CN109740154A - A kind of online comment fine granularity sentiment analysis method based on multi-task learning - Google Patents
A kind of online comment fine granularity sentiment analysis method based on multi-task learning Download PDFInfo
- Publication number
- CN109740154A CN109740154A CN201811598961.1A CN201811598961A CN109740154A CN 109740154 A CN109740154 A CN 109740154A CN 201811598961 A CN201811598961 A CN 201811598961A CN 109740154 A CN109740154 A CN 109740154A
- Authority
- CN
- China
- Prior art keywords
- text
- fine granularity
- coarseness
- emotion
- matrix
- 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.)
- Granted
Links
Abstract
The online comment fine granularity sentiment analysis method based on multi-task learning that the invention discloses a kind of, including text representation matrix sequentially input text emotion feature extractor, coarseness affective feature extraction device and fine granularity affective characteristics classifier and obtain fine granularity emotional semantic classification result；The extraction that text emotion feature extractor selects single layer CNN network to carry out text emotion information to the text representation matrix of input obtains emotion representing matrix, coarseness affective feature extraction device obtains coarseness affective characteristics vector using the extraction that multiple single layer CNN networks carry out coarseness affective characteristics to the emotion representing matrix of input, and fine granularity affective characteristics classifier carries out fine granularity emotional semantic classification using the full Connection Neural Network of multilayer to coarseness affective characteristics vector.The present invention has the advantages that classification is accurate, the training time is short, can be used for the sentiment analysis of Internet user's comment of multi-level more granularities, can be used for personalized recommendation, intelligent search or product feedback.
Description
Technical field
The invention belongs to natural language processing fields, are related to a kind of online comment fine granularity emotion based on multi-task learning
Analysis method, in particular to natural language sentiment analysis method, can be used for personalized recommendation, intelligent search or product feedback.
Background technique
With increasingly developed, online user's comment information number presentation blowout growth of e-commerce.It faces unstructured
And the huge text information of data volume, only carrying out information sifting by conventional method, not only workload is very heavy, but also is difficult
Timely and effectively obtain valuable information.How efficiently to automatically analyze and mention in time from huge user comment data
Taking viewpoint information therein, emotion information is the important subject of current text excavation applications.
The sentiment analysis of online comment is the feedback comments by analysis user after consumption to excavate user's feelings
Feel the technology of tendency.According to the angle of analysis, sentiment analysis can be divided into coarseness sentiment analysis and fine granularity sentiment analysis.
Wherein coarseness sentiment analysis refers to the total satisfactory grade according to comment and analysis user to product or consumption, without
Consider the Sentiment orientation of user for some properties of product or in a certain respect.Fine granularity sentiment analysis refers to according to user comment
User is analyzed to the satisfaction in terms of some of product or consumption.Such as the comment according to user to some restaurant, it can dig
User is excavated for the satisfaction of " attitude " in restaurant or " food mouthfeel " etc..The fine granularity emotion of online comment
Analysis has vital value for profound understanding businessman and user, excavation user feeling etc., and in internet row
Industry, which has, to be extremely widely applied.
Sentiment analysis is a subdivision research field of text mining, by with natural language processing, text analyzing and
Related computer technology extract automatically or classifying text in emotion.The existing fine granularity sentiment analysis skill based on deep learning
The general step of art is to carry out Sentiment orientation excavation respectively to each granularity: first according to fine-grained different to sentiment analysis times
Business is decomposed, by entire Task-decomposing at the sentiment analysis task in multiple granularities, and using sentiment analysis task as text
This classification task is handled, and is then carried out respectively using existing Text Mining Technology to the Sentiment orientation in each fine granularity
It excavates.But this method could only obtain efficient effect when fine granularity is fewer, in face of multi-level, various dimensions thin
Granularity sentiment analysis task, this problem of facing inefficient, low precision based on the method that granularity is excavated, which has limited bases
The practical application of sentiment analysis method is carried out in granularity.Sentiment analysis can regard a kind of special text categorization task, root as
Sentence is categorized under different emotional categories according to the content of statement.By means of the powerful ability in feature extraction of neural network,
Yoon Kim proposes a kind of file classification method Text-CNN based on convolutional neural networks, referring to " Kim
Y.Convolutional neural networks for sentence classification[J].arXiv preprint
arXiv:1408.5882,2014.".For fine granularity emotional semantic classification problem, Text-CNN can use one-dimensional convolutional Neural net
Network extracts the high-order feature of text, and the feature extracted is then inputted full Connection Neural Network, is classified.But this side
Method can not carry out while classify to multiple fine granularity emotions, equally be faced with poor efficiency, the risk of over-fitting.
Since the above-mentioned sentiment analysis method precision based on granularity is not high and inefficient, and to the emotion based on multitask point
The research of analysis method is still in blank.Therefore, a kind of more efficiently fine granularity sentiment analysis based on multi-task learning is studied
Method is the research emphasis of the art scientific and technical personnel.
Summary of the invention
It is an object of the invention to be directed to the deficiency of above-mentioned prior art, a kind of fine granularity based on multi-task learning is proposed
Sentiment analysis method reduces model complexity, extends the application range of multi-task learning to improve nicety of grading.
To achieve the above object, technical solution of the present invention includes the following:
A kind of online comment fine granularity sentiment analysis method based on multi-task learning, comprising:
Step 1: text data is segmented, is trained, being mapped and matrix construction obtains text representation matrix；
Step 2: text representation matrix sequentially inputs multitask emotional semantic classification network and obtains fine granularity emotional semantic classification result；
The multitask emotional semantic classification network include text emotion feature extractor, coarseness affective feature extraction device and
Fine granularity affective characteristics classifier；
Text emotion feature extractor selects single layer CNN network to carry out text emotion information to the text representation matrix of input
Extraction obtain emotion representing matrix, coarseness affective feature extraction device utilizes multiple single layers to the emotion representing matrix of input
The extraction that CNN network carries out coarseness affective characteristics obtains coarseness affective characteristics vector, fine granularity affective characteristics classifier pair
Coarseness affective characteristics vector carries out fine granularity emotional semantic classification using the full Connection Neural Network of multilayer.
Optionally, the text emotion feature extractor is that a variety of different size of convolution kernels are arranged in text representation square
Convolution is carried out in battle array, the Analysis On Multi-scale Features information of text is obtained, the Analysis On Multi-scale Features information of extraction is attached.
Optionally, the text emotion feature extractor by multiple convolution kernels it is of different sizes convolutional layer parallel connection and
At to extract the multiple dimensioned affective characteristics of text；
The input of text emotion feature extractor is text representation matrix, is exported as affective characteristics matrix；According to final
Intersect entropy loss and carries out backpropagation, the weight parameter of convolutional layer in training text emotion extractor.
Optionally, the coarseness affective feature extraction device, is combined similar fine granularity emotion to obtain coarse grain
Emotion is spent, the extraction of coarseness affective characteristics is carried out using multiple convolutional layers.
Optionally, each coarseness affective feature extraction device is connected parallel by multiple convolution kernels convolutional layer of different sizes
It connects, to extract multiple dimensioned affective characteristics of the text in this coarseness；
The input of coarseness affective feature extraction device is emotion representing matrix, export affective characteristics for corresponding coarseness to
Amount；Backpropagation is carried out according to final intersection entropy loss, the weight ginseng of convolutional layer in training coarseness affective feature extraction device
Number.
