CN108595602A - The question sentence file classification method combined with depth model based on shallow Model - Google Patents
The question sentence file classification method combined with depth model based on shallow Model Download PDFInfo
- Publication number
- CN108595602A CN108595602A CN201810357603.5A CN201810357603A CN108595602A CN 108595602 A CN108595602 A CN 108595602A CN 201810357603 A CN201810357603 A CN 201810357603A CN 108595602 A CN108595602 A CN 108595602A
- Authority
- CN
- China
- Prior art keywords
- question sentence
- model
- feature
- convolution
- shallow
- 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
- 238000000034 method Methods 0.000 title claims abstract description 55
- 238000013528 artificial neural network Methods 0.000 claims abstract description 37
- 230000000306 recurrent effect Effects 0.000 claims abstract description 31
- 239000000284 extract Substances 0.000 claims abstract 4
- 239000000463 material Substances 0.000 claims description 35
- 239000011159 matrix material Substances 0.000 claims description 17
- 230000006870 function Effects 0.000 claims description 16
- 238000000605 extraction Methods 0.000 claims description 7
- 230000036541 health Effects 0.000 claims description 7
- 230000007774 longterm Effects 0.000 claims description 7
- 238000010606 normalization Methods 0.000 claims description 7
- 238000007689 inspection Methods 0.000 claims description 6
- 238000012417 linear regression Methods 0.000 claims description 6
- 230000008569 process Effects 0.000 claims description 6
- 238000004364 calculation method Methods 0.000 claims description 4
- 238000001914 filtration Methods 0.000 claims description 4
- 230000009193 crawling Effects 0.000 claims description 3
- 230000002123 temporal effect Effects 0.000 claims description 3
- 210000004218 nerve net Anatomy 0.000 claims 1
- 238000003058 natural language processing Methods 0.000 abstract description 2
- 238000012549 training Methods 0.000 description 15
- 230000000694 effects Effects 0.000 description 8
- 238000012706 support-vector machine Methods 0.000 description 5
- 238000003062 neural network model Methods 0.000 description 4
- 238000012545 processing Methods 0.000 description 4
- 238000004458 analytical method Methods 0.000 description 3
- 230000008859 change Effects 0.000 description 3
- 238000013145 classification model Methods 0.000 description 3
- 238000013527 convolutional neural network Methods 0.000 description 3
- 238000013135 deep learning Methods 0.000 description 3
- 238000009826 distribution Methods 0.000 description 3
- 238000005516 engineering process Methods 0.000 description 3
- 230000008901 benefit Effects 0.000 description 2
- 238000010586 diagram Methods 0.000 description 2
- 238000010801 machine learning Methods 0.000 description 2
- 238000004519 manufacturing process Methods 0.000 description 2
- 230000001537 neural effect Effects 0.000 description 2
- ORILYTVJVMAKLC-UHFFFAOYSA-N Adamantane Natural products C1C(C2)CC3CC1CC2C3 ORILYTVJVMAKLC-UHFFFAOYSA-N 0.000 description 1
- 238000013459 approach Methods 0.000 description 1
- 238000013398 bayesian method Methods 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 238000007796 conventional method Methods 0.000 description 1
- 230000007812 deficiency Effects 0.000 description 1
- 238000011161 development Methods 0.000 description 1
- 230000018109 developmental process Effects 0.000 description 1
- 235000013399 edible fruits Nutrition 0.000 description 1
- 230000002996 emotional effect Effects 0.000 description 1
- 238000002474 experimental method Methods 0.000 description 1
- 230000010354 integration Effects 0.000 description 1
- 210000002569 neuron Anatomy 0.000 description 1
- 238000005457 optimization Methods 0.000 description 1
- 238000007781 pre-processing Methods 0.000 description 1
- 238000002360 preparation method Methods 0.000 description 1
- 230000003134 recirculating effect Effects 0.000 description 1
- 230000011218 segmentation Effects 0.000 description 1
- 238000012360 testing method Methods 0.000 description 1
- 238000005303 weighing Methods 0.000 description 1
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F40/00—Handling natural language data
- G06F40/20—Natural language analysis
- G06F40/279—Recognition of textual entities
- G06F40/289—Phrasal analysis, e.g. finite state techniques or chunking
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F40/00—Handling natural language data
- G06F40/30—Semantic analysis
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Health & Medical Sciences (AREA)
- Artificial Intelligence (AREA)
- Audiology, Speech & Language Pathology (AREA)
- Computational Linguistics (AREA)
- General Health & Medical Sciences (AREA)
- Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Machine Translation (AREA)
Abstract
The present invention relates to the question sentence file classification methods combined with depth model based on shallow Model, belong to Computer Natural Language Processing technical field.The present invention extracts the feature set of words of question sentence text first, and character pair word weight is worth to using normalized term vector after vectorization, is inputted as a part for shallow-layer linear model.Convolutional network carries out convolution using the convolution kernel of a variety of different windows sizes to question sentence text, the feature vector that the different convolution kernels for possessing equal length convolution window extract is rearranged, it is separately input to again among corresponding Recognition with Recurrent Neural Network, it finally is linked together the output of multiple Recognition with Recurrent Neural Network to obtain the syntactic-semantic feature vector of question sentence, another part as shallow-layer linear model inputs.Final shallow Model obtains the classification results of question sentence according to the input that the output by feature term vector and depth model is spliced to form.The present invention overcomes the shortcomings of single depth model, effectively improves the accuracy rate of Question Classification.
Description
Technical field
The present invention relates to the question sentence file classification methods combined with depth model based on shallow Model, belong to computer nature
Language processing techniques field.
Background technology
Question sentence text classification belongs to short text classification, is played an important role in automatically request-answering system.Question sentence text point
Class mainly classifies to question sentence by analyzing the content of question sentence.There is rule-based method in early stage, utilizes the key of question sentence
The correspondence of word or grammatical pattern and question sentence type, classifies to question sentence.This method to possess apparent interrogative or
The Question Classification effect of question sentence Based on Class Feature Word Quadric is fine, but for there is no apparent in more complex question sentence or question sentence text
Based on Class Feature Word Quadric is quite different, and the flexibility ratio of method is not high, and workload is larger, and the subjectivity of Question Classification is strong.With machine
The development of study, the Question Classification method based on machine learning become mainstream, Zhang (<26th ACM years international
Art meeting>, 2003) et al. using support vector machines (SVM), the syntactic feature for extracting sentence classifies to problem, this method
Relatively pervious method accuracy rate has obtained larger promotion.Also have in addition to this and is mutually tied rule-based with the method for machine learning
It closes, Li (<Chinese Journal of information>, 2008) etc. by interrogative and centre word rule and SVM method phases
In conjunction with making the accuracy of Question Classification further increase.Nicety of grading depends on the effect of the technologies such as syntactic analysis, but in
The syntactic analysis technology phase that the form variability of text and clause complexity cause the syntactic analysis difficulty of Chinese higher, current
To not mature enough, the order of accuarcy of question sentence text classification is affected.
