part-of-speech tagging model, wherein part-of-speech tagging training sample
Each Chinese character is labeled with corresponding part of speech label, such as verb, noun, adjective etc. in this;Server is by part of speech
Mark each Chinese character of training sample is converted into word vector, and the word vector is input to the nerve net of part-of-speech tagging model
In network model, the feature vector of Chinese character is obtained;This feature vector is input in the conditional random field models of source, in each
Chinese character carries out part-of-speech tagging, obtains part-of-speech tagging result;According to the part-of-speech tagging result of each Chinese character and originally
Part of speech label is adjusted the parameter of source conditional random field models and neural network model.
It should be understood that although each step in the flow chart of Fig. 2 is successively shown according to the instruction of arrow, this
A little steps are not that the inevitable sequence according to arrow instruction successively executes.Unless expressly state otherwise herein, these steps
It executes there is no the limitation of stringent sequence, these steps can execute in other order.Moreover, at least part in Fig. 2
Step may include that perhaps these sub-steps of multiple stages or stage are executed in synchronization to multiple sub-steps
It completes, but can execute at different times, the execution sequence in these sub-steps or stage, which is also not necessarily, successively to be carried out,
But it can be executed in turn or alternately at least part of the sub-step or stage of other steps or other steps.
In one embodiment, as shown in figure 3, providing a kind of construction device of news property identification model, comprising: mould
Type constructs module 310, neural network parameter obtains module 320, training sample obtains module 330, feature vector obtains module 340
With model training module 350, in which:
Model construction module 310, for constructing Named Entity Extraction Model, Named Entity Extraction Model includes first nerves
Network model and goal condition random field models;
Neural network parameter obtains module 320, for extracting nervus opticus net in preparatory trained part-of-speech tagging model
The neural network parameter of network model, and first nerves network model is initialized according to neural network parameter;
Training sample obtain module 330, for obtaining news corpus training sample, in news corpus training sample first
Chinese marks have corresponding label;
Feature vector obtains module 340, for converting the first word vector for the first Chinese character, and by the first word vector
It is input to first nerves network model, obtains the first eigenvector of Chinese character;
Model training module 350, for utilizing the corresponding first eigenvector of the first Chinese character and corresponding label,
Training is carried out to goal condition random field models, obtains news property identification model.
In one embodiment, model training module 350 is used to first eigenvector being input to goal condition random field
Model obtains the prediction label sequence of news corpus training sample;News language is determined according to the corresponding label of the first Chinese character
Expect the default sequence label of training sample, and calculates the cross entropy between predictive marker sequence and default sequence label;Adjust mesh
The parameter of condition random field layer is marked, so as to intersect entropy minimization.
In one embodiment, as shown in figure 4, providing a kind of construction device of news property identification model, news is real
The construction device of body identification model further includes that originating task obtains module 360;Originating task obtains module 360, for constructing part of speech mark
Injection molding type, part-of-speech tagging model include nervus opticus network model and source conditional random field models;Obtain part-of-speech tagging training
Sample, the wherein corresponding part of speech label of the second Chinese marks in part-of-speech tagging training sample;Second Chinese character is turned
The second word vector is turned to, and the second word vector is input to nervus opticus network model, obtain the second Chinese character second is special
Levy vector;Second feature vector is input in the conditional random field models of source, the prediction part of speech label of the second Chinese character is obtained;
According to the prediction part of speech label of the second Chinese character and corresponding part of speech label, pass through reverse transfer and gradient descent method pair
The parameter of nervus opticus network model and source conditional random field models is adjusted.
In one embodiment, first nerves network model includes forward recursive neural network hidden layer and backward recursive
Neural network hidden layer;Feature vector obtains module 340 and implies for the first word vector to be input to forward recursive neural network
In layer, to hidden status switch before obtaining;First word vector is input in backward recursive neural network hidden layer, to hidden after acquisition
Status switch;The first eigenvector of the first Chinese character is generated before merging to hidden status switch and backward hidden status switch.
