CN109710757A - Construction method, system and the computer readable storage medium of textual classification model - Google Patents
Construction method, system and the computer readable storage medium of textual classification model Download PDFInfo
- Publication number
- CN109710757A CN109710757A CN201811440834.9A CN201811440834A CN109710757A CN 109710757 A CN109710757 A CN 109710757A CN 201811440834 A CN201811440834 A CN 201811440834A CN 109710757 A CN109710757 A CN 109710757A
- Authority
- CN
- China
- Prior art keywords
- classification model
- textual classification
- input
- dialog information
- textual
- 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
- 238000013145 classification model Methods 0.000 title claims abstract description 146
- 238000010276 construction Methods 0.000 title claims abstract description 35
- 238000012549 training Methods 0.000 claims abstract description 82
- 238000013527 convolutional neural network Methods 0.000 claims abstract description 77
- 238000000034 method Methods 0.000 claims abstract description 63
- 230000006870 function Effects 0.000 claims description 65
- 239000013598 vector Substances 0.000 claims description 62
- 238000004590 computer program Methods 0.000 claims description 37
- 238000012545 processing Methods 0.000 claims description 34
- 238000004422 calculation algorithm Methods 0.000 claims description 29
- 230000009977 dual effect Effects 0.000 claims description 15
- 238000013507 mapping Methods 0.000 claims description 15
- 238000013519 translation Methods 0.000 claims description 14
- ORILYTVJVMAKLC-UHFFFAOYSA-N Adamantane Natural products C1C(C2)CC3CC1CC2C3 ORILYTVJVMAKLC-UHFFFAOYSA-N 0.000 claims description 11
- 238000013528 artificial neural network Methods 0.000 claims description 11
- 238000011176 pooling Methods 0.000 claims description 10
- 230000004913 activation Effects 0.000 claims description 9
- 238000005457 optimization Methods 0.000 claims description 3
- 235000013399 edible fruits Nutrition 0.000 claims description 2
- 238000003475 lamination Methods 0.000 claims 1
- 230000008569 process Effects 0.000 abstract description 30
- 238000010586 diagram Methods 0.000 description 10
- 230000007717 exclusion Effects 0.000 description 5
- 230000003993 interaction Effects 0.000 description 5
- 238000005070 sampling Methods 0.000 description 5
- 235000007926 Craterellus fallax Nutrition 0.000 description 4
- 240000007175 Datura inoxia Species 0.000 description 4
- 230000006978 adaptation Effects 0.000 description 4
- 230000015572 biosynthetic process Effects 0.000 description 4
- 238000000605 extraction Methods 0.000 description 4
- 210000002569 neuron Anatomy 0.000 description 4
- 230000009467 reduction Effects 0.000 description 4
- 230000008901 benefit Effects 0.000 description 3
- 230000008859 change Effects 0.000 description 3
- 238000005516 engineering process Methods 0.000 description 3
- 230000000694 effects Effects 0.000 description 2
- 239000000463 material Substances 0.000 description 2
- 230000000717 retained effect Effects 0.000 description 2
- 238000004364 calculation method Methods 0.000 description 1
- 238000004140 cleaning Methods 0.000 description 1
- 230000006872 improvement Effects 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000001537 neural effect Effects 0.000 description 1
Landscapes
- Image Analysis (AREA)
Abstract
The invention proposes a kind of construction method of textual classification model, the building system of textual classification model and computer readable storage mediums.Wherein, the construction method of textual classification model includes: acquisition at least three-wheel dialog information;At least three-wheel dialog information is input to convolutional neural networks textual classification model parallel;Convolutional neural networks textual classification model is trained according at least three-wheel dialog information, obtains textual classification model.Convolutional neural networks textual classification model is trained using the dialog information inputted parallel, the information inputted parallel as used in training process has the dialog information of context relation relationship, therefore, obtained training result, which can be realized, carries out text classification in conjunction with context, and then improves the accuracy of text classification.
Description
Technical field
The present invention relates to Text Classification fields, construction method, text in particular to a kind of textual classification model
The building system and computer readable storage medium of this disaggregated model.
Background technique
Multichannel convolutive neural network is applied in field of image processing more, such as applies the image recognition in human-computer interaction, or
Person realizes the quick determination of target in video object picture charge pattern.
In the related technology, what text classification was used is that (a kind of loss function is with softmax function for softmax classifier
Classifier), but softmax function have serious problem be softmax classification output probability (0~1) mutual exclusion.A certain classification
Probability height can allow other class probabilities all very low.Such as in government affairs text classification, a data is both social security classification and levies
Class of service is paid, so being difficult to judge data with single classification.Softmax classifier is originally used for the knowledge of convolutional neural networks image
Not, it needs to export 1000 label (label) and determines classification, however, in the related technology, it is pair that text classification, which is used in the process,
Single data be trained as a result, obtained training pattern cannot be associated place to the information with context relation
Reason, causes accuracy rate lower.
Therefore, a kind of construction method of textual classification model is needed, so that the model that building obtains can be realized context
Association process, and then improve the accuracy of classification.
Summary of the invention
The present invention is directed to solve at least one of the technical problems existing in the prior art or related technologies.
For this purpose, first aspect of the present invention is to propose a kind of construction method of textual classification model.
The second aspect of the invention is to propose a kind of building system of textual classification model.
The third aspect of the invention is to propose a kind of computer readable storage medium.
In view of this, according to an aspect of the present invention, it proposes a kind of construction methods of textual classification model, comprising:
Acquisition at least three-wheel dialog information;At least three-wheel dialog information is input to convolutional neural networks textual classification model parallel;Root
Convolutional neural networks textual classification model is trained according at least three-wheel dialog information, obtains textual classification model.
The construction method of textual classification model provided by the invention, acquisition at least three-wheel dialog information, wherein at least three-wheel
Dialog information can derive from the human-computer interaction scene under real scene, or be the record information of dialogue both sides, such as medical treatment
Dialog information under scene is also possible to people society office staff and handles the talk dialog information of personnel;It will acquire
Dialog information is input to parallel in CNN textual classification model, (CNN, Convolutional Neural Networks convolution mind
Through network), CNN textual classification model is trained using the dialog information inputted parallel, as used in training process
The information inputted parallel has the dialog information of context relation relationship, and therefore, obtained training result can be realized in conjunction with upper
Text classification is hereafter carried out, and then improves the accuracy of text classification.