Optionally, for each fine granularity emotional semantic classification task, using the full Connection Neural Network of multilayer, to text in phase
The emotion in fine granularity is answered to classify.
Optionally, the input of fine granularity affective characteristics classifier is the affective characteristics vector of affiliated coarseness, is exported as text
Originally belong to the probability of each emotional category；Backpropagation, all particulates of training are carried out according to final cross entropy loss function
Spend the weight parameter of affective characteristics classifier.
Optionally, the participle of the text data: true user comment text data are carried out using participle tool
Participle, obtains text sequence；
Text data cleaning: the user comment text sequence after participle is subjected to data cleansing, is stopped according to pre-set
With the stop words in vocabulary removal text；
The training of term vector: the parameters such as selected word vector dimension carry out Chinese term vector using term vector embedded technology
Training, and all words are mapped as term vector；
Mapping of the word to number: the mapping dictionary of word and number is established, word2index dictionary is denoted as, by all words
Language is mapped as the continuous number since 1；
The construction of term vector matrix: according to text to number mapping relations construct term vector matrix, refer specifically to by
Corresponding term vector, is then put into matrix by line number of the number as matrix after mapping in sequence.Wherein term vector square
The 0th row of battle array corresponds to null vector；
The standardization of text data length: according to preset sentence length threshold value to all comment texts at
Reason: the textual number sequence for being greater than threshold value for length is given up beyond part；Zero padding for curtailment；
Text representation matrix: for a comment text data, first according to the word2index dictionary of foundation by word
It is mapped as corresponding number, and carries out the standardization of text size, then the line number by number as term vector matrix carries out
Index, obtains text representation matrix.
It optionally, further include trained multitask emotional semantic classification network, using the class label of all fine granularity emotions as target
Training pattern carries out network training using the objective function of Adam algorithm optimization multitask emotional semantic classification network.
Optionally, further include trained multitask emotional semantic classification network, specifically further include following steps:
1) for some fine granularity m, the polytypic cross entropy loss function of m-th of fine granularity emotion is calculated:
Wherein, LmRepresent intersection entropy loss of the comment data under m-th of fine granularity, yiWhether ∈ { 0,1 } represents neuron
Belong to the i-th class, N represents m-th of fine-grained emotional category number, piRepresent the probability that emotional category belongs to the i-th class；
2) the Integral cross entropy loss function of multitask fine granularity emotional semantic classification network is calculated:
Wherein, λmRepresent the weight that m-th of fine-grained loss is lost in overall network, λm=1/M, M are fine granularity
Number；
3) optimize following objective function using autoadapted learning rate optimization algorithm Adam, to update the ginseng of whole network
Number, until the value of L is less than 0.01:
4) step 3) is constantly repeated, until neural network convergence, or reaches pre-set the number of iterations.
The present invention has the advantage that compared with prior art
1, the present invention is a kind of method based on multi-task learning, can obtain in user comment multi-level, more simultaneously
The Sentiment orientation of degree, and the risk of model over-fitting can be reduced to similar tasks by carrying out feature extraction simultaneously, lift scheme
Nicety of grading；
2, the present invention can automatically extract out text compared to conventional method using the method based on depth convolutional network
High-order feature reduces manual features and extracts bring influence of noise and accuracy decline problem；
3, method of the present invention due to being extracted using layered characteristic, the feature that each layer network extracts have interpretable
Property, the stability and expansibility of model also have greatly improved.
4, the multitask convolutional network that the present invention is extracted using two based on layered characteristic compares the existing granularity that is based on and carries out
The depth model of classification, training time and predicted time all greatly reduce.
Detailed description of the invention
Fig. 1 is implementation flow chart of the invention；
Fig. 2 is the structure chart for the multitask fine granularity emotional semantic classification network that the present invention designs；
Fig. 3 is the structure chart of the single task fine granularity emotional semantic classification network based on Text-CNN；
Fig. 4 is the present invention and the training time based on Text-CNN method to compare figure.
Specific embodiment
The present invention is based on the fine granularity sentiment analysis methods of multi-task learning mainly to solve existing fine granularity emotional semantic classification
The problem that precision is low, time complexity is high.The present invention is the fine granularity emotional semantic classification network based on multi-task learning, multitask
Habit, which refers to, carries out combination learning to multiple tasks, can solve single task and learns the insufficient limitation of existing use of information,
The association between multiple training missions is sufficiently excavated, the generalization ability of model is improved.In fine granularity emotion proposed by the present invention point
Mainly include three submodules in class network: it is special to extract the emotion contained in text by text emotion feature extractor first
Sign, then extracts the affective characteristics under some coarseness by coarseness affective feature extraction device, finally by fine granularity feelings
Sense classifier obtains emotional category of the text under some fine granularity.Method proposed by the present invention can solve existing fine granularity feelings
Feel the problem that the precision classified is low, time complexity is high.
The sentiment analysis of online comment is the feedback comments by analysis user after consumption to excavate user's feelings
Feel the technology of tendency.According to the angle of analysis, sentiment analysis can be divided into coarseness sentiment analysis and fine granularity sentiment analysis.Its
Middle coarseness sentiment analysis refers to the total satisfactory grade according to comment and analysis user to product or consumption, without considering user couple
Sentiment orientation in some properties of product or in a certain respect.Fine granularity sentiment analysis, which refers to, analyzes user couple according to user comment
Satisfaction in terms of some of product or consumption.Such as the comment according to user to some restaurant, user couple can be excavated
In the satisfaction of " attitude " or " food mouthfeel " etc. in restaurant.The fine granularity sentiment analysis of online comment is for depth
Understand businessman and user quarter, excavate user feeling etc. has vital value, and has extremely extensively in internet industry
General application.
Multitask fine granularity emotional semantic classification network proposed by the present invention is by text emotion feature extractor, coarseness emotion
Feature extractor and fine granularity affective characteristics classifier composition.According to real feelings classification of the text in all fine granularities with
The class probability predicted calculates the intersection entropy loss in all fine granularities, carries out backpropagation, and joint training text emotion is special
Levy the weight parameter of extractor, coarseness affective feature extraction device and fine granularity affective characteristics classifier.
Implementation step are as follows: the 1) processing of text data: carry out the word cutting of text data, obtain text sequence；Data are clear
It washes, carries out stop words and operate；Utilize word2vec training term vector；The mapping of word to number is established, and constructs term vector
Matrix；The standardization of text data length handles the length of all training datas, obtains the training input of fixed length；Structure
Text representation matrix is made, the input of network is obtained；2) multitask emotional semantic classification network is constructed, and sets multitask emotional semantic classification net
The structure and parameter of network；Training multitask emotional semantic classification network carries out the network optimization using Adam algorithm, obtains final network ginseng
Number.3) input test data obtain emotional semantic classification result of the test data in all fine granularities.
