CN108845986A - A kind of sentiment analysis method, equipment and system, computer readable storage medium - Google Patents
A kind of sentiment analysis method, equipment and system, computer readable storage medium Download PDFInfo
- Publication number
- CN108845986A CN108845986A CN201810538689.1A CN201810538689A CN108845986A CN 108845986 A CN108845986 A CN 108845986A CN 201810538689 A CN201810538689 A CN 201810538689A CN 108845986 A CN108845986 A CN 108845986A
- Authority
- CN
- China
- Prior art keywords
- memory network
- network model
- length memory
- corpus
- information
- 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
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- General Engineering & Computer Science (AREA)
- Artificial Intelligence (AREA)
- Data Mining & Analysis (AREA)
- General Physics & Mathematics (AREA)
- Physics & Mathematics (AREA)
- Computational Linguistics (AREA)
- General Health & Medical Sciences (AREA)
- Health & Medical Sciences (AREA)
- Audiology, Speech & Language Pathology (AREA)
- Life Sciences & Earth Sciences (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Evolutionary Biology (AREA)
- Evolutionary Computation (AREA)
- Information Retrieval, Db Structures And Fs Structures Therefor (AREA)
Abstract
A kind of sentiment analysis method, equipment and system, computer readable storage medium, the sentiment analysis method include：Term vector is generated according to corpus；Feature vector is generated according to the corpus, the term vector is inputted to the first length memory network model pre-established, the first information of the first length memory network model output and described eigenvector are inputted into the second length memory network model pre-established；The Sentiment orientation of the corpus is determined according to the second information that the second length memory network model exports.The scheme that this example provides, uses two layers of LSTM, it is contemplated that the long-range correlation between sentence information can more accurately reflect Sentiment orientation.
Description
Technical field
The present invention relates to a kind of sentiment analysis method, equipment and systems, computer readable storage medium.
Background technique
21 century be information technology rapid development epoch, people's lives and computer, internet are closely bound up, people and
Exchange and conmmunication mode between people has also penetrated into network.Go out using microblogging, wechat as the social media platform of representative
It is existing, more mobile interchange is rooted in the hearts of the people.By taking microblogging as an example, microblogging, the i.e. abbreviation of micro-blog are one based on customer relationship
Information share, propagate and obtain platform, user can pass through web (webpage), WAP (Wireless Application
Protocol, Wireless Application Protocol) and various client components, information is delivered with the text of 140 words or so, and realize instant
Share.Brief, the penetrating use for having attracted a large amount of public figure of microblogging, the number of fans that these public figures are driven is with ten thousand
Meter.Everyone no matter when and where can freely and easily record the drop of life, interact with friend, express viewpoint etc.
Deng.The information of each microblogging transmitting includes the personal position and emotion of publisher, it is necessary to the participant of microblogging carrying
Emotion carries out mining analysis.
Summary of the invention
A present invention at least embodiment provides a kind of sentiment analysis method, equipment and system, computer-readable storage medium
Matter.
In order to reach the object of the invention, a present invention at least embodiment provides a kind of sentiment analysis method, including：
Term vector is generated according to corpus；
Feature vector is generated according to the corpus, the term vector is inputted to the first length memory network mould pre-established
Type, the second length that the first information of the first length memory network model output and described eigenvector input are pre-established
Short memory network model；
The Sentiment orientation of the corpus is determined according to the second information that the second length memory network model exports.
A present invention at least embodiment provides a kind of sentiment analysis system, including：Data processing module, memory module and calculation
Method analysis module, wherein：
The data processing module is used for, and obtains corpus；
The memory module is used for, and stores the corpus；
The algorithm analysis module is used for, and generates feature vector according to the corpus, and term vector input is built in advance
The first vertical length memory network model, by the first length memory network model output the first information and the feature to
The second length memory network model that amount input pre-establishes；According to the second letter of the second length memory network model output
Breath determines the Sentiment orientation of the corpus.
A kind of sentiment analysis equipment of a present invention at least embodiment, including memory and processor, the memory storage
There is program, described program realizes sentiment analysis method described in any embodiment when reading execution by the processor.
A kind of computer readable storage medium of a present invention at least embodiment, the computer-readable recording medium storage have
One or more program, one or more of programs can be executed by one or more processor, to realize any reality
Apply sentiment analysis method described in example.
Scheme provided in this embodiment using two layers of LSTM, and joined characteristic information, make compared to term vector is only used
It only considered the association (long association and short association) between word for the single layer LSTM of input, single layer LSTM, can not examine completely
Consider the long-range correlation between sentence information, scheme provided in this embodiment uses two layers of LSTM, it is contemplated that between sentence information
Long-range correlation can more accurately carry out sentiment analysis.
Other features and advantages of the present invention will be illustrated in the following description, also, partly becomes from specification
It obtains it is clear that understand through the implementation of the invention.The objectives and other advantages of the invention can be by specification, right
Specifically noted structure is achieved and obtained in claim and attached drawing.
Detailed description of the invention
Attached drawing is used to provide to further understand technical solution of the present invention, and constitutes part of specification, with this
The embodiment of application technical solution for explaining the present invention together, does not constitute the limitation to technical solution of the present invention.
Fig. 1 is the sentiment analysis method flow diagram that one embodiment of the invention provides；
Fig. 2 is the first LSTM cytological map that one embodiment of the invention provides；
Fig. 3 is the 2nd LSTM cytological map that one embodiment of the invention provides；
Fig. 4 is the sentiment analysis schematic diagram that one embodiment of the invention provides；
Fig. 5 is the two-way sentiment analysis schematic diagram that one embodiment of the invention provides；
Fig. 6 is the sentiment analysis system block diagram that one embodiment of the invention provides；
Fig. 7 is the sentiment analysis method flow diagram that one embodiment of the invention provides；
Fig. 8 is the training method flow chart that one embodiment of the invention provides；
Fig. 9 is the sentiment analysis method flow diagram that one embodiment of the invention provides；
Figure 10 is the sentiment analysis method flow diagram that one embodiment of the invention provides；
Figure 11 is the sentiment analysis method flow diagram that one embodiment of the invention provides；
Figure 12 is the sentiment analysis equipment block diagram that one embodiment of the invention provides.
Specific embodiment
To make the objectives, technical solutions, and advantages of the present invention clearer, below in conjunction with attached drawing to the present invention
Embodiment be described in detail.It should be noted that in the absence of conflict, in the embodiment and embodiment in the application
Feature can mutual any combination.
Step shown in the flowchart of the accompanying drawings can be in a computer system such as a set of computer executable instructions
It executes.Also, although logical order is shown in flow charts, and it in some cases, can be to be different from herein suitable
Sequence executes shown or described step.
In the application, feelings are carried out by two layers of length memory network (Long Short-Term Memory, abbreviation LSTM)
Sense analysis.
As shown in Figure 1, one embodiment of the invention provides a kind of sentiment analysis method, including：
Step 101, term vector is generated according to corpus；
Step 102, feature vector is generated according to the corpus, the term vector is inputted to the first LSTM mould pre-established
The first information of first LSTM model output and described eigenvector are inputted the 2nd LSTM model pre-established by type；
Step 103, the Sentiment orientation of the corpus is determined according to the second information that the 2nd LSTM model exports.
Scheme provided in this embodiment, using two layers of LSTM (the first LSTM and the 2nd LSTM), compared to only using term vector
Single layer LSTM, single layer LSTM only considered the association (long association and short association) between word, can not consider sentence completely
Long-range correlation between information, scheme provided in this embodiment, it is contemplated that the long-range correlation between sentence information, it can be more accurately
Carry out sentiment analysis.
Wherein, further include before the step 101：Step 100, the target information being analysed to is pre-processed into preset format
Data as the corpus.The target information such as internet data, the internet data such as microblogging text, microblogging
Comment, the comment of wechat public platform text, wechat, commodity evaluation, body, news comment etc..
In one embodiment, in step 100, the target information being analysed to, which is pre-processed into the data of preset format, includes：
The presupposed information for including in the target information is cleaned, the presupposed information is such as picture, voice etc., by the target after cleaning
The regular structural data for preset format of information, each structural data is as a corpus.The information for needing to wash can
With pre-defined.It is of course also possible to extract information without cleaning directly from target information and generate structural data.Wherein,
Can first it judge with the presence or absence of presupposed information in target information, if it is present cleaning presupposed information；Alternatively, not making to judge, directly
It connects and target information is cleaned.
