JP5325131B2 - Pattern extraction apparatus, pattern extraction method, and program - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To automatically generate as many patterns as required which contributes to automatic sort processing, even if the number of texts to which labels are given is small. <P>SOLUTION: A statistical model which is generated by using text with a label as training data and is constituted so that probability data which shows a probability of applied optional text belonging to a predetermined set is outputted is taken as a sorting model. The sorting model is applied to the text having no label, and probability data showing the probability that the text having no label belongs to the predetermined set is generated. An index showing the degree of relevance between a first pattern which is an arbitrary pattern, and a sort result when sorting the text containing the first pattern into a set to which the text belongs is generated using the value determined from the generated probability data. <P>COPYRIGHT: (C)2011,JPO&amp;INPIT

Description

  The present invention relates to a technique for extracting a pattern of a sequence that contributes to automatic classification processing of data consisting of symbol sequences.

  There is a technique for automatically classifying a group of data consisting of symbol series (for example, document data described by character symbols, gene sequences described by gene symbols, etc.). For example, there is a technique for classifying an email group in which spam emails and other emails are mixed into a set of spam emails and a set of other emails. In such a technique, automatic data classification processing is performed based on a pattern of a certain characteristic (symbol series). For example, a series pattern having a high frequency included in data belonging to a predetermined set (for example, a set of spam mails) is prepared, and the classification target data is determined depending on whether or not the classification target data includes the pattern of the series. Are automatically classified. Therefore, in such an automatic classification technique, it is necessary to generate a sequence pattern that contributes to the automatic classification process (for example, a sequence pattern that is frequently included in data belonging to a set of spam mails) as a prerequisite.

  On the other hand, there is a technique for automatically extracting a pattern having a high frequency from a data group consisting of symbol series (see, for example, Non-Patent Document 1).

  FIG. 1 is a diagram for explaining a conventional technique for automatically extracting a pattern having a high frequency from a data group consisting of symbol sequences.

  In the following, an element (minimum unit) consisting of one or more symbols is called an item, a series consisting of one or more items is called a text, a series consisting of one or more items included in the text is called a pattern, and text This set is called a database. Examples of symbols are letters, numbers, marks, etc., and examples of items are letters, numbers, marks, words, word strings, bases, base pairs, and the like. In the example of FIG. 1, for example, “a” to “d” are items, a series such as “abcc” is one text, and a series such as “a”, “ab”, and “ac” included in the text is a pattern. It is. In the example of FIG. 1, a database consisting of five texts is handled.

  Here, consider extracting a pattern having an appearance frequency larger than ζ = 2 from the database. Note that even if the same pattern appears multiple times in one text, the number of times the pattern is counted for that text is one.

  First, the appearance frequency of an item (length 1 pattern) is calculated for a given database (referred to as “input database (IDB)”). In the example of FIG. 1, the appearance frequencies of the items “a”, “b”, “c”, and “d” are 5, 4, 4, and 2, respectively. That is, in the input database, there are three items “a”, “b”, and “c” whose appearance frequency is greater than ζ = 2. These three items “a”, “b”, “c” are stored in memory as elements of the output list (OUT) (OUT = {a, b, c}), and then each pattern “a”, “b” The appearance frequency of the pattern starting from “c” is calculated. Since the appearance frequency of the pattern starting from “d” never becomes higher than ζ = 2, the pattern starting from “d” is not subjected to subsequent processing. In this example, using a database generated by projection, the appearance frequency of patterns starting from patterns “a”, “b”, and “c” in the input database is calculated. Projection deletes each entry (initially text) in the database from the beginning to the position where a certain item (referred to as “target item”) is first found, and the rest of the entries are entered in the database. To create a new database. In the projection, entries that do not include the item of interest are excluded from the database. In the process of calculating the item appearance frequency described here, the items “a”, “b”, and “c” whose appearance frequency is greater than ζ = 2 are set as items of interest.

  First, when projection (prj (a)) is performed on the input database with “a” as the item of interest, a database (DB (a)) with entries “bcc”, “c”, “c”, “bd” Is created. Next, the appearance frequency of the item (length 1 pattern) is obtained for this database (DB (a)). Here, if there is an item whose frequency is greater than ζ = 2, a pattern that is a series in which the item of interest of the previous projection is added before the item is stored in the memory as an element of the output list (OUT). In the example of the projection (prj (a)) in FIG. 1, there are two items “b” and “c” whose appearance frequency is greater than ζ = 2. Therefore, before each item “b” “c”, the pattern “ab” “ac”, which is a series in which the attention item “a” of the past projection is added, is stored as an element of the output list (OUT). Stored (OUT = {a, b, c, ab, ac}). Further, the projection is executed again on the database (DB (a)). In the example of FIG. 1, a projection (prj (b)) is performed on the database (DB (a)) with the item “b” having a frequency greater than ζ = 2 as the item of interest. Is generated. In the database (DB (a b)), since there is no item whose frequency is higher than ζ = 2, no new element is added to the output list (OUT). Next, returning to the database (DB (a)), a projection (prj (c)) is performed with the item “c” whose frequency is greater than ζ = 2 as the target item, and the database (DB (ac)) is generated. The In the database (DB (ac)), there is no item with a frequency greater than ζ = 2, so that no new element is added to the output list (OUT). Next, returning to the input database (IDB), a projection (prj (b)) with “b” as the item of interest is performed, a database (DB (b)) is generated, and “ba” and “bc” are output lists. It is stored in memory as an element of (OUT) (OUT = {a, b, c, ab, ac, ba, bc}). After that, according to the same criteria, in the depth priority order, the item “a” “c” for the database (DB (b)) is the target item, and the item “c” for the input database (IDB) is the target item. Projection is executed, and the process ends. By recursively performing projection in this way, it is possible to efficiently obtain a pattern whose appearance frequency is greater than a certain value ζ. In the example of FIG. 1, “a, b, c, a b, a c, b a, b c” is obtained as a pattern whose appearance frequency is greater than ζ = 2.