In the present invention, the present embodiment will be used to realize any text definition of text emotion analysis for text data.Also,
The languages type of text data is not limited, for example, text data can be Chinese text or English text etc.；The present embodiment
The length of text data is not limited, for example, text data can be sentence text, be also possible to chapter text；The present embodiment
The type of text data is not limited, for example, target text can be a video display comment (abbreviation film review) or a comment on commodity,
The part text being also possible in speech draft, magazine article, literary works etc..
Step 1: the processing of text data
1a) the participle of text data: true user comment text data are subjected to Chinese point using Chinese word segmentation tool
Word obtains text sequence；
1b) text data cleans: the user comment text sequence after participle is carried out data cleansing.According to pre-set
The stop words in vocabulary removal text is deactivated, stop words is referred to without practical significance or the word unrelated with task.Such as
The Chinese symbol such as the words and ", " unrelated for sentiment analysis such as " ", " how ".
1c) the training of term vector: the parameters such as selected word vector dimension carry out Chinese term vector using word2vec technology
All words are mapped as term vector by training；As described herein in experiment, term vector dimension is set as 128, for each
Chinese word, after being handled using word2vec technology, the vector that can all obtain one 128 dimension indicates this word；
1d) mapping of the word to number: the mapping dictionary of word and number is established, word2index is denoted as, in all
Cliction language is mapped to the continuous number since 1；
1e) the construction of term vector matrix: term vector matrix is constructed according to the mapping relations of text to number, is referred specifically to
Using the number after mapping as the line number of matrix, then corresponding term vector is put into matrix in sequence.Wherein term vector
The 0th row of matrix corresponds to null vector；
1f) the standardization of text data length: since the comment text length of each user is different, and neural network needs
The input of fixed length is wanted, so handling according to preset sentence length threshold value all comment texts: big for length
Give up in the textual number sequence of threshold value beyond part；Zero padding for curtailment.Such as, it in experiment of the present invention, chooses
Average sentence length 1200 is used as threshold value, and 1200 sentence is greater than for length, is given up beyond part；For curtailment 1200
Sentence carry out zero padding.
1g) text representation matrix: for a comment text data, first according to the word2index word established in (1d)
Text is mapped as corresponding number by allusion quotation, and the standardization of data length is carried out according to (1f), then by number as word to
The line number of moment matrix is indexed, and obtains the representing matrix of text data.Such as, in experiment of the present invention, each is commented
Paper sheet, text representation matrix are all the matrix of 1200*128 size being made of corresponding term vector.
Step 2: fine-grained emotional semantic classification network of the building based on multitask
2a) establish text emotion feature extractor: selection single layer CNN network carries out the extraction of text emotion information.Setting
A variety of different convolution kernel sizes carry out convolution on text representation matrix, obtain the Analysis On Multi-scale Features information of text data, will
The feature of extraction is attached；
Establish text emotion feature extractor: selection single layer CNN network carries out the extraction of text emotion information.According to classification
The complexity of problem is arranged a variety of different size of convolution kernels and carries out convolution on text representation matrix, obtains text data
Multiple dimensioned affective characteristics, and by the text emotion feature of extraction carry out horizontal connection；Such as, in experiment of the present invention,
It is respectively [2,3,4,5] provided with four kinds of sizes, quantity is all 128 convolution kernel, for inputting the text table for being 1200*128
Show matrix, the emotion representing matrix under available four different scales, shape is all 1200*128, according to column dimension when splicing
Horizontal connection is carried out, the emotion representing matrix of 1200*512 is obtained；
Specifically, the one-dimensional convolutional neural networks of selection single layer carry out the extraction of text emotion information, four kinds of setting is different
Convolution kernel size carries out convolution on text representation matrix, obtains the Analysis On Multi-scale Features information of text data.Wherein four convolution
Core size is respectively set to 2,3,4 and 5, and each convolution kernel number is both configured to 128.Activation primitive is selected as ReLU piecewise linearity
Function, convolution step-length are set as 1.When convolution by the way of zero padding, keep the eigenmatrix dimension of output constant.It will finally mention
Four eigenmatrixes taken are attached, and obtain emotion representing matrix；
2b) establish coarseness affective feature extraction device: different fine-grained relationships are different.Such as user for
In the comment in dining room, in " whether traffic facilitates " and " whether be easy find " two fine granularities emotional category and comment in mention
And dining room " position " it is information-related.And " service " referred in " speed of serving " and comment is information-related.So by fine granularity
It is divided, similar fine granularity affective characteristics is combined, carry out coarseness affective characteristics using multiple single layer CNN networks
Extraction.
Different fine-grained relationships are different.Fine grit classification problem is sorted out first, is found associated
Then coarseness feature carries out the extraction of coarseness affective characteristics using multiple single layer CNN networks.Such as, experiment of the present invention
In, in the Sentiment orientation and comment under Sentiment orientation and fine granularity " apart from commercial circle distance " under fine granularity " whether traffic facilitates "
The feature of the coarseness " position " referred to is related.It can use the affective characteristics under CNN network extraction " traffic " coarseness, then
The emotion classifiers of fine granularity " whether traffic facilitates " and the emotion classifiers of fine granularity " apart from commercial circle distance " are inputted respectively, are obtained
To corresponding fine granularity Sentiment orientation.
For the emotion information of each coarseness, all with 2a) described in the obtained emotion of text emotion feature extractor
Representing matrix selects the one-dimensional convolutional neural networks of single layer to carry out the extraction of coarseness affective characteristics as input, is arranged four kinds
Different convolution kernel sizes carry out convolution on emotion representing matrix, obtain the Analysis On Multi-scale Features information of coarseness emotion.Wherein
Four convolution kernel sizes are respectively set to 2,3,4 and 5, and each convolution kernel number is both configured to 64.Activation primitive is selected as ReLU
Piecewise linear function, convolution step-length are set as 1.Four coarseness affective characteristics matrixes are finally carried out to horizontal connect in column dimension
It connects, obtains coarseness affective characteristics matrix.Global pool is carried out to coarseness affective characteristics matrix, obtains final coarseness feelings
Feel feature vector, other coarsenesses are also as above operated, the affective characteristics vector of all coarsenesses is obtained；
It 2c) establishes fine granularity emotion classifiers: for each fine granularity emotional semantic classification task, connecting mind entirely using multilayer
Through network, classify to emotion of the text in this fine granularity.
Specifically, for each fine granularity emotional semantic classification task, with the affective characteristics vector of the affiliated coarseness of fine granularity
As input, classified using two layers of full Connection Neural Network.First layer neuron number is set as 64, and activation primitive uses
ReLU function.Classifier over-fitting is prevented using Dropout technology, neuron dropout ratio is 0.5.Second layer neuron
Number is set as emotional category number.For example, there are four types of Sentiment orientations in each fine granularity in experiment of the invention.So
It is 4 that second layer neuron number, which is arranged,.Activation primitive is set as softmax function.
3, using the class label of all fine granularity emotions as target, training pattern optimizes multitask fine granularity emotional semantic classification
The objective function of network: using the objective function of Adam algorithm optimization network, user comment is finally obtained in all granularities
Sentiment orientation.
The present invention is the multitask depth convolutional network based on one-dimensional convolutional network construction, wherein each layer of convolutional network
It all include a variety of various sizes of convolution kernels, referring to " Convolutional Neural Networks for Sentence
Classification".Different convolution kernel sizes can extract multiple dimensioned text emotion feature.The neural network of different layers
The relevant emotion information feature of text granularity can be extracted.
