CN103729466A - Name country identification method based on WEB and GBBoosting algorithms - Google Patents

Abstract

The invention discloses a name country identification method based on WEB and GBBoosting algorithms, and belongs to the technical field of WEB data mining. The method comprises the steps of I. extracting names of scholars in universities through a WEB data extraction technology; II. constructing a GBBoosting algorithm: constructing weak classifiers, wherein each weak classifier outputs a weak classification hypothesis to an input sample, and a strong classifier is formed through weight fusion of all weak classifiers; and III. identifying countries of the names through the GBBoosting algorithm. The name country identification method based on WEB and GBBoosting algorithms disclosed by the invention effectively solves a problem on classifying names of two countries which are similar in the spelling way; and meanwhile, the method, compared to existing other classifying methods, is easier to implement, and can be better applied to engineering practices such as name country or city country semantic annotation.

Description

Name country origin recognition methods based on WEB and GBBoosting algorithm

Technical field

The invention belongs to WEB data mining technology field, be specifically related to a kind of name country origin recognition methods based on WEB and GBBoosting algorithm.

Background technology

Along with the high speed development of Internet and becoming increasingly abundant of WEB resource, in order to excavate and to need and significant data fast and accurately from the data message of magnanimity, in recent years, WEB semantic analysis technology and Text Classification are widely used at WEB Data Mining, based on being applied in of WEB, in some degree, change user's habits and customs and working method, be also subject to increasing users' welcome and appreciation.

The sorting technique such as KNN, Bayes has obtained good classifying quality in numerous classification field, for example, the people such as Xie Mei are applied to image processing field by KNN, proposed a kind of MR gradation of image nonuniformity correction dividing method based on KNN sorting algorithm (patent No.: 201010583560.6, open day: 2011.07.27); The people such as willow are applied to computer software fields by Bayes, proposed a kind of based on improve the short message intelligent classification of Bayes's classification and searching method (patent No.: 201310356056.6, open day: 2013.12.04).But the classification accuracy of above-mentioned sorting technique in name country origin classification scene needs further to be improved, and especially, in the situation that two national name spell modes are close, its classification accuracy is only higher than random conjecture.In name country origin classification application, there is great limitation in visible above-mentioned sorting algorithm.

The deficiency existing in name country origin classification problem based on above-mentioned sorting technique, the present invention proposes a kind of GBBoosting algorithm based on Boosting, be intended to solve the problem existing in name country origin classification scene, compared with other sorting algorithm, its classification accuracy and recall rate are enhanced, especially, in the close situation of two the national name spell modes of classifying, performance is outstanding.GBBoosting algorithm application, in the identification scenes such as name country origin, city country origin, is carried out to the country origin semantic tagger in name or city, and then is applied in burning hot social field, there is very important realistic meaning and wide application prospect.

Summary of the invention

In view of this, the object of the present invention is to provide a kind of name country origin recognition methods based on WEB and GBBoosting algorithm, the method is extracted the scholar of colleges and universities name by WEB Data Extraction Technology, by structure Weak Classifier, each Weak Classifier is to weak typing hypothesis of input sample output, by the weight of all Weak Classifiers, merge and form a strong classifier, finally by the country under GBBoosting algorithm identified name.

For achieving the above object, the invention provides following technical scheme:

A name country origin recognition methods based on WEB and GBBoosting algorithm, comprises the following steps: step 1: by WEB Data Extraction Technology, extract the scholar of colleges and universities name; Step 2: structure GBBoosting algorithm: structure Weak Classifier, each Weak Classifier, to weak typing hypothesis of input sample output, is merged and is formed a strong classifier by the weight of all Weak Classifiers; Step 3: by the country origin under GBBoosting algorithm identified.

Further, in step 1, by GOOGLE search engine interface, obtain institute of the colleges and universities page, then at institute's page, carry out semantic analysis and obtain the scholar of the institute place page, finally by named entity recognition technology and semantic analysis technology, obtain extracting the scholar's information in the page.