Recently as the proposition of deep learning, various deep learning frames have been widely used in image procossing and nature
In Language Processing, and the promotion breakthrough relative to conventional method is achieved, wherein in terms of sentence Document Modeling and classification, volume
Product neural network (CNN) and Recognition with Recurrent Neural Network (RNN) have become two kinds of most common deep learning neural network frameworks.
Kim(<Eprint Arxiv>, 2014) et al. the text that is originally modeled using convolutional neural networks distich Ziwen, and will obtained
Feature is used for text classification, this model structure is simple, but achieves preferable classification results, has become a kind of text classification
Baseline Methods.Tang(<Natural language processing meeting>, 2015) et al. also taken in emotional semantic classification task with Recognition with Recurrent Neural Network
Obtained preferable result.Depth model efficiently solves feature extraction complexity existing for conventional machines learning method, and transplantability is poor,
The problems such as short text character representation is sparse.But due to the too strong learning ability of depth model, Cheng (<arXiv.org>, 2016)
Et al. indicate that depth model is difficult to learn to indicate existed to effective feature vector to the lower feature of certain frequencies of occurrences
In extensive problem, and shallow Model such as SVM, linear model etc. can the feature less to occurrence number preferably learned
It practises, the perceptron network integration of Logic Regression Models and multilayer is improved the software application of Google's application shop by Cheng et al.
Recommend accuracy rate.Often occur the unbalanced problem of training data in text categorization task, single depth model to data volume compared with
Few classification is difficult that effective character representation is arrived in study.
Invention content
The present invention provides the question sentence file classification methods combined with depth model based on shallow Model, for single depth
When model faces uneven training data there are the problem of, there is using traditional shallow Model to feature the spy of stronger Memorability
Point effectively improves the accuracy rate of Question Classification.
The technical scheme is that：It is described based on the question sentence file classification method that shallow Model is combined with depth model
Method is as follows：
Step1, the question sentence language for crawling economy and finance, laws and regulations, sports, 5 health care, electronic digital classifications
Secondly material pre-processes language material text；
Step2, the feature set of words that question sentence text in question sentence language material is extracted using the method for evolution inspection CHI, will be each
Feature Words are converted into the form of term vector, and use the corresponding normalization term vector value of Feature Words as its weighted value, thus
To the part input Input1 of shallow-layer linear model；
Step3, increase question sentence keyword term vector weight, the question sentence text vector for then forming term vector matrix inputs
Into first part's convolutional network of depth model；Wherein use the convolution kernel of a variety of different windows sizes respectively to question sentence text
Convolution operation is carried out, the local phrase feature of sentence is extracted, the different convolution kernels for possessing equal length convolution window are extracted
Feature vector rearranged；
Step4, the feature vector generated in Step3 is separately input among corresponding Recognition with Recurrent Neural Network；Cycle god
The historical information of sentence can be captured by its chain structure through network, the long-term dependence characteristics of sequence data are arrived in study,
The output of the last one time step contains the characteristic information of entire sentence, and the output of multiple Recognition with Recurrent Neural Network is linked to one
The final feature as question sentence is played, another part input Input2 of shallow-layer linear model is thus obtained；
Step5, the final output Input2 of depth model in the Input1 obtained in Step2 and Step4 is spliced to form
The input of shallow Model, shallow Model part use multiple linear regression structure, finally obtain the classification results of question sentence.
The step Step1 is as follows：
Step1.1, first manual compiling crawlers, Baidu know swash take economy and finance, laws and regulations, sport fortune
Dynamic, 5 health care, electronic digital classifications question sentence language material；
Step1.2, the language material crawled, obtain unduplicated question sentence language material by filtering, duplicate removal, and it be stored in
In database；
Step1.3, the question sentence language material in database is segmented, stop words is gone to pre-process.
The step Step2 is as follows：
Step2.1, the feature set of words that question sentence text is extracted using the method for evolution inspection CHI；
Step2.2, the form for converting each Feature Words in Step2.1 to term vector, using distributed term vector
Representation method；
Step2.3, using the corresponding normalization term vector value of Feature Words as its weighted value, finally obtain question sentence text
Non- syntactic feature indicates that the part as shallow Model inputs Input1.
The step Step3 is as follows：
Step3.1, question sentence key is extracted using the method based on tf-idf that the jieba kits in python provide
The term vector of question sentence keyword is repeated once by word respectively after each word is represented as term vector in question sentence in its left and right,
Keyword shared weight in sentence just will increase at this time, thus obtain a term vector matrix；
Step3.2, the question sentence text vector that the word vector matrix obtained in Step3.1 indicates is input to depth model
First part's convolutional network in, wherein the line number of matrix be sentence in word number, columns be term vector dimension；Here make
Vertical convolution operation is carried out to question sentence with each two of the convolution kernel of 2,3,4 three kinds of different length convolution windows, is extracted
The local feature of different location in sentence, thus obtains several groups feature vector；
Step3.3, will possess identical convolution window size the extraction of different convolution kernels feature vector position in temporal sequence
Confidence breath carries out rearranging combination so that the feature vector that different convolution kernels are obtained in sentence same position convolution is spliced one
It rises.
The step Step4 is as follows：
Step4.1, by three kinds of different length convolution windows obtain in Step3.3 the feature rearranged respectively according to sentence
Son is sequentially input among corresponding three Recognition with Recurrent Neural Network；Used here as LSTM Recognition with Recurrent Neural Network, for more preferably capturing
To sentence, historical information, the long-term dependence characteristics of study to sequence data, the output of the last one time step include earlier
The characteristic information of entire question sentence；
Step4.2, the output of three Recognition with Recurrent Neural Network in Step4.1 is linked together to final feature as question sentence
It indicates, thus obtains another part input Input2 of shallow-layer linear model.
The step Step5 is as follows：
Step5.1, the final output Input2 of the Input1 obtained in Step2.3 and Step4.2 is spliced to form shallow-layer
The input of model, shallow Model is using multiple linear regression structure here, i.e., one last layer connected entirely is added with softmax
The general neural network of function；
Step5.2, the input layer content for obtaining Step5.1 pass through one layer of hidden layer, then the output of hidden layer is inputted
To obtaining final Question Classification result in sotfmax functions.
