CN109241288A - Update training method, device and the equipment of textual classification model - Google Patents
Update training method, device and the equipment of textual classification model Download PDFInfo
- Publication number
- CN109241288A CN109241288A CN201811192187.4A CN201811192187A CN109241288A CN 109241288 A CN109241288 A CN 109241288A CN 201811192187 A CN201811192187 A CN 201811192187A CN 109241288 A CN109241288 A CN 109241288A
- Authority
- CN
- China
- Prior art keywords
- training
- layer
- sample text
- classification model
- sample
- 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.)
- Withdrawn
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
- G06F40/279—Recognition of textual entities
- G06F40/289—Phrasal analysis, e.g. finite state techniques or chunking
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F40/00—Handling natural language data
- G06F40/30—Semantic analysis
Abstract
This disclosure relates to field of artificial intelligence, specifically disclose update training method, device and the equipment of a kind of textual classification model, textual classification model includes extraction of semantics layer and classification layer, before newly-increased knowledge point, the training of extraction of semantics layer and layer of classifying is completed according to the sample data of original knowledge point, comprising: the mark label for obtaining the newly-increased corresponding sample text in knowledge point and being labeled to sample text；By the feature vector for completing the extraction of semantics layer building sample text of training according to sample data；The update training of classification layer is carried out, according to the corresponding mark label of the feature vector of sample text and sample text to realize the update training of textual classification model.When needing to be updated trained to textual classification model, the update training for only carrying out classification layer realizes timely updating for textual classification model so as to substantially shorten the time that textual classification model updates training.
Description
Technical field
This disclosure relates to field of artificial intelligence, in particular to the update training method of a kind of textual classification model, dress
It sets and equipment.
Background technique
Textual classification model in existing customer service robot question answering system increases new product newly in the knowledge base of customer service robot
Behind relevant knowledge point or the relevant knowledge point of newly-increased hot issue, need to textual classification model carry out re -training, one
As one textual classification model of re -training need long time, so as to cause textual classification model update not in time, customer service
Robot can not answer the problem of newly-increased knowledge point correlation.
So while textual classification model training time long the problem of causing textual classification model to update not in time, need
It solves.
Summary of the invention
In order to solve the problems, such as present in the relevant technologies, present disclose provides a kind of update training sides of textual classification model
Method and device.
A kind of update training method of textual classification model, the textual classification model include extraction of semantics layer and classification
Layer completes the instruction of the extraction of semantics layer and the classification layer according to the sample data of original knowledge point before newly-increased knowledge point
Practice, the update training method of the textual classification model includes:
The mark label for obtaining the newly-increased corresponding sample text in knowledge point and the sample text being labeled；
By according to the sample data complete training the extraction of semantics layer building described in sample text feature to
Amount；
The classification is carried out according to the feature vector of the sample text and the corresponding mark label of the sample text
The update training of layer, to realize the update training of the textual classification model.
A kind of update training device of textual classification model, the textual classification model include extraction of semantics layer and classification
Layer completes the instruction of the extraction of semantics layer and the classification layer according to the sample data of original knowledge point before newly-increased knowledge point
Practice, the update training device of the textual classification model includes:
Module is obtained, is configured as executing: obtaining the newly-increased corresponding sample text in knowledge point and to the sample text
The mark label being labeled；
Feature vector constructs module, is configured as executing: the semanteme by completing training according to the sample data
Extract the feature vector of sample text described in layer building；
Training module is updated, is configured as executing: according to the feature vector of the sample text and the sample text
Corresponding mark label carries out the update training of the classification layer, to realize the update training of the textual classification model.
In one embodiment, described eigenvector building module includes:
Participle unit is configured as executing: the semanteme by completing training according to the sample data of original knowledge point
Extract layer segments the sample text；
Feature vector construction unit is configured as executing: according to each word in the sample text it is corresponding coding with
And the semantic weight of each word constructs the feature vector of the sample text.
In one embodiment, described device further include:
Tag along sort complementary module is configured as executing: according to the corresponding mark label supplement of the sample text
The tag along sort of classification layer；
Tag along sort set update module, is configured as executing: updating the classification layer according to the tag along sort supplemented
Tag along sort set.
In one embodiment, the update training module includes:
Tag along sort predicting unit is configured as executing: using the classification layer according to the feature of the sample text to
Amount prediction obtains tag along sort corresponding to the sample text；
Judging unit is configured as executing: carrying out the obtained tag along sort and the corresponding mark of the sample text
Infuse the consistency judgement of label；
Adjustment unit is configured as executing: if inconsistent, adjusting the parameter of the classification layer until obtained described
Tag along sort is consistent with the mark label.
In one embodiment, tag along sort predicting unit includes:
Probability prediction unit is configured as executing: according to described eigenvector predicting to obtain using the classification layer described
Feature vector corresponds to the probability of each tag along sort in the updated tag along sort set；
Tag along sort determination unit is configured as executing: the probability of traversal each tag along sort, with most probable value
Corresponding tag along sort is as the corresponding tag along sort of the sample text.
In one embodiment, described device further include:
Class test module is configured as executing: by the updated textual classification model to several test samples
Classify；
Nicety of grading computing module is configured as executing: the text after updating training is calculated according to classification results
Nicety of grading of this disaggregated model to several test samples；
Training ending module is updated, is configured as executing: if the nicety of grading reaches designated precision, terminating the text
The update training of this disaggregated model.