Referring to Fig.1, specific embodiments of the present invention are as follows:
Embodiment one:
1, treatment process introduction:
Step 1, text segments, removes stop words.
Chinese word segmentation, which refers to, is cut into individual word one by one for Chinese sequence.It is carried out using stammerer participle jieba tool
Chinese word segmentation.Stop words refers in Chinese that " ", " how " etc. does not have the word of practical significance, can also add it manually
His stop words.According to pre-set deactivated vocabulary, this partial words is deleted from data set.
Step 2, training term vector, and carry out the mapping of word to number.
Sentiment analysis technology based on deep learning needs for all words to be expressed as the dense vector of low-dimensional.Used here as
Word2vec technology carries out the training of Chinese term vector.The parameter of word2vec is arranged: window size is set as 5, minimum word frequency
2 are set as, term vector dimension is set as 128.The word2vec technology is a kind of model for generating term vector, wherein
The word that frequency of occurrence is lower than minimum word frequency 2 in entire data set will be deleted from data set.
Establish the dictionary word2index that word maps one by one to number.Wherein number is the continuous number since 1.It will
Line number of the number as word in term vector matrix, establishes term vector matrix.Such as: word " dining room " corresponds to number 12, then
Corresponding term vector is the 12nd row in term vector matrix.0th behavior full 0 vector of term vector matrix corresponds to curtailment
Text sequence in completion placeholder.
Step 3, data normalization obtains input text representation matrix.
According to word2index dictionary, user comment all in data set is mapped as Serial No..Due to nerve net
The input requirements regular length of network, so choosing the average length of user comment as input length.For being more than average length
Serial No., retain the front portion of sequence, give up beyond part.For the sequence of curtailment input length, in sequence
Tail portion carries out mending 0 operation, reaches input length.In experiment of the present invention, sentence average length is 1200.
Using Serial No. as the line number of term vector matrix, list entries is converted into text representation matrix.Text representation
The term vector of one word of each behavior of matrix.
Step 4, the network parameter of each layer of fine granularity emotional semantic classification network based on multitask is set.
4a) set text emotion feature extractor: it includes four parallel one-dimensional convolutional layers.For inputting the text of network
This representing matrix, choosing convolution kernel size is respectively 2,3,4 and 5, and convolution nuclear volume is set as 128 4 kinds of different convolution sides
Formula carries out multiple dimensioned information extraction to the emotion information in text, and activation primitive is both configured to ReLU function, and using benefit
Zero completion mode guarantees that the dimension of output is constant, obtains emotion representing matrix.
4b) set coarseness affective feature extraction device: for each coarseness, input is all emotion representing matrix, if
The convolutional network for determining the second layer all includes four parallel one-dimensional convolutional layers, and convolution kernel size is respectively 2,3,4 and 5, convolution kernel
Number is both configured to 64.Four coarseness emotion matrixes are attached, final coarseness affective characteristics matrix is obtained.Using
The method of global pool obtains final coarseness affective characteristics vector；
It 4c) sets fine granularity affective characteristics classifier: for each fine granularity emotion, all selecting two layers of fully-connected network
Classify.First layer neuron number is set as 64, and prevents over-fitting using Dropout technology, and Dropout ratio is set
It is set to 0.5.Activation primitive is set as ReLU function.Second layer neuron number is set as classification number, in experiment of the present invention,
Second layer neuron number is set as 4.Activation primitive is chosen for softmax activation primitive.
It step 5, is input with user version representing matrix, using emotional category of the user in different fine granularities as target,
The training fine-grained emotional semantic classification network of multitask.
The fine-grained emotional semantic classification network of multitask is cascade and parallel connection to be carried out by multilayer convolutional network, and connect and connect entirely
Neural network composition is connect, training carries out in the following way:
5a) for some fine granularity m, m-th of polytypic loss function of granularity emotion is defined:
Wherein, LmRepresent intersection entropy loss of the comment data under m-th of granularity, yi∈ { 0,1 } represents whether neuron belongs to
In the i-th class, N represents the emotional category number of m-th of granularity, piRepresent the probability that emotional category belongs to the i-th class.
5b) define the whole loss function of multitask fine granularity emotional semantic classification network:
Wherein, LmRepresent intersection entropy loss of the comment data under m-th of granularity, λmRepresent the loss of m-th granularity whole
The weight of volume grid loss.In the present invention, λ is setm=1/M.Wherein the meaning of M is fine granularity number.Reality of the present invention
In testing, M=20. so, fine granularity emotional semantic classification network herein is a kind of network of multi-task learning.The damage of whole network
Mistake is obtained by the loss weighted sum of multiple points of tasks.
5c) by the formula (2) in formula (1) the substitution 5b in 5a)), following loss function is obtained, adaptive learning is utilized
Rate optimization algorithm Adam optimizes following loss function, until L value is less than 0.01 or the number of iterations is more than total algebra of setting
When stop optimization, to update the parameter of whole network:
It is 128 that batch data size batchsize, which is arranged, in step 6, and the total the number of iterations of network training was 15 generations, constantly repeats to walk
Rapid 5c), until neural network convergence, or reach pre-set the number of iterations.
Step 7 carries out step 1-3 for test data set, the data that obtain that treated, inputs trained nerve net
Network obtains emotional category of the user comment in all fine granularities.
2. simulated conditions and evaluation index:
This experiment is under 14.04 system of Intel (R) Xeon (R) CPU E5-2620v4@2.10GHz Ubuntu, 128G
Memory is based under Python3.5 environment using NVIDIA Tesla P40GPU accelerans network training
It is carried out on Tensorflow1.8.0, Keras2.2.0 operation platform.
Primary evaluation index has:
1. m-th of fine-grained classification accuracy Pm: each classification fine-grained for m-th, if sample belongs to
It is such, then referred to as positive class, otherwise referred to as negative class.For the i-th class, count in this fine grit classification result actually for the i-th class and
It is predicted as the number of samples of the i-th class, is calculated in all ratios for being predicted as accounting in the i-th class sample, this ratio is known as accuracy rate.
It is averaged to the accuracy rate of all categories, obtains the accuracy rate under this fine granularity, it may be assumed that
2. m-th of fine-grained classification recall rate Rm: i-th class fine-grained for m-th counts and classifies under this fine granularity
As a result practical in is positive sample and the number of samples for being predicted as positive sample, is calculated in all ratios for being actually positive and accounting in class sample
Example.This ratio is known as recall rate.It is averaged to the recall rate of all categories, obtains the recall rate of this granularity.
3. m-th of fine-grained F1m: for the i-th class of fine granularity m, calculate F1mi=2PmiRmi/(Pmi+Rmi), it is final thin
The F1 score of granularity m is
4. whole classification F1:Network is judged for the overall performance of fine grit classification task.
Here, M indicates overall particle size number, and N indicates m-th of fine-grained classification number.
3. emulation experiment content and result
Emulate the emulation of user comment emotional semantic classification of the based on multitask fine granularity emotional semantic classification network.
