JP2017107391A - Text mining method, and text mining program - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a text mining method capable of efficiently and more accurately extracting useful information documents only required for analysis from among a large amount of information documents disclosed in various document modes by a text data format, and a text mining program.SOLUTION: A text mining method distributes an extracted one information document into a first positive example or a first negative example, and includes: a first analysis step for taking one information document of the first positive example as an analysis object of a second analysis step and eliminating one information document of the first negative example from the analysis of the second analysis step; and the second analysis step of distributing one analysis schedule information document extracted from a deemed positive example information document classified into the first positive example into the second positive example that satisfies the condition including the meaning of the evaluation of a specific product as an analysis object or the second negative example that does not satisfy this condition, and identifying one analysis schedule information document of the second positive example as the analysis object and eliminating one analysis schedule information document of the second negative example from the positive example information document.SELECTED DRAWING: Figure 1

Description

  The present invention is useful for companies, consumers, and the like from a large amount of information documents published in text data format in various document forms such as evaluation documents, advertisement documents, and documents that describe everyday situations. As information to be obtained, the present invention relates to a technique for extracting only useful documents corresponding to an analysis target by text mining.

  In recent years, with the widespread use of the Internet, it has become possible for individuals to easily send and collect information in text format via social networking services (SNS) such as blogs and Web bulletin boards. . In general, such information documents include evaluation documents related to the target specific products, etc., as information that can be useful to companies and consumers, as well as advertisement articles about the products, etc. There may be many documents such as written documents that do not require analysis for companies and consumers. Some companies use text mining equipment to extract useful information from a large number of information documents that are used to improve the production and sales of products. Some companies do document analysis. An example of such a text mining device is disclosed in Patent Document 1.

  Patent Document 1 makes it possible to display a network in which dependency relationships between words extracted from text to be analyzed are classified into a plurality of categories, and display information regarding categories, nodes, and arcs can be freely displayed on the displayed network. It is adjustable. This enables free and interactive analysis of a text document to be analyzed, from an overall overview of a wide range of dependency relationships spanning three or more words to a detailed refined display.

  Although no prior art document is disclosed, there is a known text mining method in addition to Patent Document 1. In the conventional text mining method as an example, one information document is extracted from a large amount of information documents disclosed in a text data format. Whether a sentence included in this one information document is a sentence including an evaluation corresponding to the specific product to be analyzed, whether or not it is a keyword related to the specific product by the support vector machine and this Discrimination is performed based on a feature vector composed of keywords and co-occurring words, and one information document is assigned to a positive example or a negative example.

Japanese Patent No. 48766692

  However, Patent Document 1 excludes documents such as advertisement articles that do not require analysis from various information documents disclosed in large quantities as described above, and is an evaluation document useful for companies and consumers. Just can't choose. Therefore, when a company conducts market research for new development, improvement, sales improvement, etc. for a specific product using the above-mentioned SNS etc., people use various document forms. It is necessary to go through a large amount of published information documents one by one and extract only useful evaluation documents that are linked to the specific product, and this work takes a lot of time and effort. The work efficiency is poor. In addition, in the conventional text mining method, it is not possible to determine whether the extracted information document is a positive example or a negative example with sufficient accuracy, and companies etc. develop text mining with higher determination accuracy. Wanted.

  The present invention has been made to solve the above-described problems, and only useful information documents that require analysis are selected from a large number of information documents that are disclosed in various document modes by a text data format. An object of the present invention is to provide a text mining method and a text mining program that can be extracted efficiently and with higher accuracy.

  The text mining method according to one aspect of the present invention for solving the above-described problem is that one information document is extracted from a plurality of information documents created and released in a text data format, and the one Of the sentences included in the information document, the first selected sentence, which is a selected sentence, is a sentence meaning based on subjective expression, or a sentence meaning based on objective expression for an unspecified product. The one information document corresponding to the sentence meaning by the subjective expression is a first positive example, and the one information document corresponding to the sentence meaning by the objective expression is the first information document As a negative example, the extracted one information document is discriminated by a classifier based on the feature at the end of the sentence in the last phrase of the first selected sentence, and the first positive example or the first negative document is obtained. The one corresponding to the first positive example is assigned to an example. A first analysis step in which an information document is an object to be analyzed in the next second analysis step, and the one information document corresponding to the first negative example is excluded from the analysis in the second analysis step. And the one information document is a deemed analysis case information document extracted from the deemed positive example information document, with the assumed positive example information document all classified into the first positive example as a population. Whether the second selected sentence, which is a sentence selected from the analysis scheduled information document, is a sentence including an evaluation corresponding to a specific product to be analyzed, includes the evaluation in the sentence If the one analysis schedule information document that satisfies the condition is a second positive example and the one analysis schedule information document that does not satisfy this condition is a second negative example, the extracted one analysis schedule information document A document is classified by a classifier into a key associated with the specific product. Based on a feature vector consisting of a word and a word co-occurring with the keyword, distributed to the second positive example or the second negative example, and the one corresponding to the second positive example. The second analysis step of certifying the analysis schedule information document as an analysis information document to be analyzed and excluding the one analysis schedule information document corresponding to the second negative example from the deemed positive example information document It is characterized by comprising.

  According to this aspect, at the time of text analysis, among a plurality of information documents released in various document forms, for example, an unspecified product such as an advertising document included in many negative examples. Since the information document that is the meaning of the objective expression of the product is removed in advance, only the information document with the meaning that includes the evaluation corresponding to the specific product is efficiently and efficiently used as a useful information document that needs to be analyzed. Extraction can be performed with higher accuracy.

  In said aspect, it is preferable that the object of said 1st negative example is an advertisement consisting of the content which advertises the said unspecified product.