A kind of update training equipment of textual classification model, comprising:
Processor；
Memory for storage processor executable instruction；
Wherein, the processor is configured to executing the process described above.
A kind of computer readable storage medium, is stored thereon with computer program, and the computer program is held by processor
The process described above is realized when row.
By the technical solution of the application, it has been trained according to the sample data of original knowledge point in textual classification model
On the basis of, when needing to be updated trained to textual classification model, the update training of classification layer is only carried out, realizes text point
The update training of class model realizes textual classification model so as to substantially shorten the time that textual classification model updates training
Timely update, and then can be used to carry out newly-increased knowledge point in time related for customer service robot in field of artificial intelligence etc.
The reply of problem.
It should be understood that the above general description and the following detailed description are merely exemplary, this can not be limited
It is open.
Detailed description of the invention
The drawings herein are incorporated into the specification and forms part of this specification, and shows and meets implementation of the invention
Example, and in specification together principle for explaining the present invention.
Fig. 1 is the schematic diagram of the implementation environment according to involved in the disclosure；
Fig. 2 is a kind of block diagram of server shown according to an exemplary embodiment；
Fig. 3 is a kind of flow chart of the update training method of textual classification model shown according to an exemplary embodiment；
Fig. 4 is the flow chart of the step S130 of embodiment illustrated in fig. 3；
Fig. 5 be embodiment illustrated in fig. 3 step S150 before step flow chart；
Fig. 6 is the flow chart of the step S150 of embodiment illustrated in fig. 3；
Fig. 7 is the flow chart of the step S151 of embodiment illustrated in fig. 6；
Fig. 8 be embodiment illustrated in fig. 3 step S150 after step flow chart；
Fig. 9 is a kind of block diagram of the update training device of textual classification model shown according to an exemplary embodiment；
Figure 10 is a kind of block diagram of the update training equipment of textual classification model shown according to an exemplary embodiment.
Specific embodiment
Here will the description is performed on the exemplary embodiment in detail, the example is illustrated in the accompanying drawings.Following description is related to
When attached drawing, unless otherwise indicated, the same numbers in different drawings indicate the same or similar elements.Following exemplary embodiment
Described in embodiment do not represent all embodiments consistented with the present invention.On the contrary, they be only with it is such as appended
The example of device and method being described in detail in claims, some aspects of the invention are consistent.
Fig. 1 is the schematic diagram of the implementation environment according to involved in the disclosure.The implementation environment includes: server 200 and extremely
A few terminal 100.
Wherein terminal 100 can be smart phone, tablet computer, laptop, desktop computer etc. can be with server
200 establish network connection and can run the electronic equipment of client, herein without specifically limiting.Terminal 100 and server
Wireless or wired network connection has been pre-established between 200, thus, it is realized by the client run on the terminal 100
Terminal 100 is interacted with server 200.
Based on the interaction between server 200 and terminal 100, server 200 can get user on the terminal 100
Then the sample text of input constructs the feature vector of the sample text, carries out classification prediction realization text point to feature vector
Update training of class model etc..Terminal 100 can receive the tag along sort for sample text that server 200 is returned.
It should be noted that the disclosure classification method, be not limited to dispose to handle accordingly in server 200 to patrol
Volume, it is also possible to the processing logic being deployed in other machines.For example, being deployed in the terminal device for having computing capability
The processing logic etc. of the update training of row textual classification model.
Fig. 2 is a kind of block diagram of server shown according to an exemplary embodiment.Server with this hardware configuration
It can be used for carrying out the update training of textual classification model and be deployed in implementation environment shown in FIG. 1.
It should be noted that the server is the example for adapting to the disclosure, it must not believe that there is provided to this
Any restrictions of open use scope.The server can not be construed to need to rely on or must have shown in Figure 2
One or more component in illustrative server 200.
The hardware configuration of the server can generate biggish difference due to the difference of configuration or performance, as shown in Fig. 2, clothes
Business device 200 includes: power supply 210, interface 230, at least a memory 250 and at least central processing unit (CPU, a Central
Processing Units)270。
Wherein, power supply 210 is used to provide operating voltage for each hardware device on server 200.
Interface 230 includes an at least wired or wireless network interface 231, at least a string and translation interface 233, at least one defeated
Enter output interface 235 and at least USB interface 237 etc., be used for and external device communication, such as carries out data with terminal 100
Transmission.
The carrier that memory 250 is stored as resource, can be read-only memory, random access memory, disk or CD
Deng the resource stored thereon includes operating system 251, application program 253 and data 255 etc., and storage mode can be of short duration
It stores or permanently stores.Wherein, operating system 251 is for managing and each hardware device in control server 200 and answering
It can be Windows with program 253 to realize calculating and processing of the central processing unit 270 to mass data 255
ServerTM, Mac OS XTM, UnixTM, LinuxTM, FreeBSDTM etc..Application program 253 be based on operating system 251 it
The upper computer program for completing at least one particular job, may include an at least module (being not shown in Fig. 2), each module
The series of computation machine readable instruction to server 200 can be separately included.Data 255 can be stored in disk
Sample data etc..