The construction method of above-mentioned textual classification model according to the present invention, can also have following technical characteristic:
In the above-mentioned technical solutions, it is preferable that at least three-wheel dialog information is inputted parallel by any one following mode
To convolutional neural networks textual classification model: each round will be talked at least in three-wheel dialog information using the mode of word DUAL PROBLEMS OF VECTOR MAPPING
Information is word for word mapped to vector space, generates corresponding first image, will at least corresponding first image of three-wheel dialog information is simultaneously
Row is input to convolutional neural networks textual classification model;At least three-wheel dialog information the second image will be compiled by one-hot coding
It is input to convolutional neural networks textual classification model.
In the technical scheme, each round at least three-wheel dialog information is talked with into letter in the way of word DUAL PROBLEMS OF VECTOR MAPPING
Breath is word for word mapped to vector space, generates at least three the first images, will at least corresponding first image of three-wheel dialog information is simultaneously
Row is input in CNN textual classification model, it is preferable that CNN textual classification model is multichannel convolutive neural network, and at least
Three the first images are input to multichannel convolutive neural network parallel, are trained, can be same in the way of word DUAL PROBLEMS OF VECTOR MAPPING
When at least three-wheel dialog information is mapped, and then accelerate the formation speed of the first image, reduce training sample generation
The time of process wastes;Alternatively, at least three-wheel dialog information the second image will be compiled into using one-hot coding, wherein the second image
Middle storage at least three-wheel dialog information, the second image is input in CNN textual classification model, and then reduce the figure in storage
As quantity, to be trained the management of sample, while the requirement to convolutional neural networks can reduce, be not necessarily to multi-channel mode
The sample input of context relation can be realized.
In any of the above-described technical solution, it is preferable that according at least three-wheel dialog information to convolutional neural networks text point
Class model is trained, and obtains textual classification model, is specifically included: the first image or the second image being input to convolutional layer and carried out
Convolution algorithm, and operation result is inputted in the layer of pond and carries out down-sampled processing using presetting method;By down-sampled processing result
It is input to full articulamentum, is classified by classifier, and classification results are input to optimizer and are optimized, obtains text point
Class model.
In the technical scheme, the first image or the second image are input to convolutional layer and carry out convolution algorithm, by the first figure
The feature extraction of picture or the second image comes out, and the operation result extracted is input in the layer of pond, uses presetting method
Operation result is sampled, to reduce number of samples, and sampled result is input to full articulamentum and is classified and is optimized, with
Model after being trained, and then the textual classification model obtained using above-mentioned steps is used in the training process with upper
The hereafter dialog information of incidence relation, thus, it can be combined when obtained textual classification model can classify to text upper
Hereafter related information is classified, relative to the accurate of the textual classification model raising classification for not combining contextual information training
Property.
In any of the above-described technical solution, it is preferable that presetting method max-pooling.
In the technical scheme, presetting method max-pooling, i.e., to most strong in the characteristic value being input in the layer of pond
Carry out retain, give up other weaker characteristic values, and then guarantee position and the rotational invariance of feature, furthermore it is possible to reduce
The problem of number of parameters of textual classification model, reduction model over-fitting, while input X length can be arranged as regular length
Input, so as to the quantity of determination neuron during network structure.
In any of the above-described technical solution, it is preferable that down-sampled processing result is input to full articulamentum, passes through classifier
Classify, and classification results are input to optimizer optimize and specifically include: down-sampled processing result being input to and is connected entirely
Layer is connect, is classified by sigmoid classifier, operation is iterated according to selected sigmoid loss function, until
The numerical value of sigmoid loss function is minimum.
In the technical scheme, classified using sigmoid classifier, that is, the result for avoiding the occurrence of classification can only be 0-
1 two kinds of situations are avoided the occurrence of as used mutual exclusion situation present in softmax classifier, wherein sigmoid classifier is base
In the convolutional neural networks disaggregated model of sigmoid function, it is iterated operation using sigmoid loss function, until
The numerical value of sigmoid loss function is minimum, and then the superiority and inferiority of obtained model is determined using the minimum value of loss function, works as loss
When the numerical value minimum of function to get to textual classification model reach the optimum state under the classifier, use obtained model
The accuracy classified is higher.
In any of the above-described technical solution, it is preferable that be iterated operation according to selected sigmoid loss function, directly
Numerical value minimum to sigmoid loss function specifically includes:
Operation is iterated to selected sigmoid loss function using Adam improved stochastic gradient descent algorithm,
Until the numerical value of sigmoid loss function is minimum.
In the technical scheme, being iterated in calculating process using stochastic gradient descent algorithm to loss function will loss
Inappropriate math portions are improved to Adam method in function, wherein Adam is that one kind can substitute traditional stochastic gradient descent
The first-order optimization method of process, so that the sigmoid loss function that improved stochastic gradient descent algorithm adaptation is selected, obtains
The numerical value of sigmoid loss function is minimum, improves the accuracy of textual classification model classification results.
In any of the above-described technical solution, it is preferable that word vector is to be instructed in advance using the improved Cove of Chinese-English translation training
It practises handwriting vector.
In the technical scheme, word vector is using the improved Cove pre-training word vector of Chinese-English translation training, wherein
Cove pre-training word vector is that scene vector (context vectors, i.e. Cove) carries out the word vector after pre-training, due in
More contextual informations can be generated during translator of English, therefore word vector is chosen and is selected as Chinese-English translation training in the process
Improved Cove pre-training word vector, and then it is mapped to the contextual information in the first image generated after vector space included more
The accuracy of horn of plenty, finally obtained textual classification model classification is higher.
In any of the above-described technical solution, it is preferable that after down-sampled processing result is input to full articulamentum, logical
It crosses before sigmoid classifier classified, further includes:
Down-sampled processing result will be sequentially input to dropout and relu activation.
In the technical scheme, fine in order to avoid there is the degree of fitting of training set in the training process, and collect in verifying
The middle situation for fitting difference occur, is input to dropout for down-sampled processing result, according to certain during carrying out model training
Probability to network parameter carry out stochastical sampling, the new target network that sub-network is updated as this, so that each iteration
Not will use identical sub-network, so avoid the occurrence of network by overfitting to training set the case where, utilize relu activation energy
The convergence rate for enough accelerating stochastic gradient descent algorithm, reduces the trained time.