  According to this aspect, a large number of advertisement documents that do not require analysis are excluded in the first analysis step, and the total number of documents of the assumed positive example information document that is analyzed in the second analysis step is first collected. It is possible to significantly reduce the total number of information documents. As a result, in the text mining method according to the present invention, the accuracy of extracting the information document to be analyzed is significantly higher than the case of using the conventional text mining method. To improve.

  In the above aspect, it is preferable that the feature used in the first analysis step uses morpheme.

  According to this aspect, the information document to be analyzed includes, for example, an evaluation document related to a specific product (gas stove as an example), an advertisement document including contents for promoting an unspecified product (gas stove or other product as an example). Even if it is in various document forms, such as a document that spells out everyday situations, a document that describes evaluations other than a specific product, etc., it is possible to accurately classify information documents into positive examples and negative examples. .

  In the above aspect, the specific product is preferably a gas device.

  According to this aspect, as a business, a person who handles gas equipment in-house uses, for example, an information transmission medium through SNS such as a blog or a Web bulletin board, for example, for gas equipment represented by a gas stove or the like. In market research conducted for the purpose of new development, improvement, and sales improvement, more reliable market research results can be obtained based on the classified analysis information document.

  A text mining program according to another aspect of the present invention, which has been made to solve the above problems, extracts one information document from a plurality of information documents created and published in a text data format, Among the sentences included in the information document, the first selected sentence, which is a selected sentence, is a sentence meaning based on subjective expression, or a sentence based on objective expression for an unspecified product The first information document corresponding to the sentence meaning by the subjective expression is the first positive example, and the one information document corresponding to the sentence meaning by the objective expression is the first The one information document extracted is determined by the classifier based on the feature at the end of the sentence in the last phrase of the first selected sentence, and the first positive example or the first A first step of distributing to a negative example, and the first step The one information document corresponding to the positive example is set as an object to be analyzed in the next second analysis step, and the one information document corresponding to the first negative example is set to the second analysis. A second step that is excluded from the analysis by the step, and a one that is extracted from the deemed positive example information document with the assumed positive example information document in which the one information document is all classified as the first positive example as a population. Whether or not the second selected sentence, which is a sentence selected from the one analysis planned information document, includes an evaluation corresponding to a specific product to be analyzed On the other hand, the one analysis schedule information document that satisfies the condition including the evaluation in its meaning is defined as a second positive example, and the one analysis schedule information document that does not satisfy the condition is defined as a second negative example. Then, the one analysis schedule information document extracted is classified by the classifier. A third step of determining based on a feature vector comprising a keyword related to a fixed product and a word co-occurring with the keyword, and allocating to the second positive example or the second negative example; The one analysis schedule information document corresponding to two positive examples is recognized as an analysis information document to be analyzed, and the one analysis schedule information document corresponding to the second negative example is regarded as the deemed positive example information. And a fourth step of excluding from the document.

  According to this aspect, at the time of text analysis, among a plurality of information documents released in various document forms, for example, an unspecified product such as an advertising document included in many negative examples. Since the information document that is the meaning of the objective expression of the product is removed in advance, only the information document with the meaning that includes the evaluation corresponding to the specific product is efficiently and efficiently used as a useful information document that needs to be analyzed. Extraction can be performed with higher accuracy.

  According to the text mining method and the text mining program according to the present invention, only a useful information document that needs to be analyzed is efficiently selected from a large number of information documents that are disclosed in various document modes by a text data format. Good and higher accuracy can be extracted.

It is a schematic diagram which shows the concept of the text mining method which concerns on this embodiment. It is a figure which shows the blog document mentioned as an example, respectively about the evaluation sentence and advertisement sentence which are contained in the disclosed document group. It is the 1st explanatory view showing typically the idea used for classifying an evaluation sentence and an advertising sentence at the 1st analysis step shown in FIG. It is the 2nd explanatory view explaining using the evaluation sentence illustrated in Drawing 2 about the technique which classifies the expression at the end of a sentence as a feature. FIG. 6 is a third explanatory diagram specifically showing how the sentence ending is analyzed by the feature in the first analysis step shown in FIG. 1. FIG. 6 is a fourth explanatory diagram specifically showing the state of analyzing the sentence ending with the feature following FIG. 5. FIG. 7 is a fifth explanatory diagram specifically showing a state of analyzing the sentence ending with the feature following FIG. 6. It is a figure which shows an example blog document about the evaluation sentence of the analysis object contained in the published document group, the evaluation sentence etc. which are not analysis objects, respectively. FIG. 9 is a diagram in which a focused keyword is highlighted in each blog document illustrated in FIG. 8. It is the 6th explanatory view showing typically the way of thinking used for classifying the evaluation sentence of analysis object, the evaluation sentence of non-analysis object, etc. in the 2nd analysis step shown in FIG. FIG. 10 is a seventh explanatory diagram specifically illustrating a state in which the sentence ending is analyzed in relation to the co-occurring word with respect to the specific evaluation item in the second analysis step illustrated in FIG. 1. FIG. 13 is an eighth explanatory diagram specifically showing a state in which sentence endings are analyzed in relation to co-occurring words for specific evaluation items following FIG. 11. FIG. 13 is a ninth explanatory diagram specifically showing a state in which sentence endings are analyzed in relation to co-occurring words for specific evaluation items, following FIG. 12. In the second analysis step shown in FIG. 1, in the case where the word co-occurring in the specific evaluation item is a noun, in each information document included in the published document group, the occurrence rate of the specific evaluation item in the positive example is determined. It is a figure explaining the method to calculate. It is a figure explaining the definition of a judgment standard about the analysis result of the information document classified by the text mining method. It is a table | surface which shows the result of the experiment which concerns on an Example by the text mining method which concerns on this embodiment. It is the schematic diagram which divided and showed the some information document contained in the document group searched by the keyword for every main category. It is explanatory drawing regarding the function of the classifier used with a text mining method. It is a schematic diagram which shows the concept of the conventional text mining method. It is a table | surface which shows the result of the experiment which concerns on a comparative example by the conventional text mining method.

