CN109858027A - One tool method for identifying and classifying of internet four product of electric business merchandise news - Google Patents
One tool method for identifying and classifying of internet four product of electric business merchandise news Download PDFInfo
- Publication number
- CN109858027A CN109858027A CN201910056584.7A CN201910056584A CN109858027A CN 109858027 A CN109858027 A CN 109858027A CN 201910056584 A CN201910056584 A CN 201910056584A CN 109858027 A CN109858027 A CN 109858027A
- Authority
- CN
- China
- Prior art keywords
- classification
- merchandise news
- product
- model
- identifying
- 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
Landscapes
- Information Retrieval, Db Structures And Fs Structures Therefor (AREA)
Abstract
The present invention is four product of internet electric business merchandise news, one tool method for identifying and classifying.It provides a kind of pair of internet electronic business transaction platform merchandise news and carries out four product, one tool classifying identification method, and mono- tool of Ji Sipin identifies the system building of disaggregated model.Wherein above-mentioned building and treatment process include: to compile corpus information data；Corpus data is subjected to word segmentation processing；Vectorization is carried out to the corpus data after word segmentation processing and establishes term vector library, part of speech vector library etc.；Construct the neural network structure of Classification and Identification model；The designed Classification and Identification network of training；Test and Correlative data analysis are iterated to trained identification model；The data distribution of each classification of statistical test data；It is compared according to the codomain of target function value with the test data of statistic of classification, analysis and arrangement identifies the range of the high target function value of error rate, carries out correction process.The higher classification data processing of error rate can be the classification of non-four product, one tool in this way, to improve the accuracy rate of category of model identification.
Description
Technical field
The present invention relates to technical field of data processing, the design of natural language processing (NLP) disaggregated model neural network with
Trained and data statistic analysis calculation method.
Background technique:
As the fast development of internet information and the relevant technologies are constantly updated, network application and information processing technology approach
Increasingly diversification, complication.Especially the development of transaction electron-like business platform is even more to make rapid progress, tradable commodity quantity, type
Rapid growth.Network trading commodity amount reaches hundred million or more ranks at present, network trading merchandise news type reaches 100,000 or more
Rank.This just gives the inquiry of merchandise news, supervision to propose new, requirements at the higher level.So one kind can be provided with network trading commodity
The relevant information of description reaches the class that Classification and Identification goes out tradable commodity information by the deep learning the relevant technologies of artificial intelligence
Not, this, which can inquire by classification and supervise to network commodity information, provides convenient effective processing means.
Summary of the invention:
Solution provided by the present invention is for one tool method for identifying and classifying of internet four product of electric business merchandise news
Building.Time cost can be saved for classified inquiry, the supervision of internet business merchandise news, improve supervisory efficiency.The classification
Model is food, drug, cosmetics, health care product, medical instrument referred to as " four product, one tool " for internet business merchandise news
Commodity provide identification classification processing, and internet merchandise news is facilitated to be inquired by classification and supervised；It is above-mentioned to complete to realize
Classification and Identification model construction and the method for training, the invention carry out following development process scheme:
A1, merchandise news and other text informations are compiled as corpus progress word segmentation processing；
A2, term vector calculating is carried out to the corpus after word segmentation processing, and establishes corresponding term vector library and part of speech vector library；
The deep learning network of A3, building four product, one tool identification disaggregated model；
A4, classification annotation merchandise news are as training sample, and are iterated trained network until completing identification classification mould
Type training；
A5, test analysis complete the disaggregated model discrimination situation of training；
The distribution situation of A6, the commodity data of statistical model classification error and disaggregated model objective function codomain range；
A7, building error-correction layer are filtered pretreatment to the target function value of disaggregated model identification merchandise news distribution；
Output layer output category result data after A8, classification processing.
Preferentially, merchandise news and other text informations are compiled described in A1 also to wrap as corpus progress word segmentation processing
Include: rich language material collects the covering of source various aspects and training corpus is pre-processed (corpus pretreatment: one document of a line
Or sentence, document or sentence are segmented).
