CN108984523A - A kind of comment on commodity sentiment analysis method based on deep learning model - Google Patents
A kind of comment on commodity sentiment analysis method based on deep learning model Download PDFInfo
- Publication number
- CN108984523A CN108984523A CN201810695687.3A CN201810695687A CN108984523A CN 108984523 A CN108984523 A CN 108984523A CN 201810695687 A CN201810695687 A CN 201810695687A CN 108984523 A CN108984523 A CN 108984523A
- Authority
- CN
- China
- Prior art keywords
- vector
- comment
- dictionary
- text
- word
- 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
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
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
- G06N3/084—Backpropagation, e.g. using gradient descent
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Artificial Intelligence (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- Health & Medical Sciences (AREA)
- Evolutionary Computation (AREA)
- Computational Linguistics (AREA)
- General Health & Medical Sciences (AREA)
- Life Sciences & Earth Sciences (AREA)
- Biophysics (AREA)
- Biomedical Technology (AREA)
- Audiology, Speech & Language Pathology (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Evolutionary Biology (AREA)
- Information Retrieval, Db Structures And Fs Structures Therefor (AREA)
Abstract
The present invention relates to natural language processing fields, in particular to a kind of comment on commodity sentiment analysis method based on deep learning model, the comment on commodity data of crawl, and by comment on commodity data a star and two stars evaluation be labeled as actively commenting on, four stars and five-pointed star evaluation are labeled as negative comments, comment on commodity data are divided into training set and test set, and are pre-processed;Emotional Factors dictionary collection and affective characteristics vector are constructed, the word sequence and affective characteristics vector pre-processed according to training set obtains term vector, and multiple term vectors connect to form text vector;Dynamic convolution neural network model is constructed, updates the network parameter of dynamic convolutional neural networks by BP algorithm and stochastic gradient descent algorithm using text vector as training object, finally obtain sentiment classification model and Emotion tagging is carried out to test set;Present invention combination dynamic convolutional neural networks can be improved disaggregated model generalization ability, can be realized preferable classifying quality.
Description
Technical field
The present invention relates to natural language processing field, in particular to a kind of comment on commodity emotion based on deep learning model
Analysis method.
Background technique
With the high speed development of internet and e-commerce, many e-commerce websites (such as Taobao, capital are emerged on network
East, Amazon etc.).There are a general character in these websites, are exactly also to provide one while product sale for consumer and deliver
The platform of product review.Consumer can timely show oneself to the evaluation of commodity, these comment informations can be fed back to
Businessman and potential consumer.Businessman can according to feedback improvements commodity or adjustment sales tactics, potential consumer by reference to
Correctly shopping selection is made in others' comment to commodity.But internet information enormous amount, it is commented in face of so huge
It is all very time-consuming laborious by browsing or being concluded by artificial rule one by one by text information, using artificial intelligence technology with
And the relevant technologies of natural language processing field can be largely to the Sentiment orientation progress depth automatic mining of comment text
Upper improvement and promotion user experience.Exactly because product review excavation has powerful practical application meaning and important science to grind
Study carefully value, sentiment analysis (sentiment analysis) becomes the focus of more and more researchers and industry.
The sentiment analysis method of mainstream is roughly divided into two classes at present.The first is rule-based method, main elder generation's basis
Sentiment dictionary finds out the emotion word occurred in text, then carries out simple feeling polarities statistics, according to final score and in advance
The threshold value of setting, which is compared, obtains feeling polarities conclusion.Second is the method based on conventional machines study, by a large amount of
The training of corpus is marked, emotion classifiers are generated, for classifying to test text.And rule-based method text easy to be lost
Important model under covering in notebook data, and as language phenomenon is increasing, the covering surface of this method can be more and more narrow,
It is difficult to construct more complete sentiment dictionary or relevant collocation rule, causes accuracy rate low.Meanwhile being learnt based on conventional machines
Method building sentiment classification model is mainly combined by artificial extraction feature and machine learning algorithm, however, artificial extract
Feature needs the domain knowledge of expert and a large amount of manpower and material resources, system suitability poor.
Summary of the invention
In order to solve the above-mentioned technical problem, the present invention proposes a kind of comment on commodity sentiment analysis based on deep learning model
Method, comprising:
S1, by comment on commodity data a star and two stars evaluation be labeled as actively commenting on, four stars and five-pointed star evaluation mark
For negative comments, comment on commodity data are divided into training set and test set, and are pre-processed;
S2, Emotional Factors dictionary collection and affective characteristics vector are constructed, the word sequence that is pre-processed according to training set and
Affective characteristics vector obtains term vector, and multiple term vectors connect to form text vector;
S3, building dynamic convolution neural network model pass through BP algorithm and boarding steps using text vector as training object
The network parameter that descent algorithm updates dynamic convolutional neural networks is spent, sentiment classification model is finally obtained and feelings are carried out to test set
Sense mark.
