TWI509434B - Methods and apparatus for classification - Google Patents

Description

Method and device for classification

This case is about the computer and communication fields, especially the methods and devices for classifying this article.

This paper is an important part of this paper. It refers to determining a category for each document in the document collection according to the predefined topic category. By categorizing documents through the automated paper classification system, people can better find the information and knowledge they need. In the eyes of people, classification is one of the most basic forms of cognition of information. The traditional literature classification research has rich research results and considerable practical standards. However, with the rapid growth of information in this article, especially the proliferation of information on the Internet on the Internet, this paper has become a key technology for processing and organizing a large amount of documents. Now, the classification of this paper is widely used in various fields. However, with the increasing amount of information, people's requirements for accuracy and recall of content search will become higher and higher, so the demand for classification technology is greatly increased. How to construct an effective classification system is still This article explores a major research direction.

In the field of natural language processing, the representation of this paper mainly uses Vector space model (VSM). This method thinks that each article contains some independent attributes that are expressed by concept words to reveal their contents, and each attribute can be It is regarded as a dimension of the concept space. These independent attributes are called feature items in this paper. This article can be expressed as a collection of these feature items. The closeness of the eigenvectors is usually measured by the cosine of the angle. The categories herein are then determined based on how close the vectors are to the feature vectors of the candidate categories.

In the prior art, it is necessary to calculate the degree of similarity between each vector vector and all the feature vectors of the candidate category. Each calculation needs to be measured by the angle cosine, and the calculation amount is very large, and the prior art has no constraint on the semantics of the paper, and the classification is accurate. Degree is not very good.

The embodiment of the present invention provides a method and device for classifying the present article, which is used to implement the classification, simplify the classification operation, and improve the accuracy of the classification.

A method for classification in the present invention comprises the following steps: segmenting the obtained content of the article to obtain a plurality of vocabulary; determining a vocabulary of the vocabulary in the spherical space model for each of the obtained vocabulary Vector; the vocabulary vector of the vocabulary includes the normalized word frequency obtained by normalizing the vocabulary value of the vocabulary in various categories; the spherical space model is a multi-dimensional spherical model with a radius of unit length, spherical space The dimension is equal to the number of categories, and the category corresponds to a category vector in the spherical space; for each category, the distance between the sum of the obtained vocabulary vectors of multiple vocabularies to the category vector of the category is determined; Divide into the category corresponding to the shortest distance.

An apparatus for classification in the present invention, comprising: a word segmentation module, configured to perform word segmentation on the obtained content, and obtain a plurality of vocabulary; and a query module, configured to vocabulary for each of the obtained plurality of vocabulary, Determining the vocabulary vector of the vocabulary in the spherical space model; the vocabulary vector of the vocabulary includes the normalized word frequency value obtained by normalizing the vocabulary value of the vocabulary in various categories; the spherical space model is A multi-dimensional sphere model with a radius of unit length, the dimension of the spherical space is equal to the number of categories, the category corresponds to a category vector in the spherical space, and the calculation module determines the multiple vocabulary words for each category. The distance between the sum of the sink vectors and the category vector of the category; the classification module is used to classify the text into the category corresponding to the shortest distance.

The embodiment of the present invention pre-constructs a spherical space model, and classifies the paper based on the spherical space model. In the classification process, the vector of each vocabulary in this paper and the distance from various mesh vectors are calculated to determine the points that should be classified in this paper. Category. The embodiment of the present invention achieves the classification herein, and the amount of calculation is significantly reduced relative to the angle cosine algorithm in the prior art. And in the embodiment of the present invention, the spherical space model has a radius of unit length, and the sum of the squares of the normalized lexicon vector of each vocabulary in each category is also a unit length, which is equivalent to the semantic information amount of a vocabulary. Equivalent to the unit length, the amount of semantic information is constrained, so the accuracy of the classification can be improved compared to the prior art.

The embodiment of the present invention pre-constructs a spherical space model, and classifies the paper based on the spherical space model. In the classification process, the vector of each vocabulary in this paper and the distance from various mesh vectors are calculated to determine the points that should be classified in this paper. Category. The embodiment of the present invention achieves the classification herein, and the amount of calculation is significantly reduced relative to the angle cosine algorithm in the prior art. And in the embodiment of the present invention, the spherical space model has a radius of unit length, and the sum of the squares of the normalized lexicon vector of each vocabulary in each category is also a unit length, which is equivalent to the semantic information amount of a vocabulary. Equivalent to the unit length, the amount of semantic information is constrained, so the accuracy of the classification can be improved compared to the prior art.