In any of the above-described technical solution, it is preferable that in acquisition at least after three-wheel dialog information, will at least three-wheel pair
Words information is input to parallel before convolutional neural networks textual classification model, further includes:
Data cleansing is carried out at least three-wheel dialog information.
In the technical scheme, in order to avoid repeat statement present in dialog information or the punctuation mark of mistake influence
The incidence relation of context carries out data cleansing at least three-wheel dialog information, duplicate sentence or punctuation mark is deleted,
And the dialog information that dialog length is less than three-wheel is spliced to three-wheel and is talked with, it can be tied with ensuring model in the training process
The context for closing training sample is trained, and then improves the classification accuracy for the textual classification model that training obtains.
According to the second aspect of the invention, a kind of building system of textual classification model is proposed, comprising: memory,
For storing computer program;Processor, for executing computer program with acquisition at least three-wheel dialog information;To at least three
Wheel dialog information is input to convolutional neural networks textual classification model parallel;According at least three-wheel dialog information to convolutional Neural net
Network textual classification model is trained, and obtains textual classification model.
The building system of textual classification model provided by the invention, memory store computer program;Processor executes meter
When calculation machine program, acquisition at least three-wheel dialog information, wherein at least three-wheel dialog information can derive from the people under real scene
Machine interaction scenarios, or to talk with the dialog information under the record information of both sides, such as medical scene, it is also possible to office, people society work
Make personnel and handles the talk dialog information of personnel;The dialog information that will acquire is input to CNN textual classification model parallel
In, (CNN, Convolutional Neural Networks convolutional neural networks) utilize the dialog information pair inputted parallel
CNN textual classification model is trained, and the information inputted parallel as used in training process has context relation relationship
Dialog information, therefore, obtained training result can be realized in conjunction with context carry out text classification, and then improve text classification
Accuracy.
The building system of textual classification model according to the present invention, can also have following technical characteristic:
In the above-mentioned technical solutions, it is preferable that at least three-wheel dialog information is inputted parallel by any one following mode
To convolutional neural networks textual classification model: each round will be talked at least in three-wheel dialog information using the mode of word DUAL PROBLEMS OF VECTOR MAPPING
Information is word for word mapped to vector space, generates corresponding first image, will at least to talk with corresponding first image defeated parallel for three-wheel
Enter to convolutional neural networks textual classification model;Will at least three-wheel dialog information by one-hot coding be compiled into the second image input
To convolutional neural networks textual classification model.
In the technical scheme, each round at least three-wheel dialog information is talked with into letter in the way of word DUAL PROBLEMS OF VECTOR MAPPING
Breath is word for word mapped to vector space, generates at least three the first images, will at least corresponding first image of three-wheel dialog information is simultaneously
Row is input in CNN textual classification model, it is preferable that CNN textual classification model is multichannel convolutive neural network, and at least
Three the first images are input to multichannel convolutive neural network parallel, are trained, can be same in the way of word DUAL PROBLEMS OF VECTOR MAPPING
When at least three-wheel dialog information is mapped, and then accelerate the formation speed of the first image, reduce training sample generation
The time of process wastes;Alternatively, at least three-wheel dialog information the second image will be compiled into using one-hot coding, wherein the second image
Middle storage at least three-wheel dialog information, the second image is input in CNN textual classification model, and then reduce the figure in storage
As quantity, to be trained the management of sample, while the requirement to convolutional neural networks can reduce, be not necessarily to multi-channel mode
The sample input of context relation can be realized.
In any of the above-described technical solution, it is preferable that be specifically used for execute computer program with: by the first image or second
Image is input to convolutional layer and carries out convolution algorithm, and operation result is inputted in the layer of pond and carries out down-sampled place using presetting method
Reason;
Down-sampled processing result is input to full articulamentum, is classified by classifier, and classification results are input to
Optimizer optimizes, and obtains textual classification model.
In the technical scheme, the first image or the second image are input to convolutional layer and carry out convolution algorithm, by the first figure
The feature extraction of picture or the second image comes out, and the operation result extracted is input in the layer of pond, uses presetting method
Operation result is sampled, to reduce number of samples, and sampled result is input to full articulamentum and is classified and is optimized, with
Model after being trained, and then the textual classification model obtained using above-mentioned steps is used in the training process with upper
The hereafter dialog information of incidence relation, thus, it can be combined when obtained textual classification model can classify to text upper
Hereafter related information is classified, relative to the accurate of the textual classification model raising classification for not combining contextual information training
Property.
In any of the above-described technical solution, it is preferable that presetting method max-pooling.
In the technical scheme, presetting method max-pooling, i.e., to most strong in the characteristic value being input in the layer of pond
Carry out retain, give up other weaker characteristic values, and then guarantee position and the rotational invariance of feature, furthermore it is possible to reduce
The problem of number of parameters of textual classification model, reduction model over-fitting, while input X length can be arranged as regular length
Input, so as to the quantity of determination neuron during network structure.
In any of the above-described technical solution, it is preferable that processor, be specifically used for execute computer program with: will be down-sampled
Processing result is input to full articulamentum, is classified by sigmoid classifier, according to selected sigmoid loss function into
Row iteration operation, until the numerical value of sigmoid loss function is minimum.
In the technical scheme, classified using sigmoid classifier, that is, the result for avoiding the occurrence of classification can only be 0-
1 two kinds of situations are avoided the occurrence of as used mutual exclusion situation present in softmax classifier, are carried out using sigmoid loss function
Interative computation until the numerical value of sigmoid loss function is minimum, and then determines obtained model using the minimum value of loss function
Superiority and inferiority, when the numerical value minimum of loss function to get to textual classification model reach the optimum state under the classifier, make
The accuracy classified with obtained model is higher.
In any of the above-described technical solution, it is preferable that processor, be specifically used for execute computer program with: use Adam
Improved stochastic gradient descent algorithm is iterated operation to selected sigmoid loss function, until sigmoid loses letter
Several numerical value is minimum.
In the technical scheme, being iterated in calculating process using stochastic gradient descent algorithm to loss function will loss
Inappropriate math portions are improved to Adam method in function, so that improved stochastic gradient descent algorithm adaptation was selected
Sigmoid loss function, the numerical value for obtaining sigmoid loss function is minimum, improves the accurate of textual classification model classification results
Property.