  A text mining method and a text mining program which are examples of the present invention will be described. The text mining method according to the present invention is a large amount of information documents (hereinafter referred to as “document group”) published in a text data format in various document modes such as an evaluation document, an advertisement document, and a document spelling a daily appearance. It is used to extract only useful documents corresponding to the object of analysis as information that can be useful to companies and consumers. As will be described later, the useful document is a document described in a sentence including an evaluation corresponding to a specific product to be analyzed. In this embodiment, this specific product is a gas stove belonging to a gas appliance, and the useful document is an evaluation document relating to this gas stove. One of the applicants (hereinafter referred to as “information analyst”) is a person who handles gas stoves as his own business. Like the information analyst, “ "Is also a person who handles gas stoves (product name" A ") in-house as a business.

<Documents>
First, the document group will be briefly described. FIG. 17 is a schematic diagram illustrating a plurality of information documents included in a group of documents disclosed by searching with keywords, divided into main categories. For example, in recent years, an information document created by a person himself / herself in a text data format has been freely transmitted to an information transmission medium through an SNS (social networking service) such as a blog or a Web bulletin board. The transmitted information document is stored on the information transmission medium side, and is accessible to any person who accesses the information transmission medium.

  On the other hand, when one person inputs a specific keyword to the information transmission medium, at least one information document related to this keyword is typically listed from among a plurality of information documents already transmitted. A document group is a collection of all the plurality of listed information documents. For example, the document group shown in FIG. 17 is obtained by performing a search using an information transmission medium with the keyword “gas stove * A”. However, as shown in FIG. The vast majority of a large number of documents may be occupied by documents related to advertisements.

<Text mining method according to this embodiment>
Next, the text mining method according to the present embodiment will be described. FIG. 1 is a schematic diagram showing a concept of a text mining method according to the present embodiment. In the text mining method according to the present embodiment, as shown in FIG. 1, the collected document group 1 (a plurality of information documents) is converted into the subjective document 2 (deemed positive example information document) and the objective by the first analysis step S1. It is roughly divided into Document 3 in advance. In the next second analysis step S2, the analysis information document 4 which is a useful document is extracted by further classifying the subjective document 2 divided in the first analysis step S1. The classification of the document group 1 is performed by the classifier in both the first and second analysis steps S1 and S2, and a known support vector machine (SVM) 50 is used for the classifier.

Here, the SVM 50 will be briefly described. The SVM 50 is a pattern classifier that divides an object to be identified into two classes of “positive example” and “negative example” using linearly separable input elements. In the SVM 50, as shown in the following equation (1), a function necessary for arithmetic processing is constructed by determining the direction vector (w) and the intercept (b) of the separation plane. In this function, the margin is maximized. Is done.
f (x) = w × x−b Equation (1)
w: direction vector of separation plane x: feature direction vector of case b: intercept

The margin maximization is to maximize the distance between the support vector closest to the separation plane and the separation plane. In the function of the SVM 50, the following minimization problem is solved in the following equation (2). become.
min. 1 / 2w ² Formula (2)
That is, Equation (3) is solved.
y ⁽ⁱ⁾ (w × x ⁽ⁱ⁾ −b) −1 ≧ 0 (where ∀i) (3)

  In the formula (1), when f (x) ≧ 0 is satisfied, it is classified into the “positive example” class, and when f (x) <0 is satisfied, it is classified into the “negative example” class. In general, a “positive example” is an information document that an information analyst wants to focus on, and a “negative example” is an information document other than a “positive example”. The clear definition of “positive example” and “negative example” to be used will be described later.

  FIG. 18 is an explanatory diagram regarding the function of the SVM. Before performing the task of classifying the cases with the SVM 50, the SVM 50 firstly shows a case corresponding to the “positive example” and a case corresponding to the “negative example” as shown in FIG. Learning is performed, and in the SVM 50, determination criteria for separating the “positive example” and the “negative example” are determined and set. This criterion corresponds to the direction vector of the previous separation plane. The information analyst applies to the SVM 50 the information document to be classified as unknown data. As a result, as shown in FIG. 18B, the SVM 50 determines whether the target information document corresponds to a “positive example” or a “negative example” based on a learned determination criterion. Discriminate and classify.

  In the first analysis step S1 of the text mining method according to this embodiment, one information document (for example, the first information document 10A and the second information document 10B) is selected from the document group 1 created and collected in the text data format. , Third information document 10C, fourth information document 10D, etc.) are extracted. Of the sentences included in this one information document, a first selected sentence that is a selected sentence (described in detail later, for example, a first selected sentence 11A that is one sentence in the first information document 10A). ) Is a meaning based on subjective expression or an intentional expression based on objective expression for an unspecified product.

  In this determination, when one information document extracted from the document group 1 belongs to the subjective document 2 corresponding to the meaning of subjective expression, this one information document is referred to as the above-mentioned “positive example”. In the present embodiment, it is defined as a first positive example 20A (first positive example) to be analyzed, including a document describing an evaluation related to a gas stove. On the other hand, when one information document extracted from the document group 1 belongs to the objective document 3 corresponding to the meaning of objective expression, this one information document does not correspond to the first positive example 20A. It is defined as a first negative example 20B (first negative example). That is, the first negative example 20B is an advertising document composed of contents for promoting a gas stove and other unspecified products, as exemplified in the second information document 10B, and is mainly an advertisement issued by a person other than the information analyst. It is.

  Then, one information document extracted from the document group 1 is discriminated by the SVM 50 based on the feature at the end of the sentence in the last phrase of the first selected sentence, that is, the morpheme, and the first positive example 20A or the first negative example 20B is determined. Sorted. The process of allocating to the first positive example 20A or the first negative example 20B corresponds to the first step of the text mining program according to the present embodiment.