Preferentially, described in A2 to after word segmentation processing corpus carry out term vector calculating, and establish term vector and part of speech to
Measure library processing method, which is characterized in that further include to need for original training corpus to be converted to a corpus sentence repeatedly
For device, the corpus sentence that iteration returns each time is the list that a corpus participle formats, and handles library by nature sentence
Model calculate building term vector library model object.
Preferentially, the deep learning network of the identification of four product of building, one tool described in A3 disaggregated model, the deep learning net
Network design framework technology includes: torch, caffe；The neural network design of specific building four product, one tool disaggregated model has, net
Network design packet is specific as follows comprising 6 layers of learning network:
H1: first layer convolutional neural networks (CNN), 1000 convolution kernels, convolution kernel size 3*320；
H2: second and third layer two-way long short-term memory Recognition with Recurrent Neural Network (Bi-directional LSTM RNN)；
H3: fourth, fifth layer of full articulamentum neural network (Fully Connected layer abbreviation FC)；
H4: finally by a classifier as output layer, output category result target function value.
Preferentially, by collecting classification annotation merchandise news and test analysis classification based training sample size described in A4
Ratio distribution determines the ratio value range of most suitable merchandise news textual classification model training sample.It handles number of training
Specific step is as follows for amount ratio:
I1: a large amount of corpus are collected as training sample, artificial four product, one tool classification samples is carried out and selects mark processing；
I2: corpus sample proportion, four product, one tool and non-four product, one tool sample proportion value range, 0.8 to 1.25 section, four product
Sample size ratio value range between each classification of one tool is greater than 0.2 to less than equal to 5；
I3:, can be suitably multiple in original sample size copy when the requirement of above-mentioned I2 is not achieved in corpus sample proportion
It makes and enriches sample size (sample size of copy duplication is within the scope of original 1 to 3 times), its sample proportion value is made to meet I2's
It is required that.
Preferentially, the verifying sample of four product, one tool disaggregated model described in A4 includes training set sample and test set sample two
Part, the repetitive exercise of disaggregated model corresponds to recognition accuracy with training set sample and test set sample, loss function value reaches
Index value when smaller relatively stable is the reference standard that model training is completed.
Preferentially, test analysis described in A5 completes the disaggregated model discrimination situation of training, and building one is suitable for exhibition
The statistical model for showing analysis merchandise news mistake classification situation, according to the wrong classified commodity after statistical classification model identifying processing
The probability distribution of information；" identification quantity poor " is that authentic specimen number subtracts identification quantity and obtains, and value reflects when being negative value
The Classification and Identification is easier to be influenced by the classification of other merchandise news mistakes, and the smaller influence of negative value is bigger；It should to reflect when positive value
Classification and Identification, which is less susceptible to be classified by other merchandise news mistakes, to be influenced, and the bigger influence of numerical value is smaller；" wrong identification number 1 " is
Authentic specimen number subtracts identification correct number and obtains, and value is bigger, and reflection disaggregated model is higher to the identification error rate of the classification,
It is on the contrary then smaller；The numerical value of " wrong identification number 2 " are as follows: the numerical value of " wrong identification number 1 " subtracts " identification quantity is poor " numerical value, calculates
Institute's value concentrated expression influence of the disaggregated model to commodity classification.
Preferentially, disaggregated model described in A6 is to the objective function in all classification after test sample identification classification processing
Value, on the basis of A5 step statistical analysis, statistical analysis disaggregated model is all to what is exported after merchandise news mistake Classification and Identification
The distribution relation of target function value, and according to " wrong identification number 1 " and " wrong identification number in the data of merchandise news mistake classification
The distribution statistics situation of 2 " test sample target function value, building disaggregated model mistake class object functional value and mistake are classified
Incidence relation.