In any of the above-described technical solution, it is preferable that word vector is to be instructed in advance using the improved Cove of Chinese-English translation training
It practises handwriting vector.
In the technical scheme, word vector is to train improved Cove pre-training word vector using Chinese-English translation, due to
More contextual informations can be generated during Chinese-English translation, therefore word vector is chosen and is selected as Chinese-English translation instruction in the process
Practice improved Cove pre-training word vector, and then is mapped to the contextual information in the first image generated after vector space included
The accuracy of more horn of plenty, finally obtained textual classification model classification is higher.
In any of the above-described technical solution, it is preferable that processor, be also used to execute computer program with: will be down-sampled
Processing result sequentially inputs dropout and relu activation.
In the technical scheme, fine in order to avoid there is the degree of fitting of training set in the training process, and collect in verifying
The middle situation for fitting difference occur, is input to dropout for down-sampled processing result, according to certain during carrying out model training
Probability to network parameter carry out stochastical sampling, the new target network that sub-network is updated as this, so that each iteration
Not will use identical sub-network, so avoid the occurrence of network by overfitting to training set the case where, utilize relu activation energy
The convergence rate for enough accelerating stochastic gradient descent algorithm, reduces the trained time.
In any of the above-described technical solution, it is preferable that processor, be also used to execute computer program with: at least three-wheel
Dialog information carries out data cleansing.
In the technical scheme, in order to avoid repeat statement present in dialog information or the punctuation mark of mistake influence
The incidence relation of context carries out data cleansing at least three-wheel dialog information, duplicate sentence or punctuation mark is deleted,
And the dialog information that dialog length is less than three-wheel is spliced to three-wheel and is talked with, it can be tied with ensuring model in the training process
The context for closing training sample is trained, and then improves the classification accuracy for the textual classification model that training obtains.
According to the third aspect of the present invention, it the present invention provides a kind of computer readable storage medium, is stored thereon with
Computer program realizes the building side of textual classification model in any of the above-described technical solution when computer program is executed by processor
The step of method.
A kind of computer readable storage medium provided by the invention is stored thereon with computer program, computer program quilt
The step of processor realizes the construction method of textual classification model in any of the above-described technical solution when executing, therefore there is the text
Whole technical effects of the construction method of disaggregated model, details are not described herein.
Additional aspect and advantage of the invention will become obviously in following description section, or practice through the invention
Recognize.
Detailed description of the invention
Above-mentioned and/or additional aspect of the invention and advantage will become from the description of the embodiment in conjunction with the following figures
Obviously and it is readily appreciated that, in which:
Fig. 1 shows the flow diagram of the construction method of the textual classification model of one embodiment of the present of invention;
Fig. 2 shows the flow diagrams of the construction method of the textual classification model of another embodiment of the invention;
Fig. 3 shows the flow diagram of the construction method of the textual classification model of yet another embodiment of the present invention;
Fig. 4 shows the flow diagram of the construction method of the textual classification model of another embodiment of the invention;
Fig. 5 shows the schematic block diagram of the building system of the textual classification model of one embodiment of the present of invention.
Specific embodiment
It is with reference to the accompanying drawing and specific real in order to be more clearly understood that aforementioned aspect of the present invention, feature and advantage
Applying mode, the present invention is further described in detail.It should be noted that in the absence of conflict, the implementation of the application
Feature in example and embodiment can be combined with each other.
In the following description, numerous specific details are set forth in order to facilitate a full understanding of the present invention, still, the present invention may be used also
To be implemented using other than the one described here other modes, therefore, protection scope of the present invention is not limited to following public affairs
The limitation for the specific embodiment opened.
The embodiment of first aspect present invention proposes that a kind of construction method of textual classification model, Fig. 1 show the present invention
One embodiment textual classification model construction method flow diagram.As shown in Figure 1, this method comprises:
S102, acquisition at least three-wheel dialog information;
S104 at least three-wheel dialog information will be input to convolutional neural networks textual classification model parallel;
S106 is trained convolutional neural networks textual classification model according at least three-wheel dialog information, obtains text
Disaggregated model.
The construction method of textual classification model provided by the invention, acquisition at least three-wheel dialog information, wherein at least three-wheel
Dialog information can derive from the human-computer interaction scene under real scene, or be the record information of dialogue both sides, such as medical treatment
Dialog information under scene is also possible to people society office staff and handles the talk dialog information of personnel;It will acquire
Dialog information is input to parallel in CNN textual classification model, (CNN, Convolutional Neural Networks convolution mind
Through network), CNN textual classification model is trained using the dialog information inputted parallel, as used in training process
The information inputted parallel has the dialog information of context relation relationship, and therefore, obtained training result can be realized in conjunction with upper
Text classification is hereafter carried out, and then improves the accuracy of text classification.
It is worth noting that in order to ensure model training as a result, at least three-wheel dialog information of acquisition be default problem with
Problem and answer are input to CNN text text class model and are trained by answer, can also acquire at least three-wheel dialog information training
After, be input to CNN text text class model using default problem and answer and be trained, with improve model complexity and
Fitting degree.
In the above embodiment, it is preferable that in S104 will at least three-wheel dialog information by any one following mode simultaneously
Row is input to convolutional neural networks textual classification model: using word DUAL PROBLEMS OF VECTOR MAPPING mode will at least in three-wheel dialog information it is each
Wheel dialog information is word for word mapped to vector space, generates corresponding first image, will at least three-wheel dialog information corresponding first
Image is input to convolutional neural networks textual classification model parallel;At least three-wheel dialog information by one-hot coding it will be compiled into the
Two images are input to convolutional neural networks textual classification model.
In this embodiment, by each round dialog information at least three-wheel dialog information in the way of word DUAL PROBLEMS OF VECTOR MAPPING
Word for word be mapped to vector space, generate at least three the first images, will at least corresponding first image of three-wheel dialog information it is parallel
It being input in CNN textual classification model, it is preferable that CNN textual classification model is multichannel convolutive neural network, and at least three
It opens the first image and is input to multichannel convolutive neural network parallel, be trained, it can be simultaneously in the way of word DUAL PROBLEMS OF VECTOR MAPPING
At least three-wheel dialog information is mapped, and then accelerates the formation speed of the first image, reduces training sample and generated
The time of journey wastes;Alternatively, at least three-wheel dialog information the second image will be compiled into using one-hot coding, wherein in the second image
Storage at least three-wheel dialog information, the second image is input in CNN textual classification model, and then reduces the image in storage
Quantity to be trained the management of sample, while can reduce the requirement to convolutional neural networks, be without multi-channel mode
The sample input of context relation can be achieved.