3.1. data introduction: this experiment is using the real user comment data collection of AIChallenger publication as experiment pair
As the evaluation object of data set is divided into two levels, and level one is coarseness evaluation object, is related in text " position ", " clothes
Business " etc. essential elements of evaluation, level two be fine-grained emotion object, such as " position " attribute in " whether traffic facilitates ", " whether
The fine granularities elements such as easy searching ".
Being specifically described as follows for data set, is shown in Table 1:
1 data set granularity level introduction of table
It can be seen that sharing 20 fine granularity essential elements of evaluation in data set.
There are four types of states for the Sentiment orientation of each fine granularity element: positive, negative sense, neutral, does not refer to.Sentiment orientation value and
Its meaning table of comparisons is as shown in table 2.
2 Sentiment orientation of table and its meaning
| Sentiment orientation value | -2 | -1 | 1 | 0 |
| Meaning | Positive emotion | Neutral emotion | Negative emotion | Emotion is not referred to |
Training sample shares 105000 in this data set, and test sample has 15000.
3.2. comparative experiments: the present invention is the fine granularity sensibility classification method based on multitask deep neural network, and existing
There is the network for carrying out fine granularity emotional semantic classification based on Text-CNN model to compare.Fig. 2 show multitask fine granularity emotion
The structure chart of sorter network, it can be seen that network proposed by the present invention includes three parts: text emotion feature extractor, coarse grain
Spend affective feature extraction device, fine granularity affective characteristics classifier.Emotion class of the text in all fine granularities can be exported simultaneously
Not.Fig. 3 show the network structure of the single task fine granularity emotional semantic classification network based on Text-CNN, this network every time can only
Export the emotional category in a fine granularity.For multiple fine granularity emotional semantic classification tasks, training is needed repeatedly just to obtain text
Originally the emotional category in all fine granularities.Classification accuracy, recall rate of classifying, the F1 fractional result of classification is respectively such as table 3, table
4 and table 5 shown in.
The each fine-grained classification accuracy of table 3
The each fine-grained classification recall rate of table 4
| Fine granularity number | 1 | 2 | 3 | 4 | 5 | 6 | 7 | 8 | 9 | 10 |
| Text-CNN | 0.6008 | 0.4396 | 0.6437 | 0.5233 | 0.7197 | 0.6701 | 0.6269 | 0.7220 | 0.6207 | 0.5983 |
| The present invention | 0.5963 | 0.4760 | 0.6584 | 0.4966 | 0.7391 | 0.6718 | 0.6383 | 0.7645 | 0.6615 | 0.6032 |
| Fine granularity number | 11 | 12 | 13 | 14 | 15 | 16 | 17 | 18 | 19 | 20 |
| Text-CNN | 0.6669 | 0.6800 | 0.6856 | 0.6335 | 0.6193 | 0.6191 | 0.4663 | 0.5983 | 0.5047 | 0.6014 |
| The present invention | 0.6514 | 0.7087 | 0.7283 | 0.6833 | 0.6443 | 0.6320 | 0.4652 | 0.6604 | 0.5394 | 0.6310 |
The each fine-grained F1 score of table 5
| Fine granularity number | 1 | 2 | 3 | 4 | 5 | 6 | 7 | 8 | 9 | 10 |
| Text-CNN | 0.6167 | 0.4643 | 0.6535 | 0.5456 | 0.7227 | 0.6887 | 0.6637 | 0.7244 | 0.6330 | 0.5962 |
| The present invention | 0.6120 | 0.4986 | 0.6651 | 0.5406 | 0.7484 | 0.6609 | 0.6582 | 0.7581 | 0.6853 | 0.5936 |
| Fine granularity number | 11 | 12 | 13 | 14 | 15 | 16 | 17 | 18 | 19 | 20 |
| Text-CNN | 0.6354 | 0.6924 | 0.6818 | 0.6558 | 0.6304 | 0.6251 | 0.4851 | 0.6380 | 0.5016 | 0.6138 |
| The present invention | 0.6786 | 0.7340 | 0.7275 | 0.7030 | 0.6623 | 0.6622 | 0.4988 | 0.6894 | 0.5472 | 0.6305 |
From experimental result as can be seen that the fine granularity sentiment analysis technology proposed in this paper based on multitask is better than traditional base
In the method for Text-CNN.In most of fine granularity, classification accuracy of the invention, recall rate, F1 score better than be based on
The method of Text-CNN progress emotional semantic classification.Especially in accuracy rate, accuracy rate score of the invention in all granularities all
Higher than the method based on Text-CNN.In recall rate, for fine granularity Isosorbide-5-Nitrae, on 11,17, recall rate of the invention is slightly low
In Text-CNN, but in other fine granularities, recall rate of the invention is significantly larger than Text-CNN method.Such as in fine granularity 14
On, recall rate ratio Text-CNN high about 0.05 of the invention.On each fine-grained F1 score, classification F1 of the invention points
Number is also higher than Text-CNN algorithm in most of granularities.These all absolutely prove the particulate proposed by the present invention based on multitask
Degree emotional semantic classification technology can extract the common affective characteristics between similar fine granularity, to prevent over-fitting, preferably carry out
Fine-grained emotional semantic classification.
Whole classification F1 score is as shown in table 6.
The average F1 score of all fine grit classifications of table 6
| Method | The present invention | Text-CNN |
| F1 score | 0.6477 | 0.6234 |
As can be seen from Table 6, algorithm classification performance proposed by the present invention is better than the method based on Text-CNN.
3.3. the training time compares: the present invention is the emotional semantic classification network based on multitask, can disposably be exported all
Emotional category in fine granularity, and fine granularity sensibility classification method based on Text-CNN need in each fine granularity individually into
Row emotional semantic classification, runing time are longer.The training time of the present invention and Text-CNN comparison such as Fig. 4.
From fig. 4, it can be seen that the training time of the multitask emotional semantic classification network proposed in the present invention is far smaller than based on
The fine granularity sensibility classification method of Text-CNN.It can be seen that not only classification performance is more preferable for method of the invention, but also when operation
Between it is shorter.
In conclusion the fine granularity based on multitask that the present invention is proposed for multi-level, more granularity emotional semantic classification tasks
Sensibility classification method is better than the existing single fine granularity sensibility classification method based on Text-CNN.
Claims (10)
1. a kind of online comment fine granularity sentiment analysis method based on multi-task learning characterized by comprising
Step 1: text data is segmented, is trained, being mapped and matrix construction obtains text representation matrix；
Step 2: text representation matrix sequentially inputs multitask emotional semantic classification network and obtains fine granularity emotional semantic classification result；
The multitask emotional semantic classification network includes text emotion feature extractor, coarseness affective feature extraction device and particulate
Spend affective characteristics classifier；
Text emotion feature extractor selects single layer CNN network to carry out mentioning for text emotion information to the text representation matrix of input
Emotion representing matrix is obtained, coarseness affective feature extraction device utilizes multiple single layer CNN nets to the emotion representing matrix of input
The extraction that network carries out coarseness affective characteristics obtains coarseness affective characteristics vector, and fine granularity affective characteristics classifier is to coarseness
Affective characteristics vector carries out fine granularity emotional semantic classification using the full Connection Neural Network of multilayer.