Central processing unit 270 may include the processor of one or more or more, and be set as through bus and memory
250 communications, for the mass data 255 in operation and processing memory 250.
As described in detail above, the server 200 for being applicable in the disclosure will read memory by central processing unit 270
The form of the series of computation machine readable instruction stored in 250 is trained to complete the update of textual classification model.
In the exemplary embodiment, server 200 can be by one or more application specific integrated circuit
At (Application Specific Integrated Circuit, abbreviation ASIC), digital signal processor, digital signal
Manage equipment, programmable logic device, field programmable gate array, controller, microcontroller, microprocessor or other electronic components
It realizes, for executing following file classification methods.Therefore, realize the present invention is not limited to any specific hardware circuit, software with
And the combination of the two.
Fig. 3 is a kind of flow chart of the update training method of textual classification model shown according to an exemplary embodiment.
The update training method of text disaggregated model can be executed with the server 200 of implementation environment shown in Fig. 1.Implementation shown in Fig. 3
In example, textual classification model includes extraction of semantics layer and classification layer, before newly-increased knowledge point, according to the sample number of original knowledge point
According to the training for completing extraction of semantics layer and layer of classifying, the update training method of text disaggregated model the following steps are included:
Step S110, the mark mark for obtaining the newly-increased corresponding sample text in knowledge point and sample text being labeled
Label.
The textual classification model of the application is constructed by neural network, and wherein textual classification model can pass through convolution
Neural network (CNN), Recognition with Recurrent Neural Network (RNN) etc. can carry out the neural network building of text classification, can also be by more
The neural network ensemble of seed type forms, herein without specifically limiting.
After the training for completing extraction of semantics layer and layer of classifying according to the sample data of original knowledge point, by the training,
The parameter of extraction of semantics layer and layer of classifying is determined, so that ask relevant to original knowledge point may be implemented in text disaggregated model
Topic is classified, i.e., extraction of semantics layer can construct the feature vector of text, and classification layer can be with text based feature vector pair
Text is classified.
After textual classification model completes training according to the sample data of original knowledge point, the sample number of original knowledge point
According to constituting the database of text disaggregated model.For the textual classification model in different application scene, sample data is different,
The database of corresponding textual classification model is not also identical.
Wherein increasing knowledge point newly can be the knowledge point for not including in the database of textual classification model, or for original
The knowledge point for thering is the knowledge point in database to modify, herein without specifically limiting.For newly-increased knowledge point, need to utilize
The sample text of newly-increased knowledge point and the update for carrying out textual classification model to the mark label that sample text is labeled are instructed
Practice.
For example, such as applied to the textual classification model in the customer service robot of insurance field, newly-increased knowledge point can
Corresponding to be the new insurance business developed, it is relevant to the new insurance business for increasing the sample text of knowledge point newly
Problem, such as the insurance handle process, handle material, handle the relevant issues such as condition, Claims Resolution process；Newly-increased knowledge point can be with
It is original settlement of insurance claim reschedualing, it is corresponding, increase settlement of insurance claim stream of the sample text of knowledge point as with the change newly
The problem of Cheng Xiangguan.To, the update for carrying out textual classification model after training, customer service robot can be mentioned for user
Classify out about the problem of newly-increased knowledge point, and then searches for answer according to classification results and answered to what user's presentation was searched for
Case.
The mark label of sample text is the label by manually being classified to the sample text, is being embodied
It, can be by being manually labeled to obtain mark label to sample text in example, and save the mark label of mark.
Step S130, the feature vector of the extraction of semantics layer building sample text by completing training according to sample data.
By being described above it is found that being completed after training in extraction of semantics layer according to sample data, the parameter of extraction of semantics layer
It determines, in step s 130, according to the feature vector of the extraction of semantics layer building sample text of determined parameter.And later
In step, the parameter for adjusting extraction of semantics layer again is not needed, i.e., without the update training of extraction of semantics layer.
Customer service robot especially in field of artificial intelligence, due to customer service robot carry out online service before,
Using the textual classification model of a large amount of sample data training customer service robot, the extraction of semantics layer building text of textual classification model
The perfect in shape and function of this feature vector.So before by online service after training, when increasing knowledge point newly, extraction of semantics
Layer can also construct the feature vector of text.
In one exemplary embodiment, as shown in figure 4, step S130 includes:
Step S131, by according to the sample data of original knowledge point complete training extraction of semantics layer to sample text into
Row participle.
Step S132 constructs sample according to the semantic weight of the corresponding coding of each word in sample text and each word
The feature vector of text.
Sample text is segmented, sample text is divided into several tactic phrases.Participle can use
Segmentation methods carry out, such as can be using the segmentation methods based on string matching, the segmentation methods based on understanding or based on system
The segmentation methods etc. of meter, herein without specifically limiting.
Textual classification model is completed after training according to the sample data of original knowledge point, is being constructed by sample data
The database of text disaggregated model includes the dictionary according to constructed by sample data in the database, includes in dictionary
Semantic weight corresponding to the corresponding coding of word included in sample data and word.
Semantic weight corresponding to word is for being characterized in sample text, percentage contribution of the word to sample text semanteme.