Referring to FIG. 1 , the apparatus for classification in this embodiment includes: a word segmentation module 101, a query module 102, a calculation module 103, and a classification module 104.

The word segmentation module 101 is used to segment the obtained content of the article to obtain a plurality of vocabulary.

The query module 102 is configured to determine a vocabulary vector of the vocabulary in the spherical space model for one of the obtained plurality of vocabularies. The vocabulary vector of the vocabulary includes the normalized word frequency obtained by normalizing the vocabulary value of the vocabulary in various categories; the spherical space model is a multi-dimensional spherical model with a radius of unit length, and the dimension of the spherical space is equal to The number of categories, the category corresponds to a category vector in the spherical space. The unit length may be a constant. For the convenience of calculation, the radius of the spherical space model in this embodiment is 1. The vector of each vocabulary in this paper and the distance to each category vector are linear distance or spherical distance.

The calculation module 103 is configured to determine, for each category, a vocabulary vector of a plurality of vocabulary obtained after the word segmentation and a distance to each category vector.

The classification module 104 is used to classify the text into the category corresponding to the shortest distance.

When calculating the distance between the vocabulary vector and the various mesh vectors in the text, the calculation module 103 may accumulate the normalized word frequency values of the vocabulary vectors of the plurality of vocabularies obtained after the word segmentation in the corresponding category. , get the normalized lexical vector and . The classification module 104 divides the text into categories corresponding to the largest component of the normalized lexicon vector sum.

The device further includes: an interface module 105, a filter module 106, a structure module 107, and a storage module 108, as shown in FIG.

The interface module 105 is used to obtain the text to be classified from outside the device.

The filtering module 106 is configured to filter the obtained plurality of vocabularies after obtaining a plurality of vocabulary for the word segmentation, and obtain a plurality of vocabularies that meet the filtering condition. There are various filtering conditions, such as calculating the coefficient of variation of the vocabulary according to the word frequency value of each vocabulary, and then filtering out the vocabulary whose coefficient of variation is greater than the preset coefficient of variation threshold (such as 0.5). Through the coefficient of variation, it is possible to filter out words that have little change in the frequency of words in various categories (such as you, me, etc., the frequency of words in each type of target is basically the same), while retaining the words whose frequency changes significantly in various categories ( For professional terms, the word frequency in the categories related to their profession is significantly higher than the word frequency in other categories). Words with more obvious changes in the frequency of words in various categories indicate that they mainly appear in one or several categories, and such words make more contributions to the accuracy of the classification of this paper. This embodiment considers such Words are excellent words, and should be filtered to filter out excellent words. There may be other filters, which are not listed here.

The construction module 107 is used to construct a spherical space model.

The storage module 108 is used to store the constructed spherical space model, and to store and store the various articles.

The process of constructing the spherical space model by the construction module 107 is as follows:

Let the multidimensional spherical space be S, and the dimension of S is the same as the total number of categories. The category Ci is an endpoint on the sphere and corresponds to a category vector in the spherical space, Ci={0,...,0,1,0,...,0}, which is equivalent to the sphere (equivalent to The origin point points to the end of the sphere. The i-th dimension of the category vector is 1 and the rest are 0. In this embodiment, assuming that the probability that any one of the vocabularies appears in any two of the categories Ci and Cj is probability independent, then Ci and Cj are necessarily perpendicular to each other in S, generalized to general, all category vectors { Ci} is two or two vertical.

The vocabulary vector Wm of the mth vocabulary in this embodiment is a vector in S, m=1...M, and M is the total number of vocabularies. Wm={V1, V2, ..., VN}, Vi is the normalized word frequency value on the category Ci, i=1...N, and N is the total number of categories. The normalized word frequency value points from the center of the sphere to the spherical end point, and the normalized word frequency value can be expressed as a coordinate on the category Ci. See Figure 3 for the vocabulary vector and category vector of the vocabulary. Ci, Cj and Ck represent three category vectors, O is the center of the sphere, and is also the origin (coordinates are {0,0,...,0 }).