Further, in step 2, the constitution step of Weak Classifier specifically comprises:

1) by the training text vector representation of two types, be ${\stackrel{\&RightArrow;}{V}}_1=(x_1,x_2,...,x_i,...,x_n),{\stackrel{\&RightArrow;}{V}}_2=(y_1,{\mathrm{<figure-callout\; id="2"\; label="y"\; filenames="HDA0000457910160000012.png"\; state="\{\{state\}\}">y</figure-callout>}}_2,...,y_i,...,y_n);$

2) according to formula calculate two kinds of training texts

intermediate vector

${\stackrel{\&RightArrow;}{V}}_3=(z_1{,\mathrm{<figure-callout\; id="2"\; label="z"\; filenames="HDA0000457910160000012.png"\; state="\{\{state\}\}">z</figure-callout>}}_2,...,z_i,...,z_n);$

3) according to formula

calculate intermediate vector

vertical vector

for any one test vector a _iif, (w _ia _i) > 0, by a _ilabel be+1, if (w _ia _i) < 0, by a _ilabel be-1;

Iteration Weak Classifier, its weights merge and form strong classifier, and its concrete steps are as follows:

First, given two training set D ₁=(x ₁, x ₂..., x _i..., x _n), D ₂=(y ₁, y ₂..., y _i..., y _n), a test set D _test=(z ₁, z ₂..., z _i..., z _n), by training set D1, D2, test set D _test, be expressed as vector form: $D_1=(\stackrel{\&RightArrow;}{x_1},\stackrel{\&RightArrow;}{x_2},...,\stackrel{\&RightArrow;}{x_i},...,\stackrel{\&RightArrow;}{x_n}),{\mathrm{<figure-callout\; id="2"\; label="D"\; filenames="HDA0000457910160000012.png"\; state="\{\{state\}\}">D</figure-callout>}}_2=(\stackrel{\&RightArrow;}{y_1},\stackrel{\&RightArrow;}{{\mathrm{<figure-callout\; id="2"\; label="y"\; filenames="HDA0000457910160000012.png"\; state="\{\{state\}\}">y</figure-callout>}}_2},...,\stackrel{\&RightArrow;}{y_i},...,\stackrel{\&RightArrow;}{y_n}),D_{\mathrm{Test}}=(\stackrel{\&RightArrow;}{z_1},\stackrel{\&RightArrow;}{{\mathrm{<figure-callout\; id="2"\; label="z"\; filenames="HDA0000457910160000012.png"\; state="\{\{state\}\}">z</figure-callout>}}_2},...,\stackrel{\&RightArrow;}{z_i},...,\stackrel{\&RightArrow;}{z_n}),$ And difference initialization D ₁, D ₂, D _testin sample weights;

Secondly, 1) from D ₁, D ₂in choose at random the individual sample of M (N/5<M<N) composition subset D ₁₁, D ₂₁, respectively to subset D ₁₁, D ₂₁in corresponding be added and unit obtains two vectors of vector

2), according to the construction process of linear classifier, obtain and two vectors

intermediate vector

vertical vector

generate Weak Classifier H (x) ₁; Through p circulation, obtain p different vertical vector

p Weak Classifier h (x) ₁, h (x) ₂..., h (x) _p; Final H (x)=h (x) ₁+ h (x) ₂+ ...+h (x) _p,

Further, in step 3, the scholar of colleges and universities name is gone out to scholar belonging country by GBBoosting algorithm identified.

Beneficial effect of the present invention is: the invention provides a kind of name country origin recognition methods based on WEB and GBBoosting algorithm, effectively solved unclassified problem in the situation that two national name spell modes are close; This method is more easily implemented than existing other sorting technique simultaneously, can better be applied in the engineering practices such as name country origin or city country origin semantic tagger.

Accompanying drawing explanation

In order to make object of the present invention, technical scheme and beneficial effect clearer, the invention provides following accompanying drawing and describe:

Fig. 1 is the macro flow chart of the method for the invention;

Fig. 2 is vector similarity calculating chart;

Fig. 3 is Weak Classifier structural map;

Fig. 4 is the microcosmic process flow diagram of this method.

Embodiment

Below in conjunction with accompanying drawing, the preferred embodiments of the present invention are described in detail.

Fig. 1 is the macro flow chart of the method for the invention, and as shown in the figure, this method comprises the following steps: step 1: by WEB Data Extraction Technology, extract the scholar of colleges and universities name; Step 2: structure GBBoosting algorithm: structure Weak Classifier, each Weak Classifier, to weak typing hypothesis of input sample output, is merged and is formed a strong classifier by the weight of all Weak Classifiers; Step 3: by the country origin under GBBoosting algorithm identified.

Fig. 4 is the microcosmic process flow diagram of this method, now in conjunction with Fig. 4, the concrete implementation step of this method is described.

1. by WEB Data Extraction Technology, extract the scholar of colleges and universities name

1) by GOOGLE search engine search " university+computerscience ", find institute's homepage; 2) by institute's homepage, find and comprise all scholar's information pages in this institute.In school, scholar's name generally all can be present in corresponding institute (being), as long as find the URL of corresponding institute (being) just can obtain all scholars' of school name and homepage address.In step 2, look for the URL of Computer institute of corresponding university (being), through observing the URL address of institute's (being) and two pages of the scholar of institute, can obtain two rules:

1. a rear address packet is containing previous address.