Fig. 2 shows the flow diagrams of the construction method of the textual classification model of another embodiment of the invention.Its
In, this method comprises:
S202, acquisition at least three-wheel dialog information;
S204 at least three-wheel dialog information will be input to convolutional neural networks textual classification model parallel;
First image or the second image are input to convolutional layer and carry out convolution algorithm, and operation result is inputted pond by S206
Change in layer and carries out down-sampled processing using presetting method;
Down-sampled processing result is input to full articulamentum, is classified by classifier, and classification results are defeated by S208
Enter to optimizer and optimize, obtains textual classification model.
In this embodiment, the first image or the second image are input to convolutional layer and carry out convolution algorithm, by the first image
Or second the feature extraction of image come out, and the operation result extracted is input in the layer of pond, uses presetting method pair
Operation result is sampled, and to reduce number of samples, and sampled result is input to full articulamentum and is classified and is optimized, with
Model after to training, and then the textual classification model obtained using above-mentioned steps is used in the training process with up and down
The dialog information of literary incidence relation, thus, it can be in conjunction with up and down when obtained textual classification model can classify to text
Literary related information is classified, the accuracy relative to the textual classification model raising classification for not combining contextual information training.
In the above embodiment, it is preferable that presetting method is max-pooling.
In this embodiment, presetting method max-pooling, i.e., to strongest in the characteristic value being input in the layer of pond
Retained, give up other weaker characteristic values, and then guarantee position and the rotational invariance of feature, furthermore it is possible to reduce text
The problem of number of parameters of this disaggregated model, reduction model over-fitting, while input X length can be arranged as regular length
Input, so as to the quantity of determination neuron during network structure.
Fig. 3 shows the flow diagram of the construction method of the textual classification model of yet another embodiment of the present invention.Its
In, this method comprises:
S302, acquisition at least three-wheel dialog information;
S304 at least three-wheel dialog information will be input to convolutional neural networks textual classification model parallel;
First image or the second image are input to convolutional layer and carry out convolution algorithm, and operation result is inputted pond by S306
Change in layer and carries out down-sampled processing using presetting method;
Down-sampled processing result is input to full articulamentum, is classified by sigmoid classifier, according to choosing by S308
Fixed sigmoid loss function is iterated operation, until the numerical value of sigmoid loss function is minimum, obtains text classification mould
Type.
In this embodiment, classified using sigmoid classifier, that is, the result for avoiding the occurrence of classification can only be 0-1
Two kinds of situations are avoided the occurrence of as used mutual exclusion situation present in softmax classifier, are carried out using sigmoid loss function
Interative computation until the numerical value of sigmoid loss function is minimum, and then determines obtained model using the minimum value of loss function
Superiority and inferiority, when the numerical value minimum of loss function to get to textual classification model reach the optimum state under the classifier, make
The accuracy classified with obtained model is higher.
In any of the above-described embodiment, it is preferable that operation is iterated according to selected sigmoid loss function, until
The numerical value minimum of sigmoid loss function specifically includes:
Operation is iterated to selected sigmoid loss function using Adam improved stochastic gradient descent algorithm,
Until the numerical value of sigmoid loss function is minimum.
In this embodiment, letter will be lost by being iterated in calculating process using stochastic gradient descent algorithm to loss function
Inappropriate math portions are improved to Adam method in number, so that improved stochastic gradient descent algorithm adaptation was selected
Sigmoid loss function, the numerical value for obtaining sigmoid loss function is minimum, improves the accurate of textual classification model classification results
Property.
In any of the above-described embodiment, it is preferable that word vector is using the improved Cove pre-training of Chinese-English translation training
Word vector.
In this embodiment, word vector is using the improved Cove pre-training word vector of Chinese-English translation training, due in
More contextual informations can be generated during translator of English, therefore word vector is chosen and is selected as Chinese-English translation training in the process
Improved Cove pre-training word vector, and then it is mapped to the contextual information in the first image generated after vector space included more
The accuracy of horn of plenty, finally obtained textual classification model classification is higher.
In any of the above-described embodiment, it is preferable that after down-sampled processing result is input to full articulamentum, passing through
Before sigmoid classifier is classified, further includes:
Down-sampled processing result will be sequentially input to dropout and relu activation.
In this embodiment, fine in order to avoid there is the degree of fitting of training set in the training process, and concentrated in verifying
There is the situation of fitting difference, down-sampled processing result is input to dropout, according to certain during carrying out model training
Probability carries out stochastical sampling, the new target network that sub-network is updated as this to network parameter, so that each iteration is not
Will use identical sub-network, so avoid the occurrence of network by overfitting to training set the case where, can using relu activation
The convergence rate for accelerating stochastic gradient descent algorithm, reduces the trained time.
In any of the above-described embodiment, it is preferable that after acquisition at least three-wheel dialog information, will at least three-wheel talk with
Information is input to parallel before convolutional neural networks textual classification model, further includes:
Data cleansing is carried out at least three-wheel dialog information.
In this embodiment, in order to avoid repeat statement present in dialog information or the punctuation mark of mistake influence
Incidence relation hereafter carries out data cleansing at least three-wheel dialog information, duplicate sentence or punctuation mark is deleted, and
The dialog information that dialog length is less than three-wheel is spliced to three-wheel and is talked with, can be combined with ensuring model in the training process
The context of training sample is trained, and then improves the classification accuracy for the textual classification model that training obtains.
Fig. 4 shows the flow diagram of the construction method of the textual classification model of another embodiment of the invention.Its
In, this method comprises:
S402 collects data from office, people society and medical hospital's scene, and carries out data cleansing;
S404 is mapped with the word vector of Cove method and is mapped the data after cleaning;
S406, adjustment softmax loss function are sigmoid loss function;
Three images after word DUAL PROBLEMS OF VECTOR MAPPING are input to multichannel convolutive neural network and are trained, obtain text by S408
This disaggregated model.