Preferably, comment on commodity data are divided into training set and test set, and carry out pretreatment and includes:
S11, it is used as training set by the 80% of five folding cross validations selection comment on commodity data, remaining 20% conduct test
Collection;
S12, comment on commodity data are segmented using stammerer participle tool;
S13, remove irrelevant information in comment on commodity data, including: punctuation mark, excess space, repeat it is useless
Individual character and spcial character.
Preferably, step S2 includes:
S21, Emotional Factors dictionary collection, including positive sentiment dictionary, negative sense sentiment dictionary, negative dictionary, turnover are constructed
Dictionary and opinion dictionary;
S22, the word containing Emotional Factors is captured using Emotional Factors dictionary, and pass through the affective characteristics vector pair of building word
Five kinds of Emotional Factors that Emotional Factors dictionary is concentrated carry out modeling expression;
S23, term vector sequence and affective characteristics vector are stitched together, form term vector M (wi);
S24, multiple term vector M (wi) it is spliced to form text vector S, text vector S is indicated are as follows:
Wherein,For concatenation operator, M (wi:i+j) indicate by term vector M (wi)~M (wi+j) composition feature vector square
Battle array, n indicate the length of text vector.
Preferably, step S22 includes:
S221, word w is being constructediAffective characteristics vector E (wi) when, initialization E (w firsti) it is null vector, it is expressed as E
(wi)=[0,0,0,0,0];
S222, by word wiAffective characteristics vector E (wi) in element from left to right respectively indicate in Emotional Factors dictionary
Positive sentiment dictionary, negative sense sentiment dictionary, common negative dictionary, common turnover dictionary, common advocate dictionary;
S223, by word wiIt is matched respectively with the word of five class sentiment dictionaries in Emotional Factors dictionary, successful match
When by the corresponding affective characteristics vector E (w of the dictionaryi) in element be assigned a value of 1, obtain word wiFive dimension affective characteristics vector E
(wi)=(e1,e2,e3,e4,e5)。
Preferably, step S3 includes:
S31, input layer: using text vector S as dynamic convolutional neural networks input feature vector;
S32, convolutional layer: in order to capture different features, convolution behaviour is carried out to input feature vector using m convolution filter
Make, obtains final convolution results;
S33, judge whether comprising adversative in text vector S, if comprising being split to the emotion word before and after adversative
And multistage retains maximum value to carry out segmentation pond;Otherwise k maximum pond method is utilized;
S34, using be segmented pond after feature vector P or k maximum pond after feature vector V as softmax function
Input carries out emotional semantic classification.
Preferably, the final convolution results C is indicated are as follows:
C=[c1,1,…,c1,n-h+1;c21,…,c2,n-h+1;cM, 1,…,cm,n-h+1],C∈Rm×(n-h+1);
cji=f (ωj·M(wi:i+h-1)+b);
Wherein, cjiIt indicates to pass through j-th of convolution filter ωjIth feature value in obtained characteristic pattern, n indicate text S
Length, M (wi:i+h-1) indicating term vector corresponding to i-th of word to the i-th+h-1 words in text, b is bias term, f ()
For activation primitive, h is convolution filter ωjConvolution kernel sliding window length, m indicate Feature Mapping quantity.
Preferably, if the segmentation pond in step S3 includes: Feature Mapping output cjTwo parts are divided by adversative,
Pond is carried out respectively to characteristic value front and rear part, this procedural representation are as follows:
pji=max (cji);
Wherein, 1≤j≤m and 1≤i≤n;pjiIndicate that Feature Mapping exports cjSegmentation pond, connect all pjiIt is formed and is divided
Section Text eigenvector P, m indicate the quantity of Feature Mapping.
Preferably, the k maximum pond method in step S3 includes: to be taken in all characteristic values before score using k maximum pond method
ktopFeature, and retain ktopSequencing after a feature convolution;Maximum pond number, dynamic are adjusted according to kinematic function
Function k is indicated are as follows:
Pond is carried out to final convolution results C using kinematic function k, k maximum pondization indicates are as follows:
Wherein, ktopFor top layer maximum pond number in neural network, l indicates the number of plies serial number of current convolutional layer, and L indicates volume
Total number of plies of lamination,For downward floor operation, connection is ownedForm k maximum value Text eigenvector V.
Preferably, step S34 includes: to make the feature vector V behind the feature vector P and k maximum pond behind segmentation pond
Emotional semantic classification is carried out for the input of softmax function, is indicated are as follows:
p(y|P,Wp,bp)=softmaxy(Wp·P+bp);
p(y|V,Wv,bv)=softmaxy(Wv·P+bv);
Wherein, {+1, -1 } y ∈, and actively comment is indicated when y is+1, negative comments are indicated when y is -1;WpIt indicates
The weight of feature vector P after being segmented pond, WvThe weight of feature vector V after indicating k maximum pond;bpIndicate k maximum pond
The corresponding bias term of feature vector V afterwards, bvThe corresponding bias term of feature vector V after indicating k maximum pond.