A kind of format of structural data is as shown in table 1, it may include following field：
It identifies (id), which is the sequence number of corpus；
Corpus content, the field are the content of corpus, and the content of corpus includes the information such as emoticon.Corpus content and mesh
Marking information, there may be difference, for example, eliminating picture (non-emoticon), the voice etc. in target information in corpus content
Deng.
Corpus type indicates that the corpus includes at least one of：Text, the comment of text and comment reply；
Owner, the publisher of finger speech material, the field are Optional Field；
Time, the issuing time of finger speech material, the field are Optional Field.
1 structured data fields explanation of table
By taking microblogging as an example, a microblogging text can be used as a corpus after pretreatment, and every of the microblogging text
Comment can be used as a corpus after pretreatment, and the reply of comment can be used as a corpus after pretreatment.Certainly,
Can also by microblogging text and its comment be used as a corpus, alternatively, by microblogging comment on and comment on reply as a corpus,
Etc..
Wherein, corpus can store in the database, and a plurality of associated corpus can be stored by predetermined manner.For example, just
The corresponding corpus of text, the corresponding corpus of the comment of text, the corresponding corpus of the reply of comment are sequentially stored in continuous address.
Alternatively, increasing indication field in commenting on corresponding corpus, which can carry corresponding with the associated text of the comment
Corpus ID, indicate the comment be the corpus ID instruction the corresponding text of corpus comment, it is corresponding in the reply of comment
Increase indication field in corpus, which can carry the ID of the corresponding corpus of the associated comment of reply with the comment,
The reply for indicating the comment is the reply of the corresponding comment of corpus of corpus ID instruction.
It should be noted that above structure data are merely illustrative, can according to need as other structures, for example only include
Corpus content and corpus type.
It generates term vector and refers to the vector for converting word to computer capacity understanding.In one embodiment, the step 101
In, generating term vector according to corpus includes：
One or more first term vectors will be generated after corpus participle, using first term vector as institute's predicate to
Amount；
Or, by one or more first term vectors are generated after corpus participle, based on topic belonging to the corpus or
Classification generates the second term vector, combines the first term vector and second term vector to obtain the term vector.Wherein, by first
The combination that term vector and second term vector combine to obtain the term vector, which can be, splices the two, corpus institute
The topic of category or classification can be stored in advance.
For example, GloVe etc. generates term vector using word2vec；For example, I does not believe that after segmenting as " I " " no " " phase
Letter ", the term vector of " I " are [0.59 ..., 0.70 ...], and the term vector of " no " is [0.32 ..., 0.60 ...], " it is believed that "
Term vector is [0.19 ..., 0.55 ...].Corpus (usually sentence) is divided into independent word one by one by corpus participle.
For example, corpus " I does not believe that " is belonging to it entitled " society ", after participle for " I " " no " " it is believed that ", then general
The term vector of the term vector of " I " and " society " splices, that is, the term vector that splices [my term vector, social word to
Amount] input as the first LSTM, the term vector of the term vector of " no " and " society " splice as the defeated of the first LSTM
Enter, by " it is believed that " term vector and the term vector of " society " carry out input of the splicing as the first LSTM.
In one embodiment, in step 102, the characteristic information includes at least one of：The social comment feature of extraction,
Emotion characteristic feature and macro society feature, the social comment feature indicate in the comment information of the corpus with the presence or absence of the language
Expect the reply of publisher, the emotion characteristic feature indicates first kind emotional symbol in the corpus and its comment information and the
Two class emotional symbol quantitative relations；The macro society feature indicates whether occur that other users is prompted to check this in the corpus
The prompt user information of corpus.Wherein, the comment information of the corpus i.e. comment information of the corresponding target information of the corpus；With microblogging
For, the corpus obtained by text manipulation, comment information is to be directed to the comment information of the text.
Wherein it is possible to classify to emotional symbol (such as emoticon), first kind emotional symbol and the second class emotion
Symbol, the first kind are such as positive emotional symbol, for example passive emotional symbol of the second analogy, which specific emotional symbol is
Positive emotion symbol, which emotional symbol are that Negative Affect symbol can pre-define.
By taking microblogging as an example：
Feature is commented in social activity：It analyzes microblogging text and its replys comment information, whether there is microblogging in comment information
The reply of text publisher.Under normal circumstances, text publisher can adhere to his consistent Sentiment orientation.If there is microblogging text
The reply of publisher, then the social activity comment feature be 1, if there is no microblogging text publisher reply then social activity comment spy
Sign is 0.
Emotion characteristic feature is：Microblogging text and its comment information are analyzed, the emoticon in these information, reference are collected
Expression classification group (for example emoticon is divided into two classes, positive emoticon and passive emoticon), if positive emoticon
Number amount is more than passive emoticon quantity, then this feature is 1, is on the contrary then 0.It particularly, can be with if without emoticon
It is 0 that this feature, which is arranged,.
Macro society feature：Microblogging text sequence is analyzed, if occurred in different microblogging texts in analysis length threshold value
Microblogging specifically prompts user information "@user name ", prompts microblogging publisher user concern, then this feature is 1, is then on the contrary
0。
According to characteristic information construction feature vector, for example, such as social comment feature is 1, emotion characteristic feature is 1, macro
Seeing social characteristic is 0, then feature vector is [1,1,0].Certainly, social comment feature, emotion characteristic feature and macro society are special
Sign construction feature vector, the application can also be not construed as limiting this in other sequences.
It should be noted that the value of feature is 1 or 0 merely illustrative in the application, it can according to need and be taken as other
Value.Furthermore it is possible to extract more or less features as needed.In addition, the extraction of features described above information is merely illustrative, Ke Yigen
Characteristic information is generated according to other information relevant to Sentiment orientation.
In one embodiment, the term vector includes term vector x1(0)~x1(N-1)；It is in step 102, the term vector is defeated
Enter the first length memory network model pre-established, by the first information of the first length memory network model output and institute
Stating the second length memory network model that feature vector input pre-establishes includes：
By term vector x1(0)The first length memory network model is inputted, the first length memory network model is defeated
The first information and described eigenvector out inputs the second length memory network model；
By term vector x1(0)The first length memory network model is inputted, the first length memory network model is defeated
The first information and described eigenvector out inputs the second length memory network model；
And so on, by term vector x1(N-1)The first length memory network model is inputted, first length is remembered
The first information and described eigenvector of network model output input the second length memory network model, obtain described second
Second information of length memory network model output.
Specifically, assume the term vector generated according to corpus be it is N number of, be followed successively by x1(0)~x1(N-1), the characteristic information of corpus
For d；
By x1(0)The first LSTM model is inputted, h1 is exported0, by h10The first LSTM model, output are inputted with d
h20；
By x1(1)And h10The first LSTM model is inputted, h1 is exported1, by h11、d、h20Input the 2nd LSTM mould
Type exports h21；
And so on, by x1(N-1)And h1N-2The first LSTM model is inputted, h1 is exportedN-1, by h1N-1、d、h2N-2Input
The 2nd LSTM model exports h2N-1；
In the step 103, determine that the emotion of the corpus is inclined according to the second information that the 2nd LSTM model exports
To including：
According to the h2N-1Determine the Sentiment orientation of the corpus.
In one embodiment, the cell state figure of the first LSTM model is as shown in Fig. 2, the first LSTM model is as follows：
Wherein, describedFor the state vector f1 of the forgetting door t moment of the first LSTM model(t)I-th yuan
Element, the σ is sigmod unit function, describedFor weight vectors b1fI-th of element, describedFor weight matrix
U1fThe element of i-th row, jth column, it is describedFor weight matrix W1fThe i-th row, jth column element, it is describedIt is described
The input vector x1 of the t moment of first LSTM model(t)J-th of element, i.e., t moment input j-th yuan of the term vector
Element；It is describedFor the output vector h1 at the first LSTM model t-1 moment(t-1)J-th of element；
It is describedFor the state vector g1 of the input gate t moment of the first LSTM model(t)I-th of element, it is describedFor weight vectors b1gI-th of element, it is describedFor weight matrix U1gThe element of i-th row, jth column, it is describedFor
Weight matrix W1gThe i-th row, jth column element；
It is describedFor the state vector q1 of the out gate t moment of the first LSTM model(t)I-th of element, it is describedFor weight vectors b1qI-th of element, it is describedFor weight matrix U1qThe element of i-th row, jth column, it is describedFor
Weight matrix W1qThe i-th row, jth column element；
It is describedFor intermediate state vector (i.e. cell (cell) shape of LSTM model of the first LSTM model t moment
State) s1(t)I-th of element, it is describedFor the intermediate state vector s1 at the first LSTM model t-1 moment(t-1)I-th
A element, the b1iFor i-th of element of weight vectors b1, the U1i,jFor the i-th row of weight matrix U1, jth arrange element,
The W1i,jThe element arranged for the i-th row of weight matrix W1, jth；s1(t-1)It is not shown in FIG. 2.