Preferentially, it is by A6 step to mistake that building error-correction layer described in A7, which does filtration treatment to commodity classification wrong data,
Misclassification test sample target function value distribution statistics situation constructs the error correction mould of disaggregated model mistake class object functional value
Type；The maximum two values in all target function values after error correcting model calculating parameter selection classification processing carry out subtraction calculations
It is calculated with being divided by, Definition Model function f (m0, m1), g (m0, m1)；0 > m0 > m1, and m0, m1 respectively indicate target function value
In maximum two values, then the function of error correcting model has, f (m0, m1)=m0-m1, g (m0, m1)=m0/m1；Pass through error correction
Pattern function carries out the calculating of related objective functional value, and statisticallys analyze f (m0, the m1) value and g (m0, m1) value and two after calculating
The distribution situation of numerical intervals mistake classification where kind numerical value counts the high correspondence numerical value of test sample classification error distribution and makees
The index for selection of threshold values is filtered for error correction；One threshold values is arranged to each every class and carries out judgement filtration treatment, each classification
Target function value carry out accordingly calculate after, greater than corresponding threshold value 1 or less than the descriptive labelling information of corresponding threshold value 2, be considered
It is the commodity for identifying that error rate is high, error-correction layer can be filtered processing to it, and the unified commodity by such target function value are sorted out
For general merchandise (non-four product, one tool commodity).
Preferentially, the output layer output category result data after classification processing described in A8, according to error-correction layer identifying processing
As a result it will meet optimal objective value and meet (the maximum two values progress subtraction calculations and phase of target function value of requirement described in A7
Except the result after calculating), in the range of threshold value setting, the number for being mapped as corresponding classification provides result.
Detailed description of the invention:
Fig. 1 is overall procedure of the present invention for the building of four product of internet electric business merchandise news, one tool method for identifying and classifying
Figure；
Fig. 2 is the network design process of the deep learning network of present invention building four product, one tool disaggregated model；
Fig. 3 is the processing module explanatory diagram of convolution operation in inventive network design；
Fig. 4 is the processing module explanatory diagram of Bi-LSTM sequential operation in inventive network design.
Specific embodiment
Above objects, features, and advantages in order to better illustrate the present invention, can be simpler understandable, below with reference to attached
The present invention is further illustrated with specific embodiment for figure.
Embodiment
It referring to Fig.1, is building of the present invention for four product of internet electric business merchandise news, one tool method for identifying and classifying, it is overall
Flow chart may comprise steps of:
Step A1: merchandise news and other text informations are compiled as corpus and carries out word segmentation processing；In step A1,
News, descriptive labelling information are acquired in such a way that novel data or web crawlers are collected in internet come the data in rich language material library
Information content.It is pre-processed to the corpus completed is collected, one document of a line or a sentence are divided document or sentence
Word.Used here as Chinese Academy of Sciences NLPIR Words partition system tool；Word segmentation processing result is as follows:
Original text information: { the pure face of attaining of domestic special counter certified products Lancome Lancome cyanines moistens the pure eye of eye cream 20ml gold
Frost }；
Text after word segmentation processing；{ the country/locative special counter/noun certified products/noun Lancome/noun orchid/noun
Cool/noun cyanines/noun is pure to be attained/and noun face/noun moistens/and verb/classifier frost/noun 20ml/noun gold is pure/
Noun/noun frost/noun }.
Step A2: to after word segmentation processing corpus carry out term vector calculating, and establish corresponding term vector library and part of speech to
Measure library；The step is to be calculated on the basis of the corpus of step A1 participle by the vector that training carries out participle and part of speech, and establish
Corresponding term vector and part of speech vector library are used for disaggregated model.It is the third party Python of a open source used here as Gensim
Kit；The word2vec model of realization term vector modeling in Gensim is for carrying out vector calculating；Word2vec model
Hyper parameter are as follows: Word2Vec (sg=1, sentences, size=256, window=5, min_count=3, workers=
8, iter=40) parameter declaration is as follows.