The semantic information of any vocabulary in this embodiment is the same constant. The semantic information quantity refers to the logical meaning of the existence mode and the motion state of the things perceived or expressed by the subject, and is the intrinsic meaning factor of the vocabulary. Information section. The constant is defined as the unit length, and the length of the vocabulary vector in S (ie, the distance from the end of the lexicon vector to the origin O) is also the constant. For the convenience of calculation, the constant is set to 1. The distance from the endpoint of the vocabulary vector to the origin O can be expressed as: |Wm - O| = 1 (formula 1), and further according to Wm = {V1, V2, ..., VN} Σ Vi ² =1 (Formula 2 ). It can be known from Equation 1 that the endpoints of the vocabulary vector Wm fall on the spherical surface. Since the endpoints of the vocabulary vector Wm and the category vector Ci fall on the spherical surface, the degree of semantics and category approximation of any vocabulary can be expressed by the distance between Wm and Ci, and the shorter the distance, the closer. The distance between Wm and Ci can be calculated by linear distance or spherical distance.

Since the amount of semantic information defining any vocabulary is the same constant, the normalized word frequency value is normalized after the word frequency value is obtained. And then there is Σ(Fi×k) ² =1, where Fi is the word frequency value of the vocabulary on the category Ci, and k is the preset normalization coefficient. A ^{Σ (Fi × k) 2 =} 1 can be derived (Equation 3). Then, from Vi=Fi×k (Equation 4), a conversion function (or a quantization function) of the vocabulary vector and the word frequency value can be obtained. Wm=δ(Fi)={Fi}×k (Equation 5).

Through the above description, the construction module 107 constructs a spherical space with the origin as the center of the sphere and the radius of the unit length 1 , and the end points of the vocabulary vector Wm and the category vector Ci fall on the spherical surface. The spherical space model is trained and learned through training samples to obtain a spherical space model that can be directly applied. The sample training process is similar to the classification process in this paper, and can also be implemented by other pattern recognition or manual methods.

For a paper D, D=ΣWm, Wm is the lexical vector of the mth vocabulary in this paper. The calculation module 103 calculates the distance between the ΣWm and the category vector Ci, and the category corresponding to the shortest distance is the category to be classified in this paper. Since ΣWm does not necessarily fall on the sphere, for ease of calculation, the calculation module 103 can also normalize D, multiply by the normalization coefficient k, and then calculate the distance from the category vector Ci.

Since the distance between the vocabulary vector Wm and the category vector Ci is shorter, the degree of approximation of the vocabulary vector Wm and the category vector Ci is larger. To simplify the calculation process, P={Pi}={ΣVmi} can be set. Formula 6), Pi represents the weight component of the i-th class, and the larger the Pi value is, the shorter the distance corresponding to the category vector Ci is, and the larger the ΣVmi is, the shorter the distance corresponding to the category vector Ci is. Therefore, when calculating the distance sum, the calculation module 103 may accumulate the obtained normalized component values of the plurality of vocabularies on the category to obtain the weight value of the category. The classification module 104 assigns the text to the category corresponding to the maximum weight value.

The larger the Pi value, the shorter the distance corresponding to the category vector Ci is as follows: since D=ΣWm, Wm={V1, V2, . . . , VN}, then D={ΣVm1, ΣVm2, ..... .ΣVmi,......,ΣVmn},ΣVmi is the sum of the normalized word frequency values of all vocabulary of the document on the i-th category. Let's set Pi=ΣVmi, then D={Pi}. The distance from D to Ci can be expressed as:

|D-Ci|=|{P1,P2,...Pi,...,Pn}×k-{0,0,...,0,1,0,....,0}|= k×|{P1,P2,...Pi,...,Pn}-{0,0,...,0,1/k,0,....,0}|=k×sqrt( (P1-0) ² +(P2-0) ² +...+(Pi-1/k) ² +...+(Pn-0) ² )=k×sqrt(P1 ² +P2 ² +. ..+(Pi ² -2Pi/k+1/k ² )+...+Pn ² )=k×sqrt(Σ(Pi ² )-2Pi/k+1/k ² )=sqrt(Σ(( Pi×k) ² )-2K×Pi+1)

Since Σ((Pi×k) ² )=1, so: sqrt(Σ((Pi×k) ² )-2K×Pi+1)=sqrt(1-2K×Pi+1)=sqrt(2*( 1-K×Pi)). It can be seen that the distance from D to Ci is inversely proportional to Pi, and the largest category of Pi is the closest category. Where sqrt indicates √.

The device can be located in a computer device, or each module in the device is implemented by a different computer device, and the functions of the device are performed by a plurality of computer devices. Each module in the device may be a soft body, a hardware body, or a combination of a software body and a hardware body.