The embodiment of second aspect of the present invention, proposes a kind of building system 500 of textual classification model, and Fig. 5 shows this
The schematic block diagram of the building system 500 of the textual classification model of one embodiment of invention.As shown in figure 5, textual classification model
Building system 500 include: memory 502, for storing computer program;Processor 504, for executing computer program
With: acquisition at least three-wheel dialog information;At least three-wheel dialog information is input to convolutional neural networks textual classification model parallel;
Convolutional neural networks textual classification model is trained according at least three-wheel dialog information, obtains textual classification model.
The building system 500 of textual classification model provided by the invention, memory 502 store computer program;Processor
When 504 execution computer program, acquisition at least three-wheel dialog information, wherein at least three-wheel dialog information can derive from true field
Human-computer interaction scene under scape, or to talk with the dialog information under the record information of both sides, such as medical scene, be also possible to
People society office staff and the talk dialog information for handling personnel;The dialog information that will acquire is input to CNN text parallel
In disaggregated model, (CNN, Convolutional Neural Networks convolutional neural networks) utilize the dialogue inputted parallel
Information is trained CNN textual classification model, and the information inputted parallel as used in training process is closed with context
The dialog information of connection relationship, therefore, obtained training result, which can be realized, carries out text classification in conjunction with context, and then improves text
The accuracy of this classification.
It is worth noting that in order to ensure model training as a result, at least three-wheel dialog information of acquisition be default problem with
Problem and answer are input to CNN text text class model and are trained by answer, can also acquire at least three-wheel dialog information training
After, be input to CNN text text class model using default problem and answer and be trained, with improve model complexity and
Fitting degree.
In the above embodiment, it is preferable that at least three-wheel dialog information is input to parallel by any one following mode
Convolutional neural networks textual classification model: will each round dialogue letter at least in three-wheel dialog information using the mode of word DUAL PROBLEMS OF VECTOR MAPPING
Breath is word for word mapped to vector space, generates corresponding first image, and at least corresponding first image of three-wheel dialogue is inputted parallel
To convolutional neural networks textual classification model;At least three-wheel dialog information it the second image will be compiled by one-hot coding will be input to
Convolutional neural networks textual classification model.
In this embodiment, by each round dialog information at least three-wheel dialog information in the way of word DUAL PROBLEMS OF VECTOR MAPPING
Word for word be mapped to vector space, generate at least three the first images, will at least corresponding first image of three-wheel dialog information it is parallel
It being input in CNN textual classification model, it is preferable that CNN textual classification model is multichannel convolutive neural network, and at least three
It opens the first image and is input to multichannel convolutive neural network parallel, be trained, it can be simultaneously in the way of word DUAL PROBLEMS OF VECTOR MAPPING
At least three-wheel dialog information is mapped, and then accelerates the formation speed of the first image, reduces training sample and generated
The time of journey wastes;Alternatively, at least three-wheel dialog information the second image will be compiled into using one-hot coding, wherein in the second image
Storage at least three-wheel dialog information, the second image is input in CNN textual classification model, and then reduces the image in storage
Quantity to be trained the management of sample, while can reduce the requirement to convolutional neural networks, be without multi-channel mode
The sample input of context relation can be achieved.
In any of the above-described embodiment, it is preferable that processor 504 specifically in execute computer program with: by the first image
Or second image be input to convolutional layer and carry out convolution algorithm, and operation result is inputted in the layer of pond and is dropped using presetting method
Sampling processing;
Down-sampled processing result is input to full articulamentum, is classified by classifier, and classification results are input to
Optimizer optimizes, and obtains textual classification model.
In this embodiment, processor 504 specifically in execute computer program with: the first image or the second image are inputted
Convolution algorithm, the operation knot that the feature extraction of the first image or the second image is come out, and will be extracted are carried out to convolutional layer
Fruit is input in the layer of pond, is sampled using presetting method to operation result, to reduce number of samples, and sampled result is defeated
Enter to full articulamentum and is classified and optimized, with the model after being trained, and then the text classification obtained using above-mentioned steps
Model uses the dialog information with context relation relationship in the training process, thus, obtained textual classification model
It can classify in conjunction with context relation information when can classify to text, relative to not combining contextual information training
Textual classification model improve classification accuracy.
In any of the above-described embodiment, it is preferable that presetting method max-pooling.
In this embodiment, presetting method max-pooling, i.e., to strongest in the characteristic value being input in the layer of pond
Retained, give up other weaker characteristic values, and then guarantee position and the rotational invariance of feature, furthermore it is possible to reduce text
The problem of number of parameters of this disaggregated model, reduction model over-fitting, while input X length can be arranged as regular length
Input, so as to the quantity of determination neuron during network structure.
In any of the above-described embodiment, it is preferable that processor 504, be specifically used for execute computer program with: will be down-sampled
Processing result is input to full articulamentum, is classified by sigmoid classifier, according to selected sigmoid loss function into
Row iteration operation, until the numerical value of sigmoid loss function is minimum.
In this embodiment, processor 504 be specifically used for execute computer program with: carried out using sigmoid classifier
Classification, that is, the result for avoiding the occurrence of classification can only be two kinds of situations of 0-1, avoid the occurrence of and exist as used in softmax classifier
Mutual exclusion situation, operation is iterated using sigmoid loss function, until the numerical value of sigmoid loss function is minimum, in turn
The superiority and inferiority that obtained model is determined using the minimum value of loss function, when the numerical value minimum of loss function to get the text arrived
Disaggregated model reaches the optimum state under the classifier, and the accuracy classified using obtained model is higher.
In any of the above-described embodiment, it is preferable that processor 504, be specifically used for execute computer program with: use Adam
Improved stochastic gradient descent algorithm is iterated operation to selected sigmoid loss function, until sigmoid loses letter
Several numerical value is minimum.
In the technical scheme, processor 504 be specifically used for execute computer program with: use stochastic gradient descent algorithm
Loss function is iterated in calculating process, math portions inappropriate in loss function are improved to Adam method, so as to change
The sigmoid loss function that stochastic gradient descent algorithm adaptation after is selected, obtains the numerical value of sigmoid loss function most
It is small, improve the accuracy of textual classification model classification results.
In any of the above-described embodiment, it is preferable that word vector is using the improved Cove pre-training of Chinese-English translation training
Word vector.