Table 1:
Example is as follows:
Part of speech vectors are tieed up as " country " is converted into 256 dimension term vectors and 64 after participle, specific format is in a manner of array
Omission shows as follows:
256 dimension term vectors:
0.24395664,0.16760093,0.02231296,
..., // 250 vectors are omitted here
0.45377976,0.19203474, -0.05504936 } 256 vectors of size
64 dimension part of speech vectors:
{ 1.6208177, -1.5348666, -1.288407
..., // 58 vectors are omitted here
- 1.0911843,0.95148927, -0.9333895 } 64 vectors of size
Step A3: the deep learning network of building four product, one tool identification disaggregated model；It specifically may refer to Fig. 2.Specific mind
It can be expressed as through network: INPUT- > [CONV] * 1- > [Bi-LSTM] * 2- > [FC] * 2
The detailed process including commodity identifying processing is as follows in this step:
F1, input corpus text information: { domestic special counter certified products Lancome Lancome cyanines are pure to be attained face to moisten eye cream 20ml gold pure
Eye cream }；
F2, F3 are formatted processing to input corpus text information and corpus text segments cutting process；
Feature vector assignment is carried out by participle and part of speech vector library after F4, text information participle cutting process, will be segmented
The vector that text information afterwards is converted to the dimension of feature term vector 256, part of speech vector 64 is tieed up indicates.Each descriptive labelling information is most
Take 60 words as term vector calculation expression more.Less than the vector of 60 words to mend " 0 ", more than 60 words, overage house
It abandons, does not give calculation processing.So the number of dimensions of the term vector of each final assignment of commodity are as follows: 60* (256+64)=60*320；
F5, input convolutional neural networks (CNN) carry out feature extraction, and 1000 convolution kernel 3*320 is taken to tie up.By F4 calculation processing
The vector of assignment 60*320 dimension is input to network and carries out convolution operation processing afterwards.Convolutional neural networks (CNN) take 1000 size 3*
320 convolution kernel carries out feature extraction, uses relu activation primitive, and non-linear factor is added to convolution feature extraction.Place
Process is managed referring to Fig. 3；The characteristic dimension exported after convolution operation feature extraction are as follows: ((60-3+1) * 320/320) * 1000=58*
1*1000；
Memory Neural Networks (Bi-directional LSTM) layer carries out information extraction in short-term for F6,2 two-way length.By F5
Feature vector after process of convolution is carried out Dropout2d () processing and (is ignored a part of neuron, at random using Dropout to keep away
Exempt from model over-fitting, the degree of independence between characteristic pattern can be improved) after.Be input in two layers of Bi-LSTM neural network into
The processing of row characteristic operation.It specifically may refer to Fig. 4；It is defeated after two layers of two-way length Memory Neural Networks (Bi-LSTM) processing
Series of operations treated characteristic dimension out are as follows: the feature vector of 58*1*64 dimension；
F7,2 full articulamentum (Fully Connected layer writes a Chinese character in simplified form FC) classification identifying processings.By F6, treated
Feature vector 58*64=3712 dimension, is input to calculation process in 2 full articulamentums, full articulamentum connects all features, will be defeated
Value gives classifier (such as log_softmax classifier) and carries out dimensionality reduction calculation process (full articulamentum can be special the multidimensional of output out
Sign figure (featureMap) is converted to the vector of a low-dimensional), output feature vector dimension is after first layer FC calculation process
800.It is again after 800 feature vectors are input to second layer FC calculation process by dimension, required dimension of exactly classifying is 6, input
Log_softmax classifier carries out the output processing of objective function functional value；
F8, output category result data after classification processing are carried out by classifier.
Step A4: classification annotation merchandise news is as training sample, and is iterated trained network until completing identification point
Class model training；Repetitive exercise disaggregated model network, after disaggregated model discrimination is stablized, the repetitive exercise of disaggregated model with
Reach or completes reference standard better than " table 2 " index value for model training.Complete training mission.Referring specifically to claim
18 points of description explanations.