  One information document corresponding to the first positive example 20A is an object to be analyzed in the next second analysis step S2. On the other hand, one information document corresponding to the first negative example 20B is excluded from the analysis target in the second analysis step S2. As described above, the process of assigning the target and the non-target for the analysis in the second analysis step S2 corresponds to the second step of the text mining program according to the present embodiment.

  In the second analysis step S2 performed after the first analysis step S1, the one scheduled analysis information document 6 is extracted with the subjective document 2 in which all the one information document is classified as the first positive example 20A as a population. . Of the sentences included in the one analysis schedule information document 6, a second selected sentence that is a selected sentence (described in detail later, for example, the first selected sentence 11A in the first information document 10A). Whether or not the second selected sentence 11B including) includes the evaluation corresponding to the gas stove (specific product) to be analyzed.

  In this determination, when one analysis schedule information document 6 extracted from the subjective document 2 belongs to the analysis information document 4 that satisfies the condition including the evaluation of the gas stove, this one analysis schedule information document 6 is And the handling of the second positive example 30A (second positive example). That is, the information document illustrated in the first information document 10A corresponds to the second positive example 30A (see FIG. 2). On the other hand, if the one analysis schedule information document 6 extracted from the subjective document 2 belongs to the non-analyzed document 5 that does not satisfy the condition including the evaluation of the gas stove, the second negative example 30B (first 2 negative example). As illustrated in the third information document 10C and the fourth information document 10D, the second negative example 30B is a second positive document such as a document spelling a daily situation, a document describing an evaluation other than a gas stove, or other documents. Documents other than Example 30A are targeted (see FIG. 9).

  Then, one analysis schedule information document 6 extracted from the subjective document 2 is discriminated by the SVM 50 based on a feature vector composed of a keyword related to the gas stove and a word co-occurring with the keyword, and the second positive example It is distributed to 30A or the second negative example 30B. The process of allocating to the second positive example 30A or the second negative example 30B corresponds to the third step of the text mining program according to the present embodiment.

  One information document belonging to the second positive example 30A is authorized as an analysis information document 4 corresponding to an analysis target. On the other hand, one information document corresponding to the second negative example 30B is excluded from the subjective document 2 and excluded from the analysis target. As described above, in the subjective document 2, the process of distributing the analysis information document 4 to the other information corresponds to the fourth step of the text mining program according to the present embodiment.

  Next, the text mining method according to the present embodiment will be described in detail by giving some specific examples for one information document. FIG. 2 is a diagram illustrating a blog document that is cited as an example for each of an evaluation sentence and an advertisement sentence included in the disclosed document group. FIG. 3 is a first explanatory diagram schematically showing the concept used to classify the evaluation text and the advertisement text in the first analysis step.

  As shown in FIG. 2, the first information document 10 </ b> A is an evaluation document related to the gas stove with the product name “A” handled by the former, and the second information document 10 </ b> B is an unspecified product other than the gas stove related to Shop B and Sakai. Is an advertising document described. As shown in FIG. 3, the textual meaning of the first information document 10A, which is an evaluation document, is a subjective description, while the second information document 10B, which is an advertising document, is an objective description. Tend to be. Therefore, it can be assumed that the evaluation document is an information document described in a subjective position. Further, it can be assumed that the advertisement document is an information document described from an objective standpoint.

  In the first analysis step S1, in view of the present situation where most of the disclosed document group is occupied by documents related to advertisements unnecessary for analysis, as in the document group shown in FIG. First, a large number of advertisement documents (for example, the second information document 10B) are excluded from the document group 1 (see FIG. 1). Thereby, it can be expected that the total number of documents of the subjective document 2 which is the population that performs the second analysis step S2 will be significantly smaller than the total number of documents in the document group 1, and the execution of the first analysis step S1 This contributes to improving the accuracy of extracting information documents.

  A specific method for distributing one information document extracted from the document group 1 to the first positive example 20A or the first negative example 20B will be described. FIG. 4 is a second explanatory diagram illustrating a method of classifying the expression at the end of a sentence as a feature using the evaluation sentence (first information document) illustrated in FIG. FIG. 5 is a third explanatory diagram specifically showing how the sentence ending is analyzed in the first analysis step shown in FIG. FIG. 6 is a fourth explanatory diagram specifically showing how the sentence ending is analyzed based on the features, following FIG. 5. FIG. 7 is a fifth explanatory diagram specifically showing the state of analyzing the sentence ending with the feature, following FIG. 6.

  As shown in FIG. 2 and FIG. 4, when the first selected sentence 11A included in the first information document 10A “A's A is convenient” is morphologically analyzed by the SVM 50, the final result in the first selected sentence 11A A phrase 12 “Convenient” is composed of an independent word “Convenient” and the end of the sentence “Nice”. Independent words are words of part of speech other than auxiliary verbs and particles, and the end of a sentence refers to a collection of a plurality of words in the range from the end of one sentence to the appearance of this independent word.

  In the morphological analysis, as shown in FIG. 5, when the feature to be used is a morpheme simple substance, the morpheme at the end of the sentence “I am” is the auxiliary verb “I” and the particle “Ne”. The size of the feature direction vector in each part of speech is counted as 5 for the auxiliary verb “Da” and 4 for the particle “Ne”.

  Next, the morpheme unit at the end of the sentence “Ne” is the auxiliary particle “is” counted in the size 5 of the feature direction vector and the particle “ne” counted in the size 4 of the feature direction vector. If the feature to be used is a combination of morphemes, as shown in Fig. 6, assuming that this is the combination of the morphemes at the end of the sentence, "Is + ne" counted as the feature direction vector size 2 It is added.