Such as in " which process for handling safety car owner card has " this text, the word segmentation result obtained according to step S131 is " to handle ^
The ^ process ^ that safety ^ car owner blocks ^ have ^ which ", " " " having " " which " these three words to the semantic percentage contribution of the text not
Greatly, thus three words it is corresponding in the text semantic weight it is smaller, and " handling " " safety " " car owner's card " " process "
This four words are bigger to the percentage contribution of text semanteme, so that semantic weight is also opposite in the sample text where four words
It is larger.Certainly the corresponding coding of each word and corresponding semantic weight are determining after training, in this application i.e. basis
It is determined after the training of the sample data completion extraction of semantics layer of original knowledge point.Certainly, sample data volume is bigger, and dictionary is completeer
Kind, the semantic weight of coding corresponding to word and word is also more perfect in dictionary, thus the feature of extraction of semantics layer building text
The function of vector is also more perfect.
After the word segmentation is completed, which can be constructed according to the corresponding coding of each word and the corresponding semantic weight of each word
The feature vector of this text.In a particular embodiment, generally with coding corresponding to digital representation word, and with real number representation word
Corresponding weight, so that the feature vector of constructed input text is real vector.
Step S150 carries out classification layer according to the corresponding mark label of the feature vector of sample text and sample text
Training is updated, to realize the update training of textual classification model.
The parameter of training adjustment classification layer i.e. in updating training process is updated to classification layer.According to sample text
Feature vector and sample text it is corresponding mark label carry out classification layer update training after, textual classification model can be with needle
Tag along sort corresponding to the relevant text output text in newly-increased knowledge point realizes the update instruction of textual classification model
Practice.
By the technical solution of the application, sufficiently instructed in textual classification model according to the sample data of original knowledge point
On the basis of white silk, the perfect in shape and function of the feature vector of extraction of semantics layer building text.It is needing to carry out more textual classification model
When new training, the update training of classification layer is only carried out, and extracts the feature vector of layer building sample text using primitive justice, without
The update training for carrying out primitive justice extract layer realizes text so as to substantially shorten the update training time of textual classification model
This disaggregated model timely updates.
Especially in the customer service robot of field of artificial intelligence, before customer service robot carries out online service, customer service
The sample data of the textual classification model of robot hundreds of thousands item easily, sample data volume is big, and the training time is long, text classification mould
The function of the feature vector of the extraction of semantics layer building text of type is very perfect.To be updated in textual classification model needs
When training, the update training of classification layer is only carried out, and without the update of extraction of semantics layer training, text classification is greatly shortened
The update training time of model, and ensure that textual classification model is to original knowledge point, newly-increased knowledge point after updating training
The nicety of grading of related text.It is especially less relative to the amount of original knowledge point in newly-increased knowledge point, and need to carry out text
When the update training of this disaggregated model, by the technical solution of the application, timely updating for textual classification model may be implemented, and
And it is also ensured that textual classification model nicety of grading.
In one exemplary embodiment, as shown in figure 5, before step S150 further include:
Step S010, according to the tag along sort of the corresponding mark label supplementary classification layer of sample text.
Step S030 updates the tag along sort set of classification layer according to the tag along sort supplemented.
It include the classification exportable all classification label of layer in tag along sort set.One mark label corresponds to classification
One tag along sort of layer, at newly-increased knowledge point, due to not including the sample text of newly-increased knowledge point in original knowledge point, when
So also can not the sample text to newly-increased knowledge point correctly classified.According to the corresponding mark label supplementary classification of sample text
Layer tag along sort, and update classification layer tag along sort set after, thus according to sample text carry out classification layer update
When training, the tag along sort of sample text can be determined from updated tag along sort set.
In one exemplary embodiment, as shown in fig. 6, step S150 includes:
Step S151 predicts to obtain classification corresponding to sample text according to the feature vector of sample text using classification layer
Label.
Step S152, the consistency for carrying out obtained tag along sort and the corresponding mark label of sample text judge.
Step S153, if inconsistent, the parameter of adjustment classification layer is until obtained tag along sort and mark label one
It causes.
The training of textual classification model adjusts the parameter of textual classification model in the training process, makes textual classification model
The tag along sort of output is consistent with the mark label being manually labeled.If the two is consistent, adjustment text classification mould is not needed
The parameter of type, if it is inconsistent, the parameter of adjustment textual classification model is until the two is consistent.In the technical solution of the application
In, when updating training by textual classification model, the parameter of adjustment classification layer makes the tag along sort and mark label of sample text
Unanimously.
In a particular embodiment, if passing through the tag along sort and the corresponding mark label one of sample text that layer obtains of classifying
It causes, then carries out the update training of textual classification model with next sample text.
In the prior art, the either first training or update training of textual classification model, is all extraction of semantics layer
It is trained with classification layer, i.e., in training process, if the tag along sort and sample text of classification the exported sample text of layer
Mark label it is inconsistent, then adjust extraction of semantics layer and classify layer parameter, until the two is consistent.
Since in textual classification model, the neural network structure of extraction of semantics layer is more complicated, calculating process is more complicated, transports
Calculation amount is bigger, and after the parameter for having adjusted extraction of semantics layer, extraction of semantics layer is needed according to parameter adjusted again by transporting
The feature vector for calculating building text, so the time of training text disaggregated model is long.