J. Pei, J. Han, B. Mortazavi-Asl, H. Pinto, Q. Chen, U. Dayal, and M.-C. Hsu 2001. PrefixSpan: Mining Sequential Patterns Efficiently by Prefix-Projected Pattern Growth In Proc. of the 17th ICDE, pages 215-224.

  However, in the conventional technique as described in Non-Patent Document 1, although a pattern having a high appearance frequency can be automatically extracted from the database, a pattern contributing to the automatic classification process cannot be automatically generated. For example, even if a label (for example, a label indicating whether it is spam mail or the like) that is a data representing a set to which the text belongs is assigned to each text belonging to the database, it is conventional as in Non-Patent Document 1. In technology, pattern extraction considering labels cannot be performed.

  Further, the text associated with the label (labeled) is classified for each set indicated by the label (for example, classified as spam mail), and Non-Patent Document 1 is set for each classified set. If the conventional technique as described in (1) is applied, a pattern having a high appearance frequency can be extracted separately for each set. The pattern extracted in this way is a pattern that contributes to the automatic classification process. However, the process of associating a label with a text is usually performed manually, and its cost is high. Therefore, it is difficult to prepare a large amount of text associated with labels. Furthermore, when pattern extraction is performed from a small amount of text, the number of patterns to be extracted becomes very small, and the performance of automatic classification processing using the extracted patterns is degraded.

  The present invention has been made in view of these points, and even if the number of texts with labels is small, it is possible to automatically generate a necessary number of patterns contributing to automatic classification processing. Aims to provide a new technology.

  In the present invention, an element including one or more symbols, an item including one or more items as text, a sequence including one or more items included in the text as a pattern, and data representing a set to which the text belongs Is a statistical model generated by using text associated with a label as text with label, text not associated with label as unlabeled text, and using text with label as training data, and A classification model is configured to output probability data representing the probability that an applied arbitrary text belongs to a predetermined set. The classification model is applied to unlabeled text, and probability data representing the probability that the unlabeled text belongs to a predetermined set is generated. Then, at least using a value determined from the generated probability data, the relationship between the first pattern which is an arbitrary pattern and the classification result when the text including the first pattern is classified into the set to which the text belongs. An index representing the height of is generated.

  By using the index generated in the present invention, it is possible to automatically generate a pattern having a high degree of contribution to the automatic text classification process. The index can be generated using the result of applying the classification model to unlabeled text. Therefore, in the present invention, even if the number of texts with labels is small, a necessary number of patterns contributing to the automatic classification process can be automatically generated.

The figure for demonstrating the prior art which extracts the pattern of a high frequency series automatically from the data group which consists of a symbol series. The block diagram for demonstrating the function structure of the pattern extraction apparatus of embodiment. The figure for demonstrating the process of a pattern extraction apparatus. The figure for demonstrating the process of a pattern extraction apparatus. The flowchart for demonstrating the specific example of the process of step S14. The flowchart for demonstrating the specific example of the process of step S14. The flowchart for demonstrating the specific example of the process of step S14. Pseudo code for explaining a specific example of the process of step S14. The figure for demonstrating the Example of step S14. The figure for demonstrating the Example of step S14. The figure for demonstrating the Example of step S14.

  Hereinafter, embodiments of the present invention will be described with reference to the drawings.

<Functional configuration>
FIG. 2 is a block diagram for explaining a functional configuration of the pattern extraction apparatus 1 according to the embodiment.

  As illustrated in FIG. 2, the pattern extraction apparatus 1 according to the present embodiment includes a training unit 11, a classification unit 12, a database synthesis unit 13, an extraction unit 14, a control unit 15, and storage units 16 a to 16 e. Have. The extraction unit 14 includes an index generation unit 14a, a limit value generation unit 14b, and a search unit 14c. Note that the pattern extraction apparatus 1 of the present embodiment has a predetermined program on a known computer or a dedicated computer having, for example, a CPU (central processing unit), a RAM (random-access memory), a ROM (read-only memory), and the like. It is a special device that is read and configured. That is, the training unit 11, the classification unit 12, the database synthesis unit 13, the extraction unit 14, and the control unit 15 are CPUs that execute predetermined programs, for example. The storage units 16a to 16e are, for example, an auxiliary storage device such as a hard disk device, a RAM, a register, a cache memory, or a storage area in which at least a part of these is combined. In addition, at least some of the functions of the training unit 11, the classification unit 12, the database synthesis unit 13, the extraction unit 14, and the control unit 15 may be configured by an integrated circuit, or at least some of the storage units 16a to 16e are stored. A region may be present in the integrated circuit. The pattern extraction device 1 executes each process under the control of the control unit 15. In the following, although a part of the description is omitted, data generated in each calculation is stored in each storage unit, and is read from there as needed and used for each calculation.