The depth model part is made of convolutional network layer and Recognition with Recurrent Neural Network layer；K-th of convolution window in convolutional layer
The Text Representation that the convolution nuclear convolution that mouth length is h obtains is wkh=[cki,…,ck(l-h+1)], wherein ckiIt indicates k-th
The convolution feature of convolution kernel i-th of position in question sentence text；cki=Relu (oki+ b), okiIndicate the value that convolutional calculation obtains；
oki=[xi,xi+1,…,xi+h-1]*fkh, wherein xiThe term vector of i-th of word in sentence is represented, h represents convolution kernel length of window,
[xi,xi+1,…,xi+h-1] represent in sentence the word from i-th of word to the i-th-h+1, the term vector matrix that total h word forms；fkh
Indicate that the convolution kernel that k-th of convolution length of window is h, * represent corresponding element multiplication sum operation in two matrixes；By convolutional layer
Obtained feature vector rearranges combination and then inputs three different LSTM Recognition with Recurrent Neural Network layers respectively, is formed final special
Sign vector is expressed as V=[v2,v3,v4], wherein v2,v3,v4Convolution length of window 2,3,4 is indicated respectively；The input of entire model
Layer is spliced to form by the feature term vector of shallow-layer part and the output V of depth model, and the vector for forming a m dimension indicates, X=
[wf1…wfn,V]。
The shallow Model final classification method is softmax functions.
The beneficial effects of the invention are as follows：
1, the present invention carries out term vector training using the word2vec modules of gensim, since the vector of word is the neighbour by word
What nearly word calculated, so meeting implicit semantic information in vector, suitable for semantic information extraction.Term vector is indicated
Text effectively improves the performance of model as the input of model.
2, in the preprocessing process of data, for depth model importation, increase the power of question sentence keyword term vector
Weight.Keyword in question sentence works as each word in question sentence by table to judging that the classification of sentence often has the function of bigger
It is shown as after term vector, the term vector of question sentence keyword in training corpus is repeated once respectively in its left and right, at this time keyword
Shared weight just will increase in sentence, can further increase the classification performance of model in this way.
3, the present invention is based on the question sentence textual classification model that shallow Model is combined with depth model, combine depth model with
Conventional machines learn the advantage of shallow Model respectively.Wherein depth model is by convolutional neural networks and LSTM Recognition with Recurrent Neural Network groups
It closes, in order to the syntactic-semantic feature of preferably learning text, a variety of different windows sizes is used in convolutional network
Convolution kernel carries out convolution operation to question sentence text.Simultaneously in order to overcome when the training corpus of certain class question sentence text is relatively fewer,
Depth model is difficult to learn to the validity feature vector of feature corresponding to respective classes to indicate, it is proposed that on the basis of depth model
Upper combination shallow Model has the characteristics that stronger Memorability using traditional shallow Model to feature.Training data imbalance with
In the case of balance, the accuracy rate of Question Classification achieves promotion, especially in training data imbalance, compares other models
Performance has a distinct increment.
To sum up, this question sentence file classification method combined with depth model based on shallow Model is passed through by convolutional Neural net
Network and Recognition with Recurrent Neural Network are composed depth model, and doing preferably study extraction to the syntactic-semantic feature of question sentence is used as shallow-layer
The part input of linear model, wherein depth model importation increase question sentence keyword term vector weight, and convolutional network makes
With the convolution kernel of a variety of different windows sizes.Feature term vector is inputted as another part of shallow Model, utilizes shallow Model
The advantages of, it is unbalanced in training corpus, overcome the deficiency of single depth model.Final unified model effectively carries
The accuracy rate of Question Classification is risen.
Description of the drawings
Fig. 1 is the Question Classification model structure of the present invention；
Fig. 2 is depth model part-structure figure in the present invention；
Fig. 3 is the Question Classification accuracy rate comparison diagram of different convolutional network output processing in the present invention；
Fig. 4 is different neural network models with the training increased performance change comparison diagram of iterations.
Specific implementation mode
Embodiment 1：As shown in Figs 1-4, the question sentence file classification method combined with depth model based on shallow Model, it is described
Method is as follows：
Step1, the question sentence language for crawling economy and finance, laws and regulations, sports, 5 health care, electronic digital classifications
Secondly material pre-processes language material text；
Further, the step Step1 is as follows：
Step1.1, first manual compiling crawlers, Baidu know swash take economy and finance, laws and regulations, sport fortune
Dynamic, 5 health care, electronic digital classifications question sentence language material；
Step1.2, the language material crawled, obtain unduplicated question sentence language material by filtering, duplicate removal, and it be stored in
In database；
The present invention has crawled on Baidu is known economy and finance by crawlers, laws and regulations, sports, medical treatment are defended
The problem of raw, 5 classifications of electronic digital each 5000 language materials, as first prepared language material set, i.e. balanced corpus.In addition
Health care and electronic digital are respectively removed into 3000 language materials, retain 2000 language materials, other three types language material numbers are constant, make
For second prepared corpus, i.e., uneven corpus.Each language material, which combines, takes therein 1/10th to be used as test set, remaining
Be used as training set.In view of the question sentence language material crawled is there may be repetition, these language materials increase workload, without too big
Meaning, so obtaining unduplicated question sentence corpus of text by filtering, duplicate removal on the basis of preparation language material.It is stored in database
It is in order to facilitate the management and use of data.
Step1.3, the question sentence language material in database is segmented, stop words is gone to pre-process.
The present invention is in view of by the character string forms that text dividing is multiple characters composition, can directly cause in original text
The loss of linguistic information between word, word.So to question sentence language material carry out pretreatment work, including with jieba tools into
Row Chinese word segmentation removes stop words etc., facilitates the progress of follow-up work.
Step2, the feature set of words that question sentence text in question sentence language material is extracted using the method for evolution inspection CHI, will be each
Feature Words are converted into the form of term vector, and use the corresponding normalization term vector value of Feature Words as its weighted value, thus
To the part input Input1 of shallow-layer linear model；
Further, the step Step2 is as follows：
Step2.1, the feature set of words that question sentence text is extracted using the method for evolution inspection CHI；
The present invention, using word as its essential characteristic item, does not use in terms of the Feature Selection of multiple linear structure division
Syntax grammar property, the feature selection approach evolution preferable and more commonly used using effect are examined to extract the spy of question sentence text
Word is levied, and question sentence text is indicated with feature set of words.