In this embodiment, word vector is using the improved Cove pre-training word vector of Chinese-English translation training, due in
More contextual informations can be generated during translator of English, therefore word vector is chosen and is selected as Chinese-English translation training in the process
Improved Cove pre-training word vector, and then it is mapped to the contextual information in the first image generated after vector space included more
The accuracy of horn of plenty, finally obtained textual classification model classification is higher.
In any of the above-described embodiment, it is preferable that processor 504, be also used to execute computer program with: will be down-sampled
Processing result sequentially inputs dropout and relu activation.
In this embodiment, fine in order to avoid there is the degree of fitting of training set in the training process, and concentrated in verifying
There is the situation of fitting difference, down-sampled processing result is input to dropout, according to certain during carrying out model training
Probability carries out stochastical sampling, the new target network that sub-network is updated as this to network parameter, so that each iteration is not
Will use identical sub-network, so avoid the occurrence of network by overfitting to training set the case where, can using relu activation
The convergence rate for accelerating stochastic gradient descent algorithm, reduces the trained time.
In any of the above-described embodiment, it is preferable that processor 504, be also used to execute computer program with: at least three-wheel
Dialog information carries out data cleansing.
In this embodiment, in order to avoid repeat statement present in dialog information or the punctuation mark of mistake influence
Incidence relation hereafter carries out data cleansing at least three-wheel dialog information, duplicate sentence or punctuation mark is deleted, and
The dialog information that dialog length is less than three-wheel is spliced to three-wheel and is talked with, can be combined with ensuring model in the training process
The context of training sample is trained, and then improves the classification accuracy for the textual classification model that training obtains.
The embodiment of third aspect present invention provides a kind of computer readable storage medium, is stored thereon with computer
Program realizes the step of the construction method of textual classification model in any of the above-described embodiment when computer program is executed by processor
Suddenly.
A kind of computer readable storage medium provided by the invention is stored thereon with computer program, computer program quilt
The step of processor realizes the construction method of textual classification model in any of the above-described embodiment when executing, therefore there is this article one's duty
Whole technical effects of the construction method of class model, details are not described herein.
In the description of this specification, the description of term " one embodiment ", " some embodiments ", " specific embodiment " etc.
Mean that particular features, structures, materials, or characteristics described in conjunction with this embodiment or example are contained at least one reality of the invention
It applies in example or example.In the present specification, schematic expression of the above terms are not necessarily referring to identical embodiment or reality
Example.Moreover, description particular features, structures, materials, or characteristics can in any one or more of the embodiments or examples with
Suitable mode combines.
These are only the preferred embodiment of the present invention, is not intended to restrict the invention, for those skilled in the art
For member, the invention may be variously modified and varied.All within the spirits and principles of the present invention, it is made it is any modification,
Equivalent replacement, improvement etc., should all be included in the protection scope of the present invention.
Claims (19)
1. a kind of construction method of textual classification model characterized by comprising
Acquisition at least three-wheel dialog information;
At least three-wheel dialog information is input to convolutional neural networks textual classification model parallel;
The convolutional neural networks textual classification model is trained according at least three-wheel dialog information, obtains the text
This disaggregated model.
2. the construction method of textual classification model according to claim 1, which is characterized in that
At least three-wheel dialog information is input to convolutional neural networks text classification mould by any one following mode parallel
Type:
Each round dialog information in at least three-wheel dialog information is word for word mapped to vector using the mode of word DUAL PROBLEMS OF VECTOR MAPPING
Space generates corresponding first image, corresponding first image of at least three-wheel dialog information is input to the volume parallel
Product neural network textual classification model;
At least three-wheel dialog information is compiled into the second image by one-hot coding and is input to the convolutional neural networks text
This disaggregated model.
3. the construction method of textual classification model according to claim 2, which is characterized in that
Described at least three-wheel dialog information is trained the convolutional neural networks textual classification model according to, obtains institute
Textual classification model is stated, is specifically included:
The first image or second image are input to convolutional layer and carry out convolution algorithm, and operation result is inputted into pond
Down-sampled processing is carried out using presetting method in layer;
Down-sampled processing result is input to full articulamentum, is classified by classifier, and classification results are input to optimization
Device optimizes, and obtains the textual classification model.
4. the construction method of textual classification model according to claim 3, which is characterized in that the presetting method is max-
pooling。
5. the construction method of textual classification model according to claim 3, which is characterized in that described to tie down-sampled processing
Fruit is input to full articulamentum, is classified by classifier, and classification results are input to optimizer optimize and specifically include:
Down-sampled processing result is input to full articulamentum, is classified by sigmoid classifier, according to selected
Sigmoid loss function is iterated operation, until the numerical value of the sigmoid loss function is minimum.
6. the construction method of textual classification model according to claim 5, which is characterized in that the basis was selected
Sigmoid loss function is iterated operation, until the numerical value minimum of the sigmoid loss function specifically includes:
Operation is iterated to the selected sigmoid loss function using Adam improved stochastic gradient descent algorithm,
Until the numerical value of the sigmoid loss function is minimum.
7. the construction method of textual classification model according to claim 2, which is characterized in that the word vector is in using
The improved Cove pre-training word vector of translator of English training.
8. the construction method of textual classification model according to claim 2, which is characterized in that
It is described down-sampled processing result is input to full articulamentum after, classified described by sigmoid classifier
Before, further includes:
Down-sampled processing result is sequentially input into dropout and relu activation by described.
9. the construction method of textual classification model according to claim 1, which is characterized in that talk in acquisition at least three-wheel
After information, it is described at least three-wheel dialog information is input to convolutional neural networks textual classification model parallel before,
Further include:
Data cleansing is carried out at least three-wheel dialog information.
10. a kind of building system of textual classification model characterized by comprising
Memory, for storing computer program;
Processor, for execute the computer program with:
Acquisition at least three-wheel dialog information;
At least three-wheel dialog information is input to convolutional neural networks textual classification model parallel;
The convolutional neural networks textual classification model is trained according at least three-wheel dialog information, obtains the text
This disaggregated model.
11. the building system of textual classification model according to claim 10, which is characterized in that
At least three-wheel dialog information is input to convolutional neural networks text classification mould by any one following mode parallel
Type:
Each round dialog information in at least three-wheel dialog information is word for word mapped to vector using the mode of word DUAL PROBLEMS OF VECTOR MAPPING
Space generates corresponding first image, and at least three-wheel is talked with corresponding first image and is input to the convolution mind parallel
Through Web text classification model;
At least three-wheel dialog information is compiled into the second image by one-hot coding and is input to the convolutional neural networks text
This disaggregated model.