In one embodiment, as shown in figure 5, providing a kind of construction device of news property identification model, news is real
The construction device of body identification model further includes model measurement module 370, and for obtaining news corpus test sample, news corpus is surveyed
The corresponding label of third Chinese marks in sample sheet;The word vector of third Chinese character is input to news property identification
In model, the prediction label sequence of news corpus test sample is obtained;According to the prediction label sequence of news corpus test sample
And the error rate of label Computer Corp. title recognition result of third Chinese character;If error rate is greater than preset threshold, right
Goal condition random field models in news property identification model carry out parameter adjustment.
The specific restriction of construction device about news property identification model may refer to know above for news property
The restriction of the construction method of other model, details are not described herein.Each mould in the construction device of above-mentioned news property identification model
Block can be realized fully or partially through software, hardware and combinations thereof.Above-mentioned each module can be embedded in the form of hardware or independence
In processor in computer equipment, it can also be stored in a software form in the memory in computer equipment, in order to
Processor, which calls, executes the corresponding operation of the above modules.
In one embodiment, a kind of computer equipment, including memory and processor are provided, which is stored with
Computer program, the processor perform the steps of when executing computer program
Named Entity Extraction Model is constructed, Named Entity Extraction Model includes first nerves network model and goal condition
Random field models;
The neural network parameter of nervus opticus network model in trained part-of-speech tagging model in advance is extracted, and according to mind
First nerves network model is initialized through network parameter;
News corpus training sample is obtained, the first Chinese marks in news corpus training sample have corresponding mark
Label;
The first word vector is converted by the first Chinese character, and the first word vector is input to first nerves network model,
Obtain the first eigenvector of Chinese character;
Using the corresponding first eigenvector of the first Chinese character and corresponding label, to goal condition random field models
Training is carried out, news property identification model is obtained.
In one embodiment, processor executes computer program realization and utilizes the corresponding fisrt feature of the first Chinese character
Vector and corresponding label carry out Training to goal condition random field models, obtain news property identification model
It when step, implements following steps: first eigenvector is input to goal condition random field models, obtain news corpus instruction
Practice the prediction label sequence of sample;The default label of news corpus training sample is determined according to the corresponding label of the first Chinese character
Sequence, and calculate the cross entropy between predictive marker sequence and default sequence label;The parameter of goal condition random field layer is adjusted,
So as to intersect entropy minimization.
In one embodiment, building part-of-speech tagging model is also performed the steps of when processor executes computer program,
Part-of-speech tagging model includes nervus opticus network model and source conditional random field models;Part-of-speech tagging training sample is obtained,
The corresponding part of speech label of the second Chinese marks in middle part-of-speech tagging training sample;Second is converted by the second Chinese character
Word vector, and the second word vector is input to nervus opticus network model, obtain the second feature vector of the second Chinese character;It will
Second feature vector is input in the conditional random field models of source, obtains the prediction part of speech label of the second Chinese character;According to second
The prediction part of speech label of Chinese character and corresponding part of speech label, by reverse transfer and gradient descent method to nervus opticus
The parameter of network model and source conditional random field models is adjusted.
In one embodiment, first nerves network model includes forward recursive neural network hidden layer and backward recursive
Neural network hidden layer;Processor executes computer program realization and the first word vector is input to first nerves network model, obtains
When obtaining the step of the first eigenvector of Chinese character, implements following steps: the first word vector is input to forward recursive
In neural network hidden layer, to hidden status switch before obtaining;First word vector is input to backward recursive neural network hidden layer
In, to hidden status switch after acquisition;The first Chinese character is generated to hidden status switch and backward hidden status switch before merging
First eigenvector.
In one embodiment, it is also performed the steps of when processor executes computer program and obtains news corpus test
Sample, the corresponding label of third Chinese marks in news corpus test sample;The word vector of third Chinese character is defeated
Enter into news property identification model, obtains the prediction label sequence of news corpus test sample;According to news corpus test specimens
The error rate of label Computer Corp. title recognition result of prediction label sequence originally and third Chinese character;If error rate is big
In preset threshold, then parameter adjustment is carried out to the goal condition random field models in news property identification model.
In one embodiment, a kind of computer readable storage medium is provided, computer program is stored thereon with, is calculated
Machine program performs the steps of when being executed by processor
Named Entity Extraction Model is constructed, Named Entity Extraction Model includes first nerves network model and goal condition
Random field models;
The neural network parameter of nervus opticus network model in trained part-of-speech tagging model in advance is extracted, and according to mind
First nerves network model is initialized through network parameter;
News corpus training sample is obtained, the first Chinese marks in news corpus training sample have corresponding mark
Label;
The first word vector is converted by the first Chinese character, and the first word vector is input to first nerves network model,
Obtain the first eigenvector of Chinese character;
Using the corresponding first eigenvector of the first Chinese character and corresponding label, to goal condition random field models
Training is carried out, news property identification model is obtained.