<Processing>
3 and 4 are diagrams for explaining the processing of the pattern extraction apparatus 1. Hereinafter, processing of the pattern extraction apparatus 1 will be described with reference to these drawings.

[Pre-processing]
As pre-processing, one or more pieces of labeled data including labeled text and unlabeled data including unlabeled text are stored in the storage unit 16a (FIG. 2). Examples of labeled text and unlabeled text are document data described by character symbols, gene sequences described by gene symbols, program sequences described by program codes, and the like. The labeled text is text that is associated with a label, which is data representing a set to which each text belongs, by a human in advance. The number of sets may be any number as long as it is two or more. However, in the following, two sets are set, one set being “P ₊ (positive example: +)” and the other set being “P ₋ ( Negative example:-) ". In the example of FIG. 4, the identifier (ID), the text (text with label), and the contents of the label (P ₊ , P ₋ ) (if the label represents the set P ₊ (P ₊ , P ₋ ) = (1,0) If the label represents the set P ₋ , the set (P ₊ , P ₋ ) = (0, 1)) is the labeled data. In addition, the contents of an identifier (ID) and the text (no label text) label of the (P _+, P _-) information indicating that is unknown _{((P +, P -)} = (unk, unk)) the set of Is unlabeled data.

  A predetermined attenuation parameter λ and a parameter K representing the upper limit value of the elements of the output list are stored in the storage unit 16e. The attenuation parameter λ is, for example, a value selected in advance from a range of 0 ≦ λ ≦ 1, preferably from a range of 0 <λ <1. The parameter K is, for example, a positive integer selected in advance.

[Pattern extraction processing]
First, the training unit 11 extracts the labeled text included in the labeled data from the storage unit 16a, uses the labeled text as training data, and the probability that the applied arbitrary text belongs to a predetermined set (class) A classification model is generated which is a statistical model configured to output probability data representing. The generated classification model is stored in the storage unit 16b (step S11). Any classification model may be used as long as it is configured to output probability data representing the probability that an applied arbitrary text belongs to a predetermined set. For example, a model in a well-known statistical method such as the Naive Bayes method or the maximum entropy method may be used as the classification model. Below, the classification model in the case of using the Naive Bayes method is illustrated.

[Specific Example of Step S11 (Example of Naive Bayes Method)]
In the Naive Bayes method, the probability Pr (c | x) that the text x belongs to a predetermined set (class) c is defined by the following equation.

ここで、w_iはテキストxのi番目のアイテムを表し、N(w_i,x)は、テキストxにおけるアイテムw_iの出現頻度を表す。Pr(c)は、テキストxが集合cに属することの生起確率を表し、以下の式で定義される。

Here, w _i represents the i-th item of the text x, and N (w _i , x) represents the appearance frequency of the item w _i in the text x. Pr (c) represents the occurrence probability that the text x belongs to the set c, and is defined by the following equation.

また、Pr(w_i|c)は以下の式で定義される。

Pr (w _i | c) is defined by the following equation.

ナイーブ・ベイズ法における分類モデルは、訓練データであるラベルありテキストを用いて算出（学習）された式(2)(3)の値となる。以下にナイーブ・ベイズ法において学習される分類モデルを例示する。

The classification model in the Naive Bayes method is a value of Equations (2) and (3) calculated (learned) using labeled text that is training data. The classification model learned in the Naive Bayes method is illustrated below.

In this example, the following 10 labeled texts are used as training data, and a classification model configured to output probability data representing the probability that an applied arbitrary text x belongs to the set P ₊ is generated.

この場合、訓練データであるラベルありテキストがラベルP₊で表される集合c=P₊に属する生起確率Pr(P₊)、及び、訓練データであるラベルありテキストがラベルP_-で表される集合c=P_-に属する生起確率Pr(P_-)は以下のようになる。

In this case, a training data labels have text labels P = ₊ represented collectively c P ₊ belonging probability Pr (P _+), and, the text has a label which is training data labels P _- represented by The occurrence probability Pr (P ₋ ) belonging to the set c = P ₋ is as follows.

Pr (P ₊ ) = 7/10… (4)
_{Pr (P -) = 3/} 10 ... (5)
Also, Pr (w _i | P ₊ ) for each word w _i = a to _i is as follows.

また、各単語w_i=a〜iについてのPr(w_i|P_-)は以下のようになる。

Also, Pr (w _i | P ₋ ) for each word w _i = a to _i is as follows.