2. the online comment fine granularity sentiment analysis method according to claim 1 based on multi-task learning, feature exist
In the text emotion feature extractor is that a variety of different size of convolution kernels are arranged to be rolled up on text representation matrix
Product, obtains the Analysis On Multi-scale Features information of text, the Analysis On Multi-scale Features information of extraction is attached.
3. the online comment fine granularity sentiment analysis method according to claim 1 based on multi-task learning, feature exist
In the text emotion feature extractor is connected in parallel by multiple convolution kernels convolutional layer of different sizes, to extract text
This multiple dimensioned affective characteristics；
The input of text emotion feature extractor is text representation matrix, exports the emotion representing matrix for text；According to final
Intersection entropy loss carry out backpropagation, the weight parameter of convolutional layer in training text affective feature extraction device.
4. the online comment fine granularity sentiment analysis method according to claim 1,2 or 3 based on multi-task learning, special
Sign is that similar fine granularity emotion is combined to obtain coarseness emotion by the coarseness affective feature extraction device, benefit
The extraction of coarseness affective characteristics is carried out with multiple convolutional layers.
5. the online comment fine granularity sentiment analysis method according to claim 4 based on multi-task learning, feature exist
In each coarseness affective feature extraction device is connected in parallel by multiple convolution kernels convolutional layer of different sizes, to mention
Take multiple dimensioned affective characteristics of the text in this coarseness；
The input of coarseness affective feature extraction device is the emotion representing matrix of text, is exported as the affective characteristics of corresponding coarseness
Vector；Backpropagation is carried out according to final intersection entropy loss, the weight of convolutional layer in training coarseness affective feature extraction device
Parameter.
6. the online comment fine granularity sentiment analysis method according to claim 1,2 or 3 based on multi-task learning, special
Sign is, for each fine granularity emotional semantic classification task, using the full Connection Neural Network of multilayer, to text in corresponding fine granularity
On emotion classify.
7. the online comment fine granularity sentiment analysis method according to claim 6 based on multi-task learning, feature exist
In the input of fine granularity affective characteristics classifier is the affective characteristics vector of affiliated coarseness, exports and belongs to each for text
The probability of emotional category；Backpropagation, all fine granularity affective characteristics point of training are carried out according to final cross entropy loss function
The weight parameter of full articulamentum in class device.
8. the online comment fine granularity sentiment analysis method according to claim 1,2 or 3 based on multi-task learning, special
Sign is, the participle of the text data: segments true user comment text data using participle tool, obtains
Text sequence；
Text data cleaning: the user comment text sequence after participle is subjected to data cleansing, according to pre-set stop words
Table removes the stop words in text；
The training of term vector: the parameters such as selected word vector dimension carry out the training of Chinese term vector using term vector embedded technology,
And all words are mapped as term vector；
Mapping of the word to number: the mapping dictionary of word and number is established, word2index dictionary is denoted as, all words is reflected
It penetrates as the continuous number since 1；
The construction of term vector matrix: term vector matrix is constructed according to the mapping relations of text to number, refers specifically to map
Corresponding term vector, is then put into matrix by line number of the number afterwards as matrix in sequence.Wherein term vector matrix the 0th
Row corresponds to null vector；
The standardization of text data length: all comment texts are handled according to preset sentence length threshold value: right
Give up in the textual number sequence that length is greater than threshold value beyond part；Zero padding for curtailment；
Text representation matrix: for a comment text data, word is mapped according to the word2index dictionary of foundation first
For corresponding number, and the standardization of text size is carried out, then the line number by number as term vector matrix is indexed,
Obtain the representing matrix of text.
9. the online comment fine granularity sentiment analysis method according to claim 1,2 or 3 based on multi-task learning, special
Sign is, further includes trained multitask emotional semantic classification network, using the class label of all fine granularity emotions as target training pattern,
Network training is carried out using the objective function of Adam algorithm optimization multitask emotional semantic classification network.
10. the online comment fine granularity sentiment analysis method according to claim 1,2 or 3 based on multi-task learning,
It is characterized in that, further includes trained multitask emotional semantic classification network, specifically further include following steps:
1) for some fine granularity m, the polytypic cross entropy loss function of m-th of fine granularity emotion is calculated:
Wherein, LmRepresent intersection entropy loss of the comment data under m-th of fine granularity, yi∈ { 0,1 } represents whether neuron belongs to
I-th class, N represent m-th of fine-grained emotional category number, piRepresent the probability that emotional category belongs to the i-th class；
2) the Integral cross entropy loss function of multitask fine granularity emotional semantic classification network is calculated:
Wherein, λmRepresent the weight that m-th of fine-grained loss is lost in overall network.In the present invention, λm=1/M, M are thin
Granularity number；
3) optimize following objective function using autoadapted learning rate optimization algorithm Adam, to update the parameter of whole network, directly
To L value less than 0.01:
4) step 3) is constantly repeated, until neural network convergence, or reaches pre-set the number of iterations.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811598961.1A CN109740154B (en) | 2018-12-26 | 2018-12-26 | Online comment fine-grained emotion analysis method based on multi-task learning |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811598961.1A CN109740154B (en) | 2018-12-26 | 2018-12-26 | Online comment fine-grained emotion analysis method based on multi-task learning |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN109740154A true CN109740154A (en) | 2019-05-10 |
| CN109740154B CN109740154B (en) | 2021-10-26 |
Family
ID=66361226
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201811598961.1A Active CN109740154B (en) | 2018-12-26 | 2018-12-26 | Online comment fine-grained emotion analysis method based on multi-task learning |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN109740154B (en) |
Cited By (29)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110210027A (en) * | 2019-05-30 | 2019-09-06 | 杭州远传新业科技有限公司 | Fine granularity sentiment analysis method, apparatus, equipment and medium based on integrated study |