In one embodiment, computer program is executed by processor realization using corresponding first spy of the first Chinese character
Vector and corresponding label are levied, Training is carried out to goal condition random field models, obtains news property identification model
Step when, implement following steps: first eigenvector is input to goal condition random field models, obtains news corpus
The prediction label sequence of training sample;The pre- bidding of news corpus training sample is determined according to the corresponding label of the first Chinese character
Sequence is signed, and calculates the cross entropy between predictive marker sequence and default sequence label;Adjust the ginseng of goal condition random field layer
Number, so as to intersect entropy minimization.
In one embodiment, building part-of-speech tagging mould is also performed the steps of when computer program is executed by processor
Type, part-of-speech tagging model include nervus opticus network model and source conditional random field models;Part-of-speech tagging training sample is obtained,
The corresponding part of speech label of the second Chinese marks wherein in part-of-speech tagging training sample;Is converted by the second Chinese character
Two word vectors, and the second word vector is input to nervus opticus network model, obtain the second feature vector of the second Chinese character;
Second feature vector is input in the conditional random field models of source, the prediction part of speech label of the second Chinese character is obtained;According to
The prediction part of speech label of two Chinese characters and corresponding part of speech label, by reverse transfer and gradient descent method to the second mind
Parameter through network model and source conditional random field models is adjusted.
In one embodiment, first nerves network model includes forward recursive neural network hidden layer and backward recursive
Neural network hidden layer;Computer program is executed by processor realization and the first word vector is input to first nerves network model,
When obtaining the step of the first eigenvector of Chinese character, following steps are implemented: to passing before the first word vector is input to
Return in neural network hidden layer, to hidden status switch before obtaining;First word vector is input to backward recursive neural network to imply
In layer, to hidden status switch after acquisition;The first Chinese character is generated to hidden status switch and backward hidden status switch before merging
First eigenvector.
In one embodiment, it is also performed the steps of when computer program is executed by processor and obtains news corpus survey
Sample sheet, the corresponding label of third Chinese marks in news corpus test sample;By the word vector of third Chinese character
It is input in news property identification model, obtains the prediction label sequence of news corpus test sample;It is tested according to news corpus
The error rate of label Computer Corp. title recognition result of the prediction label sequence and third Chinese character of sample;If error rate
Greater than preset threshold, then parameter adjustment is carried out to the goal condition random field models in news property identification model.
Those of ordinary skill in the art will appreciate that realizing all or part of the process in above-described embodiment method, being can be with
Relevant hardware is instructed to complete by computer program, the computer program can be stored in a non-volatile computer
In read/write memory medium, the computer program is when being executed, it may include such as the process of the embodiment of above-mentioned each method.Wherein,
To any reference of memory, storage, database or other media used in each embodiment provided herein,
Including non-volatile and/or volatile memory.Nonvolatile memory may include read-only memory (ROM), programming ROM
(PROM), electrically programmable ROM (EPROM), electrically erasable ROM (EEPROM) or flash memory.Volatile memory may include
Random access memory (RAM) or external cache.By way of illustration and not limitation, RAM is available in many forms,
Such as static state RAM (SRAM), dynamic ram (DRAM), synchronous dram (SDRAM), double data rate sdram (DDRSDRAM), enhancing
Type SDRAM (ESDRAM), synchronization link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM
(RDRAM), direct memory bus dynamic ram (DRDRAM) and memory bus dynamic ram (RDRAM) etc..
Each technical characteristic of above embodiments can be combined arbitrarily, for simplicity of description, not to above-described embodiment
In each technical characteristic it is all possible combination be all described, as long as however, the combination of these technical characteristics be not present lance
Shield all should be considered as described in this specification.