And in this application, only the parameter of adjustment classification layer, is equivalent to and is only updated training to classification layer, so as to
Substantially shorten the time that textual classification model updates training.
In actual test, tested using four public data collection, ag_news, Dbpedia, Yahoo！Answer and
Safety bank FAQ knowledge base.By Experimental comparison on four data sets, instructed using the update of the textual classification model of the application
Training time used in white silk method shorten to full textual classification model re -training and spends time taking 1/10.
In one exemplary embodiment, as shown in fig. 7, step S151 includes:
Step S210 predicts to obtain feature vector corresponding to updated tag along sort according to feature vector using classification layer
The probability of each tag along sort in set.
Step S230 traverses the probability of each tag along sort, using tag along sort corresponding to most probable value as sample
The corresponding tag along sort of text.
In one exemplary embodiment, as shown in figure 8, after step S150 further include:
Step S171 classifies to several test samples by the updated textual classification model.
The textual classification model after updating training is calculated to several surveys according to classification results in step S172
The nicety of grading of sample sheet.
Step S173 terminates the update training of textual classification model if nicety of grading reaches designated precision.
Wherein step S171-173 is used to test the nicety of grading of textual classification model after update training.Wherein test sample
It may include the relevant text of the relevant text in original knowledge point and/or newly-increased knowledge point, preferably include the text of original knowledge point
The text of this and newly-increased knowledge point.And test sample is labeled.In step S172, by textual classification model to each survey
The tag along sort of this output of sample and the mark of each test sample compare, if the two is consistent, then it is assumed that and classification is accurate,
If it is inconsistent, thinking classification error, the ratio that accurate test sample quantity of classifying accounts for total test sample, the ratio are calculated
Nicety of grading of the as updated textual classification model to several test samples.
If nicety of grading reaches designated precision, terminate the update training of textual classification model, if nicety of grading is not
Reach designated precision, then the update that repeatedly step S110, S130, S150 continues textual classification model is trained.
Following is embodiment of the present disclosure, can be used for executing the text classification that the above-mentioned server 200 of the disclosure executes
The update training method embodiment of model.For those undisclosed details in the apparatus embodiments, the disclosure is please referred to
The update training method embodiment of disaggregated model.
Fig. 9 is a kind of block diagram of the update training device of textual classification model shown according to an exemplary embodiment, should
The update training of textual classification model can be used in the server 200 of implementation environment shown in Fig. 1, execute any of the above embodiment
In textual classification model update training method all or part of step.As shown in figure 9, text disaggregated model is more
New training device includes but is not limited to: it obtains module 110, feature vector building module 130 and updates training module 150,
Middle text disaggregated model includes extraction of semantics layer and classification layer, before newly-increased knowledge point, according to the sample number of original knowledge point
According to the training for completing extraction of semantics layer and layer of classifying, which includes:
Obtain module 110, be configured as executing: obtain the newly-increased corresponding sample text in knowledge point and to sample text into
The mark label of rower note.
Feature vector constructs module 130, which connect with module 110 is obtained, be configured as executing: by according to sample
Data complete the feature vector of the extraction of semantics layer building sample text of training.
Training module 150 is updated, which connect with feature vector building module 130, be configured as executing: according to sample
The corresponding mark label of the feature vector and sample text of text carries out the update training of classification layer, to realize text classification mould
The update training of type.
In one embodiment, feature vector building module 130 includes:
Participle unit is configured as executing: the extraction of semantics by completing training according to the sample data of original knowledge point
Layer segments sample text.
Feature vector construction unit is configured as executing: according to the corresponding coding of each word in sample text and often
The feature vector of the semantic weight building sample text of a word.
In one embodiment, the update training device of textual classification model further include:
Tag along sort complementary module is configured as executing: according to the corresponding mark label supplementary classification layer of sample text
Tag along sort.
Tag along sort set update module, is configured as executing: point of classification layer is updated according to the tag along sort supplemented
Class tag set.
In one embodiment, updating training module 150 includes:
Tag along sort predicting unit is configured as executing: being measured in advance using classification layer according to the feature vector of sample text
To tag along sort corresponding to sample text, updated tag along sort set includes tag along sort corresponding to sample text.
Judging unit is configured as executing: carrying out mark label corresponding to obtained tag along sort and sample text
Consistency judgement.
Adjustment unit is configured as executing: if inconsistent, the parameter of adjustment classification layer is until obtained tag along sort
It is consistent with mark label.
In one embodiment, tag along sort predicting unit includes:
Probability prediction unit is configured as executing: being predicted to obtain feature vector correspondence according to feature vector using classification layer
The probability of each tag along sort in updated tag along sort set.
Tag along sort determination unit is configured as executing: traversing the probability of each tag along sort, institute is right with most probable value
The tag along sort answered is as the corresponding tag along sort of sample text.
In one embodiment, the update training device of textual classification model further include:
Class test module is configured as executing: by the updated textual classification model to several test samples
Classify.
Nicety of grading computing module is configured as executing: the text after updating training is calculated according to classification results
Nicety of grading of this disaggregated model to several test samples.
Training ending module is updated, is configured as executing: if nicety of grading reaches designated precision, terminating disaggregated model
Update training.