Step2.2, the form for converting each Feature Words in Step2.1 to term vector, using distribution
The term vector representation method of (distributed representation)；
During text vector, the present invention considers the limitation of tradition one-hot representation methods, selection
Distributed representation, the representation method of this term vector not only solve that one-hot dimensions are sparse to ask
Topic, and distance is close between the term vector expression of similar word, has carried certain semantic information.Term vector is indicated in this way
Text is helpful to the promotion of model performance as the input of model.The present invention is carried out using the word2vec modules of gensim
Term vector is trained.
Step2.3, using the corresponding normalization term vector value of Feature Words as its weighted value, finally obtain question sentence text
Non- syntactic feature indicates that the part as shallow Model inputs Input1.
Different weights are assigned for different Feature Words, the present invention is using most simple and effective normalization term vector value as special
Term vector weighted value is levied, the non-syntactic feature expression of question sentence text is one that Feature Words vector is denoted as subground line model
Divide input.
Step3, increase question sentence keyword term vector weight, the question sentence text vector for then forming term vector matrix inputs
Into first part's convolutional network of depth model；Wherein use the convolution kernel of a variety of different windows sizes respectively to question sentence text
Convolution operation is carried out, the local phrase feature of sentence is extracted, the different convolution kernels for possessing equal length convolution window are extracted
Feature vector rearranged；
Further, the step Step3 is as follows：
Step3.1, question sentence key is extracted using the method based on tf-idf that the jieba kits in python provide
The term vector of question sentence keyword is repeated once by word respectively after each word is represented as term vector in question sentence in its left and right,
Keyword shared weight in sentence just will increase at this time, thus obtain a term vector matrix；
In the importation of depth model, because some words in question sentence are to judging that the classification of sentence often has more your writing
With, for example " when basketball movement invents in question sentence" in, noun ' basketball ' is for differentiating that question sentence type is sport category
Play key effect.Therefore after each word in question sentence is represented as term vector, by question sentence keyword in training corpus
Term vector be repeated once respectively in its left and right, former sentence has reformed into " basketball basketball basketball movement is when to invent
's" at this time keyword shared weight in question sentence just will increase.In order to verify, keyword weight can be further in increase question sentence
Increase the classification performance of model, keyword plays the result of Question Classification crucial effect, has been one group of contrast experiment, such as
Shown in table 1：
Table 1
| Do not increase keyword weight | Increase keyword weight | |
| Accuracy rate | 0.9219 | 0.9226 |
Step3.2, the question sentence text vector that the word vector matrix obtained in Step3.1 indicates is input to depth model
First part's convolutional network in, wherein the line number of matrix be sentence in word number, columns be term vector dimension；Here make
Vertical convolution operation is carried out to question sentence with each two of the convolution kernel of 2,3,4 three kinds of different length convolution windows, is extracted
The local feature of different location in sentence, thus obtains several groups feature vector；
As shown in Fig. 2, in order to preferably learn to obtain the syntactic-semantic feature of question sentence text, used in convolutional network part
Each two of the convolution kernel of 2,3,4 three kinds of different length convolution windows carries out convolution operation to question text.Convolution length of window refers to
In each convolution operation covering sentence word quantity number.Convolution kernel is slided on sentence, is extracted different in sentence
Thus the local feature of position obtains one group of feature vector.
Step3.3, will possess identical convolution window size the extraction of different convolution kernels feature vector position in temporal sequence
Confidence breath carries out rearranging combination so that the feature vector that different convolution kernels are obtained in sentence same position convolution is spliced one
It rises.
The validity for handling and selecting different convolution windows in order to verify the present invention to convolutional network output, compares another
A kind of outer processing convolutional network exports and inputs the strategy of Recognition with Recurrent Neural Network and point of different convolution window size selection strategies
Class difference on effect.Second of link method of making comparisons is as described below, the feature after being reset to convolution, according to maximum length convolution window
Subject to the characteristic length that mouth convolution is reset, the feature obtained after other two kinds of length convolution window convolution is reset is cut therewith
Compare the part exceeded, and the feature of same position in sentence is linked together, and is input in a LSTM recirculating network.
This structure is denoted as M2:cl2,3,4.The link policy that depth part uses in shallow depth binding model of the present invention is denoted as
M1:Cl2,3,4, in addition also the model of single different length window is compared, is denoted as S respectively:Cl2, S:Cl3, S:
Cl4 indicates that window size is 2,3,4.It is tested in corpus 1, the results are shown in Figure 3.It will become apparent from the M1 of the present invention:
Cl2,3,4 tactful effects are best, and M2:The classifying quality of cl2,3,4 compares the model and M1 of single window length:Cl2,3,4,
Declining occurs in its classification accuracy, and reason may be that the feature cut away causes influence to final characteristic sequence, causes
Make LSTM that could not capture the sequence information of high quality.In addition in single length of window, when length of window is 3, classification accuracy
Highest.
Step4, the feature vector generated in Step3 is separately input among corresponding Recognition with Recurrent Neural Network；Cycle god
The historical information of sentence can be captured by its chain structure through network, the long-term dependence characteristics of sequence data are arrived in study,
The output of the last one time step contains the characteristic information of entire sentence, and the output of multiple Recognition with Recurrent Neural Network is linked to one
The final feature as question sentence is played, another part input Input2 of shallow-layer linear model is thus obtained；
Further, the step Step4 is as follows：
Step4.1, by three kinds of different length convolution windows obtain in Step3.3 the feature rearranged respectively according to sentence
Son is sequentially input among corresponding three Recognition with Recurrent Neural Network；Used here as LSTM Recognition with Recurrent Neural Network, for more preferably capturing
To sentence, historical information, the long-term dependence characteristics of study to sequence data, the output of the last one time step include earlier
The characteristic information of entire question sentence；
The present invention is considered more preferably learn the syntactic-semantic feature of sentence, be recycled in the second part of depth model refreshing
(LSTM) network is remembered through network selection shot and long term, because basic Recognition with Recurrent Neural Network model can lose sentence when sentence is longer
The information of forward portion in son, to overcome the above disadvantages, people have invented LSTM Recognition with Recurrent Neural Network models, it is relatively traditional
Neural network can preferably remember historical information earlier.
Step4.2, the output of three Recognition with Recurrent Neural Network in Step4.1 is linked together to final feature as question sentence
It indicates, thus obtains another part input Input2 of shallow-layer linear model.