  Next, as shown in FIG. 7, when the feature to be used is a group of morphemes, the grouped morphemes are used. The grouped morpheme is a morpheme that has been learned in advance by the SVM 50 and is intended for a final particle (ne, yo, na, ne, ne, ne, nei) that means confirmation / consent. If the feature to be used is a combination of morphemes, it is a verb with a feature direction vector size of 5 is "Ne" with a feature direction vector size of 4 and a feature direction vector of size 2 A particle (final particle) “ne” with a feature direction vector size 2 is added to “+ ne”.

  As described above, when the phrase 12 of the first selected sentence 11A is morphologically analyzed by the SVM 50, the first information document 10A is determined to correspond to the meaning of the subjective expression, and the analysis schedule information document 6 (one analysis schedule information) Document) is classified into the subjective document 2 which is the first positive example 20A. On the other hand, for the second information document 10B, in the same manner as the first information document 10A, when one sentence included in the second information document 10B is analyzed by the SVM 50 with the morphological analysis of the last sentence of the selected sentence, the sentence in the objective expression It is discriminated that falls within Therefore, the advertising document such as the second information document 10B is classified into the objective document 3 that is the first negative example 20B, is excluded from the analysis by the second analysis step S2, and is excluded from the document group 1.

  Next, the analysis schedule information document 6 will be described with some specific examples. FIG. 8 is a diagram illustrating an example of a blog document for each of an analysis target evaluation sentence and a non-analysis evaluation sentence included in the published document group. FIG. 9 is a diagram in which the keyword of interest is highlighted in each blog document illustrated in FIG. FIG. 10 is a sixth explanatory view schematically showing the concept used for classifying the evaluation sentence to be analyzed and the evaluation sentence not to be analyzed in the second analysis step shown in FIG.

  As described above, the first information document 10A shown in FIG. 8 is an evaluation document related to the gas stove with the product name “A” handled by A, and the analysis schedule information document is the state before the execution of the second analysis step S2 by the SVM 50. There are also six. The third information document 10 </ b> C and the fourth information document 10 </ b> D are evaluation documents other than the gas stove, and are documents that spell out everyday situations.

  In the first information document 10A, as shown in FIG. 9, the adjectives and adjective verbs (first element) “convenient, nice, wonderful, delicious, happy” (shown in a dotted frame) co-occur with these Among the nouns to be used, the noun (second element) other than the gas stove (timer function, cookie) (shown in a frame with a one-dot chain line) and the noun associated with the second element are related to the gas stove (Third element) “Instep, A, toast, grill” (shown in a solid line frame) is included. On the other hand, in the third information document 10C and the fourth information document 10D, the first element “good, fashionable, good” and the second element “flooring, appearance, free design” are included. The third element is not included.

  That is, the meaning of the first information document 10A, which is an evaluation document, is different from the meaning of the third information document 10C etc. in the usage of co-occurring adjectives and adjective verbs for nouns related to the gas stove. Can be assumed. Therefore, in the second analysis step S2, the analysis schedule information document 6 belonging to the subjective document 2 is classified based on the co-occurring adjectives and adjective verbs.

  A specific method of distributing one analysis schedule information document 6 extracted from the subjective document 2 to the second positive example 30A or the second negative example 30B will be described. FIG. 11 is a seventh explanatory diagram specifically showing how the sentence ending is analyzed in relation to the co-occurring word for the specific evaluation item in the second analysis step shown in FIG. FIG. 12 is an eighth explanatory diagram specifically showing the state of analyzing the sentence end in relation to the co-occurring words for the specific evaluation item, following FIG. FIG. 13 is a ninth explanatory diagram specifically showing how the sentence ending is analyzed in relation to the co-occurring words for the specific evaluation item, following FIG. 12. In each figure after FIG. 11, the first information document described in the table indicates the first information document 10A, and the m-th information document (m> 2) indicates other information documents.

  In the second analysis step S2, the first information document 10A is morphologically analyzed by the SVM 50 as one analysis schedule information document 6 (see FIG. 1). The morphological analysis is performed by using a noun related to the evaluation of the gas stove and the adjectives and adjective verbs included in the first information document 10A in which the noun exists as feature vectors.

  As shown in FIG. 11, the second selection sentence 11B included in the first information document 10A “The upper A is convenient. The morning toast can be burnt nicely ^^ The timer function is also great.” There are two nouns related to the evaluation: “A” having a feature direction vector size of 1 and “Convenient” having a feature direction vector size of 2 in the order of the head of the second selected sentence 11B.

  Next, although the noun “Toast” belongs to the noun related to gas stove evaluation, the word “Toast” co-occurs with “Nice”, which is classified into the above-mentioned adjectives and adjective verbs. The word “nice” does not co-occur with nouns related to gas stove evaluation. Therefore, the word “nice” has a feature direction vector size of 0 and is not counted in the morphological analysis (see FIG. 12).

  As shown in FIG. 13, it is not a combination of a noun related to gas stove evaluation and an adjective or adjective verb that co-occurs with a word by this noun. It co-occurs with “great” which is classified into the adjectives and adjective verbs mentioned. Therefore, the size 1 of the feature direction vector is counted for both the word “timer function” and the word “great”.

  FIG. 14 shows the case of the specific evaluation item in the positive example in each document included in the published document group, when the word co-occurring in the specific evaluation item is a noun in the second analysis step shown in FIG. It is a figure explaining the method of calculating an appearance rate.

  As shown in FIG. 14, the text mining method according to the present embodiment selects, as an example, 50 nouns that are keywords related to a specific product, such as nouns related to gas stove evaluation, This is performed under the condition that the number of appearing documents, which is the number of information documents in a group 1 in which a specific morpheme exists, is 10 or more. In extracting the analysis information document 4 from the document group 1, the number of keywords related to a specific product is 50 as an example, but for reference, the number of keywords selected for the 50 words is a past classification. This is due to the fact that the percentage of positive cases appears in the top 50 based on the analysis. This is because the accuracy of extracting the analysis information document 4 by the text mining method according to the present embodiment is increased by setting the number of selected keywords to about 50 words.