Table 2:
| Value type | Training set sample | Test set sample |
| Recognition accuracy | 97% | 90% |
| Loss function value | 0.25 | 0.20 |
Step A5: test analysis completes the disaggregated model discrimination situation of training；It is retouched referring specifically to 9 points of claim 1
State bright, " identification quantity poor " is that authentic specimen number subtracts identification quantity and obtains, and value reflects that the Classification and Identification is got over when being negative value
It is easy to be influenced by the classification of other merchandise news mistakes, the smaller influence of negative value is bigger；To reflect the Classification and Identification when positive value more not
It is easy to be influenced by the classification of other merchandise news mistakes, the bigger influence of numerical value is smaller；" wrong identification number 1 " is that authentic specimen number subtracts
Identification correct number is gone to obtain, value is bigger, and reflection disaggregated model is higher to the identification error rate of the classification, on the contrary then smaller；
The numerical value of " wrong identification number 2 " are as follows: the numerical value of " wrong identification number 1 " subtracts " identification quantity is poor " numerical value, and it is comprehensive to calculate institute's value
Conjunction reflects influence of the disaggregated model to commodity classification.The dimension specifically counted corresponds to table such as by taking test set sample data as an example
Under:
Table 3:
Wherein the computation rule of relevant dimension statistics is as follows:
Identify that quantity is poor: authentic specimen quantity-identification quantity (containing mistake)=identification quantity is poor
Wrong identification number 1: authentic specimen quantity-identification correct number=wrong identification number 1
Wrong identification number 2: wrong identification number 1- identifies quantity difference=wrong identification number 2
Step A6: the commodity data of statistical model classification error and the distribution feelings of disaggregated model objective function codomain range
Condition；Referring specifically to 10 points of description explanations of claim 1.
Step A7: building error-correction layer is filtered pre- place to the target function value of disaggregated model identification merchandise news distribution
Reason；Referring specifically to 11 points of descriptions explanation of claim 1, and " table 4 ", " table 5 " explanation.It is entangled after being calculated by test statistics
Each classification threshold values of staggered floor is such as shown in " table 4 ", and wherein form attributes " threshold values 1 " are the value of f (m0, m1), and " threshold values 2 " is g
The value of (m0, m1).
The result of specific commodity classification identification target function value is as follows:
(the vertical white snow flesh frost 50g thoroughly of original packing, the old and new pack white snow flesh frost 50g commodity 1:{ " Wen Biquan (WETHERM) power thoroughly
Random hair preserving moisture and protecting skin) " }；
Commodity 1 identify target function value referring specifically to " table 4 ": " 03 cosmetics " target value: -6.5803527832e-05
It is maximum in all classification function target values, all recognition results are as follows: the cosmetics of number 03；
Table 4:
| Specific name | Threshold values 1 (m0-m1) | Threshold values 2 (m0/m1) |
| 01 food | > 0.44 | < 0.66 |
| 02 drug | > 0.44 | < 0.66 |
| 03 cosmetics | > 0.44 | < 0.66 |
| 04 health care product | > 0.44 | < 0.66 |
| 05 medical instrument | > 0.48 | < 1.11 |
M0: maximum target functional value is indicated；
M1: the second target function value of sequence is indicated；
Table 5:
| Specific name | Target function value |
| 00 non-four product, one tool | -12.07728577 |
| 01 food | -18.2436924 |
| 02 drug | -29.24889755 |
| 03 cosmetics | -6.67572021484e-06 |
| 04 health care product | -13.94442844 |
| 05 medical instrument | -16.42740822 |
Threshold values 1 are as follows: f (m0, m1) indicates that m0-m1's subtracts each other numerical value；
Threshold values 2 are as follows: the numerical value of g (m0, m1) expression m0/m1 being divided by；
M0-m1=(- 6.67572021484e-06)-(- 12.0772857666) > 0.44
M0/m1=(- 6.67572021484e-06)/(- 12.0772857666) < 0.66
So classification is entered after maximum target value -6.67572021484e-06 meets error-correction layer calculating in above-mentioned calculating
Classify to A8 and exports result.
Step A8: the output layer output category result data after classification processing；By A7, treated that result data is mapped as
The cosmetics classification of " 03 " number is as output result.
Explanation described above, the merchandise news only provided by the present invention for internet electronic business transaction platform
The method for carrying out Classification and Identification model construction and training is described in detail.Used herein specific case illustrates this hair
Bright principle and implementation method carry out relevant elaboration, the explanation of above embodiment be only to aid in understanding process of the invention,
Method and core concept；At the same time, for those skilled in the art, according to the thought of the present invention, in specific embodiment
And there will be changes in application range, in conclusion the contents of this specification are not to be construed as limiting the invention.
Claims (12)
1. a kind of four product of internet electric business merchandise news, one tool method for identifying and classifying, which is characterized in that the described method includes:
A1, merchandise news and other text informations are compiled as corpus progress word segmentation processing；
A2, term vector calculating is carried out to the corpus after word segmentation processing, and establishes corresponding term vector library and part of speech vector library；
The deep learning network of A3, building four product, one tool identification disaggregated model；
A4, classification annotation merchandise news are as training sample, and are iterated trained network until completing identification disaggregated model instruction
Practice；
A5, test analysis complete the disaggregated model discrimination situation of training；
The distribution situation of A6, the commodity data of statistical model classification error and disaggregated model objective function codomain range；
A7, building error-correction layer are filtered pretreatment to the target function value of disaggregated model identification merchandise news distribution；
Output layer output category result data after A8, classification processing.
2. four product of electric business merchandise news in internet according to claim 1, one tool method for identifying and classifying, it is characterised in that: institute
It states in step A1, collects article, novel, news etc. that legal copy was issued, and acquisition reflection masses are familiar with cyber transaction and put down
Merchandise news description in platform, and training data corpus is formatted pretreatment.
3. four product of electric business merchandise news in internet according to claim 1, one tool method for identifying and classifying, it is characterised in that: institute
It states in step A2, to the iterator for needing for original training corpus to be converted to a corpus sentence, iteration is returned each time
Corpus sentence is the word list of a formatting, and calculates building term vector library model by the model that nature sentence handles library
Object.
4. four product of electric business merchandise news in internet according to claim 1, one tool method for identifying and classifying, it is characterised in that: institute
State in step A2, handled by the term vector computation model in natural language processing library, to merchandise news corpus of text carry out word to
Measure the building in library and part of speech vector library.
5. four product of electric business merchandise news in internet according to claim 1, one tool method for identifying and classifying, it is characterised in that: institute
It states in step A3, the deep learning network technology of design is combined using more frames, network is enable more to combine deep learning frame
The advantages of frame.
6. four product of electric business merchandise news in internet according to claim 1, one tool method for identifying and classifying, it is characterised in that: institute
It states in step A3, the training network layer of design will not be too deep, and no pooling layers of processing, the network of design includes 6 layers, tool
Body includes:
H1: first layer convolutional neural networks (CNN)；
H2: second and third layer two-way long short-term memory Recognition with Recurrent Neural Network (Bi-directional LSTM RNN)；
H3: fourth, fifth layer of full articulamentum neural network (Fully Connected layer abbreviation FC)；
H4: finally by a classifier as output layer, output category result target function value.
7. four product of electric business merchandise news in internet according to claim 1, one tool method for identifying and classifying, it is characterised in that: institute
It states in step A4, not instead of fixed cluster sample proportion value, by collecting classification annotation merchandise news and test analysis classification
The ratio of training samples number is distributed, and determines the ratio value range of most suitable merchandise news textual classification model training sample.
8. four product of electric business merchandise news in internet according to claim 1, one tool method for identifying and classifying, it is characterised in that: institute
It states in step A4, the verifying sample of four product, one tool disaggregated model includes training set sample and test set sample two parts, determines mould
The index that type training is completed is the stabilization by the repetitive exercise of model to reach model index value, the mark completed for model training
Standard, wherein index value is determined by the recognition accuracy and loss function value of training set sample and test set sample.
9. four product of electric business merchandise news in internet according to claim 1, one tool method for identifying and classifying, it is characterised in that: institute
It states in step A5, constructs the statistical model for showing analysis merchandise news mistake classification distribution situation, carry out statistical classification mould
The probability distribution of wrong classified commodity information after type identifying processing.
10. four product of electric business merchandise news in internet according to claim 1, one tool method for identifying and classifying, it is characterised in that:
In the step A6, sorted maximum target functional value, which is final classification, directly not to be identified to merchandise news with disaggregated model
As a result, but on the basis of the probability distribution of wrong classified commodity information after disaggregated model identifying processing statistical analysis,
Distribution relation of the statistical analysis disaggregated model to all target function values exported after merchandise news mistake Classification and Identification, and according to
Point of the test sample target function value of " wrong identification number 1 " and " wrong identification number 2 " in the data of merchandise news mistake classification
Cloth statistical conditions, the incidence relation of building disaggregated model mistake class object functional value and mistake classification.
11. four product of electric business merchandise news in internet according to claim 1, one tool method for identifying and classifying, it is characterised in that:
In the step A7, in the commodity data of statistical model classification error and the distribution situation of disaggregated model objective function codomain range
On the basis of processing is completed, building error correcting model process layer does filtering correction processing to the wrong classification data that disaggregated model identifies,
Reach the accuracy rate for improving merchandise news classification.
12. four product of electric business merchandise news in internet according to claim 1, one tool method for identifying and classifying, it is characterised in that:
In the step A8, pretreated recognition result is filtered to target function value according to error-correction layer, meets maximum target value
And meet above-mentioned 11 point requirement, in the range of threshold value setting, its target value is mapped as to the number and defeated of corresponding goods classification
Result out.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910056584.7A CN109858027A (en) | 2019-01-22 | 2019-01-22 | One tool method for identifying and classifying of internet four product of electric business merchandise news |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910056584.7A CN109858027A (en) | 2019-01-22 | 2019-01-22 | One tool method for identifying and classifying of internet four product of electric business merchandise news |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN109858027A true CN109858027A (en) | 2019-06-07 |
Family
ID=66895366
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910056584.7A Pending CN109858027A (en) | 2019-01-22 | 2019-01-22 | One tool method for identifying and classifying of internet four product of electric business merchandise news |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN109858027A (en) |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110245800A (en) * | 2019-06-19 | 2019-09-17 | 南京大学金陵学院 | A method of based on superior vector spatial model goods made to order information class indication |
Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN103646343A (en) * | 2013-12-18 | 2014-03-19 | 世纪禾光科技发展（北京）有限责任公司 | Text based commodity classification treatment method and system |
| WO2017113232A1 (en) * | 2015-12-30 | 2017-07-06 | 中国科学院深圳先进技术研究院 | Product classification method and apparatus based on deep learning |
| CN107301246A (en) * | 2017-07-14 | 2017-10-27 | 河北工业大学 | Chinese Text Categorization based on ultra-deep convolutional neural networks structural model |