この例では、式(1)から定まるPr(P₊|x)と式(4)〜(23)とが分類モデルとなる（［ステップＳ１１の具体例（ナイーブ・ベイズ法の例）］の説明終わり）。なお、ステップＳ１１の処理は、必ずしもパターン抽出処理が行われるたびに実行される必要はなく、過去に生成された分類モデルを使用できるのであれば、ステップＳ１１の処理が省略されてもよい。

In this example, Pr (P ₊ | x) determined from Expression (1) and Expressions (4) to (23) are classification models ([Specific Example of Step S11 (Example of Naive Bayes Method)]] the end). Note that the process of step S11 does not necessarily have to be executed every time the pattern extraction process is performed, and the process of step S11 may be omitted if a classification model generated in the past can be used.

  Next, the above-described unlabeled text and classification model are input to the classification unit 12. The classification unit 12 applies the classification model to unlabeled text, and generates probability data representing the probability that the unlabeled text belongs to a predetermined set (step S12). The specific method of this process differs depending on the classification model. For example, when the above-described classification model in the Naive Bayes method is used, the following probability data is generated.

[Specific Example of Step S12 (Example of Naive Bayes Method)]
When the above-described classification model in the Naive Bayes method is used, the probability Pr (c | x) that the input unlabeled text x belongs to the predetermined set c is calculated according to the equation (1). For example, when unlabeled text x = aaccddf is applied to a classification model consisting of Pr (P ₊ | x) determined from formula (1) and formulas (4) to (23), the unlabeled text x = aaccddf Is the probability Pr (P ₊ | x) belonging to the given set represented by the label P ₊

となる（［ステップＳ１２の具体例（ナイーブ・ベイズ法の例）］の説明終わり）。

(End of description of [specific example of step S12 (example of naive Bayes method)]).

The probability Pr (P ₊ | x) generated as described above is associated with the corresponding unlabeled text and stored in the storage unit 16c as data with probability data. For example, in the example of FIG. 4, an identifier (ID), text (unlabeled text), probability Pr (P ₊ | x), and probability Pr (P ₋ | x) = 1−Pr (P ₊ | x) The associated data with probability data is stored in the storage unit 16c.

Next, the database synthesizing unit 13 (FIG. 2) includes the labeled data stored in the storage unit 16a, the data with probability data stored in the storage unit 16c, and the attenuation parameter λ stored in the storage unit 16e. Entered. The database synthesizing unit 13 generates a database by synthesizing the labeled data and the data with probability data (step S13). At this time, considering that the probability Pr (P ₊ | x) and Pr (P ₋ | x) included in the data with probability data are automatically generated using the classification model, the data with probability data is Probabilities Pr (P ₊ | x) and Pr (P ₋ | x) are multiplied by the attenuation parameter λ. In the example of FIG. 4, the database is generated with the attenuation parameter λ = 0.5. The generated database is stored in the storage unit 16d.

  Next, the extraction unit 14 receives the database stored in the storage unit 16d, the parameter K and the attenuation parameter λ stored in the storage unit 16e, and contributes to the text label classification (automatic text classification process). K patterns having a high degree of contribution) are extracted, and an output list having them as elements is output (step S14). Details of the process in step S14 will be described below.

[index]
In the conventional method, based on the appearance frequency of the pattern, whether to extract a pattern included in a certain database, and whether to perform a recursive process by generating a new database by performing projection, It was decided. On the other hand, in the present embodiment, not the appearance frequency of the pattern but the degree that the pattern contributes to the label classification (that is, the pattern and the classification result when the text including the pattern is classified into the set to which the text belongs). Decide whether to extract a pattern contained in a certain database based on an index indicating the level of relevance) and whether to perform a recursive process by generating a new database by performing projection Is done.

  This index is generated by the index generation unit 14a of the extraction unit 14 using at least a value determined from the probability data generated by the classification unit 12. Below, an example of the index which index generating part 14a generates is shown.

  In this example, consider the following contingency table.

ここで、記憶部１６ｄに格納されたデータベースは、|D^L|個（|D^L|≧0）のラベルありテキストと、分類部１２で確率データがそれぞれ生成された|D^U|個（|D^U|＞0）のラベルなしテキストとを含むものとする。この例では|D^L|≧1とする。なお、|D^L|はデータベースが含むラベルありテキストの総数を表し、|D^U|はデータベースが含むラベルなしテキストの総数を表す。また、D^Lはラベルありテキストの集合を表し、D^Uはラベルなしテキストの集合を表す。

Here, the database stored in the storage unit 16d includes | D ^L | (| D ^L | ≧ 0) labeled text and | D ^U | (|| D ^U |> 0) unlabeled text. In this example, | D ^L | ≧ 1. | D ^L | represents the total number of unlabeled texts included in the database, and | D ^U | represents the total number of unlabeled texts included in the database. D ^L represents a set of labeled text, and D ^U represents a set of unlabeled text.

N is the sum of a value determined from the total number of labeled texts included in the database | D ^L | and a value determined from the total number of unlabeled texts included in the database | D ^U |. In this example, N is the sum of the total number of labeled texts included in the database | D ^L | and the product of the value determined from the total number of unlabeled texts D ^U included in the database and the attenuation parameter λ, as follows: .