It is describedFor the output vector h1 of the first LSTM model t moment(t)I-th of element；The first information
The h1 of the t moment output of the as described first LSTM model(t), tanh is hyperbolic tangent function.
Fig. 3 is the cell state figure of the 2nd LSTM model, and the 2nd LSTM model is as follows：
Wherein, describedFor the state vector f2 of the forgetting door t moment of the 2nd LSTM model(t)I-th yuan
Element, the σ is sigmod unit function, describedFor weight vectors b2fI-th of element, describedFor weight matrix
U2fThe element of i-th row, jth column, it is describedFor weight matrix W2fThe i-th row, jth column element, it is describedFor weight
Matrix V 2fThe i-th row, jth column element, it is describedFor the input vector x2 of the t moment of the 2nd LSTM model(t)'s
J-th of element, i.e. t moment are input to j-th of element of the first information of the 2nd LSTM model；It is describedFor
The output vector h2 at the 2nd LSTM model t-1 moment(t-1)J-th of element；It is describedDescribed is input to for t moment
The described eigenvector d of two LSTM models(t)J-th of element；
It is describedFor the state vector g2 of the input gate t moment of the 2nd LSTM model(t)I-th of element, it is describedFor weight vectors b2gI-th of element, describedFor weight matrix U2gThe element of i-th row, jth column, it is described
For weight matrix W2gThe i-th row, jth column element；It is describedFor weight matrix V2gThe i-th row, jth column element；
It is describedFor the state vector q2 of the out gate t moment of the 2nd LSTM model(t)I-th of element, it is describedFor weight vectors b2qI-th of element, it is describedFor weight matrix U2qThe element of i-th row, jth column, it is described
For weight matrix W2qThe i-th row, jth column element, it is describedFor weight matrix V2qThe i-th row, jth column element；
It is describedFor the intermediate state vector s2 of the 2nd LSTM model t moment(t)I-th of element, it is describedFor the intermediate state vector s2 at the 2nd LSTM model t-1 moment(t-1)I-th of element, the b2iFor weight to
Measure i-th of element of b2, the U2i,jFor the element that the i-th row of weight matrix U2, jth arrange, the W2i,jFor weight matrix W2's
The element of i-th row, jth column, the V2i,jThe element arranged for the i-th row of weight matrix V2, jth；
It is describedFor the output vector h2 of the 2nd LSTM model t moment(t)I-th of element；Second information
The h2 of the as described 2nd LSTM model t moment output(t), tanh is hyperbolic tangent function.
It should be noted that the first LSTM model and the 2nd LSTM model shown in above-mentioned formula are merely illustrative, can incite somebody to action
The LSTM model of other constructions is applied in the application.
In one embodiment, the first LSTM model and the 2nd LSTM model are based on such as under type foundation：
Training corpus is obtained, after carrying out Sentiment orientation mark to the training corpus, based on the training corpus to first
Initial LSTM model and the second initial LSTM model are trained, and obtain the first LSTM model and the 2nd LSTM mould
Type.
Wherein, the Sentiment orientation exported in step 103, which is divided to, to be positive and passive two classes, it is of course also possible to be divided into more
Multiclass, such as very positive, general positive, general passive, very passive etc., for another example actively, passive, optimistic, sadness etc.,
Alternatively, taking scoring mechanism, from passiveness to being actively 0-10 points, score is higher, represents more positive, etc..
A kind of training method is as follows：
The first LSTM initial model (also referred to as word-level LSTM) is constructed according to formula (1)-formula (5), according to formula (6)-
Formula (10) constructs the 2nd LSTM initial model (also referred to as Sentence-level LSTM)；When initial, b1, b1f,b1g,b1qCan value be 0,
U1,U1f,U1g,U1q, W1, W1f,W1g,W1qIt can use normal distribution random value, constitute parameter matrix.b2,b2f,b2g,
b2qCan value be 0, U2, U2f,U2g,U2q, W2, W2f,W2g,W2q, V2, V2f,V2g,V2qIt can use normal distribution to take at random
Value constitutes parameter matrix.It should be noted that above-mentioned initial value is merely illustrative, it can according to need and take other values, such as basis
Experience takes initial value, alternatively, being worth trained under a scene as the initial value, etc. under another scene.
Set algorithm hyper parameter, hyper parameter is as shown in table 2, and the value of hyper parameter can be set as needed.
Table 2 trains hyper parameter group field explanation
Term vector is converted by training corpus participle, the first term vector is obtained, by topic belonging to training corpus or classification
Information is converted into term vector, obtains the second term vector, will be used as the first LSTM mould after the first term vector and the combination of the second term vector
The input quantity of type；For example, term vector can be converted by word with word2vec algorithm；Characteristic information is extracted (before such as
Social the comment feature, emotion characteristic feature, macro society feature mentioned), according to characteristic information construction feature vector d；It carries out
When training, the information for needing to input is as shown in table 3, including contains training corpus and expression classification group, wherein expression classification group refers to
To classify to emoticon, one kind is positive emoticon, and one kind is passive emoticon, when extracting emotion characteristic feature,
It can be extracted according to the classification in expression classification group.
3 training corpus of table inputs explanation
Trained loss function uses cross entropy, can be defined as
WhereinFor the distribution of the Sentiment orientation of corpus i mark.It is of course also possible to use unknown losses function, this Shen
Please this is not construed as limiting.
Training corpus is divided into training set according to preset ratio, verifying collection and test set (such as training set 80%, verifying
Collection 10%, test set 10%), above-mentioned LSTM model is trained using training corpus, the parameter in model is optimized,
Terminate until trained, saves the first LSTM model and the 2nd LSTM model obtained at this time.Parameter optimization training method can adopt
With AdaDelta algorithm, AdaGrad algorithm, Adam algorithm etc..Specifically, by the term vector of training corpus (directly according to corpus
The term vector that the obtained term vector of term vector or corpus and the term vector of topic combines) input the first LSTM it is initial
Model exports the first information, and the first information and feature vector are inputted the 2nd LSTM initial model, export the second information, according to
Second information determines Sentiment orientation, Optimal Parameters, until the Sentiment orientation of output is consistent with the Sentiment orientation marked in advance.
Export Sentiment orientation P (i)=Softmax (h2i), the Sentiment orientation of output is predefined Sentiment orientation classification
(such as actively, passive, optimistic, sadness etc.).
After model training, corpus to be analyzed is inputted, (Sentiment orientation is the feelings pre-defined to output Sentiment orientation
Sense tendency type).The format of the corpus to be analyzed of input can be as shown in table 4：
The format description of the corpus to be analyzed of table 4
In one embodiment, in step 103, the Sentiment orientation for determining the corpus according to second information includes：It will
The second information input Softmax function or RELU function, by the Softmax function or the output institute predicate of RELU function
The Sentiment orientation of material.It needs, Softmax function or RELU function are merely illustrative, can according to need using other points
Class device.
Fig. 4 is the flow chart of sentiment analysis.As shown in figure 4, term vector 401 is input to the first LSTM, the first LSTM output
The first information 402, the first information 402 and feature vector d are input to the 2nd LSTM, and the 2nd LSTM exports the second information 403, by the
Two information 403 input Softmax function, and the output of Softmax function is Sentiment orientation.It should be noted that in Fig. 4, it is right
One corpus, characteristic information only one, therefore, the d in figure(t-1), d(t), d(t+1)It is identical, it is the feature letter of the corpus
Breath.h1(t-2)And h2(t-2)An initial value can be set.