Modules/unit function and the realization process of effect are specifically detailed in above-mentioned textual classification model in above-mentioned apparatus
Update training method in correspond to the realization process of step, details are not described herein.
It is appreciated that these module/units can by hardware, software, or a combination of both realize.When in hardware
When realization, these modules may be embodied as one or more hardware modules, such as one or more specific integrated circuits.When with soft
When part mode is realized, these modules may be embodied as the one or more computer journeys executed on the one or more processors
Sequence, such as the program being stored in performed by the central processing unit 270 of Fig. 2 in memory 250.
Optionally, the disclosure also provides a kind of update training equipment of textual classification model, and text sorting device can be with
It is the server 200 of implementation environment shown in Fig. 1, executes the whole in the update training method embodiment of the above textual classification model
Or part steps.As shown in Figure 10, the update training equipment of text disaggregated model includes:
Processor 1001；
Memory 1002 for 1001 executable instruction of storage processor；
Wherein, processor 1001 is configured as executing in the update training method any embodiment of the above textual classification model
All or part of step, executable instruction can be computer-readable instruction, and processor 1001 when being executed, can pass through
Communication bus/data line 1003 reads computer-readable instruction from memory 1002.
The processor of equipment in the embodiment executes the concrete mode of operation in related text disaggregated model
It updates in the embodiment of training method and performs detailed description, no detailed explanation will be given here.
In the exemplary embodiment, a kind of computer readable storage medium is additionally provided, such as can be includes instruction
Provisional and non-transitorycomputer readable storage medium.Storage Jie refers to the memory 250 that can be including instruction, above-mentioned finger
Enabling can be executed by the central processing unit 270 of server 200 to complete the update training method of above-mentioned textual classification model.
It should be understood that the present invention is not limited to the precise structure already described above and shown in the accompanying drawings, and
And various modifications and change can executed without departing from the scope.The scope of the present invention is limited only by the attached claims.
Claims (10)
1. a kind of update training method of textual classification model, the textual classification model includes extraction of semantics layer and classification layer,
Before newly-increased knowledge point, the training of the extraction of semantics layer and the classification layer is completed according to the sample data of original knowledge point,
It is characterised by comprising:
The mark label for obtaining the newly-increased corresponding sample text in knowledge point and the sample text being labeled；
By the feature vector for completing sample text described in the trained extraction of semantics layer building according to the sample data；
The classification layer is carried out according to the corresponding mark label of the feature vector of the sample text and the sample text
Training is updated, to realize the update training of the textual classification model.
2. the method according to claim 1, wherein the institute by completing training according to the sample data
Predicate justice extracts the feature vector of sample text described in layer building, comprising:
The sample text is divided by the extraction of semantics layer for completing training according to the sample data of original knowledge point
Word；
The sample text is constructed according to the semantic weight of the corresponding coding of each word in the sample text and each word
Feature vector.
3. the method according to claim 1, wherein the feature vector and institute according to the sample text
The update training that the corresponding mark label of sample text carries out the classification layer is stated, to realize the update of the textual classification model
Before training, further includes:
The tag along sort of the classification layer is supplemented according to the corresponding mark label of the sample text；
The tag along sort set of the classification layer is updated according to the tag along sort supplemented.
4. according to the method described in claim 3, it is characterized in that, the feature vector and institute according to the sample text
The update training that the corresponding mark label of sample text carries out the classification layer is stated, to realize the update of the textual classification model
Training, comprising:
It is predicted to obtain classification corresponding to the sample text according to the feature vector of the sample text using the classification layer
Label；
The consistency for carrying out the obtained tag along sort and the corresponding mark label of the sample text judges；
If inconsistent, the parameter of the classification layer is adjusted until the obtained tag along sort and the mark label one
It causes.
5. according to the method described in claim 4, it is characterized in that, it is described using the classification layer according to the sample text
Feature vector is predicted to obtain tag along sort corresponding to the sample text, comprising:
Predict to obtain described eigenvector corresponding to the updated classification according to described eigenvector using the classification layer
The probability of each tag along sort in tag set；
The probability for traversing each tag along sort, using tag along sort corresponding to most probable value as the sample text pair
The tag along sort answered.
6. the method according to claim 1, wherein according to the feature vector of the newly-increased sample and described new
Increase the corresponding mark of sample and carry out the update of the classification layer after training, further includes:
Classified by the updated textual classification model to several test samples；
Classification of the textual classification model to several test samples after updating training is calculated according to classification results
Precision；
If the nicety of grading reaches designated precision, terminate the update training of the textual classification model.
7. a kind of update training device of textual classification model, the textual classification model includes extraction of semantics layer and classification layer,
Before newly-increased knowledge point, the training of the extraction of semantics layer and the classification layer is completed according to the sample data of original knowledge point,
It is characterised by comprising:
Module is obtained, is configured as executing: obtaining the newly-increased corresponding sample text in knowledge point and the sample text is carried out
The mark label of mark；
Feature vector constructs module, is configured as executing: the extraction of semantics by completing training according to the sample data
The feature vector of sample text described in layer building；
Training module is updated, is configured as executing: is corresponding according to the feature vector of the sample text and the sample text
Mark label carry out it is described classification layer update training, with realize the textual classification model update training.