Present invention introduces Emotional Factors dictionaries, have fully considered influence of the Emotional Factors to sentiment analysis, have improved classification
Performance;Can be improved disaggregated model generalization ability in conjunction with dynamic convolutional neural networks, can be realized preferable classifying quality and
Improve model generalization ability.
Detailed description of the invention
Fig. 1 is a kind of flow chart of the comment on commodity sentiment analysis method based on deep learning model of the present invention;
Fig. 2 is the CBOW model structure in the word2vec term vector model that the present invention uses.
Specific embodiment
In order to more clearly explain the embodiment of the invention or the technical proposal in the existing technology, right below in conjunction with attached drawing
Technical solution in the embodiment of the present invention is clearly and completely described, described embodiment is only that a part of the invention is real
Example is applied, instead of all the embodiments.Based on the embodiments of the present invention, those of ordinary skill in the art are not paying creativeness
Every other embodiment obtained, shall fall within the protection scope of the present invention under the premise of labour.
The present invention proposes a kind of comment on commodity sentiment analysis method based on deep learning model, such as Fig. 1, comprising:
S1, by comment on commodity data a star and two stars evaluation be labeled as actively commenting on, four stars and five-pointed star evaluation mark
For negative comments, comment on commodity data are divided into training set and test set, and are pre-processed;
S2, Emotional Factors dictionary collection and affective characteristics vector are constructed, the word sequence that is pre-processed according to training set and
Affective characteristics vector obtains term vector, and multiple term vectors connect to form text vector;
S3, building dynamic convolution neural network model pass through BP algorithm and boarding steps using text vector as training object
The network parameter that descent algorithm updates dynamic convolutional neural networks is spent, sentiment classification model is finally obtained and feelings are carried out to test set
Sense mark.
In the implementation of the present invention, give up Samsung evaluation, comment on commodity data are divided into training set and test set, and
Carrying out pretreatment includes:
S11, it is used as training set by the 80% of five folding cross validations selection comment on commodity data, remaining 20% conduct test
Collection;
S12, comment on commodity data are segmented using stammerer participle tool;
S13, remove irrelevant information in comment on commodity data, including: punctuation mark, excess space, repeat it is useless
Individual character and spcial character.
The present invention is trained pretreated sequence of terms using Word2Vec, and wherein word2vec includes two cores
Heart model, i.e., continuous bag of words (Continuous Bag of Word, CBOW) model and Skip-gram model, the present invention
Use CBOW model, CBOW model such as Fig. 2.Assuming that existing sequence of terms is [wi-2,wi-1,wi,wi+1,wi+2];CBOW model refers to
Be in known current word wiContext [wi-2,wi-1,wi,wi+1,wi+2] under the premise of, predict current word wi.For not landed
Word use is uniformly distributed U (- 0.01,0.01) and carrys out random initializtion term vector, and the dimension of the vector can be configured as needed,
Word wiWord sequence be expressed as V (wi), if the sequence is k dimension real vector, it may be expressed as: V (wi)=(x1,x2,…xk)。
Preferably, step S2 includes:
S21, Emotional Factors dictionary collection, including positive sentiment dictionary, negative sense sentiment dictionary, negative dictionary, turnover are constructed
Dictionary and opinion dictionary;
S22, the word containing Emotional Factors is captured using Emotional Factors dictionary, and pass through the affective characteristics vector pair of building word
Five kinds of Emotional Factors that Emotional Factors dictionary is concentrated carry out modeling expression;
S23, term vector sequence and affective characteristics vector are stitched together, form term vector M (wi);
S24, multiple term vector M (wi) it is spliced to form text vector S, text vector S is indicated are as follows:
Wherein,For concatenation operator, M (wi:i+j) indicate by term vector M (wi)~M (wi+j) composition feature vector square
Battle array, n indicate the length of text vector S;Due to term vector sequence V (wi) latitude be k, affective characteristics vector E (wi) latitude be
5, therefore term vector M (wi) dimension be d=k+5, if the length of text vector be n, can be indicated with the codomain of text vector
For S ∈ Rn×d, R·Indicate dimension real number field.
For the present invention in order to capture the word in text containing Emotional Factors, analyzing influences wanting for comment text feeling polarities
Element, and Emotional Factors dictionary collection is constructed, Emotional Factors dictionary collection of the invention includes positive sentiment dictionary, negative sense emotion word
Allusion quotation, turnover dictionary, advocates the five class dictionaries such as dictionary wherein at negative dictionary, and positive sentiment dictionary and negative sense sentiment dictionary use
Positive emotion word, negative sense emotion word, positive evaluating word in " sentiment analysis with word collection (beta editions) " of HowNet and negative
4 emotion word collection of evaluating word, negative dictionary can be summed up according to the negative word often occurred in comment text without, do not have, be non-,
, be not, never, or not rather, may not, not, not, don't, not, be not desired to, not wait 15 negative words, turnover dictionary can also root
Summed up according to the observation analysis to comment text but, still but wilfully, can and but however 9 turnovers such as only
Word advocates that dictionary using the opinion word collection in " sentiment analysis with word collection (beta editions) " of HowNet, provides one in table 1
One example of sentiment dictionary.