N = | D ^L | + λ ・ | D ^U |… (25)
M is a value determined from the total number of the texts with labels included in the database that are associated with the labels indicating that they belong to a predetermined set, and the probability that the text with labels generated by the classification unit 12 belongs to the set P _+. This is the sum of the total number of values determined from the represented probability data. In this example, as follows, the total number and the label has a probability and the damping parameter text belongs to the set P ₊ Although database associated with the label has a label indicating that they belong to a predetermined set P ₊ of the text including Let M be the sum of the total number of products with λ.

なお、d_iはテキストを表し、I(d_i)はラベルありテキストd_i∈D^Lが集合P₊に属するときに1となり、それ以外のときに0となる関数を表す。また、Pr(P₊|d_j)(d_j∈D^U)は、ラベルなしテキストd_j∈D^Uが集合P₊に属する確率を表す。

Here, d _i represents a text, and I (d _i ) represents a function that is 1 when the labeled text d _i ∈D ^L belongs to the set P ₊ and 0 otherwise. Pr (P ₊ | d _j ) (d _j ∈D ^U ) represents the probability that the unlabeled text d _j ∈D ^U belongs to the set P ₊ .

y (α) is a value including the pattern α and a value determined from the total number of the texts with labels that are included in the database and that are associated with the labels indicating that they belong to a predetermined set and that include the pattern α. None represents the sum of the total number of values determined from the probability data representing the probability that the text belongs to a predetermined set. In this example, as shown below, the total number of labeled texts included in the database and including the pattern α among the labeled texts associated with the labels indicating that they belong to the set P ₊ , and the labels including the pattern α Let y (α) be the sum of the probability that none text belongs to the set P ₊ and the total number of products of the attenuation parameters λ.

なお、F(d_i,α)は、テキストd_iがパターンαを含むときに1となり、それ以外のときに0となる関数である。

Note that F (d _i , α) is a function that becomes 1 when the text d _i includes the pattern α, and 0 otherwise.

  x (α) represents the sum of a value determined from the total number of labeled texts including the pattern α and a value determined from the total number of unlabeled texts including the pattern α. In this example, x (α) is the sum of the total number of labeled texts including the pattern α and the product of the total number of unlabeled texts including the pattern α and the attenuation parameter λ as follows.

このとき、パターンαと、当該パターンαを含むテキストを当該テキストが属する集合に分類した際の分類結果と、の関連性の高さを表す指標として、以下のΧ²(α)を用いる。

At this time, the following Χ ² (α) is used as an index representing the degree of relevance between the pattern α and the classification result when the text including the pattern α is classified into the set to which the text belongs.

この例の場合、指標生成部１４ａは、記憶部１６ｄに格納されたデータベースと、記憶部１６ｅに格納された減衰パラメータλとを入力とし、式(29)にしたがって指標Χ²(α)を生成する。パターンαに対する指標Χ²(α)は記憶部１６ｅに格納される。

In the case of this example, the index generation unit 14a receives the database stored in the storage unit 16d and the attenuation parameter λ stored in the storage unit 16e, and generates the index Χ ² (α) according to the equation (29). To do. The index Χ ² (α) for the pattern α is stored in the storage unit 16e.

The extraction unit 14 outputs, as elements of the output list L, K patterns selected in descending order of contribution to the label classification based on the index. In the example of this embodiment, the above indices are sequentially generated for each pattern, and when indices corresponding to each of K or more patterns are generated, K is counted in descending order of contribution to the label classification. Let the th index be the threshold τ _K. Then, depending on whether or not each index generated thereafter satisfies the output condition based on the threshold value τ _K , whether or not the pattern corresponding to the index is an element of the output list L, and performing a projection to newly It is determined whether to generate a recursive database and perform recursive processing. For example, when using the index Χ ² (α) of Equation (29), when indices corresponding to each of K or more patterns are generated, the K-th largest index is set as the threshold τ _K. To do. Then, depending on whether or not the output condition that each generated index exceeds the threshold value τ _K satisfies the output condition, whether or not the pattern corresponding to the index is an element of the output list L, and performing a projection It is determined whether to generate a recursive database and perform recursive processing.

[Indicator limits]
However, even if an index corresponding to a certain pattern α does not satisfy the output condition, an index corresponding to a new pattern in which one or more items are added to the pattern α may satisfy the output condition. . That is, even if the index does not satisfy the output condition, a pattern that satisfies the output condition may be detected when a new database is generated by performing projection and recursive processing is performed.

Therefore, in this embodiment, the limit value generation unit 14b of the extraction unit 14 converts an arbitrary pattern that is an arbitrary series in which one or more items are added to a certain pattern α and an arbitrary text including the arbitrary pattern to the A limit value of an index representing the degree of relevance between the classification result when the text is classified into the set to which the text belongs is generated. The limit value represents the best value of the index corresponding to an arbitrary pattern in which one or more items are added to the pattern α (the value that most contributes to label classification). For example, when the index Χ ² (α) of Expression (29) is used, the following limit value Χ ² _max (α) is generated for the pattern α. Χ ² (α, y = x) represents Χ ² (α) when y = x, and Χ ² (α, y = 0) represents Χ ² (α) when y = 0. And max (ν, μ) represents a function that becomes ν when ν ≧ μ, and μ when ν <μ.