For example, for example corpus content is " I does not believe that ", in the case where not considering topic belonging to corpus, then：
A:The term vector of " I " is inputted in the first LSTM model, the first LSTM model exports h1(0), by h1(0)" I not
It is believed that " characteristic information be input to the 2nd LSTM model, the 2nd LSTM model exports h2(0)；
B:Input the term vector and h1 of " no "(0)To the first LSTM model, the first LSTM model exports h1(1), by h2(0)、h1(1)The characteristic information of " I does not believe that " is input to the 2nd LSTM model, and the 2nd LSTM model exports h2(1)；
C:Input " it is believed that " term vector and h1(1)To the first LSTM model, the first LSTM model exports h1(2), by h2(1)、
h1(2)The characteristic information of " I does not believe that " is input to the 2nd LSTM model, and the 2nd LSTM model exports h2(2)；
D:By h2(2)It is input to Softmax function, the output of Softmax function is the Sentiment orientation of " I does not believe that ".
In the case where considering topic belonging to corpus, it is assumed that entitled belonging to the corpus：" society ", then：
A:Term vector the first LSTM model of input for combining the term vector of the term vector of " I " and " society ", first
LSTM model exports h1(0), by h1(0)The characteristic information of " I does not believe that " is input to the 2nd LSTM model, the 2nd LSTM model
Export h2(0)；
B:The term vector of the term vector of " no " and " society " is combined to the term vector and h1(0)It is input to
One LSTM model, the first LSTM model export h1(1), by h2(0)、h1(1)The characteristic information of " I does not believe that " is input to second
LSTM model, the 2nd LSTM model export h2(1)；
C:By " it is believed that " term vector and " society " term vector be combined described in term vector and h1(1)It is input to
First LSTM model, the first LSTM model export h1(2), by h2(1)、h1(2)The characteristic information of " I does not believe that " is input to second
LSTM model, the 2nd LSTM model export h2(2)；
D:By h2(2)It is input to Softmax function, the output of Softmax function is the Sentiment orientation of " I does not believe that ".
It should be noted that LSTM model need to use previous output as this input (such as in step B,
Need h1(0)Input as the first LSTM model), if this is that for the first time, previous output uses an initial value,
The initial value can according to need setting.
Scheme provided in this embodiment, relative to the single layer LSTM for only using term vector, in addition to considering the association between word,
The association between sentence is also contemplated, for example, the 2nd LSTM model is according to h1 in step C(2)(word-based " I " " no " " phase
The analysis that letter " is made) and h2(1)(analysis that word-based " I " " no " makes) carries out subsequent analysis.And if only using single layer
In LSTM, step C, according to h1(1)(analysis that word-based " I " " no " makes) and " it is believed that " term vector carry out subsequent analysis,
It only considered the association between word, scheme provided by the present application, it is contemplated that the association between sentence, in addition, by extracting feature
Information joined Sentiment orientation represented by multimedia messages (such as emoticon), can more accurately reflect Sentiment orientation.
It should be noted that illustrating only three layers (the LSTM model at each moment represents one layer), the actual number of plies in figure
It is determined by the word number of the corpus inputted.
In one embodiment, described that the term vector is inputted to the first length memory network model pre-established, by institute
The second length memory that the first information and described eigenvector input for stating the output of the first length memory network model pre-establish
Network model includes：
By the term vector x1(0)~x1(N-1)From x1(0)To x1(N-1)Sequentially input the first length memory network mould
The first information of first length memory network model output and described eigenvector are inputted second length and remembered by type
Network model；Obtain term vector x1(N-1)After inputting the first length memory network model, the second length memory network mould
The positive information of the second of type output；
By the term vector x1(0)~x1(N-1)From x1(N-1)To x1(0)Sequentially input the first length memory network mould
The first information of first length memory network model output and described eigenvector are inputted second length and remembered by type
Network model；Obtain term vector x1(0)After inputting the first length memory network model, the second length memory network mould
Second reversed information of type output；
Second information according to the second length memory network model output determines the Sentiment orientation of the corpus
Including：
Second positive information and the second reversed information combination are obtained into the combined information, determined according to the combined information
The Sentiment orientation of the corpus.
As shown in figure 5, region 501 is to extract term vector according to positive sequence from corpus, it is sequentially inputted to the first LSTM model,
To extract term vector according to inverted order from corpus in region 502, it is sequentially inputted to the first LSTM model, by the in region 501
The second information is obtained after the output combination of the 2nd LSTM in the output of two LSTM models and region 502, the second information input is arrived
Softmax function, the output of Softmax function are Sentiment orientation.Combination is such as by the 2nd LSTM in region 501
The output of the 2nd LSTM is spliced in the output of model and region 502.
For example, for example corpus content is " I does not believe that ", is sequentially input in the first LSTM model in region 501
" I " " no " " it is believed that " term vector, inverted sequence is then pressed in the first LSTM model in region 502, sequentially input " it is believed that " " no "
The output of the 2nd LSTM model in region 501 and the output of the 2nd LSTM model in region 502 are carried out group by the term vector of " I "
Conjunction obtains the second information, and by the second information input to Softmax function, the output of Softmax function is " I does not believe that "
Sentiment orientation.If it is considered that topic belonging to corpus or classification, then sequentially input in the first LSTM model in region 501
Topic belonging to the combination of the term vector of topic belonging to the term vector of " I " and the corpus, the term vector of " no " and the corpus
The combination of term vector, " it is believed that " term vector and the corpus belonging to topic term vector combination, first in region 502
In LSTM model then press inverted sequence, sequentially input " it is believed that " term vector and the corpus belonging to topic term vector combination,
" no " combination of the term vector of topic belonging to term vector and the corpus, the term vector of " I " and topic belonging to the corpus
The combination of term vector.
In one embodiment, also Sentiment orientation is counted based on topic or classification, and exports each topic or classification
Synthesis Sentiment orientation accounting.It is of course also possible to directly export the Sentiment orientation of each corpus without statistics.
One embodiment of the invention provides a kind of sentiment analysis system, as shown in fig. 6, the sentiment analysis system includes：Number
According to processing module 601, memory module 602 and algorithm analysis module 603, wherein：
The data processing module 601 is used for, and generates term vector according to corpus；
The memory module 602 is used for, and stores the corpus；The memory module can be distributed file system；Such as
HDFS (Hadoop Distributed File System, Hadoop distributed file system), MongoDB etc.；
The algorithm analysis module 603 is used for, and generates term vector according to corpus, generates feature vector according to the corpus,
The term vector is inputted into the first LSTM model that pre-establishes, by the first information of the first LSTM model output and described
Feature vector inputs the 2nd LSTM model pre-established, according to the second information determination that the 2nd LSTM model exports
The Sentiment orientation of corpus.
In one embodiment, the algorithm analysis module 603 includes training unit 6031 and analytical unit 6032, wherein：
The data processing module 602 is also used to, and Sentiment orientation mark is carried out to the corpus, as training corpus；
The training unit 6031 is used for, based on the training corpus at the beginning of the preset first initial LSTM model and second
Beginning LSTM model is trained, and obtains the first LSTM model and the 2nd LSTM model；It can periodically be trained；
The analytical unit 6032 is used for, and generates feature vector according to the corpus, by term vector input described the
The first information of first LSTM model output and described eigenvector are inputted the 2nd LSTM mould by one LSTM model
Type.
In one embodiment, the analytical unit 6032 is also used to, and counts the emotion of each corpus under each topic or classification
Tendency.
In one embodiment, the system also includes display module 604, for showing the Sentiment orientation of the corpus, or
Person, shows the Sentiment orientation of topic or classification, for example, counting to the Sentiment orientation of all corpus under topic, output is not
The accounting etc. of feeling of sympathy tendency.
Specifically how to be trained and specifically how to obtain Sentiment orientation please refers to embodiment of the method, details are not described herein again.
In one embodiment, the data processing module includes data grabber unit 6011 and data pre-processing unit
6012, wherein：
The data grabber unit 6011 is used for, and obtains target information；The module can be with distributed arrangement, it can respectively
Configuration is in multiple positions.