The output of three LSTM is spliced together, the final feature vector of question sentence, i.e. V=[v are formed2,v3,v4], wherein
v2,v3,v4Indicate that convolution length of window is respectively 2,3,4 respectively.Multiwindow convolution loop combination of network depth model such as Fig. 2 institutes
Show.
Step5, the final output Input2 of depth model in the Input1 obtained in Step2 and Step4 is spliced to form
The input of shallow Model, shallow Model part use multiple linear regression structure, finally obtain the classification results of question sentence.
Further, the step Step5 is as follows：
Step5.1, the final output Input2 of the Input1 obtained in Step2.3 and Step4.2 is spliced to form shallow-layer
The input of model, shallow Model is using multiple linear regression structure here, i.e., one last layer connected entirely is added with softmax
The general neural network of function；
Step5.2, the input layer content for obtaining Step5.1 pass through one layer of hidden layer, then the output of hidden layer is inputted
To obtaining final Question Classification result in sotfmax functions.
Further, the depth model part is made of convolutional network layer and Recognition with Recurrent Neural Network layer；Kth in convolutional layer
The Text Representation that the convolution nuclear convolution that a convolution length of window is h obtains is wkh=[cki,…,ck(l-h+1)], wherein ckiTable
Show the convolution feature of k-th of convolution kernel, i-th of position in question sentence text；cki=Relu (oki+ b), okiIndicate that convolutional calculation obtains
The value arrived；oki=[xi,xi+1,…,xi+h-1]*fkh, wherein xiThe term vector of i-th of word in sentence is represented, h represents convolution kernel window
Mouth length, [xi,xi+1,…,xi+h-1] represent in sentence the word from i-th of word to the i-th-h+1, the term vector that total h word forms
Matrix；fkhIndicate that the convolution kernel that k-th of convolution length of window is h, * represent corresponding element multiplication sum operation in two matrixes；
The feature vector that convolutional layer obtains is rearranged into combination and then inputs three different LSTM Recognition with Recurrent Neural Network layers, shape respectively
It is expressed as V=[v at final feature vector2,v3,v4], wherein v2,v3,v4Convolution length of window 2,3,4 is indicated respectively；Entire mould
The input layer of type is spliced to form by the feature term vector of shallow-layer part and the output V of depth model, forms the vector table of m dimensions
Show, X=[wf1…wfn,V]。
Further, the shallow Model final classification method is softmax functions.
In order to compare shallow depth binding model and conventional machines learning model and convolutional neural networks model, cycle god
The problem of the problem of convolution loop combinational network model through network model and multiple length convolution windows classifying quality classification effect
Fruit, wherein conventional machines learning model have chosen three kinds of SVM, maximum entropy and naive Bayesian methods, in the present invention model and its
Excess-three kind neural network model is denoted as WD, CNN, RNN and M respectively:Cnn+rnn, here respectively from balanced corpus 1 and injustice
The accuracy rate of weighing apparatus corpus 2 is compared, as a result as shown in table 2, table 3.
Table 2
Table 3
By table 2, it is apparent that WD models accuracy rate in the corpus that language material balances compares other conventional machines
Model is practised, accuracy rate highest, although accuracy rate is declined in uneven language material, fall is compared to for other models
It is relatively low.
As can be seen from Table 3, the overall performance of depth model is still better than conventional model, but depth model is in unbalanced language material
Accuracy rate fall is relatively large in library, reason be in face of a certain category feature language material it is less in the case of, depth model
Too strong learning ability can increase the learning difficulty of effective characteristic of division.
In order to which further more general depth model and shallow depth binding model of the present invention, Fig. 4 are illustrated in imbalance
In corpus 2, with the increase of training iterations, the variation of respective classification accuracy.It can be seen from the figure that with model
The increase of training iterations, the Question Classification accuracy of 4 kinds of models are all increasing steadily, and iterations are at 200 times or so
When accuracy rate no longer change substantially.Shallow depth binding model is better than other three kinds of models on final classification performance.From
It can also be seen that convolutional network is slightly better than Recognition with Recurrent Neural Network on short text in figure.
In the present invention, the question sentence textual classification model that is combined with depth model based on shallow Model by shallow Model part with
Depth model part forms, and overall structure is as shown in Figure 1.
Input layer
Input layer is spliced to form by the feature term vector of shallow-layer part and the output V of depth model, formed m dimension to
Amount indicates, is denoted as X=[wf1…wfn,V]。
Softmax layers
Softmax layers are equivalent to the general neural network connected entirely for possessing one layer of hidden layer.The content of input layer is passed through
One layer of hidden layer is crossed, then being input in sotfmax functions for hidden layer is obtained into final classification results.Hidden layer is k
The neuron of a neural unit, input layer and hidden layer connects entirely.Its calculation formula：O=X*W, wherein W are m rows k row
Matrix, matrix finite element stochastic production nonzero value, then constantly with new in training.O be possess k value it is one-dimensional to
Amount.Each value represents the output valve of kth class, then is passed to softmax functions.The formula of softmax functions：
OkIndicate the output valve of neural network kth class, skRepresent the probability value that text belongs to k classifications.
In order to be trained entire model, need to define a suitable loss function, using Adam (<Computer
Science>, 2014) and optimization method minimizes or maximizes loss function and train entire model.For classification problem, one
As using cross entropy (cross-entropy) be used as its loss function.Its formula is：Hy′(y)=- ∑iyi′logyi, wherein yi′
To be true probability distribution (i.e. the class label of training corpus), yiFor the probability distribution of model prediction.Pass through minimum in this
Change Hy′(y) value trains entire model.
The specific implementation mode of the present invention is explained in detail above in conjunction with attached drawing, but the present invention is not limited to above-mentioned
Embodiment within the knowledge of a person skilled in the art can also be before not departing from present inventive concept
Put that various changes can be made.