1 emotion word dictionary example of table
Dictionary title | Emotion word example |
Positive sentiment dictionary | It praises, commend, speaking approvingly of, praising, praising, congratulating, respectfully |
Negative sense sentiment dictionary | Sadness is pitied, is grieved, sad, ridicules, ridicules, ridiculing |
It negate dictionary | Without, do not have, be non-, or not, never, rather, may not |
Turnover dictionary | But still but wilfully, can and but however |
Advocate dictionary | Discover, is striking, hear of, realize, find, feel, feeling |
For word wiWhen constructing affective characteristics vector, first by word wiAffective characteristics vector E (wi) it is initialized as E
(wi)=[0,0,0,0,0], by word wiAffective characteristics vector E (wi) in element from left to right respectively indicate Emotional Factors
Positive sentiment dictionary, negative sense sentiment dictionary in dictionary, common turnover dictionary, common advocate dictionary at common negative dictionary;Then
It will be by word wiIt is matched respectively with the word of five class sentiment dictionaries in Emotional Factors dictionary, by the dictionary pair when successful match
The affective characteristics vector E (w answeredi) in element be assigned a value of 1, obtain word wiFive dimension affective characteristics vector E (wi)=(e1,e2,
e3,e4,e5), if positive emotion word is Dpos, negative sense emotion word be Dneg, common negative word be Drev, common adversative be Dtur, often
It is D with opinion wordcla, then word wiAffective characteristics vector in five elements can respectively indicate are as follows:
Preferably, step S3 includes:
S31, input layer: using text vector S as dynamic convolutional neural networks input feature vector;
S32, convolutional layer: in order to capture different features, convolution behaviour is carried out to input feature vector using m convolution filter
Make, obtains final convolution results C;
S33, judge whether comprising adversative in text vector S, if comprising being split to the emotion word before and after adversative
And multistage retains maximum value to carry out segmentation pond, obtains segmentation Text eigenvector P;Otherwise k maximum pond method is utilized, is obtained
K maximum value Text eigenvector V;
S34, will segmentation Text eigenvector P or k maximum value Text eigenvector V as the input of softmax function into
Row emotional semantic classification.
In building dynamic convolution neural network model during being trained on the training set for being derived from the data set,
Using text vector S as dynamic convolutional neural networks input feature vector, in convolutional layer, in order to capture different features, pass through jth
The procedural representation of ith feature value in the characteristic pattern that a convolution filter obtains are as follows:
cji=f (ωj·M(wi:i+h-1)+b);
Wherein, h is the convolution kernel sliding window length of convolution filter, M (wi:i+h-1) indicate in text i-th of word to the
Term vector corresponding to i+h-1 word, cjiFor ith feature value in the characteristic pattern that is obtained by j-th of convolution filter, 1≤i
≤ n-h+1, b are bias term, and f (x) is activation primitive.
In the present embodiment, in order to accelerate the convergence rate of model, activation primitive uses Softplus-Relu activation primitive
Carry out training pattern, activation primitive f (x) may be expressed as:
F (x)=(1- α) log (1+ex)+α max (0, x), α ∈ [0,1];
Therefore, x is the variable in function, enables x=ω in the present embodimentj·M(wi:i+h-1)+b, use m convolutional filtering
Device ωj∈Rh×d(1≤j≤m) carries out convolution operation to input feature vector, obtains final convolution results C, indicates are as follows:
C=[c1,1,…,c1,n-h+1;c21,…,c2,n-h+1;cM, 1,…,cm,n-h+1],C∈Rm×(n-h+1);
Wherein, Rm×(n-h+1)Indicate the codomain of C.
It is exactly that pondization operation is carried out to the characteristic pattern of convolutional layer output, i.e. down-sampling operation, this can be with after convolution algorithm
It reduces calculation amount and prevents over-fitting.Common pond method includes: maximum pond method (max pooling), average Chi Huafa
(average pooling), segmentation Chi Huafa (chunk-max pooling), k maximum pond method (k-max pooling);This
Invention according to the extraction whether optimized comprising adversative during pond to notable feature in text, if in the text comprising turn
Roll over word, then by the emotion word before and after adversative be split and multistage retain maximum value come carry out segmentation pond;If in the text
Not comprising adversative, then k maximum pond method is utilized.
For the sentence comprising adversative, the emotion word before and after adversative is split and multistage retains maximum value, if
Feature Mapping exports cjTwo parts are divided by adversative, pond are carried out respectively to characteristic value front and rear part, this procedural representation
Are as follows:
pji=max (cji);
Wherein, 1≤j≤m and 1≤i≤n;pjiIndicate that Feature Mapping exports cjSegmentation pond, connect all pjiIt is formed and is divided
Section Text eigenvector P, m indicate the quantity of Feature Mapping.