2. in a rear address, also comprise " people, faculty, faculty & Advisors " feature.

Only need to travel through the all-links in School of Computer Science's (being), filter out and in link, meet above-mentioned two rules and link the URL that corresponding word is " faculty or people ", found through experiments and generally can filter out two URL addresses, why occur that two URL are owing to generally containing people menu in institute, and faculty belongs to the submenu link of people, second URL is only the link needing, so select second URL address when there is two URL, otherwise select first address.Finally input filters out URL can obtain all scholars' name and the personal homepage of correspondence.3) by the faculty of School of Computer Science's (being) page, extract all scholars' name and homepage.Whether extract all links of the faculty of School of Computer Science's (being) page, find text corresponding to link, be name by named entity technical Analysis text.

2. realize GBBoosting algorithm: structure Weak Classifier, each Weak Classifier, to weak typing hypothesis of input sample output, is merged and formed a strong classifier by the weight of all Weak Classifiers.

The structure of Weak Classifier is to be the inner product of vectors size that judges two class texts by simple space vector similarity, asks two vectorial corner dimensions.As shown in Figure 2, two texts are more similar, and the angle of corresponding vector is less, and the cosine value of angle is larger.As shown in Figure 3, Weak Classifier improves on the basis of simple space vector similarity, constructs a simple linear classifier.Its concrete steps are as follows:

Step 1: the training text vector representation of given two types ${\stackrel{\&RightArrow;}{V}}_1=(x_1,x_2,...,x_i,...,x_n),{\stackrel{\&RightArrow;}{V}}_2=(y_1,{\mathrm{<figure-callout\; id="2"\; label="y"\; filenames="HDA0000457910160000012.png"\; state="\{\{state\}\}">y</figure-callout>}}_2,...,y_i,...,y_n);$ Step 2: 1) according to formula

calculate two kinds of training texts

intermediate vector

2) according to formula

calculate intermediate vector vertical vector

$\stackrel{\&RightArrow;}{V}=(m_1,{\mathrm{<figure-callout\; id="2"\; label="m"\; filenames="HDA0000457910160000012.png"\; state="\{\{state\}\}">m</figure-callout>}}_2,...,m_i,...,m_n).$

Step 3: the vector that has a d dimension

with threshold value 0, for any one test vector a _iif, (w _ia _i) > 0, by a _ilabel be+1, if (w _ia _i) < 0, by a _ilabel be-1.

By Weak Classifier, be the basis of realizing GBBoosting algorithm, each Weak Classifier, to weak typing hypothesis of input sample output, is merged and is formed a strong classifier by the weight of all Weak Classifiers.Given two training set D ₁=(x ₁, x ₂..., x _i..., x _n), D ₂=(y ₁, y ₂..., y _i..., y _n).Respectively from D ₁, D ₂in choose at random M sample, generate two vectors

by calculating the intermediate vector vectorial with two

vertical vector by test set D _test=(z ₁, z ₂..., z _i..., z _n) in each sample and vectorial V do dot product, the classification of the positive and negative judgement sample by dot product result, its concrete steps are as follows:

Step 1: two training set D ₁=(x ₁, x ₂..., x _i..., x _n), D ₂=(y ₁, y ₂..., y _i..., y _n), a test set D _test=(z ₁, z ₂..., z _i..., z _n), by training set D1, D2, test set D _test, be expressed as vector form: $D_1=(\stackrel{\&RightArrow;}{x_1},\stackrel{\&RightArrow;}{x_2},...,\stackrel{\&RightArrow;}{x_i},...,\stackrel{\&RightArrow;}{x_n}),{\mathrm{<figure-callout\; id="2"\; label="D"\; filenames="HDA0000457910160000012.png"\; state="\{\{state\}\}">D</figure-callout>}}_2=(\stackrel{\&RightArrow;}{y_1},\stackrel{\&RightArrow;}{{\mathrm{<figure-callout\; id="2"\; label="y"\; filenames="HDA0000457910160000012.png"\; state="\{\{state\}\}">y</figure-callout>}}_2},...,\stackrel{\&RightArrow;}{y_i},...,\stackrel{\&RightArrow;}{y_n}),D_{\mathrm{Test}}=(\stackrel{\&RightArrow;}{z_1},\stackrel{\&RightArrow;}{{\mathrm{<figure-callout\; id="2"\; label="z"\; filenames="HDA0000457910160000012.png"\; state="\{\{state\}\}">z</figure-callout>}}_2},...,\stackrel{\&RightArrow;}{z_i},...,\stackrel{\&RightArrow;}{z_n}),$ And difference initialization D ₁, D ₂, D _testin sample weights.

