CN108920545A - The Chinese affective characteristics selection method of sentiment dictionary and Ka Fang model based on extension - Google Patents
The Chinese affective characteristics selection method of sentiment dictionary and Ka Fang model based on extension Download PDFInfo
- Publication number
- CN108920545A CN108920545A CN201810610017.7A CN201810610017A CN108920545A CN 108920545 A CN108920545 A CN 108920545A CN 201810610017 A CN201810610017 A CN 201810610017A CN 108920545 A CN108920545 A CN 108920545A
- Authority
- CN
- China
- Prior art keywords
- class
- value
- item
- feature
- characteristic item
- 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
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F40/00—Handling natural language data
- G06F40/20—Natural language analysis
- G06F40/205—Parsing
- G06F40/216—Parsing using statistical methods
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Computational Linguistics (AREA)
- Artificial Intelligence (AREA)
- Audiology, Speech & Language Pathology (AREA)
- Health & Medical Sciences (AREA)
- General Health & Medical Sciences (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Probability & Statistics with Applications (AREA)
- Machine Translation (AREA)
- Information Retrieval, Db Structures And Fs Structures Therefor (AREA)
Abstract
The invention discloses the Chinese affective characteristics selection method of a kind of sentiment dictionary based on extension and Ka Fang model, the present invention is directed to current Research Challenges and deficiency, proposes the Chinese affective characteristics selection method of sentiment dictionary and Ka Fang model based on extension.First combine Hownet and it is improved based on the polarity value calculating method of word frequency to emotion word each in dictionary carry out polarity number calculating, establish have feeling polarities value dictionary；Then by negative word in comment text clause and its number occur and detecting, antisense conversion process carried out to the affective characteristic words in negative word confining spectrum, in effective solution emotion text the problem of polarity reversion；Finally the feeling polarities value of characteristic item and its two factor of frequency feature in class are dissolved into card side's feature selection module (CHI), substantially have modified the defect of card side's model, so that there are improved model better affective characteristics to select performance, text emotion classifying quality can be effectively improved.
Description
Technical field
The present invention relates to a kind of natural language processing technique more particularly to a kind of sentiment dictionaries and Ka Fangmo based on extension
The Chinese affective characteristics selection method of type.
Background technique
In recent years, sentiment analysis oneself become natural language processing in a hot issue, in Market Forecast Analysis, the people
The numerous areas such as meaning investigation, intelligent shopping guide, public comment suffer from wide application space and development prospect.Emotion comment text
After feature vector, various dimensions disaster may be generated, is unfavorable for model training, therefore, text feature selection is especially heavy
It wants.Emotion text feature selecting is generally divided into two classes, and one kind is directly to pass through look-up table using sentiment dictionary to extract affective characteristics
Word [1], advantage be simple and efficient, the disadvantage is that do not consider affective characteristics to the influence degree of model, meanwhile, have ignored sentiment dictionary
Outer feature；Another kind of is the feature selection approach based on statistics, such as card side's model, IG method, WLLR method, MI method, TF-
IDF method etc. calculates training text characteristic item word frequency and document frequency using statistical method, so as to find out each feature
The weight coefficient of item, using the high characteristic item of coefficient as the characteristic item of final choice.Currently, second method is more commonly used, it is preceding
People obtains card release side's model by experimental study and IG method is the conclusion of one of current most effective feature selecting algorithm, especially
Other methods are substantially better than in category distribution relative equilibrium.Therefore the defect and deficiency for inquiring into and correcting both methods, mention
The efficiency of high feature selecting has very important practical significance.
In recent years, some scholars have made some improvements [2-4] for the deficiency of IG algorithm and Ka Fang model algorithm.
TFIG algorithm [2] using the characteristic item frequency of occurrences (including characteristic item do not occur, occur it is primary, there are multiple three kinds of situations) carry out
The calculating of information gain, this method are pretty good to long text classifying quality, but apply in short text sentiment analysis, then effect one
As.CHI_LF algorithm [3] is by the information such as frequency disribution and characteristic item and class between position, class in frequency in the class of characteristic item, class
The other positive and negative degree of correlation incorporates in card side's model, and effect is preferable under class deflection condition, but applies relatively equal in category distribution
Performance boost is limited in the short text of weighing apparatus.
There are one emotion comment texts, and distinguishing feature is：Generally existing negative word in comment text.The appearance of negative word
The feeling polarities in its confining spectrum are inverted, can be brought to the supervised learning sorting algorithm based on bag of words certain negative
It influences.Therefore it needs to handle negative word before carrying out feature selecting.Xia [5] et al. is detected using negative word and feature
Word statistical method obtains corresponding converting text to the affective characteristic words conversion of text, and the text after making full use of extension carries out
Pairs of training and test.The method reduce the dependences to additional corpus data, to the emotion text adaptability in each field
It is relatively strong, but this method, because to invert to most of affective characteristic words outside negative word confining spectrum, there are some features
Noise, if the emotion text bad to Chinese word segmentation effect, has certain negative effect.
Summary of the invention
The object of the invention is that providing a kind of sentiment dictionary and Ka Fang based on extension to solve the above-mentioned problems
The Chinese affective characteristics selection method of model.