12. the building system of textual classification model according to claim 11, which is characterized in that
The processor, be specifically used for executing the computer program with: the first image or the second image are input to volume
Lamination carries out convolution algorithm, and operation result is inputted in the layer of pond and carries out down-sampled processing using presetting method;
Down-sampled processing result is input to full articulamentum, is classified by classifier, and classification results are input to optimization
Device optimizes, and obtains the textual classification model.
13. the building system of textual classification model according to claim 12, which is characterized in that the presetting method is
max-pooling。
14. the building system of textual classification model according to claim 12, which is characterized in that
The processor, be specifically used for executing the computer program with: down-sampled processing result is input to full articulamentum, is led to
It crosses sigmoid classifier to classify, operation is iterated according to selected sigmoid loss function, until the sigmoid
The numerical value of loss function is minimum.
15. the building system of textual classification model according to claim 14, which is characterized in that
The processor, be specifically used for executing the computer program with: use the improved stochastic gradient descent algorithm of Adam
Operation is iterated to selected sigmoid loss function, until the numerical value of the sigmoid loss function is minimum.
16. the building system of textual classification model according to claim 11, which is characterized in that the word vector is to use
The improved Cove pre-training word vector of Chinese-English translation training.
17. the building system of textual classification model according to claim 11, which is characterized in that the processor is also used
In execute the computer program with: down-sampled processing result sequentially input into dropout and relu activate described.
18. the building system of textual classification model according to claim 10, which is characterized in that
The processor, be also used to execute the computer program with: at least three-wheel dialog information carry out data cleansing.
19. a kind of computer readable storage medium, is stored thereon with computer program, which is characterized in that the computer program
The step of construction method of textual classification model as claimed in any one of claims 1-9 wherein is realized when being executed by processor.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201811440834.9A CN109710757A (en) | 2018-11-29 | 2018-11-29 | Construction method, system and the computer readable storage medium of textual classification model |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201811440834.9A CN109710757A (en) | 2018-11-29 | 2018-11-29 | Construction method, system and the computer readable storage medium of textual classification model |
Publications (1)
Publication Number | Publication Date |
---|---|
CN109710757A true CN109710757A (en) | 2019-05-03 |
Family
ID=66255257
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201811440834.9A Pending CN109710757A (en) | 2018-11-29 | 2018-11-29 | Construction method, system and the computer readable storage medium of textual classification model |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN109710757A (en) |
Cited By (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN111400452A (en) * | 2020-03-16 | 2020-07-10 | 腾讯科技(深圳)有限公司 | Text information classification processing method, electronic device and computer readable storage medium |
Citations (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN107025284A (en) * | 2017-04-06 | 2017-08-08 | 中南大学 | The recognition methods of network comment text emotion tendency and convolutional neural networks model |
CN108170848A (en) * | 2018-01-18 | 2018-06-15 | 重庆邮电大学 | A kind of session operational scenarios sorting technique towards China Mobile's intelligent customer service |
-
2018
- 2018-11-29 CN CN201811440834.9A patent/CN109710757A/en active Pending
Patent Citations (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN107025284A (en) * | 2017-04-06 | 2017-08-08 | 中南大学 | The recognition methods of network comment text emotion tendency and convolutional neural networks model |
CN108170848A (en) * | 2018-01-18 | 2018-06-15 | 重庆邮电大学 | A kind of session operational scenarios sorting technique towards China Mobile's intelligent customer service |
Cited By (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN111400452A (en) * | 2020-03-16 | 2020-07-10 | 腾讯科技(深圳)有限公司 | Text information classification processing method, electronic device and computer readable storage medium |
CN111400452B (en) * | 2020-03-16 | 2023-04-07 | 腾讯科技(深圳)有限公司 | Text information classification processing method, electronic device and computer readable storage medium |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN110070183B (en) | Neural network model training method and device for weakly labeled data | |
CN108536679B (en) | Named entity recognition method, device, equipment and computer readable storage medium | |
CN110597991B (en) | Text classification method and device, computer equipment and storage medium | |
CN112541501B (en) | Scene character recognition method based on visual language modeling network | |
CN109785833A (en) | Human-computer interaction audio recognition method and system for smart machine | |
CN110083700A (en) | A kind of enterprise's public sentiment sensibility classification method and system based on convolutional neural networks | |
CN116168352B (en) | Power grid obstacle recognition processing method and system based on image processing | |
CN108549658A (en) | A kind of deep learning video answering method and system based on the upper attention mechanism of syntactic analysis tree | |
CN109977428A (en) | A kind of method and device that answer obtains | |
CN109840322A (en) | It is a kind of based on intensified learning cloze test type reading understand analysis model and method | |
CN109711356B (en) | Expression recognition method and system | |
CN111597341B (en) | Document-level relation extraction method, device, equipment and storage medium | |
CN108509833A (en) | A kind of face identification method, device and equipment based on structured analysis dictionary | |
CN116311483B (en) | Micro-expression recognition method based on local facial area reconstruction and memory contrast learning | |
CN114841151B (en) | Medical text entity relation joint extraction method based on decomposition-recombination strategy | |
CN107832721A (en) | Method and apparatus for output information | |
CN116994188A (en) | Action recognition method and device, electronic equipment and storage medium | |
CN115455194A (en) | Knowledge extraction and analysis method and device for railway faults | |
CN111126155A (en) | Pedestrian re-identification method for generating confrontation network based on semantic constraint | |
CN109710757A (en) | Construction method, system and the computer readable storage medium of textual classification model | |
CN113362852A (en) | User attribute identification method and device | |
CN111522923A (en) | Multi-round task type conversation state tracking method | |
CN116758379A (en) | Image processing method, device, equipment and storage medium | |
CN116311493A (en) | Two-stage human-object interaction detection method based on coding and decoding architecture | |
CN115116444A (en) | Processing method, device and equipment for speech recognition text and storage medium |
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 | ||
WD01 | Invention patent application deemed withdrawn after publication | ||
WD01 | Invention patent application deemed withdrawn after publication |
Application publication date: 20190503 |