The data pre-processing unit 6012 is used for, and target information is pre-processed into the data of preset format as institute's predicate
Material, is stored in the memory module；Memory module can be distributed file system.Start by set date sentiment analysis task is (certainly,
Can also be started with the non-timed, for example, starting after receiving enabled instruction, alternatively, starting, etc. after meeting trigger condition), it is loaded into
Resulting model file after algorithm training, inputs corpus data, obtains the Sentiment orientation of every corpus, counts by unit of topic
Sentiment orientation, and be shown.
The application is further illustrated below by specific example.
Example one
Interconnection abundant can be generated in public microblogging used in daily life, wechat circle of friends and internet site
Net data.In order to realize that the supervision of the social security on network and public opinion guidance perform effectively, analysis in real time and tracking are public
Focus and society's dynamic of public opinion in real time are very important.
As shown in fig. 7, the present embodiment provides a kind of sentiment analysis methods, including：
Step 701, grab the comment under content of microblog and microblogging using distributed reptile, wechat circle of friends content and
Comment；
Step 702, the data grabbed are pre-processed, pretreated data structured are stored in HDFS,
Each structural data is as a corpus.A kind of format of storage is as shown in table 5.
Step 703, Sentiment orientation mark is carried out to part corpus, as training corpus.
A kind of notation methods are as shown in table 6 below.It should be noted that the notation methods in table 6 are merely illustrative, it can basis
Need to be labeled as more Sentiment orientations.
Step 704, the first LSTM initial model is established, training corpus is inputted the first LSTM by the 2nd LSTM initial model
Initial model and the 2nd LSTM initial model, are trained, and obtain the first LSTM model and the 2nd LSTM model, the super ginseng of model
Number is as shown in table 7.
In one embodiment, theme belonging to corpus can be by directly being obtained by microblogging classification information, such as star, science and technology,
Society etc..The expression information of input can presort according to the expression that microblogging carries.Loss function is cross entropy, training platform
Select Caffe2.
5 data memory format of table
6 Sentiment orientation of table mark
Table is arranged in 7 example of table, one hyper parameter
| Parameter | Type | Description |
| dropout_rate | double | It is set as 0.5 |
| batch_size | int | It is set as 3 |
| word_embedding_dim | int | It is set as 256 |
| length_training_text | int | It is set as 5 |
It should be noted that above-mentioned hyper parameter value is merely illustrative, it can according to need and be set as other values.
Step 705, the corpus being analysed to inputs, the first LSTM model and the 2nd LSTM obtained in invocation step 704
Model carries out sentiment analysis, and analysis result is presented；It is exemplified below table 8.
Table 8 analyzes result
| Content | Sentiment orientation |
| It is preferential to buy：News comment says that she does not know how everybody sees making a show before seeing | Negative sense |
| Wave Jack mark：Well warm elder sister, the strength of emotional affection | It is positive |
In one embodiment, it can be counted for topic in step 705, count each microblogging under some topic and its comment
The Sentiment orientation of opinion.
In one embodiment, as shown in figure 8, the training process in step 704 includes：
Step 801：Corpus content sentence is segmented, converts term vector for word；And the word with topic belonging to corpus
Vector is spliced, and input vector is formed.
It should be noted that in another embodiment, the term vector that can also be directly segmented using corpus content is not used
The term vector of topic；
For example, term vector can be converted by word with word2vec algorithm；
Step 802：The first LSTM, parameter b1, b1 therein are constructed using above-mentioned formula (1) (2) (3) (4) (5)f、b1g、
b1qIt is initialized as the vector that all elements are 0, U1, U1f,U1g,U1q,W1,W1f,W1g,W1qIt is random to can use normal distribution
Value constitutes parameter matrix；Export the first information；
Step 803：Feature is extracted, according to three kinds of features (social activity comment feature, emotion characteristic feature, macroscopical society
Meeting feature), extract latent structure feature vector d；
Step 804：The 2nd LSTM is constructed using formula (6) (7) (8) (9) (10), wherein the input quantity X2 is step
The first information of 802 outputs, d are the feature vector in step 803.Parameter b2, b2f,b2g,b2qBeing initialized as all elements is 0
Vector, U2, U2f,U2g,U2q, W2, W2f,W2g,W2q, V2, V2f,V2g,V2qUsing normal distribution random value, parameter is constituted
Matrix；
Step 805, it is as shown in table 7 to define hyper parameter value, shown in trained loss function such as formula (11), by training language
Material is divided into training set according to preset ratio, and verifying collection and test set are trained, according to defeated after the specified step number of training method iteration
Algorithm model out, the specified step number is such as 1000 steps, certainly, merely illustrative herein, can according to need other steps of iteration
Number.
Scheme provided in this embodiment can accurately analyze netizen to the attitude of topic, track society's dynamic of public opinion in real time.
Example two
The commodity that the performance of e-commerce pushes more and more commodity producers directly to be produced are carried out on the net
It sells, such as automobile, household electrical appliances and food etc..After consumer buys commodity, can directly it make comments in commodity page.
As shown in figure 9, including：
Step 901：For specific commodity, it is collected respectively in different electric business platforms (such as Taobao, Jingdone district, Suning's electricity
Quotient, when equal) buyer's comment；
Step 902：Since data volume is huge, can use distributed deployment data processing module, respective pretreatment data,
Distributed data base HBase is stored in by data are regular.
In the present embodiment, Sentiment orientation mark can be carried out to comment according to the marking situation of user, for example, being more than or equal to
Samsung is then to like, conversely, being then not like；
Step 903：For electric quotient data data, algorithm is deployed on distributed tensorflow platform, with GPU plus
It carries and calculates, carry out off-line training；
Establish the first initial LSTM model and the second initial LSTM model respectively, in the present embodiment, hyper parameter setting such as table 9
Shown, loss function is cross entropy.Training obtains the first LSTM model and the 2nd LSTM model on tensorflow platform.
9 example of table, two hyper parameter value
| Parameter | Type | Description |
| dropout_rate | double | It is set as 0.5 |
| batch_size | int | It is set as 5 |
| word_embedding_dim | int | It is set as 256 |
| length_training_text | int | It is set as 5 |
Step 904：The corpus being analysed to is inputted according to the format of table 4, the first LSTM mould obtained in invocation step 903
Type and the 2nd LSTM analyze the Sentiment orientation of corpus, and result (for example, counting according to type of merchandize) is analyzed in statistic of classification, generate
Report.
It should be noted that in another embodiment, can also directly export the comment of each commodity without statistics
Sentiment orientation.
In one embodiment, the step 903 includes：
Step 9031：By corpus content sentence segment, calculated with word2vec algorithm, by word be converted into word to
Amount, referred to as content term vector；Corresponding type of merchandize will be commented on as affiliated topic, such as washing machine, mobile phone, micro-wave oven
Deng according to topic generation topic term vector；Content term vector and topic term vector are spliced to the input as the first LSTM；It needs
Illustrate, in another embodiment, content term vector can also be only used, does not use topic term vector；
Step 9032：The first LSTM, parameter b1, b1 therein are constructed using formula (1) (2) (3) (4) (5)f,b1g,b1q
It is initialized as the vector that all elements are 0, U1, U1f,U1g,U1q, W1, W1f,W1g,W1qUtilize normal distribution random value, structure
At parameter matrix；Term vector is inputted into word-level LSTM；Certainly, above-mentioned initial value is merely illustrative, can according to need and takes it
He is worth, such as can be using the value of the first LSTM model trained under other scenes as the initial value under this scene.
Step 9033：Feature is extracted, (social activity comment feature, emotion characteristic feature are macro than three kinds of features as previously mentioned
See social characteristic), construction feature vector d.
It should be noted that in the comment reply in Jingdone district store, there are the replies of many systems or customer service standardization to reply,
These contents can be filtered when carrying out feature extraction.Feature is commented on for social activity, the additional comment of original text author can be considered as
The reply content of author；In comment on commodity be not present expression information when, can first a customized sentiment dictionary, by what is wherein occurred
Emotion vocabulary is extracted as expression information；
Step 9034：The 2nd LSTM is constructed using above-mentioned formula (6) (7) (8) (9) (10), wherein input quantity x2 is step
The output of 9032 the oneth LSTM, d are feature vector described in step 9033.Parameter b2, b2 thereinf,b2g,b2qInitially
Turn to the vector that all elements are 0, U2, U2f,U2g,U2q, W2, W2f,W2g,W2q, V2, V2f,V2g,V2qUsing normal distribution with
Machine value constitutes parameter matrix；Certainly, above-mentioned initial value is merely illustrative, can according to need and takes other values, such as can be by it
The value of trained 2nd LSTM model is as the initial value under this scene under his scene.