8. device according to claim 7, which is characterized in that described eigenvector constructs module and includes:
Participle unit is configured as executing: the extraction of semantics by completing training according to the sample data of original knowledge point
Layer segments the sample text；
Feature vector construction unit is configured as executing: according to the corresponding coding of each word in the sample text and often
The semantic weight of a word constructs the feature vector of the sample text.
9. a kind of update training equipment of textual classification model characterized by comprising
Processor；
Memory for storage processor executable instruction；
Wherein, the processor is configured to executing such as method described in any one of claims 1 to 6.
10. a kind of computer readable storage medium, is stored thereon with computer program, which is characterized in that the computer program
Such as method described in any one of claims 1 to 6 is realized when being executed by processor.
Priority Applications (2)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811192187.4A CN109241288A (en) | 2018-10-12 | 2018-10-12 | Update training method, device and the equipment of textual classification model |
| PCT/CN2018/125250 WO2020073531A1 (en) | 2018-10-12 | 2018-12-29 | Update training method, apparatus and device for text classification model |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811192187.4A CN109241288A (en) | 2018-10-12 | 2018-10-12 | Update training method, device and the equipment of textual classification model |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN109241288A true CN109241288A (en) | 2019-01-18 |
Family
ID=65052732
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201811192187.4A Withdrawn CN109241288A (en) | 2018-10-12 | 2018-10-12 | Update training method, device and the equipment of textual classification model |
Country Status (2)
| Country | Link |
|---|---|
| CN (1) | CN109241288A (en) |
| WO (1) | WO2020073531A1 (en) |
Cited By (10)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109994103A (en) * | 2019-03-26 | 2019-07-09 | 北京博瑞彤芸文化传播股份有限公司 | A kind of training method of intelligent semantic Matching Model |
| CN110472665A (en) * | 2019-07-17 | 2019-11-19 | 新华三大数据技术有限公司 | Model training method, file classification method and relevant apparatus |
| CN110633476A (en) * | 2019-09-27 | 2019-12-31 | 北京百度网讯科技有限公司 | Method and device for acquiring knowledge annotation information |
| CN110717023A (en) * | 2019-09-18 | 2020-01-21 | 平安科技（深圳）有限公司 | Method and device for classifying interview answer texts, electronic equipment and storage medium |
| CN110851546A (en) * | 2019-09-23 | 2020-02-28 | 京东数字科技控股有限公司 | Verification method, model training method, model sharing method, system and medium |
| CN111522570A (en) * | 2020-06-19 | 2020-08-11 | 杭州海康威视数字技术股份有限公司 | Target library updating method and device, electronic equipment and machine-readable storage medium |
| CN111737472A (en) * | 2020-07-01 | 2020-10-02 | 携程计算机技术（上海）有限公司 | Method and system for updating text classification model, electronic device and storage medium |
| CN112148874A (en) * | 2020-07-07 | 2020-12-29 | 四川长虹电器股份有限公司 | Intention identification method and system capable of automatically adding potential intention of user |
| WO2021119949A1 (en) * | 2019-12-16 | 2021-06-24 | 深圳市欢太科技有限公司 | Text classification model training method, text classification method and apparatus, and electronic device |
| CN116881464A (en) * | 2023-09-06 | 2023-10-13 | 北京睿企信息科技有限公司 | Method for model training based on newly added label and storage medium |
Families Citing this family (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN115186780B (en) * | 2022-09-14 | 2022-12-06 | 江西风向标智能科技有限公司 | Discipline knowledge point classification model training method, system, storage medium and equipment |
Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108009589A (en) * | 2017-12-12 | 2018-05-08 | 腾讯科技（深圳）有限公司 | Sample data processing method, device and computer-readable recording medium |
| CN108062331A (en) * | 2016-11-08 | 2018-05-22 | 南京理工大学 | Increment type naive Bayesian file classification method based on Lifelong Learning |
| CN108090178A (en) * | 2017-12-15 | 2018-05-29 | 北京锐安科技有限公司 | A kind of text data analysis method, device, server and storage medium |
| CN108520030A (en) * | 2018-03-27 | 2018-09-11 | 深圳中兴网信科技有限公司 | File classification method, Text Classification System and computer installation |
Family Cites Families (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN104166706B (en) * | 2014-08-08 | 2017-11-03 | 苏州大学 | Multi-tag grader construction method based on cost-sensitive Active Learning |
| US9858262B2 (en) * | 2014-09-17 | 2018-01-02 | International Business Machines Corporation | Information handling system and computer program product for identifying verifiable statements in text |
| CN108509484B (en) * | 2018-01-31 | 2022-03-11 | 腾讯科技（深圳）有限公司 | Classifier construction and intelligent question and answer method, device, terminal and readable storage medium |
-
2018
- 2018-10-12 CN CN201811192187.4A patent/CN109241288A/en not_active Withdrawn
- 2018-12-29 WO PCT/CN2018/125250 patent/WO2020073531A1/en active Application Filing
Patent Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108062331A (en) * | 2016-11-08 | 2018-05-22 | 南京理工大学 | Increment type naive Bayesian file classification method based on Lifelong Learning |