For not including the sentence of adversative, k before score is taken in all characteristic values using k maximum pond methodtopFeature,
And retain ktopSequencing after a feature convolution;Maximum pond number is adjusted according to kinematic function, kinematic function k is indicated
Are as follows:
Wherein, ktopFor top layer maximum pond number in neural network, l indicates the number of plies serial number of current convolutional layer, and L indicates volume
Total number of plies of lamination,For downward floor operation.
Vector after most terminal cistern may be expressed as:
It connects all in text vectorK maximum value Text eigenvector V is formed, m indicates the quantity of Feature Mapping;Its
The codomain of middle k maximum value Text eigenvector V is expressed as V ∈ Rkm。
In neural network, using be segmented pond after feature vector P or k maximum pond after feature vector V as
The input of softmax function carries out emotional semantic classification, indicates are as follows:
p(y|P,Wp,bp)=softmaxy(Wp·P+bp);
p(y|V,Wv,bv)=softmaxy(Wv·P+bv);
Wherein, {+1, -1 } y ∈, and when y indicates actively comment for+1, when y is -1 expression negative comments;Wp∈RP,Wv∈
RV;bpIndicate the bias term in vector P, bvIndicate the bias term in vector V.
It is entirely training, is being trained using Adam batch gradient decline (Mini-batch Gradient Descent) algorithm,
Each layer parameter is adjusted with BP algorithm, over-fitting is prevented using Dropout and L2 canonical when training.
It completes neural network after training, among test set input model, the word in test set can be carried out
Emotion tagging.
Those of ordinary skill in the art will appreciate that all or part of the steps in the various methods of above-described embodiment is can
It is completed with instructing relevant hardware by program, which can be stored in a computer readable storage medium, storage
Medium may include: ROM, RAM, disk or CD etc..
Embodiment provided above has carried out further detailed description, institute to the object, technical solutions and advantages of the present invention
It should be understood that embodiment provided above is only the preferred embodiment of the present invention, be not intended to limit the invention, it is all
Any modification, equivalent substitution, improvement and etc. made for the present invention, should be included in the present invention within the spirit and principles in the present invention
Protection scope within.
Claims (9)
1. a kind of comment on commodity sentiment analysis method based on deep learning model characterized by comprising
S1, by comment on commodity data a star and two stars evaluation be labeled as actively commenting on, four stars and five-pointed star evaluation be labeled as disappearing
Comment on commodity data are divided into training set and test set, and pre-processed by pole comment;
S2, Emotional Factors dictionary collection and affective characteristics vector are constructed, the word sequence and emotion pre-processed according to training set
Feature vector obtains term vector, and multiple term vectors connect to form text vector;
S3, building dynamic convolution neural network model, pass through under BP algorithm and stochastic gradient using text vector as training object
The network parameter that algorithm updates dynamic convolutional neural networks is dropped, sentiment classification model is finally obtained and emotion mark is carried out to test set
Note.
2. a kind of comment on commodity sentiment analysis method based on deep learning model according to claim 1, feature exist
In, comment on commodity data are divided into training set and test set, and carry out pretreatment include:
S11, by five folding cross validations choose comment on commodity data 80% be used as training set, remaining 20% be used as test set;
S12, comment on commodity data are segmented using stammerer participle tool;
S13, remove irrelevant information in comment on commodity data, including: punctuation mark, repeats useless list at excess space
Word and spcial character, obtain word sequence.
3. a kind of comment on commodity sentiment analysis method based on deep learning model according to claim 1, feature exist
In step S2 includes:
S21, Emotional Factors dictionary collection, including positive sentiment dictionary, negative sense sentiment dictionary, negative dictionary, turnover dictionary are constructed
With opinion dictionary;
S22, the word containing Emotional Factors is captured using Emotional Factors dictionary, and by the affective characteristics vector of building word to emotion
Five kinds of Emotional Factors that element dictionary is concentrated carry out modeling expression;
S23, term vector sequence and affective characteristics vector are stitched together, form term vector M (wi);
S24, multiple term vector M (wi) it is spliced to form text vector S, text vector S is indicated are as follows:
Wherein,For concatenation operator, M (wi:i+j) indicate by term vector M (wi)~M (wi+j) composition eigenvectors matrix, n
Indicate the length of text vector S.