The several embodiments of the application above described embodiment only expresses, the description thereof is more specific and detailed, but simultaneously
It cannot therefore be construed as limiting the scope of the patent.It should be pointed out that coming for those of ordinary skill in the art
It says, without departing from the concept of this application, various modifications and improvements can be made, these belong to the protection of the application
Range.Therefore, the scope of protection shall be subject to the appended claims for the application patent.
Claims (10)
1. a kind of construction method of news property identification model, which comprises
Named Entity Extraction Model is constructed, the Named Entity Extraction Model includes first nerves network model and goal condition
Random field models;
The neural network parameter of nervus opticus network model in trained part-of-speech tagging model in advance is extracted, and according to the mind
The first nerves network model is initialized through network parameter;
News corpus training sample is obtained, the first Chinese marks in the news corpus training sample have corresponding mark
Label;
The first word vector is converted by first Chinese character, and the first word vector is input to the first nerves net
Network model obtains the first eigenvector of the Chinese character;
Using the corresponding first eigenvector of first Chinese character and corresponding label, to the goal condition random field
Model carries out Training, obtains news property identification model.
2. the method according to claim 1, wherein described utilize corresponding first spy of first Chinese character
Vector and corresponding label are levied, Training is carried out to the goal condition random field models, obtains news property identification
The step of model, comprising:
The first eigenvector is input to goal condition random field models, obtains the prediction of the news corpus training sample
Sequence label;
The default sequence label of the news corpus training sample is determined according to the corresponding label of first Chinese character, and is counted
Calculate the cross entropy between the predictive marker sequence and the default sequence label;
The parameter of the goal condition random field layer is adjusted, so that the intersection entropy minimization.
3. the method according to claim 1, wherein described extract in advance the in trained part-of-speech tagging model
Before the step of neural network parameter of two neural network models, comprising:
Part-of-speech tagging model is constructed, the part-of-speech tagging model includes nervus opticus network model and source condition random field mould
Type;
Part-of-speech tagging training sample is obtained, wherein the corresponding word of the second Chinese marks in the part-of-speech tagging training sample
Property label;
The second word vector is converted by second Chinese character, and the second word vector is input to nervus opticus network mould
Type obtains the second feature vector of second Chinese character;
The second feature vector is input in the source conditional random field models, the prediction of second Chinese character is obtained
Part of speech label;
According to the prediction part of speech label of second Chinese character and corresponding part of speech label, pass through reverse transfer and gradient
Descent method is adjusted the parameter of nervus opticus network model and source conditional random field models.
4. the method according to claim 1, wherein the first nerves network model includes forward recursive nerve
Network hidden layer and backward recursive neural network hidden layer;
It is described that the first word vector is input to the first nerves network model, obtain the fisrt feature of the Chinese character
The step of vector, comprising:
The first word vector is input in the forward recursive neural network hidden layer, to hidden status switch before obtaining;
The first word vector is input in the backward recursive neural network hidden layer, to hidden status switch after acquisition;
Merge the hidden status switch of the forward direction and the backward hidden status switch generates the first spy of first Chinese character
Levy vector.
5. the method according to claim 1, which is characterized in that described to obtain news property identification model
The step of after, further includes:
Obtain news corpus test sample, the corresponding label of third Chinese marks in the news corpus test sample;
The word vector of the third Chinese character is input in news property identification model, the news corpus test specimens are obtained
This prediction label sequence;
According to the prediction label sequence of the news corpus test sample and label Computer Corp. of the third Chinese character
The error rate of title recognition result;
If the error rate be greater than preset threshold, to the goal condition random field models in the news property identification model into
The adjustment of row parameter.
6. a kind of construction device of news property identification model, which is characterized in that described device includes:
Model construction module, for constructing Named Entity Extraction Model, the Named Entity Extraction Model includes first nerves net
Network model and goal condition random field models;
Neural network parameter obtains module, for extracting nervus opticus network model in preparatory trained part-of-speech tagging model
Neural network parameter, and the first nerves network model is initialized according to the neural network parameter;
Training sample obtain module, for obtaining news corpus training sample, in the news corpus training sample first in
Chinese character is labeled with corresponding label;
Feature vector obtains module, for converting the first word vector for first Chinese character, and by first word to
Amount is input to the first nerves network model, obtains the first eigenvector of the Chinese character;
Model training module, it is right for utilizing the corresponding first eigenvector of first Chinese character and corresponding label
The goal condition random field models carry out Training, obtain news property identification model.