The internal structure and function of the classification device are described above. The implementation process of this classification is introduced below.

Referring to FIG. 4, the main method of the method in this embodiment is as follows: Step 401: Perform word segmentation on the obtained content to obtain multiple vocabularies.

Step 402: Determine a vocabulary vector of the vocabulary in the spherical space model for each of the obtained plurality of vocabularies. The vocabulary vector of the vocabulary includes the normalized word frequency obtained by normalizing the vocabulary value of the vocabulary in various categories; the spherical space model is a multi-dimensional spherical model with a radius of unit length, and the dimension of the spherical space is equal to The number of categories, the category corresponds to a category vector in the spherical space.

Step 403: For each category, determine the distance from the sum of the obtained vocabulary vectors of the plurality of vocabularies to the category vector of the category.

Step 404: Divide the article into the category corresponding to the shortest distance.

In this embodiment, the classification can be performed by distance, and the classification can be performed according to the sum of the lexicon vectors in the present paper. The following describes the process of classification in detail for the two cases.

Referring to FIG. 5, the method for passing the distance and performing the classification in this embodiment is as follows: Step 501: Perform word segmentation on the obtained content to obtain multiple vocabularies.

Step 502: Filter the obtained multiple vocabularies to obtain a plurality of vocabularies that meet the filtering conditions. The filtering module 106 can filter the vocabulary according to the word frequency value. There are various filtering conditions, such as the word frequency average value remaining on all categories is greater than the preset number; for example, the maximum component (ie, the maximum normalized word frequency value) in the normalized vocabulary vector of the vocabulary is greater than the preset word frequency threshold. Hui et al., not listed here.

Step 503: Query the normalized word frequency value of the vocabulary on various categories for each vocabulary that meets the filtering condition. Among them, the normalized word frequency value of all vocabulary in various categories is pre-stored. If the stored vocabulary is not enough, the vocabulary of the vocabulary in each category cannot be queried. Both are 0. If there is no normalized word frequency value of the vocabulary in various categories, but the word frequency value of the vocabulary in various categories is stored, the query module 102 can query the word frequency value and normalize the word frequency value. To obtain the normalized word frequency value, the specific implementation can be found in Equation 4. This step can filter out interference vocabulary (such as uncharted vocabulary and common vocabulary, etc.) and try to filter out professional vocabulary to improve the accuracy of the classification.

Step 504: For each category, determine a vocabulary vector of a plurality of vocabularies that meet the filtering condition and a distance to each category vector. The distance between the vocabulary vector and the category vector is a linear distance or a spherical distance.

Prior to step 504, the resulting lexicon vector sum may also be normalized such that the normalized vocabulary vector sum falls within the spherical space. The normalized vocabulary vector and distance to the category are then determined in step 504.

Step 505: Divide the article into the category corresponding to the shortest distance.

This article can be further saved in the database by category.

Referring to FIG. 6, the method for normalizing the vocabulary vector and performing the classification in this embodiment is as follows: Step 601: Perform word segmentation on the obtained content to obtain multiple vocabularies.

Step 602: Filter the obtained multiple vocabularies to obtain a plurality of vocabularies that meet the filtering conditions.

Step 603: Query the normalized word frequency value of the vocabulary on various categories for each vocabulary that meets the filtering condition. Among them, the normalized word frequency value of all vocabulary in various categories is pre-stored.

Step 604: For each category, accumulate the obtained normalized word frequency values of the plurality of vocabularies on the category to obtain a normalized vocabulary vector sum. See Equation 6 for specific implementation.

Step 605: Divide the text into categories corresponding to the largest component of the normalized lexicon vector sum.

The software for implementing the embodiments of the present invention can be stored in a storage medium such as a floppy disk, a hard disk, a compact disk, and a flash memory.

The embodiment of the present invention improves the VSM, pre-constructs a spherical space model, and classifies the paper based on the spherical space model. In the classification process, the distance between the sum of the vocabulary vectors and the category vector is calculated to determine the The category to be classified. The embodiment of the present invention achieves the classification herein, and the amount of calculation is significantly reduced relative to the angle cosine algorithm in the prior art. And in the embodiment of the present invention, the spherical space model has a radius of unit length, and the squared sum of the normalized lexical vectors of a normalized vocabulary in various categories is also a unit length, which is equivalent to a vocabulary. The semantic information quantity is equivalent to the unit length, and the amount of semantic information is constrained. Therefore, the accuracy of the classification can be improved compared with the prior art.