4. a kind of comment on commodity sentiment analysis method based on deep learning model according to claim 3, feature exist
In step S22 includes:
S221, word w is being constructediAffective characteristics vector E (wi) when, initialization E (w firsti) it is null vector, it is expressed as E (wi)
=[0,0,0,0,0];
S222, by word wiAffective characteristics vector E (wi) in element from left to right respectively indicate in Emotional Factors dictionary just
To sentiment dictionary, negative sense sentiment dictionary, common negative dictionary, common turnover dictionary, commonly uses and advocate dictionary;
S223, by word wiIt is matched respectively with the word of five class sentiment dictionaries in Emotional Factors dictionary, when successful match should
The corresponding affective characteristics vector E (w of dictionaryi) in element be assigned a value of 1, obtain word wiFive dimension affective characteristics vector E (wi)=
(e1,e2,e3,e4,e5)。
5. a kind of comment on commodity sentiment analysis method based on deep learning model according to claim 1, feature exist
In step S3 includes:
S31, the input layer that convolutional neural networks are inputted using text vector S as dynamic convolutional neural networks input feature vector;
S32, in the convolutional layer of neural network, convolution operation is carried out to input feature vector using m convolution filter, is obtained final
Convolution results C;
S33, in the pond layer of convolutional neural networks model, whether judge in text vector S comprising adversative, if comprising to turning
Emotion word before and after folding word is split and multistage retains maximum value to carry out segmentation pond, obtains segmentation Text eigenvector P;
Otherwise k maximum pond method is utilized, k maximum value Text eigenvector V is obtained;
S34, Text eigenvector P or k maximum value Text eigenvector V will be segmented as the input of softmax function progress feelings
Sense classification.
6. a kind of comment on commodity sentiment analysis method based on deep learning model according to claim 5, feature exist
In the final convolution results C is indicated are as follows:
C=[c1,1,…,c1,n-h+1;c21,…,c2,n-h+1;cM, 1,…,cm,n-h+1];
cji=f (ωj·M(wi:i+h-1)+b);
Wherein, cjiIt indicates to pass through j-th of convolution filter ωjIth feature value in obtained characteristic pattern, n indicate the length of text S
Degree, M (wi:i+h-1) indicating term vector corresponding to i-th of word to the i-th+h-1 words in text, b is bias term, and f () is sharp
Function living, h are convolution filter ωjConvolution kernel sliding window length, m indicate Feature Mapping quantity.
7. a kind of comment on commodity sentiment analysis method based on deep learning model according to claim 5, feature exist
In if the segmentation pond in step S33 includes: Feature Mapping output cjTwo parts are divided by adversative, before characteristic value
Part carries out pond, this procedural representation respectively afterwards are as follows:
pji=max (cji);
Wherein, 1≤j≤m and 1≤i≤n;pjiIndicate that Feature Mapping exports cjSegmentation pond, connect all pjiForm segmentation text
Eigen vector P, m indicate the quantity of Feature Mapping.
8. a kind of comment on commodity sentiment analysis method based on deep learning model according to claim 5, feature exist
It include: to take in all characteristic values k before score using k maximum pond method in, the k maximum pond method in step S33topFeature, and
And retain ktopSequencing between a feature convolution;Maximum pond number is adjusted according to kinematic function, kinematic function k is indicated are as follows:
Pond is carried out to final convolution results C using kinematic function k, k maximum pondization indicates are as follows:
Wherein, ktopFor top layer maximum pond number in neural network, l indicates the number of plies serial number of current convolutional layer, and L indicates convolutional layer
Total number of plies, n indicate text vector S length,For downward floor operation, connection is ownedIt is special to form k maximum pond text
Levy vector V.
9. a kind of comment on commodity sentiment analysis method based on deep learning model according to claim 5, feature exist
In step S34 includes: using segmentation pond Text eigenvector P or k maximum pond Text eigenvector V as softmax function
Input carry out emotional semantic classification, indicate are as follows:
p(y|P,Wp,bp)=softmaxy(Wp·P+bp);
p(y|V,Wv,bv)=softmaxy(Wv·P+bv);
Wherein, {+1, -1 } y ∈, and actively comment is indicated when y is+1, negative comments are indicated when y is -1;WpIndicate segmentation pond
The weight of feature vector P after change, WvIndicate the weight of k maximum pond Text eigenvector V;bpSpy after indicating k maximum pond
Levy the corresponding bias term of vector V, bvThe corresponding bias term of feature vector V after indicating k maximum pond.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201810695687.3A CN108984523A (en) | 2018-06-29 | 2018-06-29 | A kind of comment on commodity sentiment analysis method based on deep learning model |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201810695687.3A CN108984523A (en) | 2018-06-29 | 2018-06-29 | A kind of comment on commodity sentiment analysis method based on deep learning model |
Publications (1)
Publication Number | Publication Date |
---|---|
CN108984523A true CN108984523A (en) | 2018-12-11 |
Family
ID=64539080
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201810695687.3A Pending CN108984523A (en) | 2018-06-29 | 2018-06-29 | A kind of comment on commodity sentiment analysis method based on deep learning model |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN108984523A (en) |
Cited By (18)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN109684634A (en) * | 2018-12-17 | 2019-04-26 | 北京百度网讯科技有限公司 | Sentiment analysis method, apparatus, equipment and storage medium |
CN110059183A (en) * | 2019-03-22 | 2019-07-26 | 重庆邮电大学 | A kind of automobile industry User Perspective sensibility classification method based on big data |
CN110083785A (en) * | 2019-04-29 | 2019-08-02 | 清华大学 | The Sex, Age method of discrimination and device of record are searched for based on user |
CN110232123A (en) * | 2019-05-28 | 2019-09-13 | 第四范式(北京)技术有限公司 | The sentiment analysis method and device thereof of text calculate equipment and readable medium |
CN110297902A (en) * | 2019-05-22 | 2019-10-01 | 哈尔滨理工大学 | A kind of commodity evaluation analysis method based on multitask deep learning |
CN110442857A (en) * | 2019-06-18 | 2019-11-12 | 平安科技(深圳)有限公司 | Emotion intelligent determination method, device and computer readable storage medium |
CN111274402A (en) * | 2020-02-07 | 2020-06-12 | 南京邮电大学 | E-commerce comment emotion analysis method based on unsupervised classifier |
CN111309859A (en) * | 2020-01-21 | 2020-06-19 | 上饶市中科院云计算中心大数据研究院 | Scenic spot network public praise emotion analysis method and device |
CN111709267A (en) * | 2020-03-27 | 2020-09-25 | 吉林大学 | Electroencephalogram signal emotion recognition method of deep convolutional neural network |
CN111858939A (en) * | 2020-07-27 | 2020-10-30 | 上海五节数据科技有限公司 | Text emotion classification method based on context information and convolutional neural network |
CN111985216A (en) * | 2020-08-25 | 2020-11-24 | 武汉长江通信产业集团股份有限公司 | Emotional tendency analysis method based on reinforcement learning and convolutional neural network |
CN112017632A (en) * | 2020-09-02 | 2020-12-01 | 浪潮云信息技术股份公司 | Automatic conference record generation method |
CN112926737A (en) * | 2021-03-01 | 2021-06-08 | 创新奇智(上海)科技有限公司 | Model training method, data processing method and device and electronic equipment |
CN112988921A (en) * | 2019-12-13 | 2021-06-18 | 北京四维图新科技股份有限公司 | Method and device for identifying map information change |
CN113010667A (en) * | 2019-12-20 | 2021-06-22 | 王道维 | Training method for machine learning decision model by using natural language corpus |
CN113159831A (en) * | 2021-03-24 | 2021-07-23 | 湖南大学 | Comment text sentiment analysis method based on improved capsule network |
CN113705243A (en) * | 2021-08-27 | 2021-11-26 | 电子科技大学 | Emotion analysis method |
CN116340511A (en) * | 2023-02-16 | 2023-06-27 | 深圳市深弈科技有限公司 | Public opinion analysis method combining deep learning and language logic reasoning |
Citations (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN107025284A (en) * | 2017-04-06 | 2017-08-08 | 中南大学 | The recognition methods of network comment text emotion tendency and convolutional neural networks model |
-
2018
- 2018-06-29 CN CN201810695687.3A patent/CN108984523A/en active Pending
Patent Citations (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN107025284A (en) * | 2017-04-06 | 2017-08-08 | 中南大学 | The recognition methods of network comment text emotion tendency and convolutional neural networks model |
Non-Patent Citations (4)
Title |
---|
NAL KALCHBRENNER,EDWARD GREFENSTETTE,PHIL BLUNSOM: "《A Convolutional Neural Network for Modelling Sentences》", 《DEPARTMENT OF COMPUTER SCIENCE UNIVERSITY OF OXFORD》 * |
RIE JOHNSON,TONG ZHANG: "《Effective Use of Word Order for Text Categorization with Convolutional Neural Networds》", 《TO APPEAR IN NAACL HLT 2015》 * |
喻涛等: "利用动态多池卷积神经网络的情感分析模型", 《计算机科学与探索》 * |
鲁新新等: "L2-SVM下的短文本情感分类动态CNN模型", 《计算机应用与软件》 * |
Cited By (24)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN109684634A (en) * | 2018-12-17 | 2019-04-26 | 北京百度网讯科技有限公司 | Sentiment analysis method, apparatus, equipment and storage medium |
CN109684634B (en) * | 2018-12-17 | 2023-07-25 | 北京百度网讯科技有限公司 | Emotion analysis method, device, equipment and storage medium |
CN110059183A (en) * | 2019-03-22 | 2019-07-26 | 重庆邮电大学 | A kind of automobile industry User Perspective sensibility classification method based on big data |
CN110059183B (en) * | 2019-03-22 | 2022-08-23 | 重庆邮电大学 | Automobile industry user viewpoint emotion classification method based on big data |
CN110083785A (en) * | 2019-04-29 | 2019-08-02 | 清华大学 | The Sex, Age method of discrimination and device of record are searched for based on user |
CN110297902A (en) * | 2019-05-22 | 2019-10-01 | 哈尔滨理工大学 | A kind of commodity evaluation analysis method based on multitask deep learning |
CN110232123A (en) * | 2019-05-28 | 2019-09-13 | 第四范式(北京)技术有限公司 | The sentiment analysis method and device thereof of text calculate equipment and readable medium |