Step 9035：It is as shown in table 9 to define hyper parameter value, trained loss function such as formula (11)；
Step 9036：Training corpus is divided into training set according to preset ratio, verifying collection and test set to the first LSTM and
2nd LSTM is trained, according to output algorithm model after the specified step number of training method iteration；For example, 1000 step of iteration.
The present embodiment can carry out consumer's sentiment analysis for particular commodity, more accurately to confirm consumer to production
The hobby of product adjusts product quality and sales tactics.
Example three
In smart city system, the policy information moment of government's publication affects the production and operation life of society.This
In example, after the publication of orientation analysis specific subject policy, the common people facilitate government and adjust and improve its policy to its Sentiment orientation
Strategy.
As shown in Figure 10, including：
Step 1001：The questionnaire of government policy information is collected, including network online investigation questionnaire and visiting and investigating is asked
Volume；
Step 1002：By the questionnaire input system；
Step 1003：Data preprocessing module pre-processes the data of the questionnaire of typing, by the regular deposit data of data
Library MySQL；It should be noted that database MySQL is merely illustrative herein, it can according to need and use other databases.
Step 1004：The first LSTM model and the 2nd LSTM model are established in starting algorithm training.In the present embodiment,
The LSTM model of pyTorch platform construction first and second LSTM model, loss function is cross entropy, hyper parameter value such as table
Shown in 10.It should be noted that hyper parameter value is merely illustrative, it can according to need and take other values.
Step 1005：Sentiment analysis is carried out to corpus to be analyzed, counts Sentiment orientation, and feeds back to relevant policies publication
Department.
10 example of table, three hyper parameter value
| Parameter | Type | Description |
| dropout_rate | double | It is set as 0.5 |
| batch_size | int | It is set as 3 |
| word_embedding_dim | int | It is set as 128 |
| length_training_text | int | It is set as 3 |
In one embodiment, the step 1004 includes：
Step 10041：Corpus content sentence is segmented, converts term vector for word with word2vec；The present embodiment
In, topic is government affairs information classification, such as forestry, health, medical treatment, house etc., by the term vector of topic and content term vector
Input quantity after being spliced as the first LSTM；It should be noted that in another embodiment, lexical word can also be only used
Vector does not use topic term vector；
Step 10042：The first initial LSTM model, b1, b1 are constructed using formula (1) (2) (3) (4) (5)f、b1g、b1qJust
Begin to turn to the vector that all elements are 0, U1, U1f,U1g,U1q,W1,W1f,W1g,W1qIt can use normal distribution random value,
Constitute parameter matrix；It is above-mentioned term vector that it, which is inputted,；
Step 10043：Characteristic information is extracted, (social activity comment feature, emotion characteristic feature are macro for example, extracting three kinds of features
See social characteristic), according to characteristic information construction feature vector d；
Step 10044：The 2nd LSTM is constructed using formula (6) (7) (8) (9) (10), wherein input quantity X2 is step
The output of 10042 the oneth LSTM, d are the feature vector in step 10043.Parameter b2, b2f,b2g,b2qIt is initialized as all members
The vector that element is 0, U2, U2f,U2g,U2q, W2, W2f,W2g,W2q, V2, V2f,V2g,V2qUtilize normal distribution random value, structure
At parameter matrix；
Step 10045：It is as shown in table 10 to define hyper parameter value, shown in trained loss function such as formula (11).
Step 10046：Training corpus is divided into training set according to preset ratio, verifying collection and test set are trained, root
The first LSTM model and the 2nd LSTM model are exported after specifying step number according to training method iteration.
Example four
Internet news information is often the flashpoint of public opinion, therefore pays close attention to social influence caused by particular news
Become of crucial importance.In the present embodiment, effectively sentiment analysis is carried out to news and news comment, filtering and emphasis monitoring cause
The news information of great society's repercussion.
As shown in figure 11, including：
Step 1101：Collect the content and comment information of particular news；
Step 1102：It is pre-processed, carries out abstract extraction for the content of news, convert short essay for long text news
This information；
Step 1103：Combination News abstract and comment；That is, news comment is carried out with news in brief corresponding；
Step 1104：Carry out algorithm training.In the present embodiment, in the initial LSTM of tensorflow platform construction first and
Two initial LSTM, loss function are cross entropy, and hyper parameter setting is as shown in table 11.
The setting of 11 example of table, four hyper parameter
| Parameter | Type | Description |
| dropout_rate | double | It is set as 0.5 |
| batch_size | int | It is set as 5 |
| word_embedding_dim | int | It is set as 256 |
| length_training_text | int | It is set as 5 |
Step 1105：Sentiment analysis is carried out to corpus to be analyzed using the first LSTM model and the 2nd LSTM model, point
Class statistic analysis result generates report.
In one embodiment, the step 1104 includes：
Step 11041：Corpus (being herein news in brief) sentence is segmented, is calculated with word2vec algorithm, is obtained interior
Hold term vector；The corresponding classifying content of news, will as affiliated topic, such as sport, finance and economics, science and technology, society, tourism etc.
Topic shift is topic term vector, and after being spliced with content term vector, the input quantity as the first LSTM；It needs to illustrate
It is that in another embodiment, content term vector can also be only used, does not use topic term vector；
Step 11042：The first LSTM, parameter b1, b1 therein are constructed using formula (1) (2) (3) (4) (5)f、b1g、b1q
It is initialized as the vector that all elements are 0, U1, U1f,U1g,U1q,W1,W1f,W1g,W1qIt can use normal distribution to take at random
Value constitutes parameter matrix；
Step 11043：Characteristic information is extracted, for example (social activity comment feature, emotion characteristic feature are macro for three kinds of features of extraction
See social characteristic), it is based on extracted characteristic information construction feature vector d；
Step 11044：The 2nd LSTM is constructed using formula (6) (7) (8) (9) (10), input quantity X2 is step 11042
The output of first LSTM, d are the feature vector in step 11043.Parameter b2, b2f,b2g,b2qIt is initialized as all members
The vector that element is 0, U2, U2f,U2g,U2q, W2, W2f,W2g,W2q, V2, V2f,V2g,V2qUtilize normal distribution random value, structure
At parameter matrix；
Step 11045：It is as shown in table 11 to define hyper parameter value, shown in trained loss function such as formula (11)；
Step 11046：Training corpus is divided into training set according to preset ratio, verifying collection and test set are trained, defeated
Algorithm model out, i.e. the first LSTM model of output and the 2nd LSTM model.
The implementation for carrying out sentiment analysis under several scenes using the application is presented above, it should be noted that this
Apply without being limited thereto, can be used for the sentiment analysis under other scenes, for example carry out emotion point for the usage experience to APP
Analysis, etc..Alternatively, it is also possible to not carry out algorithm training to each scene, same model can be shared with multiple scenes.
One embodiment of the invention provides a kind of sentiment analysis equipment, including memory and processor, the memory storage
There is program, described program realizes sentiment analysis method described in any of the above-described embodiment when reading execution by the processor.
One embodiment of the invention provides a kind of computer readable storage medium, and the computer-readable recording medium storage has
One or more program, one or more of programs can be executed by one or more processor, to realize above-mentioned
Sentiment analysis method described in one embodiment.
The computer readable storage medium includes：It is USB flash disk, read-only memory (ROM, Read-Only Memory), random
Access memory (RAM, Random Access Memory), mobile hard disk, magnetic or disk etc. are various to can store program
The medium of code.
Although disclosed herein embodiment it is as above, the content only for ease of understanding the present invention and use
Embodiment is not intended to limit the invention.Technical staff in any fields of the present invention is taken off not departing from the present invention
Under the premise of the spirit and scope of dew, any modification and variation, but the present invention can be carried out in the form and details of implementation
Scope of patent protection, still should be subject to the scope of the claims as defined in the appended claims.