With a more accurate classification of the paper, it is beneficial to improve the accuracy of the classification storage and the classification search (or search) in this paper.

It will be obvious to those skilled in the art that various changes and modifications may be made in the present invention without departing from the spirit and scope of the invention. In this way, if the modifications and variations of the present invention fall within the scope of the patent application and its equivalents, the present invention is intended to include such modifications and variations.

101. . . Word segmentation module

102. . . Query module

103. . . Computing module

104. . . Classification module

105. . . Interface module

106. . . Filter module

107. . . Construction module

108. . . Storage module

Figure 1 is a main structural view of the device in the embodiment of the present invention;

Figure 2 is a detailed structural view of the device in the embodiment of the present invention;

Figure 3 is a schematic view of a spherical space in the embodiment of the present invention;

4 is a flow chart of the main method of classification in the embodiment of the present invention;

5 is a flow chart of a method for passing the distance and performing the classification in the embodiment of the present invention;

FIG. 6 is a flow chart of a method for classifying a vocabulary vector and performing the classification in the embodiment of the present invention.

Claims (10)

A method for classification according to the present invention, comprising the steps of: segmenting the obtained content of the article, obtaining a plurality of vocabulary; filtering the obtained plurality of vocabularies to obtain a plurality of vocabularies meeting the filtering condition, the filtering The method comprises: calculating a coefficient of variation of the vocabulary according to a word frequency value of each vocabulary; and filtering out a vocabulary whose coefficient of variation is greater than a threshold of a predetermined coefficient of variation; and for a vocabulary of the plurality of vocabularies obtained, Determining the vocabulary vector of the vocabulary in the spherical space model, wherein the dimension of the space contained in the spherical model is equal to the number of categories, and the category corresponds to a category vector in the spherical space; for each category, the determined number is determined The distance from the vocabulary vector of the vocabulary to the category vector of the category; and the text is classified into the category corresponding to the shortest distance according to the determined distance. The method of claim 1, wherein the distance from the vocabulary vector to the category vector is a linear distance or a spherical distance. The method of claim 1, wherein the vocabulary vector of the vocabulary includes a normalized word frequency obtained by normalizing the vocabulary value of each vocabulary; and the spherical space model is A multidimensional sphere model with a radius of unit length. The method of claim 4, wherein the unit The length is 1. The method of claim 1, wherein the step of determining the distance between the sum of the obtained vocabulary vectors of the plurality of vocabularies to the categories of the destination categories includes: pre-existing all vocabulary in each category The normalized word frequency value, the cumulative vocabulary vector of the plurality of vocabulary is accumulated in the normalized word frequency value of the category to obtain a normalized vocabulary vector sum; and the paper is divided into the shortest distance The steps of the corresponding category include: dividing the text into categories corresponding to the largest component of the normalized lexicon vector sum. An apparatus for classification in the present invention, comprising: a word segmentation module, configured to perform segmentation on the obtained content, and obtain a plurality of vocabulary; and a filtering module, configured to filter the obtained plurality of vocabulary to obtain The multiple vocabulary that meets the filtering condition calculates the coefficient of variation of the vocabulary according to the word frequency value of each vocabulary, and filters out the vocabulary whose coefficient of variation is greater than the preset coefficient of variation coefficient; the query module is used to obtain a vocabulary of a plurality of vocabulary, determining a vocabulary vector of the vocabulary in the spherical space model, wherein the spherical space model includes a dimension equal to the number of categories, and the category corresponds to a category vector in the spherical space; a calculation module, configured to determine, for each category, a distance from a sum of the vocabulary vectors of the plurality of vocabularies obtained to the category vector of the category; and a classification module configured to divide the document according to the determined distance Shortest distance From the corresponding category. The device of claim 6, wherein the distance from the vocabulary vector to the category is a linear distance or a spherical distance. The device according to claim 6, wherein the vocabulary vector of the vocabulary includes a normalized word frequency obtained by normalizing the vocabulary value of each vocabulary; and the spherical space model is A multidimensional sphere model with a radius of unit length. The device of claim 8, wherein the unit length is one. The device of claim 6, wherein the computing module pre-stores the normalized word frequency values of all the vocabulary in various categories, and the obtained vocabulary vectors of the plurality of vocabularies are on the category. The normalized word frequency values are accumulated to obtain a normalized vocabulary vector sum; and the classification module divides the text into categories corresponding to the largest component of the normalized lexicon vector sum.