| CN110287318A (en) * | 2019-06-06 | 2019-09-27 | 秒针信息技术有限公司 | The detection method and device of business operation, storage medium, electronic device |
| 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 |
| CN110348877A (en) * | 2019-05-27 | 2019-10-18 | 上海大学 | IN service proposed algorithm based on big data, computer readable storage medium |
| CN110458233A (en) * | 2019-08-13 | 2019-11-15 | 腾讯云计算（北京）有限责任公司 | Combination grain object identification model training and recognition methods, device and storage medium |
| CN110473546A (en) * | 2019-07-08 | 2019-11-19 | 华为技术有限公司 | A kind of media file recommendation method and device |
| CN110472115A (en) * | 2019-08-08 | 2019-11-19 | 东北大学 | A kind of social networks text emotion fine grit classification method based on deep learning |
| CN110488973A (en) * | 2019-07-23 | 2019-11-22 | 清华大学 | A kind of virtual interactive message leaving system and method |
| CN110598786A (en) * | 2019-09-09 | 2019-12-20 | 京东方科技集团股份有限公司 | Neural network training method, semantic classification method and semantic classification device |
| CN110718211A (en) * | 2019-09-26 | 2020-01-21 | 东南大学 | Keyword recognition system based on hybrid compressed convolutional neural network |
| CN110888971A (en) * | 2019-11-29 | 2020-03-17 | 支付宝(杭州)信息技术有限公司 | Multi-round interaction method and device for robot customer service and user |
| CN110887240A (en) * | 2019-12-03 | 2020-03-17 | 美的集团股份有限公司 | Water heater temperature control method and device, water heater and electronic equipment |
| CN111078881A (en) * | 2019-12-09 | 2020-04-28 | 携程计算机技术（上海）有限公司 | Fine-grained emotion analysis method and system, electronic equipment and storage medium |
| CN111339752A (en) * | 2020-02-18 | 2020-06-26 | 哈尔滨工业大学 | Evaluation object-oriented emotion analysis method for multi-task joint learning |
| CN111353042A (en) * | 2020-02-27 | 2020-06-30 | 浙江大学 | Fine-grained text viewpoint analysis method based on deep multi-task learning |
| CN111401226A (en) * | 2020-03-13 | 2020-07-10 | 电子科技大学 | Rapid identification method for radiation source |
| CN111475615A (en) * | 2020-03-12 | 2020-07-31 | 哈尔滨工业大学(深圳)(哈尔滨工业大学深圳科技创新研究院) | Fine-grained emotion prediction method, device and system for emotion enhancement and storage medium |
| CN111814453A (en) * | 2020-07-09 | 2020-10-23 | 重庆大学 | Fine-grained emotion analysis method based on BiLSTM-TextCNN |
| 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 |
| CN112199501A (en) * | 2020-10-13 | 2021-01-08 | 华中科技大学 | Scientific and technological information text classification method |
| US11010564B2 (en) * | 2019-02-05 | 2021-05-18 | International Business Machines Corporation | Method for fine-grained affective states understanding and prediction |
| CN113127604A (en) * | 2021-05-10 | 2021-07-16 | 齐鲁工业大学 | Comment text-based fine-grained item recommendation method and system |
| CN113221551A (en) * | 2021-05-28 | 2021-08-06 | 复旦大学 | Fine-grained emotion analysis method based on sequence generation |
| CN113268667A (en) * | 2021-05-28 | 2021-08-17 | 汕头大学 | Chinese comment emotion guidance-based sequence recommendation method and system |
| CN113297352A (en) * | 2021-06-07 | 2021-08-24 | 苏州大学 | Attribute-level emotion classification method and device based on multitask network |
| CN113920313A (en) * | 2021-09-29 | 2022-01-11 | 北京百度网讯科技有限公司 | Image processing method, image processing device, electronic equipment and storage medium |
| CN114091026A (en) * | 2021-11-25 | 2022-02-25 | 云南电网有限责任公司信息中心 | Integrated learning-based network abnormal intrusion detection method and system |
| CN115862875A (en) * | 2023-02-27 | 2023-03-28 | 四川大学华西医院 | Postoperative pulmonary complication prediction method and system based on multi-type feature fusion |
| CN117390141A (en) * | 2023-12-11 | 2024-01-12 | 江西农业大学 | Agricultural socialization service quality user evaluation data analysis method |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102236636A (en) * | 2010-04-26 | 2011-11-09 | 富士通株式会社 | Method and device for analyzing emotional tendency |
| US20170344880A1 (en) * | 2016-05-24 | 2017-11-30 | Cavium, Inc. | Systems and methods for vectorized fft for multi-dimensional convolution operations |
| CN108038107A (en) * | 2017-12-22 | 2018-05-15 | 东软集团股份有限公司 | Sentence sensibility classification method, device and its equipment based on convolutional neural networks |
-
2018
- 2018-12-26 CN CN201811598961.1A patent/CN109740154B/en active Active
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102236636A (en) * | 2010-04-26 | 2011-11-09 | 富士通株式会社 | Method and device for analyzing emotional tendency |
| US20170344880A1 (en) * | 2016-05-24 | 2017-11-30 | Cavium, Inc. | Systems and methods for vectorized fft for multi-dimensional convolution operations |
| CN108038107A (en) * | 2017-12-22 | 2018-05-15 | 东软集团股份有限公司 | Sentence sensibility classification method, device and its equipment based on convolutional neural networks |
Non-Patent Citations (1)
| Title |
|---|
| 刘龙飞等: "基于卷积神经网络的微博情感倾向性分析", 《中文信息学报》 * |
Cited By (50)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US11132511B2 (en) | 2019-02-05 | 2021-09-28 | International Business Machines Corporation | System for fine-grained affective states understanding and prediction |
| US11010564B2 (en) * | 2019-02-05 | 2021-05-18 | International Business Machines Corporation | Method for fine-grained affective states understanding and prediction |
| 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 |
| CN110348877A (en) * | 2019-05-27 | 2019-10-18 | 上海大学 | IN service proposed algorithm based on big data, computer readable storage medium |
| CN110348877B (en) * | 2019-05-27 | 2023-11-14 | 上海大学 | Intelligent service recommendation algorithm based on big data and computer readable storage medium |
| CN110210027A (en) * | 2019-05-30 | 2019-09-06 | 杭州远传新业科技有限公司 | Fine granularity sentiment analysis method, apparatus, equipment and medium based on integrated study |
| CN110287318A (en) * | 2019-06-06 | 2019-09-27 | 秒针信息技术有限公司 | The detection method and device of business operation, storage medium, electronic device |
| CN110287318B (en) * | 2019-06-06 | 2021-09-17 | 秒针信息技术有限公司 | Service operation detection method and device, storage medium and electronic device |
| CN110473546B (en) * | 2019-07-08 | 2022-05-31 | 华为技术有限公司 | Media file recommendation method and device |
| CN110473546A (en) * | 2019-07-08 | 2019-11-19 | 华为技术有限公司 | A kind of media file recommendation method and device |
| CN110488973B (en) * | 2019-07-23 | 2020-11-10 | 清华大学 | Virtual interactive message leaving system and method |
| CN110488973A (en) * | 2019-07-23 | 2019-11-22 | 清华大学 | A kind of virtual interactive message leaving system and method |
| CN110472115B (en) * | 2019-08-08 | 2022-08-02 | 东北大学 | Social network text emotion fine-grained classification method based on deep learning |
| CN110472115A (en) * | 2019-08-08 | 2019-11-19 | 东北大学 | A kind of social networks text emotion fine grit classification method based on deep learning |
| CN110458233B (en) * | 2019-08-13 | 2024-02-13 | 腾讯云计算（北京）有限责任公司 | Mixed granularity object recognition model training and recognition method, device and storage medium |
| CN110458233A (en) * | 2019-08-13 | 2019-11-15 | 腾讯云计算（北京）有限责任公司 | Combination grain object identification model training and recognition methods, device and storage medium |
| CN110598786A (en) * | 2019-09-09 | 2019-12-20 | 京东方科技集团股份有限公司 | Neural network training method, semantic classification method and semantic classification device |