CN110442857A (en) * | 2019-06-18 | 2019-11-12 | 平安科技(深圳)有限公司 | Emotion intelligent determination method, device and computer readable storage medium |
CN110442857B (en) * | 2019-06-18 | 2024-05-10 | 平安科技(深圳)有限公司 | Emotion intelligent judging method and device and computer readable storage medium |
CN112988921A (en) * | 2019-12-13 | 2021-06-18 | 北京四维图新科技股份有限公司 | Method and device for identifying map information change |
CN113010667A (en) * | 2019-12-20 | 2021-06-22 | 王道维 | Training method for machine learning decision model by using natural language corpus |
CN111309859A (en) * | 2020-01-21 | 2020-06-19 | 上饶市中科院云计算中心大数据研究院 | Scenic spot network public praise emotion analysis method and device |
CN111274402A (en) * | 2020-02-07 | 2020-06-12 | 南京邮电大学 | E-commerce comment emotion analysis method based on unsupervised classifier |
CN111274402B (en) * | 2020-02-07 | 2022-09-23 | 南京邮电大学 | E-commerce comment emotion analysis method based on unsupervised classifier |
CN111709267B (en) * | 2020-03-27 | 2022-03-29 | 吉林大学 | Electroencephalogram signal emotion recognition method of deep convolutional neural network |
CN111709267A (en) * | 2020-03-27 | 2020-09-25 | 吉林大学 | Electroencephalogram signal emotion recognition method of deep convolutional neural network |
CN111858939A (en) * | 2020-07-27 | 2020-10-30 | 上海五节数据科技有限公司 | Text emotion classification method based on context information and convolutional neural network |
CN111985216A (en) * | 2020-08-25 | 2020-11-24 | 武汉长江通信产业集团股份有限公司 | Emotional tendency analysis method based on reinforcement learning and convolutional neural network |
CN112017632A (en) * | 2020-09-02 | 2020-12-01 | 浪潮云信息技术股份公司 | Automatic conference record generation method |
CN112926737A (en) * | 2021-03-01 | 2021-06-08 | 创新奇智(上海)科技有限公司 | Model training method, data processing method and device and electronic equipment |
CN113159831A (en) * | 2021-03-24 | 2021-07-23 | 湖南大学 | Comment text sentiment analysis method based on improved capsule network |
CN113705243A (en) * | 2021-08-27 | 2021-11-26 | 电子科技大学 | Emotion analysis method |
CN116340511A (en) * | 2023-02-16 | 2023-06-27 | 深圳市深弈科技有限公司 | Public opinion analysis method combining deep learning and language logic reasoning |
CN116340511B (en) * | 2023-02-16 | 2023-09-15 | 深圳市深弈科技有限公司 | Public opinion analysis method combining deep learning and language logic reasoning |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN108984523A (en) | A kind of comment on commodity sentiment analysis method based on deep learning model | |
CN109299268A (en) | A kind of text emotion analysis method based on dual channel model | |
CN105608477B (en) | Method and system for matching portrait with job position | |
CN105046515B (en) | Method and device for sorting advertisements | |
CN108763216A (en) | A kind of text emotion analysis method based on Chinese data collection | |
CN106383815A (en) | Neural network sentiment analysis method in combination with user and product information | |
CN109740154A (en) | A kind of online comment fine granularity sentiment analysis method based on multi-task learning | |
CN110929034A (en) | Commodity comment fine-grained emotion classification method based on improved LSTM | |
CN107391483A (en) | A kind of comment on commodity data sensibility classification method based on convolutional neural networks | |
CN109408823B (en) | A kind of specific objective sentiment analysis method based on multi-channel model | |
CN108710680A (en) | It is a kind of to carry out the recommendation method of the film based on sentiment analysis using deep learning | |
CN105183717B (en) | A kind of OSN user feeling analysis methods based on random forest and customer relationship | |
CN110245229A (en) | A kind of deep learning theme sensibility classification method based on data enhancing | |
CN108984530A (en) | A kind of detection method and detection system of network sensitive content | |
CN107609009A (en) | Text emotion analysis method, device, storage medium and computer equipment | |
CN107025284A (en) | The recognition methods of network comment text emotion tendency and convolutional neural networks model | |
CN109902177A (en) | Text emotion analysis method based on binary channels convolution Memory Neural Networks | |
CN110825850B (en) | Natural language theme classification method and device | |
CN106815192A (en) | Model training method and device and sentence emotion identification method and device | |
CN105022754A (en) | Social network based object classification method and apparatus | |
Hong et al. | How Good is the Crowd at “real” WSD? | |
CN108038205A (en) | For the viewpoint analysis prototype system of Chinese microblogging | |
CN110489523A (en) | A kind of fine granularity sentiment analysis method based on online shopping evaluation | |
CN109800413A (en) | Recognition methods, device, equipment and the readable storage medium storing program for executing of media event | |
CN108062421A (en) | A kind of extensive picture multiscale semanteme search 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 | ||
RJ01 | Rejection of invention patent application after publication |
Application publication date: 20181211 |
|
RJ01 | Rejection of invention patent application after publication |