Claims (13)
1. a kind of sentiment analysis method, including：
Term vector is generated according to corpus；
Feature vector is generated according to the corpus, the term vector is inputted to the first length memory network model pre-established,
The first information of first length memory network model output and described eigenvector are inputted into the second length pre-established
Memory network model；
The Sentiment orientation of the corpus is determined according to the second information that the second length memory network model exports.
2. sentiment analysis method according to claim 1, which is characterized in that before the generation feature vector according to corpus also
Including the target information being analysed to is pre-processed into the data of preset format as the corpus.
3. sentiment analysis method according to claim 1, which is characterized in that the corpus includes corpus content, alternatively, language
Expect content and corpus type；The corpus type indicates that the corpus includes at least one of：Text, text comment and comment
The reply of opinion.
4. sentiment analysis method according to claim 1, which is characterized in that described to generate term vector packet according to the corpus
It includes：
One or more first term vectors will be generated after corpus participle, using first term vector as the term vector；
Or, being generated one or more first term vectors are generated after corpus participle based on topic belonging to the corpus or classification
Second term vector combines first term vector and second term vector to obtain the term vector.
5. sentiment analysis method according to claim 1, which is characterized in that described to generate feature vector according to the corpus
Including：According to the corpus extract characteristic information, according to the characteristic information generate feature vector, the characteristic information include with
It is at least one lower：Social activity comment feature, emotion characteristic feature and macro society feature, the social comment feature indicate the corpus
Comment information in whether there is corpus publisher reply；The emotion characteristic feature indicates the corpus and its comment letter
In breath include first kind emotional symbol and the second class emotional symbol quantitative relation；Described in the macro society feature instruction
Whether the prompt user information that prompts other users check the corpus is occurred in corpus.
6. sentiment analysis method according to claim 1, which is characterized in that
The term vector includes term vector x1(0)~x1(N-1)；
It is described that the term vector is inputted to the first length memory network model pre-established, by the first length memory network
The first information and described eigenvector of model output input the second length memory network model pre-established and include：
By term vector x1(0)The first length memory network model is inputted, by the first length memory network model output
The first information and described eigenvector input the second length memory network model；
By term vector x1(0)The first length memory network model is inputted, by the first length memory network model output
The first information and described eigenvector input the second length memory network model；
And so on, by term vector x1(N-1)The first length memory network model is inputted, by the first length memory network
The first information and described eigenvector of model output input the second length memory network model, obtain second length
Second information of memory network model output.
7. sentiment analysis method according to any one of claims 1 to 6, which is characterized in that the first length memory network
Model and the second length memory network model are based under type such as and establish：
Sentiment orientation mark is carried out to training corpus, based on the training corpus to the preset first initial length memory network mould
Type and the second initial length memory network model are trained, and obtain the first length memory network model and described
Two length memory network models.
8. sentiment analysis method according to any one of claims 1 to 6, which is characterized in that the first length memory network
Model is as follows：
Wherein, describedIt is i-th yuan of the state vector of the forgetting door t moment of the first length memory network model
Element, the σ is sigmod unit function, describedFor weight vectors b1fI-th of element, describedFor weight matrix
U1fThe element of i-th row, jth column, it is describedFor weight matrix W1fThe i-th row, jth column element, it is describedIt is described
J-th yuan of the term vector of j-th of element of the input vector of the t moment of one length memory network model, i.e. t moment input
Element, it is describedFor the output vector h1 at the first length memory network model t-1 moment(t-1)J-th of element；
It is describedIt is described for i-th of element of the state vector of the input gate t moment of the first length memory network modelFor weight vectors b1gI-th of element, it is describedFor weight matrix U1gThe element of i-th row, jth column, it is describedFor
Weight matrix W1gThe i-th row, jth column element；
It is describedIt is described for i-th of element of the state vector of the out gate t moment of the first length memory network modelFor weight vectors b1qI-th of element, it is describedFor weight matrix U1qThe element of i-th row, jth column, it is describedFor
Weight matrix W1qThe i-th row, jth column element；
It is describedFor the intermediate state vector s1 of the first length memory network model t moment(t)I-th of element, it is describedFor the intermediate state vector s1 at the first length memory network model t-1 moment(t-1)I-th of element, the b1i
For i-th of element of weight vectors b1, the U1i,jFor the element that the i-th row of weight matrix U1, jth arrange, the W1i,jFor weight
The element that i-th row of matrix W 1, jth arrange；
It is describedFor the output vector h1 of the first length memory network model t moment(t)I-th of element, described first
Information is the h1 of the first length memory network model t moment output(t), tanh is hyperbolic tangent function.
9. sentiment analysis method according to any one of claims 1 to 6, which is characterized in that the second length memory network
Model is as follows：
Wherein, describedFor the second length memory network model forgetting door t moment state vector i-th of element,
The σ is sigmod unit function, describedFor weight vectors b2fI-th of element, it is describedFor weight matrix U2fThe
The element of i row, jth column, it is describedFor weight matrix W2fThe i-th row, jth column element, it is describedFor weight matrix
V2fThe i-th row, jth column element, it is describedFor the input vector x2 of the t moment of the second length memory network model(t)
J-th of element, i.e. t moment j-th of element being input to the first information of the second length memory network model, institute
It statesFor the output vector h2 at the second length memory network model t-1 moment(t-1)J-th of element, it is describedFor
T moment is input to j-th of element of the described eigenvector of the second length memory network model；
It is describedIt is described for i-th of element of the state vector of the input gate t moment of the second length memory network modelFor weight vectors b2gI-th of element, describedFor weight matrix U2gThe element of i-th row, jth column, it is describedFor
Weight matrix W2gThe i-th row, jth column element；It is describedFor weight matrix V2gThe i-th row, jth column element；
It is describedIt is described for i-th of element of the state vector of the out gate t moment of the second length memory network modelFor weight vectors b2qI-th of element, it is describedFor weight matrix U2qThe element of i-th row, jth column, it is described
For weight matrix W2qThe i-th row, jth column element, it is describedFor weight matrix V2qThe i-th row, jth column element；
It is describedFor the intermediate state vector s2 of the second length memory network model t moment(t)I-th of element, it is describedFor the intermediate state vector s2 at the second length memory network model t-1 moment(t-1)I-th of element, the b2i
For i-th of element of weight vectors b2, the U2i,jFor the element that the i-th row of weight matrix U2, jth arrange, the W2i,jFor weight
The element that i-th row of matrix W 2, jth arrange, the V2i,jThe element arranged for the i-th row of weight matrix V2, jth；
It is describedFor the output vector h2 of the second length memory network model t moment(t)I-th of element；Described second
Information is the h2 of the second length memory network model t moment output(t), tanh is hyperbolic tangent function.
10. sentiment analysis method according to any one of claims 1 to 6, which is characterized in that
The term vector includes term vector x1(0)~x1(N-1)；
It is described that the term vector is inputted to the first length memory network model pre-established, by the first length memory network
The first information and described eigenvector of model output input the second length memory network model pre-established and include：
By the term vector x1(0)~x1(N-1)From x1(0)To x1(N-1)The first length memory network model is sequentially input, by institute
The first information and described eigenvector for stating the output of the first length memory network model input the second length memory network mould
Type；Obtain term vector x1(N-1)After inputting the first length memory network model, the second length memory network model is obtained
The positive information of the second of output；
By the term vector x1(0)~x1(N-1)From x1(N-1)To x1(0)The first length memory network model is sequentially input, by institute
The first information and described eigenvector for stating the output of the first length memory network model input the second length memory network mould
Type；Obtain term vector x1(0)After inputting the first length memory network model, the second length memory network model is obtained
Second reversed information of output；
Second information according to the second length memory network model output determines that the Sentiment orientation of the corpus includes：
Described second positive information and the second reversed information are combined acquisition combined information, according to the combined information
Determine the Sentiment orientation of the corpus.
11. a kind of sentiment analysis system, which is characterized in that including：Data processing module, memory module and algorithm analysis module,
Wherein：
The data processing module is used for, and obtains corpus；
The memory module is used for, and stores the corpus；
The algorithm analysis module is used for, and generates feature vector according to the corpus, and term vector input is pre-established
First length memory network model, the first information of the first length memory network model output and described eigenvector is defeated
Enter the second length memory network model pre-established；The second information exported according to the second length memory network model is true
The Sentiment orientation of the fixed corpus.