  The positive case appearance ratio is a quotient obtained by dividing the total frequency of positive examples by the sum of the total frequency of positive examples and the total frequency of negative examples. The total frequency of the positive examples is obtained for each keyword of each information document such as the first information document 10A for the analysis information document 4 determined as the second positive example 30A by the second analysis step S2. , The sum of the magnitudes of the feature direction vectors. The total frequency of the negative examples is calculated based on the non-analyzed document 5 determined to be the second negative example 30B by the second analysis step S2, and each one of the third information document 10C, the fourth information document 10D, etc. Is the sum of the magnitudes of the feature direction vectors for the same keywords as in the positive example.

<Investigation experiment>
Next, in extracting a useful document from the document group 1, the text mining method according to this embodiment is compared with the text mining method according to this embodiment and the conventional text mining method mentioned in the background art. A survey experiment was conducted to confirm the presence or absence of significance. In the investigation experiment, two experiments were performed: an example in which a useful document was extracted by the text mining method according to the present embodiment, and a comparative example in which a useful document was extracted by a conventional text mining method.

<Experiment according to Examples>
An information analyst uses the text mining method according to the present embodiment to extract a gas stove evaluation document (analysis information document 4) from a document group 1 consisting of 88 blog documents. The experiment concerning was conducted.

  FIG. 19 is a schematic diagram showing the concept of a conventional text mining method. In the conventional text mining method, a gas stove evaluation document that is a sentence including an evaluation corresponding to a gas stove is set as a positive example 40A, and other documents other than the positive example 40A (an advertisement document, a document spelling a daily situation) A positive example and a negative example are defined as a negative example 40B. In the conventional text mining method, as shown in FIG. 19, the information documents extracted from the collected document group 1 are classified into positive examples 40A and negative examples 40B by the SVM 50, and are useful as analysis targets. It is a one-stage classification method that directly extracts documents. Specifically, in the conventional text mining method, when the information document is classified by the SVM 50, one sentence included in the information document includes four parts of speech of "noun / adjective / adverb / auxiliary verb". Based on the appearance frequency of these parts of speech, the morphological analysis is performed.

<Experiment according to comparative example>
In addition, the information analyst uses a conventional text mining method to select a gas stove evaluation document (corresponding to the analysis information document 4) from the same document group 1 (blog document with 88 documents) used in the experiment according to the embodiment. An experiment related to a comparative example was conducted for the purpose of extracting the.

<Preparation before experiment>
In this embodiment, in this embodiment, the information analyst searches the information document using the information transmission medium using the keywords “gas stove / purchase, table stove, built-in stove, Toho gas / stove, gas stove / evaluation”. As a result, a group of documents disclosing all of the plurality of information documents listed was collected. “Toho Gas” used as a keyword is one of the applicants. The collected document group has a total number 146 in which the number of “negative examples” is 73 and the number of “positive examples” is also 73 in accordance with this number. The contents of all the information documents 146 constituting the document group are all confirmed in advance by the operator. Learning to the SVM 50 was performed using 73 documents for both the “negative example” and the “positive example”.

  The “positive example” is intended only for documents that describe evaluations related to gas stoves. The breakdown of information documents belonging to “Negative Examples” includes 34 documents for advertising documents consisting of contents that promote gas stoves and other unspecified products, 29 documents for documents that describe everyday situations, and evaluations other than for gas stoves Is 9 documents, and other documents are 1 document.

<Common conditions for the two experiments>
As described above, the information analyst performs an experiment related to the blog document (information document) having the number of documents 88 as the main contents by detailed examination from among the collected document group of 146 documents. Both of the experiments related to the comparative example are listed in the document group 1 (see FIGS. 1 and 19) as a population.

<Evaluation method for two experiments>
FIG. 15 is a diagram illustrating the definition of the determination criterion for the analysis result of the document classified by the text mining method. As shown in FIG. 15, for the information document belonging to the “positive example” confirmed in advance by the operator, in the experiment related to the example, it is correctly determined as the “second positive example 30A”, and in the experiment related to the comparative example, When it is correctly determined as “positive example 40A”, it is treated as “TP”. In addition, regarding an information document belonging to the “positive example” confirmed in advance by the operator, it was erroneously determined as the “second negative example 30B” in the experiment related to the example, and in the experiment related to the comparative example, “negative” When it is determined as “Example 40B”, it is treated as “TN”. In addition, an information document belonging to the “negative example” confirmed in advance by the operator is erroneously determined as “second positive example 30A” in the experiment related to the example, and in the experiment related to the comparative example, the “positive example” is erroneously determined. When it is determined as “Example 40A”, it is treated as “FP”. In addition, regarding the information document belonging to the “negative example” confirmed in advance by the operator, it is correctly determined as the “second negative example 30B” in the experiment according to the example, and the “negative example 40B” is correctly determined in the experiment according to the comparative example. ”Is treated as text“ FN ”.

  The correct answer rate A represents the ratio of “positive example” and “negative example” correctly determined in each experiment (model), and is a quotient of (TP + FN) / (TP + TN + FP + FN). The correct answer rate TR is an index indicating how comprehensively the “correct example” is extracted from the population (document group 1) on the model, and the quotient of TP / (TP + TN). It is represented by The negative example correct answer FR is an index indicating how exhaustively “negative examples” are extracted from the model population (document group 1), and is a quotient of FN / (FP + FN). It is represented by

<Experimental result>
FIG. 16 is a table showing the results of the experiment according to the example by the text mining method according to the present embodiment. FIG. 20 is a table showing the results of an experiment according to a comparative example using a conventional text mining method. In the experiment according to the example, as shown in FIG. 16, the correct answer rate TR was 0.81. The negative example correct answer rate FR was 0.84. The average accuracy rate A was 0.83. On the other hand, in the experiment according to the comparative example, as shown in FIG. 20, the correct case correct answer rate TR was 0.57. The negative example correct answer rate FR was 0.78. The average accuracy rate A was 0.73.