Χ ² _max (α) = max (Χ ² (α, y = x), Χ ² (α, y = 0))… (30)
Then, depending on whether or not the limit value satisfies a predetermined search condition, it is determined whether to perform projection to generate a new database and perform recursive processing.

[Search process]
The search unit 14c of the extraction unit 14 uses the above-described index and limit value generated for the determination target pattern, determines whether the determination target pattern is an element of the output list, and performs projection. Decide whether to generate a new database and perform recursive processing.

(I) That is, when the index generated for the first pattern α satisfies a predetermined first output condition (for example, Χ ² (α)> τ _K ), the search unit 14c outputs the first pattern α. While outputting as an element of the list, the index generation unit 14a outputs a second pattern that is a series in which one or more items are added to the first pattern α, and a second text including the second pattern. A second index is generated that indicates the degree of relevance between the classification result when the text is classified into the set to which the text belongs. If the second index satisfies a predetermined second output condition by subsequent recursive processing, the second pattern is output as an element of the output list.

(II) Further, when the limit value generation unit 14b does not satisfy the first output condition when the index corresponding to the first pattern α satisfies the predetermined search condition (for example, Χ ² (α) ≦ τ _K And Χ ² _max (α)> τ _K ), the search unit 14c outputs one or more items in the first pattern α without outputting the first pattern α as an element of the output list. A third index representing the high degree of relevance between the third pattern that is a series to which the third text is added and the classification result when the third text including the third pattern is classified into the set to which the third text belongs. Generate. Then, if the third index satisfies a predetermined third output condition by subsequent recursive processing, the third pattern is output as an element of the output list.

(III) When the index corresponding to the first pattern α does not satisfy the first output condition and the limit value does not satisfy the predetermined search condition (for example, Χ ² (α) ≦ τ _K and Χ ² _max (α ) ≦ τ _K ), the first pattern α is not an element of the output list, and the third index is not generated.

In other words, the processing of the extraction unit 14 is as follows when the index Χ ² (α) in Expression (29) and the limit value Χ ² _max (α) in Expression (30) are used.

If (I) Χ ² (α)> τ _K , the pattern α is output as an element of the output list, and projection is performed, and a sequence in which one or more items are added to the pattern α is set as a new pattern α. Perform recursive processing.

(II) If Χ ² (α) ≤ τ _K and Χ ² _max (α)> τ _K , projection is performed without using pattern α as an element of the output list, and one or more items are added to pattern α. The sequence is set as a new pattern α, and recursive processing is performed.

(III) If Χ ² (α) ≦ τ _K and Χ ² _max (α) ≦ τ _K , pattern α is not an element of the output list, projection is not performed, and one or more items are added to pattern α Does not process the series.

[Specific Example of Processing in Step S14]
5 to 7 are flowcharts for explaining a specific example of the process of step S14. FIG. 8 is a pseudo code for explaining a specific example of the processing in step S14. Hereinafter, a specific example of the process of step S14 will be described with reference to these drawings.

First, the search unit 14c of the extraction unit 14 sets the pattern α to be empty (α = []), sets R as the database stored in the storage unit 16d, and sets the threshold τ _K = nan (a value indicating that it is undetermined). (Step S1401). These values are stored in the storage unit 16e.

  Next, the search unit 14c determines whether R is an empty set (R = {}) (step S1402). Here, if R = {}, the process ends. On the other hand, if R = {} is not set, the search unit 14c sets RR to an empty set (RR ← {}) (step S1403), and determines whether α is empty (α = []). (Step S1404).

  If it is determined that α is empty, the search unit 14c sets RR to R (RR ← R) (step S1405), and sets a set of items included in R to β (β ← itemset (R)). (Step S1406), and the process of step S1415 described later is executed.

  On the other hand, if it is determined that α is not empty, a sequence entry trans (transεR) including items included in R is selected (step S1408). When the projection has never been executed and R is a database stored in the storage unit 16d, the text included in the database stored in the storage unit 16d corresponds to the entry trans. Further, when projection has been executed in the past, the remaining series excluding items deleted by the previous projection from the text corresponds to the entry trans. Next, the search unit 14c deletes the selected entry trans from the beginning to the position where the item α is first found, and sets the remaining series as subseq (subseq ← postseq (last (α), trans)) is executed (step S1409). Next, the search unit 14c determines whether subseq is not empty (subseq ≠ []) (step S1410). Here, if subseq is not empty, the search unit 14c sets a value obtained by adding subseq as an entry to RR as a new RR (RR ← append (RR, subseq)) (step S1411), and processing in step S1412 Execute. On the other hand, if subseq is empty, the search unit 14c executes the process of step S1412 without executing the process of step S1411. In the process of step S1412, the search unit 14c determines whether the process has been completed for all entries transεR (step S1412). If it is determined that processing has not been completed for all entries transεR, the process returns to step S1408. On the other hand, if it is determined that the processing has been completed for all entries transεR, the search unit 14c sets the set of items included in the RR as β (β ← itemset (RR)) (step S1413). The process of step S1415 is executed. Note that the processing from step S1408 to S1413 corresponds to projection.

In the process of step S1415, the search unit 14c selects an item (itemεβ) belonging to β (step S1415). Next, the search unit 14c generates a sequence in which the item item selected in step S1415 is added to the end of α as a new pattern α (α ← append (α, [item])) (step S1416). Next, the search unit 14c determines whether the number of elements | L | of the patterns belonging to the output list L is less than the parameter K stored in the storage unit 16e (| L | <K) (step S1417). . If it is determined that | L | <K, the index generation unit 14a generates an index Χ ² (α) for the pattern α according to the equation (29), and the search unit 14c determines the pattern α and the index. set of the ^{Χ 2 (α) [α,} Χ 2 (α)] was added as an element of the output list L and updates the output list L. The updated output list L (L ← append (L, [α, Χ ² (α)])) is stored in the storage unit 16e (step S1418). Next, the search unit 14c determines whether or not the number of elements | L | of the patterns belonging to the output list L is equal to the parameter K stored in the storage unit 16e (| L | = K) (Step S14). S1419). If it is determined that | L | = K is not satisfied, processing in step S1422 described later is executed. On the other hand, if it is determined that | L | = K, the search unit 14c newly reorders the elements [α, Χ ² (α)] of the output list L in descending order of the index Χ ² (α). The output list L (L = sort (L)) is stored in the storage unit 16e (step S1420). Next, the search unit 14c sets the index of the Kth element (the index of the smallest value) of the output list L as the threshold τ _K (τ _K = Χ ² (L [K])), and updates the threshold τ _K. Is stored in the storage unit 16e (step S1421), and the process of the next step S1422 is executed.

  In the process of step S1422, the extraction unit 14 recursively executes the processes from step S1402 to S1431 (call WTPS (α, RR, K)) with respect to the current (α, RR, K) (step S1422). ). Thereafter, the process of step S1431 described later is executed.

On the other hand, if it is determined in step S1418 that | L | <K is not satisfied, the index generation unit 14a generates an index Χ ² (α) for the pattern α according to the equation (29), and stores it in the storage unit 16e. The search unit 14c determines whether or not τ ² (α)> τ _K is satisfied using the threshold value τ _K stored in the storage unit 16e (step S1423). When the threshold is not yet determined (τ _K = nan), it is assumed that Χ ² (α)> τ _K is not satisfied. Here, when it is determined that ( ² (α)> τ _K is satisfied, the search unit 14 c uses the last element (most index Χ ² (α) of the output list L stored in the storage unit 16 e. Is deleted, and a new output list L (L = lastdel (L)) including the remaining elements is generated and stored in the storage unit 16e (step S1424). Next, the search unit 14c adds the set [α, Χ ² (α)] of the pattern α and the index Χ ² (α) as an element of the output list L, and updates the output list L. The updated output list L (L ← append (L, [α, Χ ² (α)])) is stored in the storage unit 16e (step S1425). Next, the search unit 14c converts the elements [α, Χ ² (α)] of the output list L into a new output list L (L = sort (L)) in the descending order of the index Χ ² (α). And stored in the storage unit 16e (step S1426). Next, the search unit 14c sets the index of the Kth element (the index of the smallest value) of the output list L as the threshold τ _K (τ _K = Χ ² (L [K])), and updates the threshold τ _K. And stored in the storage unit 16e (step S1427). Next, the extraction unit 14 recursively executes the processing from step S1402 to S1431 (call WTPS (α, RR, K)) for the current (α, RR, K) (step S1428). Thereafter, the process of step S1431 described later is executed.

On the other hand, if it is determined in step S1423 that Χ ² (α)> τ _K is not satisfied, the limit value generating unit 14b sets the limit value Χ ² _max (α) for the pattern α to the equation (30). Are generated and stored in the storage unit 16e, and the search unit 14c uses the threshold value τ _K stored in the storage unit 16e to determine whether or not Χ ² _max (α)> τ _K is satisfied (step S1429). . Here, if it is determined that Χ ² _max (α)> τ _K is satisfied, the extraction unit 14 performs the processing from step S1402 to S1431 (call WTPS (α, RR, K)) is recursively executed (step S1430). Thereafter, the following step S1431 is executed. On the other hand, when it is determined that Χ ² _max (α)> τ _K is not satisfied, the process of step S1431 is executed without executing the process of step S1430.

  In the process of step S1431, the search unit 14c determines whether the process has been completed for all items itemεβ (step S1431). Here, if the processing is completed for all items itemεβ, the processing in step S14 is completed. On the other hand, if the process has not been completed for all items itemεβ, the process returns to step S1415.