Claims (8)
1. the question sentence file classification method combined with depth model based on shallow Model, it is characterised in that：The method it is specific
Steps are as follows：
Step1, the question sentence language material for crawling economy and finance, laws and regulations, sports, 5 health care, electronic digital classifications,
Secondly language material text is pre-processed；
Step2, the feature set of words that question sentence text in question sentence language material is extracted using the method for evolution inspection CHI, by each feature
Word is converted into the form of term vector, and uses the corresponding normalization term vector value of Feature Words as its weighted value, thus obtains shallow
The part input Input1 of layer linear model；
Step3, increase question sentence keyword term vector weight, the question sentence text vector that term vector matrix forms then is input to depth
It spends in first part's convolutional network of model；Wherein question sentence text is carried out respectively using the convolution kernel of a variety of different windows sizes
Convolution operation extracts the local phrase feature of sentence, will possess the spy of the different convolution kernels extraction of equal length convolution window
Sign vector is rearranged；
Step4, the feature vector generated in Step3 is separately input among corresponding Recognition with Recurrent Neural Network；Recycle nerve net
Network can capture the historical information of sentence by its chain structure, and the long-term dependence characteristics of study to sequence data are last
The output of one time step contains the characteristic information of entire sentence, and the output of multiple Recognition with Recurrent Neural Network is linked together work
For the final feature of question sentence, another part input Input2 of shallow-layer linear model is thus obtained；
Step5, the final output Input2 of depth model in the Input1 obtained in Step2 and Step4 is spliced to form shallow-layer
The input of model, shallow Model part use multiple linear regression structure, finally obtain the classification results of question sentence.
2. the question sentence file classification method according to claim 1 combined with depth model based on shallow Model, feature
It is：The step Step1 is as follows：
Step1.1, first manual compiling crawlers, Baidu know swash take economy and finance, laws and regulations, sports,
The question sentence language material of 5 health care, electronic digital classifications；
Step1.2, the language material crawled, obtain unduplicated question sentence language material by filtering, duplicate removal, and it be stored in data
In library；
Step1.3, the question sentence language material in database is segmented, stop words is gone to pre-process.
3. the question sentence file classification method according to claim 1 combined with depth model based on shallow Model, feature
It is：The step Step2 is as follows：
Step2.1, the feature set of words that question sentence text is extracted using the method for evolution inspection CHI；
Step2.2, the form for converting each Feature Words in Step2.1 to term vector are indicated using distributed term vector
Method；
Step2.3, using the corresponding normalization term vector value of Feature Words as its weighted value, finally obtain the non-sentence of question sentence text
Method character representation, the part input Input1 as shallow Model.
4. the question sentence file classification method according to claim 1 combined with depth model based on shallow Model, feature
It is：The step Step3 is as follows：
Step3.1, question sentence keyword is extracted using the method based on tf-idf that the jieba kits in python provide, when
Each word is represented as after term vector in question sentence, the term vector of question sentence keyword is repeated once respectively in its left and right, at this time
Keyword shared weight in sentence just will increase, and thus obtain a term vector matrix；
Step3.2, the question sentence text vector that the word vector matrix obtained in Step3.1 indicates is input to the of depth model
In a part of convolutional network, wherein the line number of matrix is the number of word in sentence, and columns is the dimension of term vector；Used here as 2,
Each two of the convolution kernel of 3,4 three kinds of different length convolution windows carries out vertical convolution operation to question sentence, extracts sentence
Thus the local feature of middle different location obtains several groups feature vector；
Step3.3, by the feature vector for the different convolution kernels extraction for possessing identical convolution window size, position is believed in temporal sequence
Breath carries out rearranging combination so that the feature vector that different convolution kernels are obtained in sentence same position convolution is stitched together.
5. the question sentence file classification method according to claim 1 combined with depth model based on shallow Model, feature
It is：The step Step4 is as follows：
It is Step4.1, three kinds of different length convolution windows obtain in Step3.3 the feature rearranged is suitable according to sentence respectively
Sequence is input among corresponding three Recognition with Recurrent Neural Network；Used here as LSTM Recognition with Recurrent Neural Network, for more preferably capturing sentence
Son historical information earlier, the long-term dependence characteristics of study to sequence data, the output of the last one time step contain whole
The characteristic information of a question sentence；
Step4.2, the output of three Recognition with Recurrent Neural Network in Step4.1 is linked together to final mark sheet as question sentence
Show, thus obtains another part input Input2 of shallow-layer linear model.
6. the question sentence file classification method according to claim 1 combined with depth model based on shallow Model, feature
It is：The step Step5 is as follows：
Step5.1, the final output Input2 of the Input1 obtained in Step2.3 and Step4.2 is spliced to form shallow Model
Input, here shallow Model use multiple linear regression structure, i.e., one last layer connected entirely is added with softmax functions
General neural network；
Step5.2, the input layer content for obtaining Step5.1 pass through one layer of hidden layer, then the output of hidden layer are input to
Final Question Classification result is obtained in sotfmax functions.
7. the question sentence file classification method according to claim 1 combined with depth model based on shallow Model, feature
It is：The depth model part is made of convolutional network layer and Recognition with Recurrent Neural Network layer；K-th of convolution window is long in convolutional layer
The Text Representation that the convolution nuclear convolution that degree is h obtains is wkh=[cki,…,ck(l-h+1)], wherein ckiIndicate k-th of convolution
The convolution feature of core i-th of position in question sentence text；cki=Relu (oki+ b), okiIndicate the value that convolutional calculation obtains；oki=
[xi,xi+1,…,xi+h-1]*fkh, wherein xiThe term vector of i-th of word in sentence is represented, h represents convolution kernel length of window, [xi,
xi+1,…,xi+h-1] represent in sentence the word from i-th of word to the i-th-h+1, the term vector matrix that total h word forms；fkhTable
Show that the convolution kernel that k-th of convolution length of window is h, * represent corresponding element multiplication sum operation in two matrixes；Convolutional layer is obtained
To feature vector rearrange combination then respectively input three different LSTM Recognition with Recurrent Neural Network layers, form final feature
Vector is expressed as V=[v2,v3,v4], wherein v2,v3,v4Convolution length of window 2,3,4 is indicated respectively；The input layer of entire model
It is spliced to form by the feature term vector of shallow-layer part and the output V of depth model, the vector for forming a m dimension indicates, X=
[wf1…wfn,V]。
8. the question sentence file classification method according to claim 6 combined with depth model based on shallow Model, feature
It is：The shallow Model final classification method is softmax functions.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810357603.5A CN108595602A (en) | 2018-04-20 | 2018-04-20 | The question sentence file classification method combined with depth model based on shallow Model |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810357603.5A CN108595602A (en) | 2018-04-20 | 2018-04-20 | The question sentence file classification method combined with depth model based on shallow Model |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN108595602A true CN108595602A (en) | 2018-09-28 |
Family
ID=63613629
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201810357603.5A Pending CN108595602A (en) | 2018-04-20 | 2018-04-20 | The question sentence file classification method combined with depth model based on shallow Model |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN108595602A (en) |
Cited By (21)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109271493A (en) * | 2018-11-26 | 2019-01-25 | 腾讯科技（深圳）有限公司 | A kind of language text processing method, device and storage medium |
| CN109739956A (en) * | 2018-11-08 | 2019-05-10 | 第四范式（北京）技术有限公司 | Corpus cleaning method, device, equipment and medium |
| CN109857852A (en) * | 2019-01-24 | 2019-06-07 | 安徽商贸职业技术学院 | A kind of the screening judgment method and system of electric business online comment training set feature |
| CN109871904A (en) * | 2019-03-11 | 2019-06-11 | 广东工业大学 | Inscriptions on bones or tortoise shells word identification model and training method, system, equipment, computer media |
| CN109918507A (en) * | 2019-03-08 | 2019-06-21 | 北京工业大学 | One kind being based on the improved file classification method of TextCNN |
| CN110009027A (en) * | 2019-03-28 | 2019-07-12 | 腾讯科技（深圳）有限公司 | Comparison method, device, storage medium and the electronic device of image |
| CN110046233A (en) * | 2019-02-12 | 2019-07-23 | 阿里巴巴集团控股有限公司 | Problem distributing method and device |
| CN110110372A (en) * | 2019-04-09 | 2019-08-09 | 华东师范大学 | A kind of user's timing behavior automatic segmentation prediction technique |
| CN110245353A (en) * | 2019-06-20 | 2019-09-17 | 腾讯科技（深圳）有限公司 | Natural language representation method, device, equipment and storage medium |
| CN110298036A (en) * | 2019-06-06 | 2019-10-01 | 昆明理工大学 | A kind of online medical text symptom identification method based on part of speech increment iterative |
| CN110309860A (en) * | 2019-06-06 | 2019-10-08 | 昆明理工大学 | The method classified based on grade malignancy of the convolutional neural networks to Lung neoplasm |
| CN110442720A (en) * | 2019-08-09 | 2019-11-12 | 中国电子技术标准化研究院 | A kind of multi-tag file classification method based on LSTM convolutional neural networks |
| CN110516070A (en) * | 2019-08-28 | 2019-11-29 | 上海海事大学 | A kind of Chinese Question Classification method based on text error correction and neural network |
| CN110991161A (en) * | 2018-09-30 | 2020-04-10 | 北京国双科技有限公司 | Similar text determination method, neural network model obtaining method and related device |
| CN111382244A (en) * | 2018-12-29 | 2020-07-07 | 深圳市优必选科技有限公司 | Deep retrieval matching classification method and device and terminal equipment |
| TWI717826B (en) * | 2019-02-13 | 2021-02-01 | 開曼群島商創新先進技術有限公司 | Method and device for extracting main words through reinforcement learning |
| CN112992356A (en) * | 2021-03-30 | 2021-06-18 | 太原理工大学 | Heart failure prediction method and device based on convolutional layer feature rearrangement and SVM |
| CN112989052A (en) * | 2021-04-19 | 2021-06-18 | 北京建筑大学 | Chinese news text classification method based on combined-convolutional neural network |
| CN113553844A (en) * | 2021-08-11 | 2021-10-26 | 四川长虹电器股份有限公司 | Domain identification method based on prefix tree features and convolutional neural network |
| CN113869458A (en) * | 2021-10-21 | 2021-12-31 | 成都数联云算科技有限公司 | Training method of text classification model, text classification method and related device |
| WO2024045247A1 (en) * | 2022-08-31 | 2024-03-07 | 福建天甫电子材料有限公司 | Production management and control system for ammonium fluoride production and control method therefor |
Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN101320374A (en) * | 2008-07-10 | 2008-12-10 | 昆明理工大学 | Field question classification method combining syntax structural relationship and field characteristic |
| CN105912528A (en) * | 2016-04-18 | 2016-08-31 | 深圳大学 | Question classification method and system |
| CN107038480A (en) * | 2017-05-12 | 2017-08-11 | 东华大学 | A kind of text sentiment classification method based on convolutional neural networks |
| CN107608999A (en) * | 2017-07-17 | 2018-01-19 | 南京邮电大学 | A kind of Question Classification method suitable for automatically request-answering system |
| CN107832312A (en) * | 2017-01-03 | 2018-03-23 | 北京工业大学 | A kind of text based on deep semantic discrimination recommends method |
-
2018
- 2018-04-20 CN CN201810357603.5A patent/CN108595602A/en active Pending
Patent Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN101320374A (en) * | 2008-07-10 | 2008-12-10 | 昆明理工大学 | Field question classification method combining syntax structural relationship and field characteristic |
| CN105912528A (en) * | 2016-04-18 | 2016-08-31 | 深圳大学 | Question classification method and system |
| CN107832312A (en) * | 2017-01-03 | 2018-03-23 | 北京工业大学 | A kind of text based on deep semantic discrimination recommends method |
| CN107038480A (en) * | 2017-05-12 | 2017-08-11 | 东华大学 | A kind of text sentiment classification method based on convolutional neural networks |
| CN107608999A (en) * | 2017-07-17 | 2018-01-19 | 南京邮电大学 | A kind of Question Classification method suitable for automatically request-answering system |
Non-Patent Citations (2)
| Title |
|---|
| CHUHAN WU 等: ""THU NGN at NAACL-2018 Metaphor Shared Task: Neural Metaphor Detecting with CNN-LSTM Model"", 《HTTPS://WWW.RESEARCHGATE.NET》 * |
| RONG ZHANG 等: ""Deep and Shallow Model for Insurance Churn Prediction Service"", 《2017 IEEE 14TH INTERNATIONAL CONFERENCE ON SERVICES COMPUTING》 * |
Cited By (31)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110991161A (en) * | 2018-09-30 | 2020-04-10 | 北京国双科技有限公司 | Similar text determination method, neural network model obtaining method and related device |
| CN110991161B (en) * | 2018-09-30 | 2023-04-18 | 北京国双科技有限公司 | Similar text determination method, neural network model obtaining method and related device |