7. device according to claim 6, which is characterized in that the model training module is used for the fisrt feature
Vector is input to goal condition random field models, obtains the prediction label sequence of the news corpus training sample;According to described
The corresponding label of first Chinese character determines the default sequence label of the news corpus training sample, and calculates the pre- mark
Remember the cross entropy between sequence and the default sequence label;The parameter for adjusting the goal condition random field layer, so that described
Intersect entropy minimization.
8. device according to claim 6, which is characterized in that further include that originating task obtains module;
The originating task obtains module, and for constructing part-of-speech tagging model, the part-of-speech tagging model includes nervus opticus network
Model and source conditional random field models;Part-of-speech tagging training sample is obtained, wherein the in the part-of-speech tagging training sample
The corresponding part of speech label of two Chinese marks;The second word vector is converted by second Chinese character, and by described second
Word vector is input to nervus opticus network model, obtains the second feature vector of second Chinese character;It is special by described second
Sign vector is input in the source conditional random field models, obtains the prediction part of speech label of second Chinese character;According to institute
The prediction part of speech label and corresponding part of speech label for stating the second Chinese character, by reverse transfer and gradient descent method to
The parameter of two neural network models and source conditional random field models is adjusted.
9. a kind of computer equipment, including memory and processor, the memory are stored with computer program, feature exists
In the step of processor realizes any one of claims 1 to 5 the method when executing the computer program.
10. a kind of computer readable storage medium, is stored thereon with computer program, which is characterized in that the computer program
The step of method described in any one of claims 1 to 5 is realized when being executed by processor.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201811089168.9A CN109446514A (en) | 2018-09-18 | 2018-09-18 | Construction method, device and the computer equipment of news property identification model |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201811089168.9A CN109446514A (en) | 2018-09-18 | 2018-09-18 | Construction method, device and the computer equipment of news property identification model |
Publications (1)
Publication Number | Publication Date |
---|---|
CN109446514A true CN109446514A (en) | 2019-03-08 |
Family
ID=65532977
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201811089168.9A Pending CN109446514A (en) | 2018-09-18 | 2018-09-18 | Construction method, device and the computer equipment of news property identification model |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN109446514A (en) |
Cited By (23)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN109992782A (en) * | 2019-04-02 | 2019-07-09 | 深圳市华云中盛科技有限公司 | Legal documents name entity recognition method, device and computer equipment |
CN110348012A (en) * | 2019-07-01 | 2019-10-18 | 北京明略软件系统有限公司 | Determine method, apparatus, storage medium and the electronic device of target character |
CN110516251A (en) * | 2019-08-29 | 2019-11-29 | 秒针信息技术有限公司 | A kind of construction method, construction device, equipment and the medium of electric business entity recognition model |
CN110532570A (en) * | 2019-09-10 | 2019-12-03 | 杭州橙鹰数据技术有限公司 | A kind of method and apparatus of method and apparatus and model training that naming Entity recognition |
CN110598213A (en) * | 2019-09-06 | 2019-12-20 | 腾讯科技(深圳)有限公司 | Keyword extraction method, device, equipment and storage medium |
CN110717331A (en) * | 2019-10-21 | 2020-01-21 | 北京爱医博通信息技术有限公司 | Neural network-based Chinese named entity recognition method, device, equipment and storage medium |
CN110866402A (en) * | 2019-11-18 | 2020-03-06 | 北京香侬慧语科技有限责任公司 | Named entity identification method and device, storage medium and electronic equipment |
CN111159200A (en) * | 2019-12-31 | 2020-05-15 | 华中科技大学鄂州工业技术研究院 | Data storage method and device based on deep learning |
CN111291566A (en) * | 2020-01-21 | 2020-06-16 | 北京明略软件系统有限公司 | Event subject identification method and device and storage medium |
CN111368544A (en) * | 2020-02-28 | 2020-07-03 | 中国工商银行股份有限公司 | Named entity identification method and device |
CN111444723A (en) * | 2020-03-06 | 2020-07-24 | 深圳追一科技有限公司 | Information extraction model training method and device, computer equipment and storage medium |
CN111583911A (en) * | 2020-04-30 | 2020-08-25 | 深圳市优必选科技股份有限公司 | Speech recognition method, device, terminal and medium based on label smoothing |
CN111651995A (en) * | 2020-06-07 | 2020-09-11 | 上海建科工程咨询有限公司 | Accident information automatic extraction method and system based on deep circulation neural network |
CN111737560A (en) * | 2020-07-20 | 2020-10-02 | 平安国际智慧城市科技股份有限公司 | Content search method, field prediction model training method, device and storage medium |
CN111753506A (en) * | 2020-05-15 | 2020-10-09 | 北京捷通华声科技股份有限公司 | Text replacement method and device |
CN111832291A (en) * | 2020-06-02 | 2020-10-27 | 北京百度网讯科技有限公司 | Entity recognition model generation method and device, electronic equipment and storage medium |
CN111859948A (en) * | 2019-04-28 | 2020-10-30 | 北京嘀嘀无限科技发展有限公司 | Language identification, language model training and character prediction method and device |
CN111885000A (en) * | 2020-06-22 | 2020-11-03 | 网宿科技股份有限公司 | Network attack detection method, system and device based on graph neural network |
CN111950277A (en) * | 2019-04-30 | 2020-11-17 | 中移(苏州)软件技术有限公司 | Business situation entity determining method, device and storage medium |
CN112101041A (en) * | 2020-09-08 | 2020-12-18 | 平安科技(深圳)有限公司 | Entity relationship extraction method, device, equipment and medium based on semantic similarity |
CN112699683A (en) * | 2020-12-31 | 2021-04-23 | 大唐融合通信股份有限公司 | Named entity identification method and device fusing neural network and rule |
CN113449113A (en) * | 2020-03-27 | 2021-09-28 | 京东数字科技控股有限公司 | Knowledge graph construction method and device, electronic equipment and storage medium |
CN113836928A (en) * | 2021-09-28 | 2021-12-24 | 平安科技(深圳)有限公司 | Text entity generation method, device, equipment and storage medium |
Citations (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20160171974A1 (en) * | 2014-12-15 | 2016-06-16 | Baidu Usa Llc | Systems and methods for speech transcription |
CN106557462A (en) * | 2016-11-02 | 2017-04-05 | 数库(上海)科技有限公司 | Name entity recognition method and system |
CN107203511A (en) * | 2017-05-27 | 2017-09-26 | 中国矿业大学 | A kind of network text name entity recognition method based on neutral net probability disambiguation |
CN107622050A (en) * | 2017-09-14 | 2018-01-23 | 武汉烽火普天信息技术有限公司 | Text sequence labeling system and method based on Bi LSTM and CRF |
CN107797992A (en) * | 2017-11-10 | 2018-03-13 | 北京百分点信息科技有限公司 | Name entity recognition method and device |
-
2018
- 2018-09-18 CN CN201811089168.9A patent/CN109446514A/en active Pending
Patent Citations (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20160171974A1 (en) * | 2014-12-15 | 2016-06-16 | Baidu Usa Llc | Systems and methods for speech transcription |
CN106557462A (en) * | 2016-11-02 | 2017-04-05 | 数库(上海)科技有限公司 | Name entity recognition method and system |
CN107203511A (en) * | 2017-05-27 | 2017-09-26 | 中国矿业大学 | A kind of network text name entity recognition method based on neutral net probability disambiguation |
CN107622050A (en) * | 2017-09-14 | 2018-01-23 | 武汉烽火普天信息技术有限公司 | Text sequence labeling system and method based on Bi LSTM and CRF |
CN107797992A (en) * | 2017-11-10 | 2018-03-13 | 北京百分点信息科技有限公司 | Name entity recognition method and device |
Non-Patent Citations (3)
Title |
---|
JOHN M. GIORGI ET AL: "Transfer learning for biomedical named entity recognition with neural networks", 《BIOINFORMATICS》, vol. 34, no. 23, 1 June 2018 (2018-06-01), pages 4087 - 4094 * |
ZHILIN YANG ET AL: "Transfer Learning For Sequence Tagging With Hierarchical Recurrent Networks", 《ARXIV》, 18 March 2017 (2017-03-18), pages 1 - 10 * |