The present invention is achieved through the following technical solutions above-mentioned purpose：
1, the present invention includes the following steps：
(1) card side's feature selection module：Assume initially that Feature Words with classification are directly incoherent, if divided using card side
It is bigger that the calculated test value of cloth deviates threshold value, then more confident negative null hypothesis, receives the standby of null hypothesis and then assume：Feature
Word and classification have the very high degree of association；The calculation formula of card side's model is as follows：
Wherein, AijIt indicates to include characteristic item tiAnd belong to class cjNumber of documents, BijIt indicates to include characteristic item tiBut do not belong to
In class cjNumber of documents, CijIt indicates not including characteristic item tiBut belong to class cjNumber of documents, DijIt indicates not including characteristic item ti
And it is not belonging to class cjNumber of documents, N indicate total number of documents amount；
(2) polarity number of affective characteristics item is incorporated into card side's model：One by one by the affective characteristics item in emotion comment text
It is compared with the dictionary with feeling polarities value, if in dictionary, the value of this feature item in dictionary is made for the affective characteristics item
For the feeling polarities value of this feature item, such as " outstanding " corresponding value in dictionary is 0.919, then is used as " outstanding " for 0.919
Feeling polarities value；If this feature item does not appear in dictionary, calculated using formula (6), such as " bright in luster ", is led to
Crossing formula and calculating its feeling polarities value is 0.705；The final polarity number for obtaining all affective characteristics items in text；
From the definition of card side's model, it can be found that, matrix A indicates that the document comprising characteristic item belongs to the quantity of each class, is special
Most important one in sign selection, its value directly determines the final value of feature selecting；In all identical situation of other situations,
It include characteristic item t in matrix AiAnd belong to class cjValue AijIt is bigger, then characteristic item tiTo class cjContribution margin it is higher, characteristic item
tiWith classification cjJust there is the deeper degree of association, at this time by characteristic item tiChoose the accuracy rate that can be improved text classification；Together
When, other matrix Bs, C, D are calculated by matrix A；
But matrix A only considered number of documents, not consider to the feeling polarities of characteristic item itself, and feeling polarities value is high
Characteristic item in emotional semantic classification have biggish effect；The present invention considers this influence of two parts factor to classification, to feelings
Sense polarity number is merged with matrix A, and the number of documents comprising characteristic item is enclosed to the feeling polarities value of this feature item, matrix
A develops into A'ij=Aij×Eij；Such as in all similar situation of other situations, characteristic item " adjusting ", " liking " is in commendation class
The number of files of middle appearance is respectively 5,5, and original matrix A judges the two words to the important journey of commendation class simply by document frequency
Degree, such case are then difficult to distinguish which word more important to commendation class；And for modified matrix A ', due to the pole of the two words
Property value be respectively 0.1,0.9, then its matrix A ' in it is corresponding value be 5.5,9.5, be equivalent to the feature high for feeling polarities value
Item provides an increment, larger can will possibly occur less but feeling polarities value in training corpus in the feature selecting stage
Higher feature is chosen；
Wherein, EijIt is an emotional intensity matrix, row indicates that affective characteristics item, column indicate class, and first row indicates derogatory sense class,
Secondary series indicates commendation class, eiIt is the characteristic item emotional value obtained using sentiment dictionary and formula (6), wherein working as eiIt, should when >=0
Characteristic item is commendatory term, then EijIn value in two classes be respectively：
From formula (8) as can be seen that when this feature item is commendatory term, A'ijBe equivalent to in matrix A include characteristic item ti
And belong to class c1The value A of commendation classi1It has been multiplied by a big weight 1+ei, include characteristic item t in matrix AiAnd belong to class c0Derogatory sense
The value A of classi0It remains unchanged；
Work as eiWhen < 0, this feature item is derogatory term, then EijTwo class intermediate values be respectively：
From formula (9) as can be seen that when this feature item is derogatory term, A'ijBe equivalent to in matrix A include characteristic item ti
And belong to class c0The value A of derogatory sense classi0It has been multiplied by a big weight 1+ | ei|, it include characteristic item t in matrix AiAnd belong to class c1It praises
The value A of adopted classi1It remains unchanged；
After finding out A' according to above-mentioned formula, in conjunction with feature quantity N, B', C', D' are found out, finally obtains new characteristic value：
(3) frequency information and card side's models coupling between characteristic item class：Card side's model consider always comprising characteristic item or
Number of documents without characteristic item, not in view of the frequency that characteristic item occurs in class, but characteristic frequency is also feature selecting
One key factor；Such as characteristic item tiAnd tkIt is worth relatively by card side's model is calculated, if but characteristic item tiIn class
cjThe frequency of middle appearance is very big, and tkThe frequency of appearance is seldom, then tiIn class cjExpressive ability obviously than tjBy force, but card side
Model does not reflect the difference of this respect but；TF (t is utilized in the present inventioni,cj) indicate characteristic item tiIn classification cjIn word frequency
Number, formula are：
Wherein tfk(ti,cj) indicate feature tiBelonging to class cjDocument dkIn word frequency number, tf (ti,cj)minIndicate feature
tiBelonging to class cjSingle document in the minimum word frequency number that occurs, f (ti,cj)maxIndicate feature tiBelonging to class cjSingle document in
The maximum word frequency number of appearance, | cj| expression belongs to class cjNumber of documents；
In view of influence of the number of documents to it between different classes of, now it is normalized：
As characteristic item normalizes word frequency in each class；
(4) affective characteristics select formula：Frequency between the polarity number of affective characteristics item and the class of characteristic item is incorporated into Ka Fangmo
In type, final formula is as follows：
According to formula (13), the amendment chi-square value of each affective characteristics can be acquired, to all values by descending
Sequence arranges, and takes out affective characteristics corresponding to certain amount of maximum modified chi-square value as final affective characteristics, then
Feature vector is carried out to emotion text according to selected affective characteristics, text classifier is recycled to classify.
The beneficial effects of the present invention are：
The present invention is the Chinese affective characteristics selection method of a kind of sentiment dictionary based on extension and Ka Fang model, and existing
Technology is compared, and the present invention is directed to current Research Challenges and deficiency, is proposed in sentiment dictionary and Ka Fang model based on extension
Literary affective characteristics selection method.First combine Hownet and it is improved based on the polarity value calculating method of word frequency to feelings each in dictionary
Feel word and carry out polarity number calculating, establishes the dictionary for having feeling polarities value；Then by negative word in comment text clause and
There is number and is detected in it, and handles the emotion word in negative word confining spectrum, to effectively limit negative
The negative effect of word bring；Finally word frequency between characteristic item feeling polarities value and relevant feature class is melted with card side's model
It closes, substantially has modified the defect of card side's model, so that there are improved model better affective characteristics to select performance, it can
It is effective to improve text emotion classifying quality.
Detailed description of the invention
Fig. 1 is the dictionary Establishing process figure with feeling polarities value of the invention；
Fig. 2 is feeling polarities judging nicety rate curve graph of the invention；
Fig. 3 is hotel's data classification precision curve graph of the invention；
Fig. 3 (a) is to be classified using Nae Bayesianmethod to hotel's Chinese comment data, and Fig. 3 (b) is to utilize support
Vector machine classifies to hotel's data；
Fig. 4 is book data nicety of grading curve graph of the invention；
Fig. 4 (a) is to be classified using Nae Bayesianmethod to books Chinese comment data, and Fig. 4 (b) is to utilize support
Vector machine classifies to book data；
Fig. 5 is notebook data nicety of grading curve graph of the invention；
Fig. 5 (a) is to be classified using Nae Bayesianmethod to notebook Chinese comment data, and Fig. 5 (b) is to utilize branch
Vector machine is held to classify to notebook data.
Specific embodiment
The present invention will be further explained below with reference to the attached drawings：
Establish the dictionary with feeling polarities value
The polarity number calculating of affective characteristics item is mainly the following method：One is semantic-based method, using knowing
Adopted elite tree is netted to calculate affective characteristics item and pass judgement on the similarity between adopted benchmark word, using the similarity as the affective characteristics item
Polarity number[8-9], the feeling polarities value based on Hownet, which calculates, depends on benchmark word, can not calculate simultaneously for some unregistered words
Its feeling polarities value；One is Statistics-Based Methods[10], pass through the co-occurrence information between statistics affective characteristics item and benchmark word
Obtaining its polarity number, this method calculates simple, but the support of Big-corpus is needed, while also relying on benchmark word, when
Between complexity it is high；There are also one is the affection computation methods based on word frequency[11], the individual character using affective characteristics item passing judgement on adopted feelings
The word frequency information that occurs obtains its feeling polarities value in sense dictionary, and this method does not depend on benchmark word, while not needing a large amount of
Corpus carries out co-occurrence search, calculates simple, and time complexity is low, but it compares the vocabulary dependent on sentiment dictionary.
Present invention combination Hownet and affection computation method based on word frequency, and related improvement has been carried out to it, it obtains final
Feeling polarities value.
Lexicon extension and pretreatment：
(1) present invention combines the sentiment dictionary of Hownet publication and Taiwan Univ.'s natural language processing laboratory to provide simplified
Chinese sentiment dictionary (NTUSD) is pre-processed (merging and deletion including emotion word), the emotion being expanded to the dictionary
Dictionary.
(2) word each in sentiment dictionary is resolved into individual character, then counts all words and occurs in passing judgement on adopted sentiment dictionary
Quantity, formed have word frequency dictionary.
Similarity calculation based on Hownet：
Hownet be one it is fairly perfect, using concept representated by Chinese and english as object, to disclose concept and general
The commonsense knowledge base based on relation on attributes between thought.
Chinese vocabulary description in Hownet is all based on " justice is former " this basic conception.Due in Chinese a word not
Different concepts may be given expression in same context, therefore, word is organized into the set of several concepts by Hownet, each general
Read all by one group of description justice it is former represented by.As shown in table 1, vocabulary " disagreeable " is there are three concept, and two of them are adjectives, and one
It is verb, each concept is indicated by multiple justice are former, such as " aValue | attribute value ", " easiness | difficulty or ease ".Liu Qun[6]Et al.
The calculation method of two semantic similarities of the word based on Hownet is described in detail.
The former representation of the justice of 1 vocabulary of table
Polarity value calculating method of the emotion word based on Hownet is a kind of method based on sentiment dictionary, it is to choose some praise
Then derogatory sense benchmark word calculates emotion word with the tightness degree for passing judgement on adopted benchmark word according to emotion word.The present invention uses
40 pairs are passed judgement on adopted benchmark word[8], wherein commendation benchmark vocabulary is shown as pi(i takes 1~40), derogatory sense benchmark vocabulary are shown as nj(j takes 1~
40).Specific formula for calculation is as follows：
Wherein, sim (pi, w) and indicate the emotion word and commendation benchmark word piSimilarity, sim (nj, w) and indicate the emotion word
With derogatory sense benchmark word njSimilarity, ehownet(w) polarity number of the emotion word is indicated.Work as ehownet(w) 0 >, then by the emotion
Word is classified as commendatory term, ehownet(w) 0 <, then be classified as derogatory term for the emotion word, and the emotion that numerical values recited represents emotion word is strong
Degree.
Again to ehownet(w) minimax normalized, the as polarity number based on Hownet of the emotion word are carried out.2.3
Similarity calculation based on word frequency
The meaning of Chinese terms is expressed by forming the Chinese character of the word, therefore can be by the polarity number group of emotion word
It is indicated at the tendency degree function of all Chinese characters of the word.And the tendency degree of Chinese character, by calculating the Chinese character in sentiment dictionary
The frequency of appearance obtains[11]。
If emotion word W is made of k Chinese character, it is expressed as W=w1w2...wk, Chinese character wiThe Sentiment orientation degree of (i takes 1~k)
For Ti, then the polarity number e of emotion wordtf(w) shown in calculating such as formula (2).
Wherein, λiIndicate composition Chinese character wiWeight,The meaning of emotion word is codetermined by each Chinese character that forms
, here equal weight can be set by each Chinese character.TiIndicate Chinese character wiSentiment orientation degree, can be by formula (3) (4) (5)
It obtains jointly.
Wherein,Indicate the initial tendency degree of Chinese character, fPi、fNiRespectively indicate Chinese character wiIn commendation dictionary and derogatory sense dictionary
The frequency of appearance, n, m are illustrated respectively in the quantity of different Chinese character in commendation dictionary and derogatory sense dictionary.σ is the normalized factor,
Its formula is：
According to the calculation method of formula (3), if the Chinese character occurs in commendation dictionary but number is indefinite, but in derogatory term
Do not occur in allusion quotation, then the Sentiment orientation degree calculated is all 1, this is not obviously consistent with practical, Chinese character Sentiment orientation degree be usually with
Pass judgement on the increase of frequency of occurrences difference in adopted dictionary and increase, therefore need to be to the T in formula (3)iIt is weighted amendment, such as
Shown in formula (5)：
Wherein, λ indicates the absolute value for the frequency difference that the Chinese character occurs in passing judgement on adopted dictionary.
According to above-mentioned formula, polarity number of the emotion word based on word frequency can be obtained.For any one emotion word W, tendency
Spend etf(w) value range is -1~1, etf(w) being greater than 0 indicates that the vocabulary reaches positive emotion, and what is expressed less than 0 is negative
Emotion, etf(w) absolute value is bigger, and the emotion for illustrating that the vocabulary reaches is stronger.
Establish the dictionary for having feeling polarities value：
Present invention combination Hownet and emotion value calculating method based on word frequency, and both are obtained by linear regression model (LRM)
The best weights weight values of method, obtain the final dictionary with feeling polarities value, and specific establishment process is as shown in Figure 1.
Wherein, feature of present invention word polarity number can be obtained by formula (1) and formula (2) weighting.Calculation formula such as formula (6) institute
Show.
E=α × ehownet+β×etf (6)
Wherein, the value range of α and β is 0~1, and alpha+beta=1.The present invention is using linear regression model (LRM) to sentiment dictionary
All posting terms of middle known polarity are tested, and curve graph such as Fig. 2 institute that its feeling polarities judging nicety rate changes with α is obtained
Show.As α=0, the polarity of emotion word is completely by etfIt determines, as α=1, the polarity of emotion word is completely by ehownetIt determines, pole
Property judging nicety rate is lower.As α=0.451, β=0.549, accuracy rate reaches peak.
All Feature Words polarity numbers in sentiment dictionary are calculated using formula (6), if Feature Words polarity judges incorrectly,
Its polarity number is then set as 0, if polarity correct judgment, calculated result is the feeling polarities value of the word.
Processing to negative word：
In the machine learning classification to emotion comment text, bag of words are current most widely used feature vectors
Mode, but bag of words ignore the sequence of word in sentence, destroy syntactic structure, cannot be disposed to negative word.Such as
There is negative word in fruit comment text, the sentence feeling polarities in its confining spectrum can be converted to its opposite polarity, such as：
" I/like/this/book " and " I do not like/or not/this/book ", there is negative word in second comment sentence, by the comment sentence
Polarity inverted, and in bag of words, this phenomenon is not reflected, thus by machine learning classification, easily by second
Sentence is mistakenly classified as commendation class.In order to reduce such case, the present invention is before carrying out feature selecting, first to negative word
And its text in confining spectrum is handled.
Establish antisense dictionary：
In emotion comment text, the emotion description to things is generally described with verb, adjective, adverbial word.Therefore, greatly
Part verb, adjective, adverbial word feeling polarities with respect to other parts of speech for, it is obvious.Based on this feature, the present invention couple
All emotion word carries out part of speech judgement in training set, by belong to verb, adjective, adverbial word emotion word extract and be included in word
In allusion quotation, the emotion word polarity in dictionary is calculated by formula (6), then presses the commendation word set of acquisition and derogatory sense word set
It is ranked up according to the sequence of emotional intensity from big to small.If PiFor the commendation emotion word of commendation polarity number ranking i-th bit, NiTo demote
The derogatory sense emotion word of adopted polarity number ranking i-th bit can form and pass judgement on emotion word one-to-one relationship (Pi,Ni), wherein i≤min
(len(Pi),len(Ni)), len (Pi) indicate commendation word set quantity, len (Ni) indicate derogatory sense word set quantity.For not having
The emotion word for forming corresponding relationship is substantially negligible its polarity number, so in antonym due to these word polarity number very littles
Directly this partial feeling word is deleted in allusion quotation.
Detection and processing negative word：
It is detected using all negative words of the look-up table to emotion comment text training set and test set, such as table 2.It determines
The position that negative word occurs.
Table 2 negates vocabulary
For the range that negative word defines, negative word is directly started that (emotion comment text is by making pauses in reading unpunctuated ancient writings to clause by the present invention
Punctuate is separated into multiple clauses) confining spectrum of the range as negative word between ending word[5].In the range that negative word defines
Interior, the emotion word that will appear in antisense dictionary is all converted to its corresponding antonym, other emotion words remain unchanged, and will
Negative word removes.Such as：Become after " I do not like/or not/this/book " conversion " I/disagreeable/this/book ".
Occurs even number negative word in if clause, such as：" I do not think/or not/this/hotel/or not clean ", " there is no/mono-
A/people/or not satisfied " etc., this generic clause is double negative sentences[12], there is no quilts for the emotion word polarity in negative word confining spectrum
Reversion, therefore to the clause for even number negative word occur, directly negative word is removed, the emotion word in negative word confining spectrum is protected
It holds constant；To the clause for odd number negative word occur, the emotion word in negative word confining spectrum is inverted, while will negative
Word removes.
Feature selecting is carried out by negative word after treatment, then to all affective characteristic words.
Improved feature selection approach：
Card side's feature selection module：
The basic thought of " hypothesis testing " in statistics is utilized in card side's model：Assume initially that Feature Words are direct with classification
It is incoherent, if bigger using the calculated test value deviation threshold value of chi square distribution, more confident negative null hypothesis,
Receive the standby of null hypothesis then to assume：Feature Words and classification have the very high degree of association.The calculation formula of card side's model is as follows：
Wherein, AijIt indicates to include characteristic item tiAnd belong to class cjNumber of documents, BijIt indicates to include characteristic item tiBut do not belong to
In class cjNumber of documents, CijIt indicates not including characteristic item tiBut belong to class cjNumber of documents, DijIt indicates not including characteristic item ti
And it is not belonging to class cjNumber of documents, N indicate total number of documents amount.
The polarity number of affective characteristics item is incorporated into card side's model：
Affective characteristics item in emotion comment text is compared with the dictionary with feeling polarities value one by one, if the emotion
Characteristic item is in dictionary, then using the value of this feature item in dictionary as the feeling polarities value of this feature item, such as " outstanding " in word
Corresponding value is 0.919 in allusion quotation, then by the 0.919 feeling polarities value as " outstanding ".If this feature item does not appear in dictionary
In, then it is calculated using formula (6), such as " bright in luster ", calculating its feeling polarities value by formula is 0.705.Finally obtain
Obtain the polarity number of all affective characteristics items in text.
From the definition of card side's model, it can be found that, matrix A indicates that the document comprising characteristic item belongs to the quantity of each class, is special
Most important one in sign selection, its value directly determines the final value of feature selecting.In all identical situation of other situations,
It include characteristic item t in matrix AiAnd belong to class cjValue AijIt is bigger, then characteristic item tiTo class cjContribution margin it is higher, characteristic item
tiWith classification cjJust there is the deeper degree of association, at this time by characteristic item tiChoose the accuracy rate that can be improved text classification.Together
When, other matrix Bs, C, D are calculated by matrix A.
But matrix A only considered number of documents, not consider to the feeling polarities of characteristic item itself, and feeling polarities value is high
Characteristic item in emotional semantic classification have biggish effect.The present invention considers this influence of two parts factor to classification, to feelings
Sense polarity number is merged with matrix A, and the number of documents comprising characteristic item is enclosed to the weight (feeling polarities of this feature item
Value), matrix A develops into A'ij=Aij×Eij.Such as in all similar situation of other situations, characteristic item " adjusting ", " liking "
The number of files occurred in commendation class is respectively 5,5, and original matrix A judges the two words to commendation class simply by document frequency
Significance level, such case is then difficult to distinguish which word more important to commendation class.And for modified matrix A ', due to this two
The polarity number of a word is respectively 0.1,0.9, then its matrix A ' in it is corresponding value be 5.5,9.5, be equivalent to as feeling polarities value
High characteristic item provides an increment, larger can will possibly occur less but feelings in training corpus in the feature selecting stage
The sense higher feature of polarity number is chosen.
Wherein, EijIt is an emotional intensity matrix, row indicates that affective characteristics item, column indicate class, and first row indicates derogatory sense class,
Secondary series indicates commendation class, eiIt is the characteristic item emotional value obtained using sentiment dictionary and formula (6), wherein working as eiIt, should when >=0
Characteristic item is commendatory term, then EijIn value in two classes be respectively：
From formula (8) as can be seen that when this feature item is commendatory term, A'ijBe equivalent to in matrix A include characteristic item ti
And belong to class c1The value A of (commendation class)i1It has been multiplied by a big weight (1+ei), it include characteristic item t in matrix AiAnd belong to class c0
The value A of (derogatory sense class)i0It remains unchanged.
Work as eiWhen < 0, this feature item is derogatory term, then EijTwo class intermediate values be respectively：
From formula (9) as can be seen that when this feature item is derogatory term, A'ijBe equivalent to in matrix A include characteristic item ti
And belong to class c0The value A of (derogatory sense class)i0Be multiplied by a big weight (1+ | ei|), it include characteristic item t in matrix AiAnd belong to class
c1The value A of (commendation class)i1It remains unchanged.
After finding out A' according to above-mentioned formula, in conjunction with feature quantity N, B', C', D' are found out, finally obtains new characteristic value：
Frequency information and card side's models coupling between characteristic item class：
What card side's model considered always includes characteristic item or the number of documents without characteristic item, not in view of characteristic item exists
The frequency occurred in class, but characteristic frequency is also a key factor of feature selecting.Such as characteristic item tiAnd tkBy card side's model
It is calculated to be worth relatively, if but characteristic item tiIn class cjThe frequency of middle appearance is very big, and tkThe frequency of appearance is seldom,
Then tiIn class cjExpressive ability obviously than tjBy force, but card side's model does not reflect the difference of this respect.It is sharp in the present invention
With TF (ti,cj) indicate characteristic item tiIn classification cjIn word frequency number, formula is：
Wherein tfk(ti,cj) indicate feature tiBelonging to class cjDocument dkIn word frequency number, tf (ti,cj)minIndicate feature
tiBelonging to class cjSingle document in the minimum word frequency number that occurs, f (ti,cj)maxIndicate feature tiBelonging to class cjSingle document in
The maximum word frequency number of appearance, | cj| expression belongs to class cjNumber of documents.
In view of influence of the number of documents to it between different classes of, now it is normalized：
As characteristic item normalizes word frequency in each class.
Affective characteristics of the present invention select formula：
The present invention incorporates frequency between the polarity number of affective characteristics item and the class of characteristic item in card side's model, final formula
As follows：
According to formula (13), the amendment chi-square value of each affective characteristics can be acquired, to all values by descending
Sequence arranges, and takes out affective characteristics corresponding to certain amount of maximum modified chi-square value as final affective characteristics, then
Feature vector is carried out to emotion text according to selected affective characteristics, text classifier is recycled to classify.
Description of test：
Experimental data：
The present invention using Tan Song wave acquisition hotel, books, three fields of notebook Chinese comment data, each field
It respectively include 2000 fronts and 2000 negative reviews
3 experimental data of table
Algorithm experimental：
(base is calculated for main contrast forefathers' algorithm document [3], document [5] and basic IG feature selecting algorithm of the present invention
Method), to verify the validity of inventive algorithm.
Five folding cross validations are carried out to each data set：It is 5 equal portions by data set random division, wherein four parts are used as training
Data, it is in addition a to be used as test data, data set is trained and is tested by this mode, and utilizes naive Bayesian point
Two kinds of classification methods of class and support vector machines are classified.Obtain 5 numerical value are averaged and obtain final result.Experiment
As a result as follows：
1. pair hotel's Chinese comment data carry out the result of naive Bayesian and support vector cassification as shown in figure 3, from
Fig. 3 can be seen that inventive algorithm during feature quantity is 200~3000 all better than other algorithms.
2. pair books Chinese comment data carries out naive Bayesian and the result of support vector cassification is as shown in Figure 4：
From fig. 4, it can be seen that inventive algorithm is better than other algorithms using effect when Naive Bayes Classifier, utilize
When support vector machine classifier, inventive algorithm reaches F maximum value 0.917 when feature quantity is 2600, and document [5], text
[3], base algorithm are offered when feature quantity is 2600, F value is 0.911,0.906,0.893 respectively.Inventive algorithm is than other
Algorithm will be got well.
3. pair notebook Chinese comment data carries out naive Bayesian and the result of support vector cassification is as shown in Figure 5：
From fig. 5, it can be seen that inventive algorithm is substantially better than other algorithms during feature quantity is 200~3000.
When inventive algorithm utilizes Naive Bayes Classification, when feature quantity is 1800, reach F maximum value 0.922, document [5], text
Offer [3], base algorithm is 0.907,0.897,0.896 when feature quantity is 1800, inventive algorithm is obviously better than other
Algorithm.When using support vector cassification, the fluctuation of document [5] algorithm is too big, document [3] although, base algorithm fluctuation it is small,
Maximum F value is smaller, respectively 0.909,0.882, and inventive algorithm lies substantially in the peak of all algorithms, and F value is up to
0.92, it is better than other algorithms.
The present invention is tested by the Chinese comment data to hotel, books, notebook three fields and verifies this hair
Bright algorithm.Naive Bayesian is utilized from being can be seen that after the present invention carries out affective characteristics selection to each data set in Fig. 3~5
When classification, relative to basic algorithm, F value improves 4% or so, and relative to existing innovatory algorithm, F value improves 2% or so；Benefit
When with support vector cassification, relative to basic algorithm, F value improves 3% or so, and relative to existing innovatory algorithm, F value is improved
1% or so.Illustrate that inventive algorithm further improves the effect of text emotion feature selecting compared to existing algorithm.
Conclusion：
Feeling polarities value and word frequency information are dissolved into card side's model on the basis of forefathers study and carry out feelings by the present invention
The selection for feeling characteristic item, is optimized feature selection approach, this method is additionally contemplates that negative word to text emotion pole simultaneously
The influence of property, is detected and is judged to negative word, and handled the emotion word in negative word confining spectrum, further mentioned
The high accuracy of emotion text classification.Future can continue to explore the pass of feature vector weight in feeling polarities value and classifier
Connection optimizes emotion text classifier using feeling polarities value, improves the nicety of grading that classifier is directed to emotion text.
Bibliography：
[1] emotional orientation analysis computer age .2017,3 of Yang Kui, the Duan Qiongjin based on sentiment dictionary method:10.
[2]Xu Y,Chen L.Term-frequency Based Feature Selection Methods for
Text Categorization[C]//Fourth International Conference on Genetic and
Evolutionary Computing.IEEE Computer Society,2010:280-283.
[3] Ah antelope Song, Liu Haifeng, Liu keep optimization CHI text feature selection method of the raw based on position and word frequency information
[J].Computer Science\s&\sapplication,2015,05(9):322-330.
[4] Shi Hui, Jia Daiping, seedling train improvement information gain text feature selection algorithm computer of the based on word frequency information
Using 2014,34 (11):3279.
[5]Xia R,Xu F,Zong C,et al.Dual Sentiment Analysis:Considering Two
Sides of One Review[J].IEEE Transactions on Knowledge&Data Engineering,2015,
27(8):2120-2133.
[6] Liu Qun, Li Su build and are based on《Hownet》Similarity of Words calculate [J] Chinese computing linguistics, 2002.
[7]Lee S Y M.Sentiment Classification and Polarity Shifting.[C]//
International Conference on Computational Linguistics.Association for
Computational Linguistics,2010:635-643.
[8] Wang Zhenyu, Wu Zeheng, Hu Fangtao calculate [J] computer work based on the word feeling polarities of HowNet and PMI
Journey, 2012,38 (15):187-189.
[9] a kind of Beijing semantic similarity calculation method Information technology based on HowNet of Fan Hongyi, Zhang Yangsen
College journal (natural science edition), 2014,29 (4):42.
[10] Wu Jinyuan, Ji Junzhong, Zhao Xuewu, Wu Chensheng, Du Fang China are based on the emotion word weight of Feature Selection
Calculate Beijing University of Technology journal, 2016,42 (1):142.
[11] Xu Xiaodan, Duan Zhengjie, it is old in educate emotion method for digging mountain of the based on expanding sentiment dictionary and characteristic weighing
Eastern college journal (engineering version), 2014,6 (44):15-19.
[12] Liu Yujiao, jade pendant are taken root, Wu Shaomei, Chinese text emotion point of the Soviet Union Chong based on emotion dictionary in conjunction with conjunction
Class Sichuan University journal (natural science edition), 2015,52 (1):57-62.
Basic principles and main features and advantages of the present invention of the invention have been shown and described above.The technology of the industry
Personnel are it should be appreciated that the present invention is not limited to the above embodiments, and the above embodiments and description only describe this
The principle of invention, without departing from the spirit and scope of the present invention, various changes and improvements may be made to the invention, these changes
Change and improvement all fall within the protetion scope of the claimed invention.The claimed scope of the invention by appended claims and its
Equivalent thereof.
Claims (1)
1. a kind of Chinese affective characteristics selection method of sentiment dictionary and Ka Fang model based on extension, which is characterized in that including
Following steps：
(1) card side's feature selection module：Assume initially that Feature Words with classification are directly incoherent, if utilizing chi square distribution meter
It is bigger that the test value of calculating deviates threshold value, then more confident negative null hypothesis, receives the standby of null hypothesis and then assume：Feature Words with
Classification has the very high degree of association；The calculation formula of card side's model is as follows：
Wherein, AijIt indicates to include characteristic item tiAnd belong to class cjNumber of documents, BijIt indicates to include characteristic item tiBut it is not belonging to class cj
Number of documents, CijIt indicates not including characteristic item tiBut belong to class cjNumber of documents, DijIt indicates not including characteristic item tiAnd not
Belong to class cjNumber of documents, N indicate total number of documents amount；
(2) polarity number of affective characteristics item is incorporated into card side's model：By the affective characteristics item in emotion comment text one by one with band
Have emotion polarity number dictionary compare, if the affective characteristics item in dictionary, using the value of this feature item in dictionary be used as this
The feeling polarities value of characteristic item, such as " outstanding " corresponding value in dictionary is 0.919, then by 0.919 feelings as " outstanding "
Feel polarity number；If this feature item does not appear in dictionary, calculated using formula (6), such as " bright in luster ", passes through public affairs
It is 0.705 that formula, which calculates its feeling polarities value,；The final polarity number for obtaining all affective characteristics items in text；
From the definition of card side's model, it can be found that, matrix A indicates that the document comprising characteristic item belongs to the quantity of each class, is feature choosing
Most important one in selecting, its value directly determines the final value of feature selecting；In all identical situation of other situations, matrix
It include characteristic item t in AiAnd belong to class cjValue AijIt is bigger, then characteristic item tiTo class cjContribution margin it is higher, characteristic item tiWith
Classification cjJust there is the deeper degree of association, at this time by characteristic item tiChoose the accuracy rate that can be improved text classification；Meanwhile
Other matrix Bs, C, D are calculated by matrix A；
But matrix A only considered number of documents, not consider to the feeling polarities of characteristic item itself, and the spy that feeling polarities value is high
Levy item has biggish effect in emotional semantic classification；This influence of two parts factor to classification is considered herein, to feeling polarities
Value is merged with matrix A, and the number of documents comprising characteristic item is enclosed to the feeling polarities value of this feature item, and matrix A develops
For A'ij=Aij×Eij；Such as in all similar situation of other situations, characteristic item " adjusting ", " liking " occurs in commendation class
Number of files be respectively 5,5, original matrix A judge simply by document frequency the two words to the significance level of commendation class, this
Kind of situation is then difficult to distinguish which word more important to commendation class；And for modified matrix A ', due to the polarity number of the two words
Respectively 0.1,0.9, then its matrix A ' in it is corresponding value be 5.5,9.5, be equivalent to and mentioned for the high characteristic item of feeling polarities value
For an increment, it is higher larger can will possibly to occur less but feeling polarities value in training corpus in the feature selecting stage
Feature to being chosen；
Wherein, EijIt is an emotional intensity matrix, row indicates affective characteristics item, and arranging indicates class, first row expression derogatory sense class, second
Column indicate commendation class, eiIt is the characteristic item emotional value obtained using sentiment dictionary and formula (6), wherein working as eiWhen >=0, this feature
Item is commendatory term, then EijIn value in two classes be respectively：
From formula (8) as can be seen that when this feature item is commendatory term, A'ijBe equivalent to in matrix A include characteristic item tiAnd belong to
In class c1The value A of commendation classi1It has been multiplied by a big weight 1+ei, include characteristic item t in matrix AiAnd belong to class c0Derogatory sense class
Value Ai0It remains unchanged；
Work as eiWhen < 0, this feature item is derogatory term, then EijTwo class intermediate values be respectively：
From formula (9) as can be seen that when this feature item is derogatory term, A'ijBe equivalent to in matrix A include characteristic item tiAnd belong to
In class c0The value A of derogatory sense classi0It has been multiplied by a big weight 1+ | ei|, it include characteristic item t in matrix AiAnd belong to class c1Commendation class
Value Ai1It remains unchanged；
After finding out A' according to above-mentioned formula, in conjunction with feature quantity N, B', C', D' are found out, finally obtains new characteristic value：
(3) frequency information and card side's models coupling between characteristic item class：What card side's model considered always includes characteristic item or is free of
The number of documents of characteristic item, not in view of the frequency that characteristic item occurs in class, but characteristic frequency is also one of feature selecting
Key factor；Such as characteristic item tiAnd tkIt is worth relatively by card side's model is calculated, if but characteristic item tiIn class cjIn
The frequency of appearance is very big, and tkThe frequency of appearance is seldom, then tiIn class cjExpressive ability obviously than tjBy force, but Ka Fangmo
Type does not reflect the difference of this respect but；TF (t is utilized hereini,cj) indicate characteristic item tiIn classification cjIn word frequency number, it is public
Formula is：
Wherein tfk(ti,cj) indicate feature tiBelonging to class cjDocument dkIn word frequency number, tf (ti,cj)minIndicate feature ti?
Belong to class cjSingle document in the minimum word frequency number that occurs, f (ti,cj)maxIndicate feature tiBelonging to class cjSingle document in occur
Maximum word frequency number, | cj| expression belongs to class cjNumber of documents；
In view of influence of the number of documents to it between different classes of, now it is normalized：
As characteristic item normalizes word frequency in each class；
(4) affective characteristics select formula：Frequency between the polarity number of affective characteristics item and the class of characteristic item is incorporated in card side's model,
Final formula is as follows：
According to formula (13), the amendment chi-square value of each affective characteristics can be acquired, descending sequence is pressed to all values
Arrangement, takes out affective characteristics corresponding to certain amount of maximum modified chi-square value as final affective characteristics, then basis
Selected affective characteristics carry out feature vector to emotion text, and text classifier is recycled to classify.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810610017.7A CN108920545B (en) | 2018-06-13 | 2018-06-13 | Chinese emotion feature selection method based on extended emotion dictionary and chi-square model |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810610017.7A CN108920545B (en) | 2018-06-13 | 2018-06-13 | Chinese emotion feature selection method based on extended emotion dictionary and chi-square model |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN108920545A true CN108920545A (en) | 2018-11-30 |
| CN108920545B CN108920545B (en) | 2021-07-09 |
Family
ID=64419529
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201810610017.7A Expired - Fee Related CN108920545B (en) | 2018-06-13 | 2018-06-13 | Chinese emotion feature selection method based on extended emotion dictionary and chi-square model |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN108920545B (en) |
Cited By (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109933793A (en) * | 2019-03-15 | 2019-06-25 | 腾讯科技（深圳）有限公司 | Text polarity identification method, apparatus, equipment and readable storage medium storing program for executing |
| CN110895557A (en) * | 2019-11-27 | 2020-03-20 | 广东智媒云图科技股份有限公司 | Text feature judgment method and device based on neural network and storage medium |
| CN112214527A (en) * | 2020-09-25 | 2021-01-12 | 桦蓥(上海)信息科技有限责任公司 | Financial object abnormal factor screening and analyzing method based on failure response code |
| CN112464646A (en) * | 2020-11-23 | 2021-03-09 | 中国船舶工业综合技术经济研究院 | Text emotion analysis method for defense intelligence library in national defense field |
Citations (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US5887069A (en) * | 1992-03-10 | 1999-03-23 | Hitachi, Ltd. | Sign recognition apparatus and method and sign translation system using same |
| CN102945268A (en) * | 2012-10-25 | 2013-02-27 | 北京腾逸科技发展有限公司 | Method and system for excavating comments on characteristics of product |
| CN103955451A (en) * | 2014-05-15 | 2014-07-30 | 北京优捷信达信息科技有限公司 | Method for judging emotional tendentiousness of short text |
| CN104484815A (en) * | 2014-12-18 | 2015-04-01 | 刘耀强 | Product-oriented emotion analysis method and system based on fuzzy body |
| CN104778240A (en) * | 2015-04-08 | 2015-07-15 | 重庆理工大学 | Micro blog text data classification method on basis of multi-feature fusion |
| CN105468731A (en) * | 2015-11-20 | 2016-04-06 | 成都科来软件有限公司 | Preprocessing method of text sentiment analysis characteristic verification |
| CN105512311A (en) * | 2015-12-14 | 2016-04-20 | 北京工业大学 | Chi square statistic based self-adaption feature selection method |
| CN107908715A (en) * | 2017-11-10 | 2018-04-13 | 中国民航大学 | Microblog emotional polarity discriminating method based on Adaboost and grader Weighted Fusion |
-
2018
- 2018-06-13 CN CN201810610017.7A patent/CN108920545B/en not_active Expired - Fee Related
Patent Citations (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US5887069A (en) * | 1992-03-10 | 1999-03-23 | Hitachi, Ltd. | Sign recognition apparatus and method and sign translation system using same |
| CN102945268A (en) * | 2012-10-25 | 2013-02-27 | 北京腾逸科技发展有限公司 | Method and system for excavating comments on characteristics of product |
| CN103955451A (en) * | 2014-05-15 | 2014-07-30 | 北京优捷信达信息科技有限公司 | Method for judging emotional tendentiousness of short text |
| CN104484815A (en) * | 2014-12-18 | 2015-04-01 | 刘耀强 | Product-oriented emotion analysis method and system based on fuzzy body |
| CN104778240A (en) * | 2015-04-08 | 2015-07-15 | 重庆理工大学 | Micro blog text data classification method on basis of multi-feature fusion |
| CN105468731A (en) * | 2015-11-20 | 2016-04-06 | 成都科来软件有限公司 | Preprocessing method of text sentiment analysis characteristic verification |
| CN105512311A (en) * | 2015-12-14 | 2016-04-20 | 北京工业大学 | Chi square statistic based self-adaption feature selection method |
| CN107908715A (en) * | 2017-11-10 | 2018-04-13 | 中国民航大学 | Microblog emotional polarity discriminating method based on Adaboost and grader Weighted Fusion |
Non-Patent Citations (3)
| Title |
|---|
| CHANGLI ZHANG ET AL.: "Sentiment Analysis of Chinese Documents: From Sentence to Document Level", 《JOURNAL OF THE AMERICAN SOCIETY FOR INFORMATION SCIENCE AND TECHNOLOGY》 * |
| 刘玉娇 等: "基于情感字典与连词结合的中文文本情感分类", 《四川大学学报（自然科学版）》 * |
| 董祥和: "基于情感特征向量空间模型的中文商品评论倾向分类算法", 《计算机应用与软件》 * |
Cited By (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109933793A (en) * | 2019-03-15 | 2019-06-25 | 腾讯科技（深圳）有限公司 | Text polarity identification method, apparatus, equipment and readable storage medium storing program for executing |
| CN109933793B (en) * | 2019-03-15 | 2023-01-06 | 腾讯科技（深圳）有限公司 | Text polarity identification method, device and equipment and readable storage medium |
| CN110895557A (en) * | 2019-11-27 | 2020-03-20 | 广东智媒云图科技股份有限公司 | Text feature judgment method and device based on neural network and storage medium |
| CN110895557B (en) * | 2019-11-27 | 2022-06-21 | 广东智媒云图科技股份有限公司 | Text feature judgment method and device based on neural network and storage medium |
| CN112214527A (en) * | 2020-09-25 | 2021-01-12 | 桦蓥(上海)信息科技有限责任公司 | Financial object abnormal factor screening and analyzing method based on failure response code |
| CN112464646A (en) * | 2020-11-23 | 2021-03-09 | 中国船舶工业综合技术经济研究院 | Text emotion analysis method for defense intelligence library in national defense field |
Also Published As
| Publication number | Publication date |
|---|---|
| CN108920545B (en) | 2021-07-09 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Dahou et al. | Word embeddings and convolutional neural network for arabic sentiment classification | |
| Ibrahim et al. | Sentiment analysis for modern standard arabic and colloquial | |
| Dalal et al. | Opinion mining from online user reviews using fuzzy linguistic hedges | |
| CN108920545A (en) | The Chinese affective characteristics selection method of sentiment dictionary and Ka Fang model based on extension | |
| CN108763214B (en) | Automatic construction method of emotion dictionary for commodity comments | |
| CN110728153A (en) | Multi-category emotion classification method based on model fusion | |
| Al-Ghadhban et al. | Arabic sarcasm detection in Twitter | |
| Ljajić et al. | Improving sentiment analysis for twitter data by handling negation rules in the Serbian language | |
| Ofek et al. | Improving sentiment analysis in an online cancer survivor community using dynamic sentiment lexicon | |
| CN107423371A (en) | A kind of positive and negative class sensibility classification method of text | |
| Chang et al. | A METHOD OF FINE-GRAINED SHORT TEXT SENTIMENT ANALYSIS BASED ON MACHINE LEARNING. | |
| Sousa et al. | Word sense disambiguation: an evaluation study of semi-supervised approaches with word embeddings | |
| Ibrahim et al. | Automatic expandable large-scale sentiment lexicon of Modern Standard Arabic and Colloquial | |
| Al-Saqqa et al. | Stemming effects on sentiment analysis using large arabic multi-domain resources | |
| Bashir et al. | Automatic Hausa LanguageText Summarization Based on Feature Extraction using Naïve Bayes Model | |
| Kotelnikova et al. | Lexicon-based methods and BERT model for sentiment analysis of Russian text corpora | |
| Hegde et al. | Findings of the shared task on sentiment analysis in tamil and tulu code-mixed text | |
| Reda et al. | A hybrid arabic text summarization approach based on transformers | |
| Tungthamthiti et al. | Recognition of sarcasm in microblogging based on sentiment analysis and coherence identification | |
| Kaddoura et al. | EnhancedBERT: A feature-rich ensemble model for Arabic word sense disambiguation with statistical analysis and optimized data collection | |
| CN111694960A (en) | E-commerce comment emotion analysis model based on combination of part-of-speech characteristics and viewpoint characteristics and convolutional neural network | |
| Liu et al. | Chinese comparative sentence identification based on the combination of rules and statistics | |
| Feng et al. | Product feature extraction via topic model and synonym recognition approach | |
| Batanović | Semantic similarity and sentiment analysis of short texts in Serbian | |
| Dinsoreanu et al. | Unsupervised Twitter Sentiment Classification. |
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 | ||
| CF01 | Termination of patent right due to non-payment of annual fee | ||
| CF01 | Termination of patent right due to non-payment of annual fee |
Granted publication date: 20210709 |