<Experimental considerations>
When the experimental results according to the example and the experimental results according to the comparative example are compared, the experimental results according to the example are compared for any of the correct answer rate TR, the negative answer rate FR, and the average correct answer rate A. It exceeded the experimental results of the examples. In particular, there was a difference of 10% in the average accuracy rate A, and the average accuracy rate A exceeded 80%. This is thought to be due to the following reasons. That is, the text mining method according to the present embodiment is divided into a first analysis step S1 and a second analysis step S2, unlike a conventional text mining method that extracts a useful document targeted for analysis by a one-step classification method. ing. When the information analyst extracts only useful documents corresponding to the object of analysis from the document group 1 as information that can be useful to companies, consumers, etc., the text mining method according to this embodiment uses the first In the analysis step S1, advertisement documents that are considered to be numerous are first excluded from the document group 1. In addition, since the total number of documents of the subjective document 2 that is a population for performing the second analysis step S2 is significantly reduced from the total number of documents of the document group 1 collected first, in the second analysis step S2, noise and It is considered that the advertisement document is eliminated, the information document is easily analyzed, and the analysis accuracy is improved.

  Next, operations and effects of the text mining method and the text mining program according to the present embodiment will be described. In the text mining program according to the present embodiment, one information document (first information document 10A, second information document 10B, etc.) is extracted from the document group 1 created and published in a text data format. Of the sentences included in the information document, the first selected sentence (for example, the first selected sentence 11A), which is a selected sentence, is a sentence meaning based on subjective expression, or for an unspecified product. One information document corresponding to the meaning of subjective expression is set as the first positive example 20A, and one information document corresponding to the meaning of objective expression is Assuming that the first negative example 20B is used, the extracted one information document is determined by the SVM 50 based on the feature at the end of the sentence in the final phrase (for example, phrase 12) of the first selected sentence, and the first positive example 20A or First to distribute to the first negative example 20B And one information document corresponding to the first positive example 20A is set as an object to be analyzed in the next second analysis step, and one information document corresponding to the first negative example 20B is set to the second A second step to be excluded from the analysis by the analysis step, and one analysis planned information document extracted from the subjective document 2 with the subjective document 2 in which all one information document is classified as the first positive example 20A as a population 6, the second selected sentence (for example, the second selected sentence 11B) which is a sentence selected from the one analysis schedule information document 6 is a sentence including an evaluation corresponding to a specific product to be analyzed. Whether or not there is one analysis schedule information document 6 that satisfies the condition including the wording of evaluation as the second positive example 30A, and one analysis schedule information document 6 that does not satisfy this condition is the second negative As an example 30B, the extracted one analysis schedule information document 6 is represented by SVM5. To determine based on a feature vector consisting of a keyword related to a specific product and a word that co-occurs with the keyword, and to distribute to the second positive example 30A or the second negative example 30B, One analysis schedule information document 6 corresponding to the positive example 30A is recognized as the analysis information document 4 to be analyzed, and one analysis schedule information document 6 corresponding to the second negative example 30B is excluded from the subjective document 2. And a fourth step.

  In the text mining method according to the present embodiment, one information document (first information document 10A, second information document 10B, etc.) is extracted from the document group 1 created and published in the text data format, A first selected sentence (for example, the first selected sentence 11A) that is a selected sentence among sentences included in one information document is a sentence meaning based on subjective expression or an unspecified product The first positive example 20A is one information document corresponding to the meaning of the subjective expression, and one piece of information corresponding to the intention of the objective expression. Assuming that the document is the first negative example 20B, the extracted one information document is determined by the SVM 50 based on the feature at the end of the sentence in the last phrase (for example, phrase 12) of the first selected sentence. Sort to 20A or the first negative example 20B, One information document corresponding to one positive example 20A is set as an object of analysis performed in the next second analysis step S2, and one information document corresponding to the first negative example 20B is excluded from the analysis in the second analysis step S2. The first analysis step S1 and the subject information document 6 in which one information document is all classified into the first positive example 20A as a population, and one analysis scheduled information document 6 extracted from the subjective document 2, one analysis Whether the second selected sentence (for example, the second selected sentence 11B), which is a sentence selected from the schedule information document 6, is a sentence including an evaluation corresponding to a specific product to be analyzed If one analysis schedule information document 6 that satisfies the condition including evaluation in its meaning is defined as a second positive example 30A, and one analysis schedule information document 6 that does not satisfy this condition is defined as a second negative example 30B, extraction is performed. The one analysis schedule information document 6 that has been It discriminates on the basis of a feature vector composed of a keyword related to a certain product and a word co-occurring with this keyword, and is assigned to the second positive example 30A or the second negative example 30B, and corresponds to the second positive example 30A. A second analysis step S2 for identifying one analysis schedule information document 6 as an analysis information document 4 to be analyzed and excluding one analysis schedule information document 6 corresponding to the second negative example 30B from the subjective document 2; It is characterized by comprising.

  Because of these features, advertising documents that are included in large numbers as negative examples are removed in advance from a large number of document groups 1 released in various document modes at the time of text analysis. As a simple information document, only a textual information document including an evaluation corresponding to a specific product can be extracted efficiently and with higher accuracy.

  In addition, the text mining method according to the present embodiment is characterized in that the target of the first negative example 20B is an advertisement composed of contents for promoting an unspecified product, and therefore, many advertising documents that do not require analysis. Is eliminated in the first analysis step S1, and the total number of documents of the subjective document 2 to be analyzed in the second analysis step S2 can be significantly reduced from the total number of documents in the document group 1 collected first. As a result, in the text mining method according to the present embodiment, the accuracy of extracting the information document to be analyzed is higher than that in the case of using the conventional text mining method. The average accuracy rate is significantly improved, such as an increase of 10%.