| CN109739956A (en) * | 2018-11-08 | 2019-05-10 | 第四范式（北京）技术有限公司 | Corpus cleaning method, device, equipment and medium |
| CN109739956B (en) * | 2018-11-08 | 2020-04-10 | 第四范式（北京）技术有限公司 | Corpus cleaning method, apparatus, device and medium |
| CN109271493A (en) * | 2018-11-26 | 2019-01-25 | 腾讯科技（深圳）有限公司 | A kind of language text processing method, device and storage medium |
| CN111382244B (en) * | 2018-12-29 | 2023-04-14 | 深圳市优必选科技有限公司 | Deep retrieval matching classification method and device and terminal equipment |
| CN111382244A (en) * | 2018-12-29 | 2020-07-07 | 深圳市优必选科技有限公司 | Deep retrieval matching classification method and device and terminal equipment |
| CN109857852A (en) * | 2019-01-24 | 2019-06-07 | 安徽商贸职业技术学院 | A kind of the screening judgment method and system of electric business online comment training set feature |
| CN109857852B (en) * | 2019-01-24 | 2021-02-23 | 安徽商贸职业技术学院 | Method and system for screening and judging characteristics of E-commerce online comment training set |
| CN110046233A (en) * | 2019-02-12 | 2019-07-23 | 阿里巴巴集团控股有限公司 | Problem distributing method and device |
| TWI717826B (en) * | 2019-02-13 | 2021-02-01 | 開曼群島商創新先進技術有限公司 | Method and device for extracting main words through reinforcement learning |
| CN109918507A (en) * | 2019-03-08 | 2019-06-21 | 北京工业大学 | One kind being based on the improved file classification method of TextCNN |
| CN109871904A (en) * | 2019-03-11 | 2019-06-11 | 广东工业大学 | Inscriptions on bones or tortoise shells word identification model and training method, system, equipment, computer media |
| CN110009027A (en) * | 2019-03-28 | 2019-07-12 | 腾讯科技（深圳）有限公司 | Comparison method, device, storage medium and the electronic device of image |
| CN110110372A (en) * | 2019-04-09 | 2019-08-09 | 华东师范大学 | A kind of user's timing behavior automatic segmentation prediction technique |
| CN110110372B (en) * | 2019-04-09 | 2023-04-18 | 华东师范大学 | Automatic segmentation prediction method for user time sequence behavior |
| CN110309860A (en) * | 2019-06-06 | 2019-10-08 | 昆明理工大学 | The method classified based on grade malignancy of the convolutional neural networks to Lung neoplasm |
| CN110298036A (en) * | 2019-06-06 | 2019-10-01 | 昆明理工大学 | A kind of online medical text symptom identification method based on part of speech increment iterative |
| CN110298036B (en) * | 2019-06-06 | 2022-07-22 | 昆明理工大学 | Online medical text symptom identification method based on part-of-speech incremental iteration |
| CN110245353A (en) * | 2019-06-20 | 2019-09-17 | 腾讯科技（深圳）有限公司 | Natural language representation method, device, equipment and storage medium |
| CN110245353B (en) * | 2019-06-20 | 2022-10-28 | 腾讯科技（深圳）有限公司 | Natural language expression method, device, equipment and storage medium |
| CN110442720A (en) * | 2019-08-09 | 2019-11-12 | 中国电子技术标准化研究院 | A kind of multi-tag file classification method based on LSTM convolutional neural networks |
| CN110516070A (en) * | 2019-08-28 | 2019-11-29 | 上海海事大学 | A kind of Chinese Question Classification method based on text error correction and neural network |
| CN112992356B (en) * | 2021-03-30 | 2022-04-26 | 太原理工大学 | Heart failure prediction method and device based on convolutional layer feature rearrangement and SVM |
| CN112992356A (en) * | 2021-03-30 | 2021-06-18 | 太原理工大学 | Heart failure prediction method and device based on convolutional layer feature rearrangement and SVM |
| CN112989052B (en) * | 2021-04-19 | 2022-03-08 | 北京建筑大学 | Chinese news long text classification method based on combination-convolution neural network |
| CN112989052A (en) * | 2021-04-19 | 2021-06-18 | 北京建筑大学 | Chinese news text classification method based on combined-convolutional neural network |
| CN113553844A (en) * | 2021-08-11 | 2021-10-26 | 四川长虹电器股份有限公司 | Domain identification method based on prefix tree features and convolutional neural network |
| CN113553844B (en) * | 2021-08-11 | 2023-07-25 | 四川长虹电器股份有限公司 | Domain identification method based on prefix tree features and convolutional neural network |
| CN113869458A (en) * | 2021-10-21 | 2021-12-31 | 成都数联云算科技有限公司 | Training method of text classification model, text classification method and related device |
| WO2024045247A1 (en) * | 2022-08-31 | 2024-03-07 | 福建天甫电子材料有限公司 | Production management and control system for ammonium fluoride production and control method therefor |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN108595602A (en) | The question sentence file classification method combined with depth model based on shallow Model | |
| CN109376242B (en) | Text classification method based on cyclic neural network variant and convolutional neural network | |
| Rajayogi et al. | Indian food image classification with transfer learning | |
| CN105975573B (en) | A kind of file classification method based on KNN | |
| CN108829818A (en) | A kind of file classification method | |
| CN110442684A (en) | A kind of class case recommended method based on content of text | |
| CN110188272B (en) | Community question-answering website label recommendation method based on user background | |
| CN107239529A (en) | A kind of public sentiment hot category classification method based on deep learning | |
| CN110083700A (en) | A kind of enterprise's public sentiment sensibility classification method and system based on convolutional neural networks | |
| CN107516110A (en) | A kind of medical question and answer Semantic Clustering method based on integrated convolutional encoding | |
| CN107169035A (en) | A kind of file classification method for mixing shot and long term memory network and convolutional neural networks | |
| CN110222163A (en) | A kind of intelligent answer method and system merging CNN and two-way LSTM | |
| CN107291822A (en) | The problem of based on deep learning disaggregated model training method, sorting technique and device | |
| CN109635108A (en) | A kind of remote supervisory entity relation extraction method based on human-computer interaction | |
| CN107766324A (en) | A kind of text coherence analysis method based on deep neural network | |
| CN110321563A (en) | Text emotion analysis method based on mixing monitor model | |
| CN108121702A (en) | Mathematics subjective item reads and appraises method and system | |
| CN108052504A (en) | Mathematics subjective item answers the structure analysis method and system of result | |
| CN110413791A (en) | File classification method based on CNN-SVM-KNN built-up pattern | |
| CN109101584A (en) | A kind of sentence classification improved method combining deep learning with mathematical analysis | |
| CN110263174A (en) | - subject categories the analysis method based on focus | |
| CN110298036A (en) | A kind of online medical text symptom identification method based on part of speech increment iterative | |
| CN110188195A (en) | A kind of text intension recognizing method, device and equipment based on deep learning | |
| CN107480141A (en) | It is a kind of that allocating method is aided in based on the software defect of text and developer's liveness | |
| Qian | Exploration of machine algorithms based on deep learning model and feature extraction |
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 | ||
| RJ01 | Rejection of invention patent application after publication | ||
| RJ01 | Rejection of invention patent application after publication |
Application publication date: 20180928 |