12. a kind of sentiment analysis equipment, which is characterized in that including memory and processor, the memory is stored with program, institute
Program is stated when reading execution by the processor, realizes the sentiment analysis method as described in claims 1 to 10 is any.
13. a kind of computer readable storage medium, which is characterized in that the computer-readable recording medium storage have one or
Multiple programs, one or more of programs can be executed by one or more processor, to realize such as claims 1 to 10
Any sentiment analysis method.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810538689.1A CN108845986A (en) | 2018-05-30 | 2018-05-30 | A kind of sentiment analysis method, equipment and system, computer readable storage medium |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810538689.1A CN108845986A (en) | 2018-05-30 | 2018-05-30 | A kind of sentiment analysis method, equipment and system, computer readable storage medium |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN108845986A true CN108845986A (en) | 2018-11-20 |
Family
ID=64209997
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201810538689.1A Pending CN108845986A (en) | 2018-05-30 | 2018-05-30 | A kind of sentiment analysis method, equipment and system, computer readable storage medium |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN108845986A (en) |
Cited By (9)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110297907A (en) * | 2019-06-28 | 2019-10-01 | 谭浩 | Generate method, computer readable storage medium and the terminal device of interview report |
| CN111191677A (en) * | 2019-12-11 | 2020-05-22 | 北京淇瑀信息科技有限公司 | User characteristic data generation method and device and electronic equipment |
| CN111354354A (en) * | 2018-12-20 | 2020-06-30 | 深圳市优必选科技有限公司 | Training method and device based on semantic recognition and terminal equipment |
| CN112083806A (en) * | 2020-09-16 | 2020-12-15 | 华南理工大学 | Self-learning emotion interaction method based on multi-modal recognition |
| CN112333708A (en) * | 2020-10-27 | 2021-02-05 | 广东工业大学 | Telecommunication fraud detection method and system based on bidirectional gating circulation unit |
| CN112463947A (en) * | 2020-11-26 | 2021-03-09 | 上海明略人工智能(集团)有限公司 | Marketing scheme iteration method, marketing scheme iteration system, computer equipment and readable storage medium |
| US11120214B2 (en) | 2018-06-29 | 2021-09-14 | Alibaba Group Holding Limited | Corpus generating method and apparatus, and human-machine interaction processing method and apparatus |
| CN113518023A (en) * | 2021-09-13 | 2021-10-19 | 深圳小小小科技有限公司 | Control method and device for household appliance |
| CN113762343A (en) * | 2021-08-04 | 2021-12-07 | 德邦证券股份有限公司 | Method, device and storage medium for processing public opinion information and training classification model |
Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107544957A (en) * | 2017-07-05 | 2018-01-05 | 华北电力大学 | A kind of Sentiment orientation analysis method of business product target word |
| CN107862343A (en) * | 2017-11-28 | 2018-03-30 | 南京理工大学 | The rule-based and comment on commodity property level sensibility classification method of neutral net |
-
2018
- 2018-05-30 CN CN201810538689.1A patent/CN108845986A/en active Pending
Patent Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107544957A (en) * | 2017-07-05 | 2018-01-05 | 华北电力大学 | A kind of Sentiment orientation analysis method of business product target word |
| CN107862343A (en) * | 2017-11-28 | 2018-03-30 | 南京理工大学 | The rule-based and comment on commodity property level sensibility classification method of neutral net |
Non-Patent Citations (2)
| Title |
|---|
| MINLIE HUANG 等: "Modeling Rich Contexts for Sentiment Classification with LSTM", 《CORNELL UNIVERSITY: COMPUTER SCIENCE: COMPUTATION AND LANGUAGE》 * |
| SHALINI GHOSH 等: "Contextual LSTM (CLSTM) models for Large scale NLP tasks", 《CORNELL UNIVERSITY: COMPUTER SCIENCE: COMPUTATION AND LANGUAGE》 * |
Cited By (12)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US11120214B2 (en) | 2018-06-29 | 2021-09-14 | Alibaba Group Holding Limited | Corpus generating method and apparatus, and human-machine interaction processing method and apparatus |
| CN111354354A (en) * | 2018-12-20 | 2020-06-30 | 深圳市优必选科技有限公司 | Training method and device based on semantic recognition and terminal equipment |
| CN111354354B (en) * | 2018-12-20 | 2024-02-09 | 深圳市优必选科技有限公司 | Training method, training device and terminal equipment based on semantic recognition |
| CN110297907A (en) * | 2019-06-28 | 2019-10-01 | 谭浩 | Generate method, computer readable storage medium and the terminal device of interview report |
| CN111191677A (en) * | 2019-12-11 | 2020-05-22 | 北京淇瑀信息科技有限公司 | User characteristic data generation method and device and electronic equipment |
| CN111191677B (en) * | 2019-12-11 | 2023-09-26 | 北京淇瑀信息科技有限公司 | User characteristic data generation method and device and electronic equipment |
| CN112083806A (en) * | 2020-09-16 | 2020-12-15 | 华南理工大学 | Self-learning emotion interaction method based on multi-modal recognition |
| CN112333708A (en) * | 2020-10-27 | 2021-02-05 | 广东工业大学 | Telecommunication fraud detection method and system based on bidirectional gating circulation unit |
| CN112463947A (en) * | 2020-11-26 | 2021-03-09 | 上海明略人工智能(集团)有限公司 | Marketing scheme iteration method, marketing scheme iteration system, computer equipment and readable storage medium |
| CN113762343A (en) * | 2021-08-04 | 2021-12-07 | 德邦证券股份有限公司 | Method, device and storage medium for processing public opinion information and training classification model |
| CN113762343B (en) * | 2021-08-04 | 2024-03-15 | 德邦证券股份有限公司 | Method, device and storage medium for processing public opinion information and training classification model |
| CN113518023A (en) * | 2021-09-13 | 2021-10-19 | 深圳小小小科技有限公司 | Control method and device for household appliance |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN108845986A (en) | A kind of sentiment analysis method, equipment and system, computer readable storage medium | |
| CN107341145B (en) | A kind of user feeling analysis method based on deep learning | |
| Kaiser et al. | Mining consumer dialog in online forums | |
| US9710829B1 (en) | Methods, systems, and articles of manufacture for analyzing social media with trained intelligent systems to enhance direct marketing opportunities | |
| CN110750987B (en) | Text processing method, device and storage medium | |
| CN104484336B (en) | A kind of Chinese comment and analysis method and its system | |
| Chyrun et al. | Content monitoring method for cut formation of person psychological state in social scoring | |
| CN110750648A (en) | Text emotion classification method based on deep learning and feature fusion | |
| CN104077417A (en) | Figure tag recommendation method and system in social network | |
| Saranya et al. | A Machine Learning-Based Technique with IntelligentWordNet Lemmatize for Twitter Sentiment Analysis. | |
| CN116955591A (en) | Recommendation language generation method, related device and medium for content recommendation | |
| Gao et al. | Chatbot or Chat-Blocker: Predicting chatbot popularity before deployment | |
| Lyras et al. | Modeling Credibility in Social Big Data using LSTM Neural Networks. | |
| Al-Otaibi et al. | Finding influential users in social networking using sentiment analysis | |
| Biswas et al. | A new ontology-based multimodal classification system for social media images of personality traits | |
| Cetinkaya et al. | Twitter account classification using account metadata: organizationvs. individual | |
| CN116703515A (en) | Recommendation method and device based on artificial intelligence, computer equipment and storage medium | |
| Wang et al. | CA-CD: context-aware clickbait detection using new Chinese clickbait dataset with transfer learning method | |
| Lakshmi et al. | Sentiment analysis of twitter data | |
| Sarigiannidis et al. | A novel lexicon-based approach in determining sentiment in financial data using learning automata | |
| Wang et al. | Prediction of perceived utility of consumer online reviews based on lstm neural network | |
| Sujatha et al. | Text-based conversation analysis techniques on social media using statistical methods | |
| Gudumotu et al. | A Survey on Deep Learning Models to Detect Hate Speech and Bullying in Social Media | |
| Roshchina et al. | Evaluating the similarity estimator component of the TWIN personality-based recommender system | |
| Naseri et al. | A two-stage deep neural model with capsule network for personality identification |
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 |