  The text mining method according to the present embodiment is characterized in that the feature used in the first analysis step S1 uses morphemes. Therefore, the information document to be analyzed is, for example, an evaluation document relating to a specific product (gas stove) as exemplified in the first information document 10A, and a content promoting an unspecified product as exemplified in the second information document 10B. Even if it is in various document modes, such as a document spelling out a daily appearance or a document describing an evaluation other than a gas stove, as exemplified in the advertisement document made up of, the third information document 10C and the fourth information document 10D It is possible to classify information documents into positive examples and negative examples with high accuracy.

  In addition, the text mining method according to the present embodiment is characterized in that the specific product is a gas appliance, so that an information analyst who is a person handling a gas stove in-house as a business, for example, a blog or a Web bulletin board In market research conducted for the purpose of new development, improvement, sales improvement, etc., targeting gas stoves using information transmission media through SNS, etc., based on the classified analysis information document 4, more reliable High market research results can be obtained.

  As is clear from the above description, the text mining method and the text mining program according to the present invention require analysis from a large amount of information documents published in various document modes according to the text data format. Only useful information documents can be extracted efficiently and with higher accuracy.

  Note that the above-described embodiment is merely an example, and does not limit the present invention in any way. It goes without saying that various improvements and modifications can be made without departing from the scope of the invention.

  For example, in the present embodiment, the classifier is SVM50. However, the classifier used in the text mining method and the text mining program according to the present invention is not limited to SVM50. As long as it is a pattern discriminator that can be divided into two classes, “” and “negative example”.

1 Document group (multiple information documents)
2. Subjective document (deemed positive example information document)
5 Analysis Information Document 6 Analysis Schedule Information Document (One Analysis Schedule Information Document)
10A First information document (one information document)
10B Second information document (one information document)
10C Third information document (one information document)
10D 4th information document (one information document)
11A First selected sentence (first selected sentence)
11B Second selected sentence (second selected sentence)
20A First positive example (first positive example)
20B First negative example (first negative example)
30A Second positive example (second positive example)
30B Second negative example (second negative example)
50 SVM (classifier)
S1 First analysis step (first analysis step)
S2 Second analysis step (second analysis step)

Claims (5)

One information document is extracted from a plurality of information documents created and published in a text data format, and the first selection is a selected sentence among sentences included in the one information document The one information document corresponding to the sentence meaning by the subjective expression, whether the sentence is the sentence meaning by the subjective expression or the intention by the objective expression for the unspecified product. Is the first positive example, and the one information document corresponding to the meaning of the objective expression is the first negative example,
The extracted one information document is determined by the classifier based on the feature of the end of the sentence in the last phrase of the first selected sentence, and is distributed to the first positive example or the first negative example, The one information document corresponding to the first positive example is set as an object of analysis performed in the second analysis step, and the one information document corresponding to the first negative example is set as the second information document. A first analysis step excluded from the analysis by the analysis step;
The deemed information example document in which the one information document is all classified as the first positive example is defined as a population, and the one analysis schedule information document is extracted from the assumed positive example information document. Whether or not the second selected sentence, which is a sentence selected from the information document, is a sentence including an evaluation corresponding to a specific product to be analyzed, a condition including the evaluation as a sentence When the satisfied one analysis schedule information document is a second positive example, and the one analysis schedule information document that does not satisfy this condition is a second negative example,
The one analysis schedule information document extracted is determined by a classifier based on a feature vector including a keyword related to the specific product and a word co-occurring with the keyword, and the second positive example or The one analysis schedule information document corresponding to the second positive example is identified as the analysis information document to be analyzed, and the one corresponding to the second negative example is assigned to the second negative example. A second analysis step of excluding the analysis scheduled information document from the deemed positive example information document;
Consisting of,
Text mining method characterized by The text mining method according to claim 1,
The target of the first negative example is an advertisement composed of content promoting the unspecified product,
Text mining method characterized by In the text mining method according to claim 1 or claim 2,
The feature used in the first analysis step is a morpheme,
Text mining method characterized by In the text mining method according to any one of claims 1 to 3,
The specific product is a gas appliance;
Text mining method characterized by One information document is extracted from a plurality of information documents created and published in a text data format, and the first selection is a selected sentence among sentences included in the one information document The one information document corresponding to the sentence meaning by the subjective expression, whether the sentence is the sentence meaning by the subjective expression or the intention by the objective expression for the unspecified product. Is the first positive example, and the one information document corresponding to the meaning of the objective expression is the first negative example,
The first information document extracted is discriminated by the classifier based on the feature at the end of the sentence in the final phrase of the first selected sentence, and is distributed to the first positive example or the first negative example. And the steps
The one information document corresponding to the first positive example is set as an object to be analyzed in the second analysis step, and the one information document corresponding to the first negative example is set to the first information document. A second step excluded from analysis by the two analysis steps;
The deemed information example document in which the one information document is all classified as the first positive example is defined as a population, and the one analysis schedule information document is extracted from the assumed positive example information document. Whether or not the second selected sentence, which is a sentence selected from the information document, is a sentence including an evaluation corresponding to a specific product to be analyzed, a condition including the evaluation as a sentence When the satisfied one analysis schedule information document is a second positive example, and the one analysis schedule information document that does not satisfy this condition is a second negative example,
The one analysis schedule information document extracted is determined by a classifier based on a feature vector including a keyword related to the specific product and a word co-occurring with the keyword, and the second positive example or A third step of allocating to the second negative example;
The one analysis schedule information document corresponding to the second positive example is recognized as an analysis information document to be analyzed, and the one analysis schedule information document corresponding to the second negative example is regarded as the deemed positive document. A fourth step of excluding from the example information document;
Consisting of,
A text mining program characterized by

Images