冯蕴天;张宏军;郝文宁;陈刚;: "基于深度信念网络的命名实体识别", 计算机科学, no. 04, 15 April 2016 (2016-04-15), pages 224 - 230 * |
Cited By (37)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN109992782A (en) * | 2019-04-02 | 2019-07-09 | 深圳市华云中盛科技有限公司 | Legal documents name entity recognition method, device and computer equipment |
CN109992782B (en) * | 2019-04-02 | 2023-07-07 | 深圳市华云中盛科技股份有限公司 | Legal document named entity identification method and device and computer equipment |
CN111859948A (en) * | 2019-04-28 | 2020-10-30 | 北京嘀嘀无限科技发展有限公司 | Language identification, language model training and character prediction method and device |
CN111950277A (en) * | 2019-04-30 | 2020-11-17 | 中移(苏州)软件技术有限公司 | Business situation entity determining method, device and storage medium |
CN110348012A (en) * | 2019-07-01 | 2019-10-18 | 北京明略软件系统有限公司 | Determine method, apparatus, storage medium and the electronic device of target character |
CN110348012B (en) * | 2019-07-01 | 2022-12-09 | 北京明略软件系统有限公司 | Method, device, storage medium and electronic device for determining target character |
CN110516251B (en) * | 2019-08-29 | 2023-11-03 | 秒针信息技术有限公司 | Method, device, equipment and medium for constructing electronic commerce entity identification model |
CN110516251A (en) * | 2019-08-29 | 2019-11-29 | 秒针信息技术有限公司 | A kind of construction method, construction device, equipment and the medium of electric business entity recognition model |
CN110598213A (en) * | 2019-09-06 | 2019-12-20 | 腾讯科技(深圳)有限公司 | Keyword extraction method, device, equipment and storage medium |
CN110532570A (en) * | 2019-09-10 | 2019-12-03 | 杭州橙鹰数据技术有限公司 | A kind of method and apparatus of method and apparatus and model training that naming Entity recognition |
CN110717331B (en) * | 2019-10-21 | 2023-10-24 | 北京爱医博通信息技术有限公司 | Chinese named entity recognition method, device and equipment based on neural network and storage medium |
CN110717331A (en) * | 2019-10-21 | 2020-01-21 | 北京爱医博通信息技术有限公司 | Neural network-based Chinese named entity recognition method, device, equipment and storage medium |
CN110866402B (en) * | 2019-11-18 | 2023-11-28 | 北京香侬慧语科技有限责任公司 | Named entity identification method and device, storage medium and electronic equipment |
CN110866402A (en) * | 2019-11-18 | 2020-03-06 | 北京香侬慧语科技有限责任公司 | Named entity identification method and device, storage medium and electronic equipment |
CN111159200B (en) * | 2019-12-31 | 2023-10-17 | 华中科技大学鄂州工业技术研究院 | Data storage method and device based on deep learning |
CN111159200A (en) * | 2019-12-31 | 2020-05-15 | 华中科技大学鄂州工业技术研究院 | Data storage method and device based on deep learning |
CN111291566B (en) * | 2020-01-21 | 2023-04-28 | 北京明略软件系统有限公司 | Event main body recognition method, device and storage medium |
CN111291566A (en) * | 2020-01-21 | 2020-06-16 | 北京明略软件系统有限公司 | Event subject identification method and device and storage medium |
CN111368544B (en) * | 2020-02-28 | 2023-09-19 | 中国工商银行股份有限公司 | Named entity identification method and device |
CN111368544A (en) * | 2020-02-28 | 2020-07-03 | 中国工商银行股份有限公司 | Named entity identification method and device |
CN111444723A (en) * | 2020-03-06 | 2020-07-24 | 深圳追一科技有限公司 | Information extraction model training method and device, computer equipment and storage medium |
CN113449113A (en) * | 2020-03-27 | 2021-09-28 | 京东数字科技控股有限公司 | Knowledge graph construction method and device, electronic equipment and storage medium |
CN111583911A (en) * | 2020-04-30 | 2020-08-25 | 深圳市优必选科技股份有限公司 | Speech recognition method, device, terminal and medium based on label smoothing |
CN111583911B (en) * | 2020-04-30 | 2023-04-14 | 深圳市优必选科技股份有限公司 | Speech recognition method, device, terminal and medium based on label smoothing |
CN111753506B (en) * | 2020-05-15 | 2023-12-08 | 北京捷通华声科技股份有限公司 | Text replacement method and device |
CN111753506A (en) * | 2020-05-15 | 2020-10-09 | 北京捷通华声科技股份有限公司 | Text replacement method and device |
CN111832291B (en) * | 2020-06-02 | 2024-01-09 | 北京百度网讯科技有限公司 | Entity recognition model generation method and device, electronic equipment and storage medium |