Step 2: 1) from D ₁, D ₂in choose at random the individual sample of M (N/5<M<N) composition subset D ₁₁, D ₂₁, respectively to subset D ₁₁, D ₂₁in corresponding be added and unit obtains two vectors of vector

2), according to the construction process of linear classifier, obtain and two vectors

intermediate vector

vertical vector

generate Weak Classifier H (x) ₁.Through p circulation, obtain p different vertical vector

p Weak Classifier h (x) ₁, h (x) ₂..., h (x) _p.

Step 3: H (x)=h (x) ₁+ h (x) ₂+ ...+h (x) _p,

Finally explanation is, above preferred embodiment is only unrestricted in order to technical scheme of the present invention to be described, although the present invention is described in detail by above preferred embodiment, but those skilled in the art are to be understood that, can to it, make various changes in the form and details, and not depart from the claims in the present invention book limited range.

Claims (4)

1. the name country origin recognition methods based on WEB and GBBoosting algorithm, is characterized in that: comprise the following steps: step 1: by WEB Data Extraction Technology, extract the scholar of colleges and universities name;

Step 2: structure GBBoosting algorithm: structure Weak Classifier, each Weak Classifier, to weak typing hypothesis of input sample output, is merged and is formed a strong classifier by the weight of all Weak Classifiers;

Step 3: by the country origin under GBBoosting algorithm identified.

2. the name country origin recognition methods based on WEB and GBBoosting algorithm according to claim 1, it is characterized in that: in step 1, by GOOGLE search engine interface, obtain institute of the colleges and universities page, then at institute's page, carry out semantic analysis and obtain the scholar of the institute place page, finally by named entity recognition technology and semantic analysis technology, obtain extracting the scholar's information in the page.

3. the name country origin recognition methods based on WEB and GBBoosting algorithm according to claim 1, is characterized in that: in step 2, the constitution step of Weak Classifier specifically comprises:

1) by the training text vector representation of two types, be ${\stackrel{\&RightArrow;}{V}}_1=(x_1,x_2,...,x_i,...,x_n),{\stackrel{\&RightArrow;}{V}}_2=(y_1,y_2,...,y_i,...,y_n);$

2) according to formula

calculate two kinds of training texts

intermediate vector

${\stackrel{\&RightArrow;}{V}}_3=(z_1{,z}_2,...,z_i,...,z_n);$

3) according to formula

calculate intermediate vector

vertical vector

for any one test vector a _iif, (w _ia _i) > 0, by a _ilabel be+1, if (w _ia _i) < 0, by a _ilabel be-1;

Iteration Weak Classifier, its weights merge and form strong classifier, and its concrete steps are as follows:

First, given two training set D ₁=(x ₁, x ₂..., x _i..., x _n), D ₂=(y ₁, y ₂..., y _i..., y _n), a test set D _test=(z ₁, z ₂..., z _i..., z _n), by training set D1, D2, test set D _test, be expressed as vector form: $D_1=(\stackrel{\&RightArrow;}{x_1},\stackrel{\&RightArrow;}{x_2},...,\stackrel{\&RightArrow;}{x_i},...,\stackrel{\&RightArrow;}{x_n}),D_2=(\stackrel{\&RightArrow;}{y_1},\stackrel{\&RightArrow;}{y_2},...,\stackrel{\&RightArrow;}{y_i},...,\stackrel{\&RightArrow;}{y_n}),D_{\mathrm{Test}}=(\stackrel{\&RightArrow;}{z_1},\stackrel{\&RightArrow;}{z_2},...,\stackrel{\&RightArrow;}{z_i},...,\stackrel{\&RightArrow;}{z_n}),$ And difference initialization D ₁, D ₂, D _testin sample weights;

Secondly, 1) from D ₁, D ₂in choose at random the individual sample of M (N/5<M<N) composition subset D ₁₁, D ₂₁, respectively to subset D ₁₁, D ₂₁in corresponding be added and unit obtains two vectors of vector 2), according to the construction process of linear classifier, obtain and two vectors

intermediate vector

vertical vector

generate Weak Classifier H (x) ₁; Through p circulation, obtain p different vertical vector p Weak Classifier h (x) ₁, h (x) ₂..., h (x) _p; Final H (x)=h (x) ₁+ h (x) ₂+ ...+h (x) _p,

4. the name country origin recognition methods based on WEB and GBBoosting algorithm according to claim 1, is characterized in that: in step 3, the scholar of colleges and universities name is gone out to scholar belonging country by GBBoosting algorithm identified.

Images