| US20180144246A1 (en) * | 2016-11-16 | 2018-05-24 | Indian Institute Of Technology Delhi | Neural Network Classifier |
| CN108595418A (en) * | 2018-04-03 | 2018-09-28 | 上海透云物联网科技有限公司 | A kind of commodity classification method and system |
-
2019
- 2019-01-22 CN CN201910056584.7A patent/CN109858027A/en active Pending
Patent Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN103646343A (en) * | 2013-12-18 | 2014-03-19 | 世纪禾光科技发展（北京）有限责任公司 | Text based commodity classification treatment method and system |
| WO2017113232A1 (en) * | 2015-12-30 | 2017-07-06 | 中国科学院深圳先进技术研究院 | Product classification method and apparatus based on deep learning |
| US20180144246A1 (en) * | 2016-11-16 | 2018-05-24 | Indian Institute Of Technology Delhi | Neural Network Classifier |
| CN107301246A (en) * | 2017-07-14 | 2017-10-27 | 河北工业大学 | Chinese Text Categorization based on ultra-deep convolutional neural networks structural model |
| CN108595418A (en) * | 2018-04-03 | 2018-09-28 | 上海透云物联网科技有限公司 | A kind of commodity classification method and system |
Non-Patent Citations (1)
| Title |
|---|
| 冯园园: "短文本分类技术及其场景应用研究 ——基于某某宝交易数据", 《中国优秀硕士学位论文全文数据库 (经济与管理科学辑)》 * |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110245800A (en) * | 2019-06-19 | 2019-09-17 | 南京大学金陵学院 | A method of based on superior vector spatial model goods made to order information class indication |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109522556B (en) | Intention recognition method and device | |
| CN108899051B (en) | Speech emotion recognition model and recognition method based on joint feature representation | |
| CN111737458A (en) | Intention identification method, device and equipment based on attention mechanism and storage medium | |
| CN110083700A (en) | A kind of enterprise's public sentiment sensibility classification method and system based on convolutional neural networks | |
| CN109740154A (en) | A kind of online comment fine granularity sentiment analysis method based on multi-task learning | |
| CN108920445A (en) | A kind of name entity recognition method and device based on Bi-LSTM-CRF model | |
| CN106997341B (en) | A kind of innovation scheme matching process, device, server and system | |
| CN112418653A (en) | Number portability and network diver identification system and method based on machine learning algorithm | |
| CN112328900A (en) | Deep learning recommendation method integrating scoring matrix and comment text | |
| CN110377913B (en) | Emotion analysis method and device, electronic equipment and storage medium | |
| CN116541911B (en) | Packaging design system based on artificial intelligence | |
| CN108319734A (en) | A kind of product feature structure tree method for auto constructing based on linear combiner | |
| CN106844632A (en) | Based on the product review sensibility classification method and device that improve SVMs | |
| CN108764194A (en) | A kind of text method of calibration, device, equipment and readable storage medium storing program for executing | |
| CN109344399A (en) | A kind of Text similarity computing method based on the two-way lstm neural network of stacking | |
| CN110472040A (en) | Extracting method and device, storage medium, the computer equipment of evaluation information | |
| CN109087205A (en) | Prediction technique and device, the computer equipment and readable storage medium storing program for executing of public opinion index | |
| CN110297888A (en) | A kind of domain classification method based on prefix trees and Recognition with Recurrent Neural Network | |
| CN110633467A (en) | Semantic relation extraction method based on improved feature fusion | |
| CN110134845A (en) | Project public sentiment monitoring method, device, computer equipment and storage medium | |
| CN107357785A (en) | Theme feature word abstracting method and system, feeling polarities determination methods and system | |
| CN109829478A (en) | One kind being based on the problem of variation self-encoding encoder classification method and device | |
| CN110222192A (en) | Corpus method for building up and device | |
| CN112905739A (en) | False comment detection model training method, detection method and electronic equipment | |
| CN114942974A (en) | E-commerce platform commodity user evaluation emotional tendency classification method |
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 | ||
| DD01 | Delivery of document by public notice |
Addressee: Wu Wei Document name: Notice of the First Examination Opinion |
|
| DD01 | Delivery of document by public notice | ||
| DD01 | Delivery of document by public notice |
Addressee: Wu Wei Document name: Deemed withdrawal notice |
|
| DD01 | Delivery of document by public notice | ||
| WD01 | Invention patent application deemed withdrawn after publication |
Application publication date: 20190607 |
|
| WD01 | Invention patent application deemed withdrawn after publication |