CN111832291A (en) * | 2020-06-02 | 2020-10-27 | 北京百度网讯科技有限公司 | Entity recognition model generation method and device, electronic equipment and storage medium |
CN111651995A (en) * | 2020-06-07 | 2020-09-11 | 上海建科工程咨询有限公司 | Accident information automatic extraction method and system based on deep circulation neural network |
CN111885000A (en) * | 2020-06-22 | 2020-11-03 | 网宿科技股份有限公司 | Network attack detection method, system and device based on graph neural network |
CN111885000B (en) * | 2020-06-22 | 2022-06-21 | 网宿科技股份有限公司 | Network attack detection method, system and device based on graph neural network |
CN111737560A (en) * | 2020-07-20 | 2020-10-02 | 平安国际智慧城市科技股份有限公司 | Content search method, field prediction model training method, device and storage medium |
WO2021121198A1 (en) * | 2020-09-08 | 2021-06-24 | 平安科技(深圳)有限公司 | Semantic similarity-based entity relation extraction method and apparatus, device and medium |
CN112101041A (en) * | 2020-09-08 | 2020-12-18 | 平安科技(深圳)有限公司 | Entity relationship extraction method, device, equipment and medium based on semantic similarity |
CN112699683A (en) * | 2020-12-31 | 2021-04-23 | 大唐融合通信股份有限公司 | Named entity identification method and device fusing neural network and rule |
CN113836928A (en) * | 2021-09-28 | 2021-12-24 | 平安科技(深圳)有限公司 | Text entity generation method, device, equipment and storage medium |
CN113836928B (en) * | 2021-09-28 | 2024-02-27 | 平安科技(深圳)有限公司 | Text entity generation method, device, equipment and storage medium |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN109446514A (en) | Construction method, device and the computer equipment of news property identification model | |
CN109522393A (en) | Intelligent answer method, apparatus, computer equipment and storage medium | |
CN110704633A (en) | Named entity recognition method and device, computer equipment and storage medium | |
CN109145315A (en) | Text interpretation method, device, storage medium and computer equipment | |
CN110162627A (en) | Data increment method, apparatus, computer equipment and storage medium | |
CN109977234A (en) | A kind of knowledge mapping complementing method based on subject key words filtering | |
KR20190085098A (en) | Keyword extraction method, computer device, and storage medium | |
CN109271646A (en) | Text interpretation method, device, readable storage medium storing program for executing and computer equipment | |
CN108628974A (en) | Public feelings information sorting technique, device, computer equipment and storage medium | |
CN110399484A (en) | Sentiment analysis method, apparatus, computer equipment and the storage medium of long text | |
CN111462751B (en) | Method, apparatus, computer device and storage medium for decoding voice data | |
CN109815331A (en) | Construction method, device and the computer equipment of text emotion disaggregated model | |
CN110334179A (en) | Question and answer processing method, device, computer equipment and storage medium | |
CN111428448B (en) | Text generation method, device, computer equipment and readable storage medium | |
CN112633423B (en) | Training method of text recognition model, text recognition method, device and equipment | |
CN110968725B (en) | Image content description information generation method, electronic device and storage medium | |
Kamal et al. | Textmage: The automated bangla caption generator based on deep learning | |
CN109977394A (en) | Text model training method, text analyzing method, apparatus, equipment and medium | |
CN112016271A (en) | Language style conversion model training method, text processing method and device | |
CN109885830A (en) | Sentence interpretation method, device, computer equipment | |
CN114528394B (en) | Text triple extraction method and device based on mask language model | |
CN109992770A (en) | A kind of Laotian name entity recognition method based on combination neural net | |
CN115098722B (en) | Text and image matching method and device, electronic equipment and storage medium | |
CN112905763B (en) | Session system development method, device, computer equipment and storage medium | |
CN114638229A (en) | Entity identification method, device, medium and equipment of record data |
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 |