| CN108009589A (en) * | 2017-12-12 | 2018-05-08 | 腾讯科技（深圳）有限公司 | Sample data processing method, device and computer-readable recording medium |
| CN108090178A (en) * | 2017-12-15 | 2018-05-29 | 北京锐安科技有限公司 | A kind of text data analysis method, device, server and storage medium |
| CN108520030A (en) * | 2018-03-27 | 2018-09-11 | 深圳中兴网信科技有限公司 | File classification method, Text Classification System and computer installation |
Cited By (15)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109994103A (en) * | 2019-03-26 | 2019-07-09 | 北京博瑞彤芸文化传播股份有限公司 | A kind of training method of intelligent semantic Matching Model |
| CN110472665A (en) * | 2019-07-17 | 2019-11-19 | 新华三大数据技术有限公司 | Model training method, file classification method and relevant apparatus |
| CN110717023B (en) * | 2019-09-18 | 2023-11-07 | 平安科技（深圳）有限公司 | Method and device for classifying interview answer text, electronic equipment and storage medium |
| CN110717023A (en) * | 2019-09-18 | 2020-01-21 | 平安科技（深圳）有限公司 | Method and device for classifying interview answer texts, electronic equipment and storage medium |
| CN110851546B (en) * | 2019-09-23 | 2021-06-29 | 京东数字科技控股有限公司 | Verification method, model training method, model sharing method, system and medium |
| CN110851546A (en) * | 2019-09-23 | 2020-02-28 | 京东数字科技控股有限公司 | Verification method, model training method, model sharing method, system and medium |
| CN110633476A (en) * | 2019-09-27 | 2019-12-31 | 北京百度网讯科技有限公司 | Method and device for acquiring knowledge annotation information |
| CN110633476B (en) * | 2019-09-27 | 2024-04-05 | 北京百度网讯科技有限公司 | Method and device for acquiring knowledge annotation information |
| WO2021119949A1 (en) * | 2019-12-16 | 2021-06-24 | 深圳市欢太科技有限公司 | Text classification model training method, text classification method and apparatus, and electronic device |
| CN111522570A (en) * | 2020-06-19 | 2020-08-11 | 杭州海康威视数字技术股份有限公司 | Target library updating method and device, electronic equipment and machine-readable storage medium |
| CN111522570B (en) * | 2020-06-19 | 2023-09-05 | 杭州海康威视数字技术股份有限公司 | Target library updating method and device, electronic equipment and machine-readable storage medium |
| CN111737472A (en) * | 2020-07-01 | 2020-10-02 | 携程计算机技术（上海）有限公司 | Method and system for updating text classification model, electronic device and storage medium |
| CN112148874A (en) * | 2020-07-07 | 2020-12-29 | 四川长虹电器股份有限公司 | Intention identification method and system capable of automatically adding potential intention of user |
| CN116881464A (en) * | 2023-09-06 | 2023-10-13 | 北京睿企信息科技有限公司 | Method for model training based on newly added label and storage medium |
| CN116881464B (en) * | 2023-09-06 | 2023-11-24 | 北京睿企信息科技有限公司 | Method for model training based on newly added label and storage medium |
Also Published As
| Publication number | Publication date |
|---|---|
| WO2020073531A1 (en) | 2020-04-16 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109241288A (en) | Update training method, device and the equipment of textual classification model | |
| US11720822B2 (en) | Gradient-based auto-tuning for machine learning and deep learning models | |
| US11488055B2 (en) | Training corpus refinement and incremental updating | |
| CN103679190B (en) | Sorter, classification method and electronic equipment | |
| CN111639516B (en) | Analysis platform based on machine learning | |
| CN111602148A (en) | Regularized neural network architecture search | |
| US11379718B2 (en) | Ground truth quality for machine learning models | |
| CN110598869B (en) | Classification method and device based on sequence model and electronic equipment | |
| US11550707B2 (en) | Systems and methods for generating and executing a test case plan for a software product | |
| CN113204938B (en) | Time delay characteristic improvement method and device of integrated circuit and storage medium | |
| CN111259647A (en) | Question and answer text matching method, device, medium and electronic equipment based on artificial intelligence | |
| CN112420125A (en) | Molecular attribute prediction method and device, intelligent equipment and terminal | |
| CN114144770A (en) | System and method for generating data sets for model retraining | |
| US20230004486A1 (en) | Software testing using machine learning | |
| CN113947086A (en) | Sample data generation method, training method, corpus generation method and apparatus | |
| CN111723222A (en) | Image search and training system | |
| CN113312899A (en) | Text classification method and device and electronic equipment | |
| CN113515625A (en) | Test result classification model training method, classification method and device | |
| CN109376243A (en) | File classification method and device | |
| Mat Jizat et al. | Evaluation of the transfer learning models in wafer defects classification | |
| CN108733702B (en) | Method, device, electronic equipment and medium for extracting upper and lower relation of user query | |
| US20220292393A1 (en) | Utilizing machine learning models to generate initiative plans | |
| CN113609018A (en) | Test method, training method, device, apparatus, medium, and program product | |
| CN106156077A (en) | The method and apparatus selected for mixed model | |
| US11609936B2 (en) | Graph data processing method, device, and computer program product |
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 | ||
| WW01 | Invention patent application withdrawn after publication | ||
| WW01 | Invention patent application withdrawn after publication |
Application publication date: 20190118 |