| US11934790B2 (en) | 2019-09-09 | 2024-03-19 | Boe Technology Group Co., Ltd. | Neural network training method and apparatus, semantic classification method and apparatus and medium |
| CN110598786B (en) * | 2019-09-09 | 2022-01-07 | 京东方科技集团股份有限公司 | Neural network training method, semantic classification method and semantic classification device |
| CN110718211A (en) * | 2019-09-26 | 2020-01-21 | 东南大学 | Keyword recognition system based on hybrid compressed convolutional neural network |
| CN110718211B (en) * | 2019-09-26 | 2021-12-21 | 东南大学 | Keyword recognition system based on hybrid compressed convolutional neural network |
| CN110888971B (en) * | 2019-11-29 | 2022-05-24 | 支付宝(杭州)信息技术有限公司 | Multi-round interaction method and device for robot customer service and user |
| CN110888971A (en) * | 2019-11-29 | 2020-03-17 | 支付宝(杭州)信息技术有限公司 | Multi-round interaction method and device for robot customer service and user |
| CN110887240A (en) * | 2019-12-03 | 2020-03-17 | 美的集团股份有限公司 | Water heater temperature control method and device, water heater and electronic equipment |
| CN111078881A (en) * | 2019-12-09 | 2020-04-28 | 携程计算机技术（上海）有限公司 | Fine-grained emotion analysis method and system, electronic equipment and storage medium |
| CN111078881B (en) * | 2019-12-09 | 2023-04-07 | 携程计算机技术（上海）有限公司 | Fine-grained sentiment analysis method and system, electronic equipment and storage medium |
| CN111339752A (en) * | 2020-02-18 | 2020-06-26 | 哈尔滨工业大学 | Evaluation object-oriented emotion analysis method for multi-task joint learning |
| CN111353042A (en) * | 2020-02-27 | 2020-06-30 | 浙江大学 | Fine-grained text viewpoint analysis method based on deep multi-task learning |
| CN111353042B (en) * | 2020-02-27 | 2023-06-30 | 浙江大学 | Fine granularity text viewpoint analysis method based on deep multitasking learning |
| CN111475615A (en) * | 2020-03-12 | 2020-07-31 | 哈尔滨工业大学(深圳)(哈尔滨工业大学深圳科技创新研究院) | Fine-grained emotion prediction method, device and system for emotion enhancement and storage medium |
| CN111475615B (en) * | 2020-03-12 | 2023-06-20 | 哈尔滨工业大学(深圳)(哈尔滨工业大学深圳科技创新研究院) | Fine granularity emotion prediction method, device and system for emotion enhancement and storage medium |
| CN111401226A (en) * | 2020-03-13 | 2020-07-10 | 电子科技大学 | Rapid identification method for radiation source |
| CN111401226B (en) * | 2020-03-13 | 2022-11-01 | 电子科技大学 | Rapid identification method for radiation source |
| CN111814453A (en) * | 2020-07-09 | 2020-10-23 | 重庆大学 | Fine-grained emotion analysis method based on BiLSTM-TextCNN |
| CN111814453B (en) * | 2020-07-09 | 2023-08-08 | 重庆大学 | Fine granularity emotion analysis method based on BiLSTM-textCNN |
| 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 |
| CN112199501A (en) * | 2020-10-13 | 2021-01-08 | 华中科技大学 | Scientific and technological information text classification method |
| CN112199501B (en) * | 2020-10-13 | 2024-03-19 | 华中科技大学 | Scientific and technological information text classification method |
| CN113127604B (en) * | 2021-05-10 | 2022-10-04 | 齐鲁工业大学 | Comment text-based fine-grained item recommendation method and system |
| CN113127604A (en) * | 2021-05-10 | 2021-07-16 | 齐鲁工业大学 | Comment text-based fine-grained item recommendation method and system |
| CN113268667A (en) * | 2021-05-28 | 2021-08-17 | 汕头大学 | Chinese comment emotion guidance-based sequence recommendation method and system |
| CN113221551A (en) * | 2021-05-28 | 2021-08-06 | 复旦大学 | Fine-grained emotion analysis method based on sequence generation |
| CN113297352A (en) * | 2021-06-07 | 2021-08-24 | 苏州大学 | Attribute-level emotion classification method and device based on multitask network |
| CN113297352B (en) * | 2021-06-07 | 2024-01-16 | 苏州大学 | Attribute-level emotion classification method and device based on multitasking network |
| CN113920313A (en) * | 2021-09-29 | 2022-01-11 | 北京百度网讯科技有限公司 | Image processing method, image processing device, electronic equipment and storage medium |
| CN114091026A (en) * | 2021-11-25 | 2022-02-25 | 云南电网有限责任公司信息中心 | Integrated learning-based network abnormal intrusion detection method and system |
| CN115862875A (en) * | 2023-02-27 | 2023-03-28 | 四川大学华西医院 | Postoperative pulmonary complication prediction method and system based on multi-type feature fusion |
| CN115862875B (en) * | 2023-02-27 | 2024-02-09 | 四川大学华西医院 | Postoperative pulmonary complications prediction method and system based on multi-type feature fusion |
| CN117390141A (en) * | 2023-12-11 | 2024-01-12 | 江西农业大学 | Agricultural socialization service quality user evaluation data analysis method |
| CN117390141B (en) * | 2023-12-11 | 2024-03-08 | 江西农业大学 | Agricultural socialization service quality user evaluation data analysis method |
Also Published As
| Publication number | Publication date |
|---|---|
| CN109740154B (en) | 2021-10-26 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109740154A (en) | A kind of online comment fine granularity sentiment analysis method based on multi-task learning | |
| CN109189925B (en) | Word vector model based on point mutual information and text classification method based on CNN | |
| CN109977232B (en) | Graph neural network visual analysis method based on force guide graph | |
| CN106815369B (en) | A kind of file classification method based on Xgboost sorting algorithm | |
| CN111563164B (en) | Specific target emotion classification method based on graph neural network | |
| CN109933670B (en) | Text classification method for calculating semantic distance based on combined matrix | |
| CN108038205B (en) | Viewpoint analysis prototype system for Chinese microblogs | |
| CN108460089A (en) | Diverse characteristics based on Attention neural networks merge Chinese Text Categorization | |
| CN109558487A (en) | Document Classification Method based on the more attention networks of hierarchy | |
| CN107301171A (en) | A kind of text emotion analysis method and system learnt based on sentiment dictionary | |
| CN110472042B (en) | Fine-grained emotion classification method | |
| CN110765260A (en) | Information recommendation method based on convolutional neural network and joint attention mechanism | |
| CN110245229A (en) | A kind of deep learning theme sensibility classification method based on data enhancing | |
| CN110083700A (en) | A kind of enterprise's public sentiment sensibility classification method and system based on convolutional neural networks | |
| CN109165383B (en) | Data aggregation, analysis, mining and sharing method based on cloud platform | |
| CN106445919A (en) | Sentiment classifying method and device | |
| CN108614875A (en) | Chinese emotion tendency sorting technique based on global average pond convolutional neural networks | |
| CN107292696A (en) | A kind of automobile intelligent purchase guiding system and implementation method | |
| CN110728153A (en) | Multi-category emotion classification method based on model fusion | |
| CN110297888B (en) | Domain classification method based on prefix tree and cyclic neural network | |
| CN111680225B (en) | WeChat financial message analysis method and system based on machine learning | |
| CN112925908A (en) | Attention-based text classification method and system for graph Attention network | |
| CN110717330A (en) | Word-sentence level short text classification method based on deep learning | |
| CN110046943B (en) | Optimization method and optimization system for network consumer subdivision | |
| CN112231477A (en) | Text classification method based on improved capsule network |
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 | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |