CN106202372A - A kind of method of network text information emotional semantic classification - Google Patents

Abstract

A kind of method that the invention discloses network text information emotional semantic classification, comprises the steps: step one, first determines whether whether document belongs to news, if belonging to news, only extracting title and carrying out emotional semantic classification, the most whole document being carried out emotional semantic classification；Step 2, the document classifying needs carry out pretreatment；Step 3, according to text length, document is classified: use TF IDF to calculate feature weight more than the document of 140 characters length and then utilize the logistic regression grader trained to classify；Artificial emotion classifying rules is used to classify less than the document of 140 characters length.Compared with prior art, the positive effect of the present invention is: the inventive method is for long text, the different characteristics of short text, use machine learning algorithm to build grader and formulate the technology path that characteristic of division combines with domain expert, it is possible to accurately and timely find reaction information, sensitive information and the negative report related in network public-opinion.

Description

A kind of method of network text information emotional semantic classification

Technical field

The invention belongs to natural language processing field, a kind of method relating to network text information emotional semantic classification.

Background technology

Network, as a kind of New Media, has been given play to the unimpeded will of the people more and more, has been expressed demand, supervises by public opinionus, joins The effect that political affairs are discussed political affairs and other major and important matters, increasing people utilizes the Internet to express oneself demand at the aspect such as interests, politics, expresses oneself Attitude or view to hot spot of societys such as the people's livelihood, the administration of justice, anti-corruption.Especially when groupment, unexpected incidents occur, People are often transmitted by the Internet in the very first time or obtain information.But some hostile forces utilize the hidden of network simultaneously Cover environment, manufacture the network public opinion deviating from main flow political culture, destroy steady politics and social harmony；The net that minority is unique The people utilize the facility of network to have a mind to spread dummy message even rumour, to reaching the individual purpose of oneself；Some netizens give vent to feelings Thread, delivers vulgar, the speech of Lycoperdon polymorphum Vitt on network.And tradition relies on artificial method to be difficult to tackle the collection of online magnanimity information With study and judge, it is therefore desirable to network public sentiment information is judged by the sensibility classification method of automatization, it is achieved to reaction, sensitivity, negative The emphasis of the public feelings informations such as face finds.

Summary of the invention

In order to overcome the shortcoming of prior art, a kind of method that the invention provides network text information emotional semantic classification.

The technical solution adopted in the present invention is: a kind of method of network text information emotional semantic classification, comprises the steps:

Step one, first determine whether whether document belongs to news, if belonging to news, only extracting title and carrying out emotional semantic classification, The most whole document is carried out emotional semantic classification；

Step 2, the document classifying needs carry out pretreatment；

Step 3, according to text length, document is classified:

(1) then use TF-IDF to calculate feature weight more than the document of 140 characters length utilizes train to patrol Collect recurrence grader to classify；

(2) artificial emotion classifying rules is used to classify less than the document of 140 characters length.

Compared with prior art, the positive effect of the present invention is: the inventive method is for long text, the different spies of short text Point, uses machine learning algorithm to build grader and formulates the technology path that characteristic of division combines with domain expert, it is possible to accurately Find reaction information, sensitive information and the negative report related in network public-opinion in time.

Accompanying drawing explanation

Examples of the present invention will be described by way of reference to the accompanying drawings, wherein:

Fig. 1 is network text information classification schematic diagram；

Fig. 2 is logistic regression classification based training process schematic；

Fig. 3 is artificial formulation emotional semantic classification rule schematic diagram.

Detailed description of the invention

A kind of method of network text information emotional semantic classification, as depicted in figs. 1 and 2, comprises the steps:

Step one, first determine whether whether document belongs to news, if belonging to news, only extracting title and carrying out emotional semantic classification, The most whole document is carried out emotional semantic classification；

Step 2, the document needing classification is carried out pretreatment:

Pretreatment refers to use Chinese lexical analysis system ICTCLAS that text is carried out participle, then filters stop words:

Destination document carried out emotional semantic classification and to first have to carry out data prediction, mainly include participle and remove stop words two Part.The ICTCLAS Chinese word segmentation system of Inst. of Computing Techn. Academia Sinica's exploitation is based on stacking hidden Markov model (Hierarchical Hidden Markov Model), it is possible to effectively by the text success participle of input, and follow part of speech closely Export, and support to import the dictionary that user provides.By ICTCLAS, input text is carried out participle and part-of-speech tagging, Then filter and disable the word in vocabulary, thus remove the word without practical significance, it is achieved data prediction.

Step 3, according to text length, document is classified:

(1) then use TF-IDF to calculate feature weight more than the document of 140 characters length utilizes train to patrol Collect recurrence grader to classify:

(1) feature extraction: by calculating χ²Statistical value selection sort feature, TF-IDF calculates feature weight:

χ²Statistics also referred to as CHI statistics, is the statistics of the dependence for two variablees of tolerance conventional in statistics Amount, characteristic item t and the χ of classification c²Statistical value is calculated as follows:

${\mathrm{\χ}}^2(t,c)=\frac{N{(AD-BC)}^2}{(A+C)(B+D)(A+B)(C+D)}$

Wherein N represents that total number of documents, A represent and comprises characteristic item t and belong to the number of files of classification c, and B represents and comprises feature T and be not belonging to the number of files of classification c, C represents and does not comprises characteristic item t and belong to the number of files of classification c, and D represents and do not wraps Containing characteristic item t and the number of files that is not belonging to classification c.

Obtaining t χ in each category²The t χ to all categories can be calculated according to following equation after statistical value²System Meter meansigma methods.

${{\mathrm{\χ}}^2}_{AVG}\left(t\right)=\underset{i=1}{\overset{M}{\Σ}}P(t,c_i){\mathrm{\χ}}^2(t,c_i)$

Wherein, P (t, c_i) represent that characteristic item t belongs to classification c_iProbability.

When carrying out feature selection, set threshold values, only retain χ²Statistical value is more than the lexical item of threshold values as characteristic item.

To give weights, the spy high to discrimination to different characteristic item by TF-IDF method after screening obtains characteristic item Levy item and give bigger weights.TF-IDF weight computing formula is as follows, and wherein tf refers to characteristic item word frequency, and what n represented is bag Containing the number of files of current signature item, N is total number of documents.

$W=tf\×idf=tf\×log\frac{N}{n}$

(2) training LR grader: use logistic regression algorithm to build grader

Can be the N-dimensional vector in vector space model by text representation by data prediction and feature extraction, will This N-dimensional vector differentiates that result exports as grader as the input of logistic regression grader, document emotion, carries out grader Training.In logistic regression sorting algorithm, using cross entropy cost function to carry out the degree of predictive metrics mistake, computing formula is such as Under, wherein y is desired output, and a is the output that grader is actual, and for avoiding training over-fitting, arranging parameter c is 200, arranges It is 0.01 that gradient descent algorithm learns the step-length δ parameter of J (x) minima.

$J\left(x\right)=-\frac{1}{n}\underset{x}{\Σ}\[y\log a+(1-y)log(1-a)\]+c{\left(w\right)}^2$

Weight vectors w and amount of bias b can be obtained through training, can be according to given input by formula below Example x is calculated its conditional probability distribution P belonging to a certain classification (Y | X), and the maximum classification of probit is then belonging to it Classification.

$P\left(Y\right|X)=\frac{e^{(w\·x+b)}}{1+e^{(w\·x+b)}}$

(2) artificial emotion classifying rules is used to classify less than the document of 140 characters length:

Domain expert builds seed dictionary of all categories respectively, then extends seed by Chinese thesaurus with word2vector Dictionary, forms feature database of all categories；The impact on word emotion value of COMPREHENSIVE CALCULATING negative word, sentence pattern, it is judged that document emotion value with The artificial size setting emotion threshold values of all categories.

As it is shown on figure 3, use the artificial process formulating text message emotional semantic classification rule to specifically include that

(1) seed words obtains

Weight vectors w can be obtained by machine learning algorithm training text this process of information emotion classifiers, pass through The Feature Words that screening weighted value is bigger can get part classifying seed words, the most comprehensive Northeastern University, Taiwan Univ., China The emotional semantic classification dictionary of Zhi Wangdeng mechanism, obtains final emotional semantic classification seed words.

(2) synonym extension

" Chinese thesaurus extended edition " that Harbin Institute of Technology's social computing is issued with Research into information retrieval center comprises altogether 77, Article 343, word, organizes all entries included together according to tree-shaped hierarchical structure, and provides Pyatyi to encode.According to " with Justice word word woods extended edition " the synonym phrase that provides carries out synonym extension to seed words.

(3) Word2Vector semantic extension

Word2Vector is a efficient work that word is characterized as real number value vector that Google increased income in year in 2013 Tool, utilizes the thought that the degree of depth learns, and the contextual information of bluebeard compound solves the similarity of text semantic.Word2Vector uses The term vector representation of DistributedRepresentation, uses the neutral net of three layers to build language model Mould, obtains word expression in vector space simultaneously.Utilize Word2Vector that existing emotional semantic classification word is carried out semantic similitude Degree calculates, and extracts the higher word of similarity and adds emotional semantic classification dictionary.

(4) according to setting classifying rules, document is classified:

Emotion word in emotional semantic classification dictionary is given different weights, by the relative position information of negative word Yu emotion word Calculate the emotion value of content of the sentence, then in conjunction with the different sentence pattern such as exclamative sentence, confirmative question, interrogative sentence to whole word emotion value shadow Ring, calculate the overall emotion value of sentence, the emotion value of all categories of all for document sentences is added the document emotion value that obtains and sets The threshold values of fixed each emotional category compares, thus judges document emotion.

The inventive method is carried out following accuracy test: obtained by web crawlers and include Sina's microblogging, forum, news Websites etc. 200,000 data is tested, and has been randomly chosen the observation of corresponding number in the result that all algorithms judge (random JAVA code is (int) (Math.random () * n) to sample, and during wherein n is 200,000 test data, algorithm judges Number of samples for respective type).5 emotions of record are studied and judged personnel and are judged the result of same observation sample simultaneously, finally take people The legitimate reading that result is sample that number is most.Data in table 1 are test result, it can be seen that the present invention is for reaction, quick The accuracy that sense, front, negative text emotion are classified is the highest.

Table 1 test of heuristics result table

Emotional category	Algorithm marks	Algorithm mark pair	Accuracy
				Negatively	502	385	0.766932271
Front	505	364	0.720792079
				Sensitive	715	540	0.755244755
Reaction	39	33	0.846153846

Claims (9)

1. the method for a network text information emotional semantic classification, it is characterised in that: comprise the steps:

Step one, first determine whether whether document belongs to news, if belonging to news, only extracting title and carrying out emotional semantic classification, otherwise Then whole document is carried out emotional semantic classification；

Step 2, the document classifying needs carry out pretreatment；

Step 3, according to text length, document is classified:

(1) then use TF-IDF to calculate feature weight more than the document of 140 characters length utilizes the logic trained to return Grader is returned to classify；

(2) artificial emotion classifying rules is used to classify less than the document of 140 characters length.

The method of a kind of network text information emotional semantic classification the most according to claim 1, it is characterised in that: described in step 2 Pretreatment refer to input text carry out participle and part-of-speech tagging, be then filtered off stop words.

The method of a kind of network text information emotional semantic classification the most according to claim 2, it is characterised in that: use Chinese word Method is analyzed system ICTCLAS and text is carried out pretreatment.

The method of a kind of network text information emotional semantic classification the most according to claim 1, it is characterised in that: described in step 3 The method using TF IDF calculating feature weight then to utilize the logistic regression grader trained to carry out classifying is:

(1) χ of characteristic item t and classification c it is calculated as follows²Statistical value:

${\mathrm{\χ}}^2(t,c)=\frac{N{(AD-BC)}^2}{(A+C)(B+D)(A+B)(C+D)}$

Wherein N represents that total number of documents, A represent and comprises characteristic item t and belong to the number of files of classification c, and B represents and comprises characteristic item t And being not belonging to the number of files of classification c, C represents and do not comprises characteristic item t and belong to the number of files of classification c, D represents and does not comprises spy Levy a t and be not belonging to the number of files of classification c；

(2) t χ to all categories it is calculated as follows²Assembly average:

${{\mathrm{\χ}}^2}_{AVG}\left(t\right)=\underset{i=1}{\overset{M}{\Σ}}P(t,c_i){\mathrm{\χ}}^2(t,c_i);$

(3) filter out more than the χ setting threshold values²The lexical item of statistical value is as characteristic item；

(4) use TF IDF method to come different characteristic item to give weights, give bigger weights to the characteristic item that discrimination is high；

(5) training LR grader:

It is the N-dimensional vector in vector space model by text representation, using vectorial for this N-dimensional input as logistic regression grader, Document emotion differentiates that result exports as grader, is trained grader obtaining weight vectors w and amount of bias b；

(6) it is calculated its conditional probability distribution belonging to a certain classification according to given input example x The classification of probit maximum is then its generic.

The method of a kind of network text information emotional semantic classification the most according to claim 4, it is characterised in that: (4th) step institute Stating weights uses equation below to calculate:

$W=tf\×idf=tf\×log\frac{N}{n}$

Wherein tf is characterized a word frequency, and n is the number of files comprising current signature item, and N is total number of documents.

The method of a kind of network text information emotional semantic classification the most according to claim 4, it is characterised in that: at logistic regression In sorting algorithm, following cross entropy cost function is used to carry out the degree of predictive metrics mistake:

$J\left(x\right)=-\frac{1}{n}\underset{x}{\Σ}\[y\log a+(1-y)log(1-a)\]+c{\left(w\right)}^2$

Wherein y is desired output, and a is the output that grader is actual, and c is 200.

The method of a kind of network text information emotional semantic classification the most according to claim 1, it is characterised in that: described in step 3 The method using artificial emotion classifying rules to carry out classifying is:

(1) seed words obtains；

(2) seed words is carried out synonym extension；

(3) utilize Word2Vector that existing emotional semantic classification word is carried out semantic extension；

(4) according to setting classifying rules, document is classified.

The method of a kind of network text information emotional semantic classification the most according to claim 7, it is characterised in that: described seed words Acquisition methods be: obtain weight vectors by machine learning algorithm training text information emotion classifiers, then by screening The Feature Words that weighted value is bigger gets part classifying seed words；The machines such as the most comprehensive Northeastern University, Taiwan Univ., middle National IP Network The emotional semantic classification dictionary of structure, obtains final emotional semantic classification seed words.

The method of a kind of network text information emotional semantic classification the most according to claim 7, it is characterised in that: divide according to setting The method that document is classified by rule-like is: the emotion word in emotional semantic classification dictionary gives different weights, passes through negative word Relative position information with emotion word calculates the emotion value of content of the sentence, then in conjunction with differences such as exclamative sentence, confirmative question, interrogative sentences Whole word emotion value is affected by sentence pattern, calculates the overall emotion value of sentence, the emotion value of all categories of all for document sentences is added The document emotion value obtained compares with the threshold values of each emotional category of setting, thus judges document emotion.