[Example of Step S14]
9 to 11 are diagrams for explaining the embodiment of step S14. Hereinafter, an example of step S14 will be described with reference to these drawings. In this embodiment, the database shown in FIG. 3 (multiplied by the attenuation parameter λ = 0.5) is used as the initial database stored in the storage unit 16d. Further, K = 4 and λ = 0.5. M = 1.8 and N = 3.5, which are constants in the initial database. Also, the upper right subscripts of the items a, b, c, and d of the nodes of the tree structure shown in FIGS. 10 and 11 are the indices Χ ² ( α), and the lower right subscript represents the limit value Χ ² _max (α) of the pattern α. Further, x shown in FIGS. 10 and 11 is a new pattern α (α ← append (α, [item])) obtained by adding an item item which is an element of β to the end of the pattern α, and the new For pattern α, x (α) generated according to equation (28) is represented. Further, y shown in FIGS. 10 and 11 is a new pattern α (α ← append (α, [item])) obtained by adding an item item, which is an element of β, to the end of the pattern α. For pattern α, y (α) generated according to equation (27) is represented.

  1. First, α = [], RR ← R, β = {a, b, c, d} are set (steps S1401 to S1406 / FIG. 10A).

2. Next, α = a is set (steps S1415 and S1416). Since | L | <K (k = 4), a pair [a, 0] of a and index Χ ² (a) = 0 is added to the elements of the output list L (steps S1417, S1418 / FIG. 9). (A)). At this time, the threshold τ _K = nan. Further, projection is performed (step S1422), and β = {b, c, d} is set in the recursive processing (call WTPS (α, RR, K) / steps S1402 to S1431) (FIG. 10 ( B)).

In the recursive processing of 3.2, α = ab is set (steps S1415 and S1416 / FIG. 10B). Since | L | <K (k = 4), a set [ab, 0.5] of ab and index Χ ² (ab) = 0.5 is added to the elements of the output list L (steps S1417, S1418 / FIG. 9). (B)). At this time, the threshold τ _K = nan. Further, projection is performed (step S1422), and β = {c, d} is set in the recursive process (FIG. 10B).

In the recursive process of 4.3, c of β = {c, d} is selected as item, and α = abc is set (steps S1415, S1416 / FIG. 10B). Since | L | <K (k = 4), a pair [abc, 1.3] of abc and index Χ ² (abc) = 1.3 is added to the elements of the output list L (steps S1417, S1418 / FIG. 9). (C)). At this time, the threshold τ _K = nan. Further, projection is performed (step S1422), and β = {c} is set in the recursive processing (FIG. 10B).

5. In the recursive processing of 4, α = abcc is set (steps S1415, S1416 / FIG. 10B). Since | L | <K (k = 4), a pair [abcc, 1.3] of abcc and the index Χ ² (abcc) = 1.3 is added to the elements of the output list L (steps S1417 and S1418).

6. Since | L | = K (k = 4) in the recursive processing of 4, the elements of the output list L are rearranged based on the index Χ ² (α), and the threshold τ _K = 0 is obtained. (Steps S1419 to S1421 / FIG. 9D).

  7. Since no more projections are possible (yes in step S1431 of 4 recursive processing), 4 recursive processing ends, and 3 recursive processing returns. Here, α = a b, β = {c, d}, but c of β = {c, d} has been processed.

In the recursive process of 8.3, d is selected as item out of β = {c, d}, and α = abd is set (step 1415, S1416 / FIG. 10B). Since | L | <K (k = 4) is not satisfied (step S1417), and the index Χ ² (abd) = 0.2 for α = abd exceeds the threshold τ _K = 0 (step S1423), from the output list L, [a, 0] is deleted, the pair [abd, 0.2] of abd and index Χ ² (abd) = 0.2 is added to the elements of output list L, the elements of output list L are rearranged, and the threshold is τ _K = It is updated to 0.2 (steps S1424 to S1427 / FIG. 9E).

  9. Since no more projection is possible (yes in step S1431 of the 3 recursive process), the 3 recursive process ends and the process returns to the 2 recursive process. Here, α = a, β = {b, c, d}, but b of β = {b, c, d} has been processed (FIG. 10B).

In the recursive processing of 10.2, c is selected as item out of β = {b, c, d}, and α = ac is set (step 1415, S1416 / FIG. 10B). Since | L | <K (k = 4) is not satisfied (step S1417) and the index Χ ² (ac) = 2.1 for α = ac exceeds the threshold value τ _K = 0.2 (step S1423), [abd, 0.2] is deleted, a pair [ac, 2.1] of ac and the index Χ ² (ac) = 2.1 is added to the elements of the output list L, the elements of the output list L are rearranged, and the threshold is τ _K = It is updated to 0.5 (steps S1424 to S1427 / FIG. 9F). Further, projection is performed (step S1428), and β = {c} is set in the recursive process (FIG. 10B).

11. In 10 recursive processes, c of β = {c} is selected as item, and α = acc is set (step 1415, S1416 / FIG. 10B). Since | L | <K (k = 4) is not satisfied (step S1417) and the index Χ ² (acc) = 1.3 for α = acc exceeds the threshold τ _K = 0.5 (step S1423), [ab, 0.5] is deleted, the set [ac, 2.1] with acc and index Χ ² (acc) = 1.3 is added to the elements of the output list L, the elements of the output list L are rearranged, and the threshold is τ _K = Updated to 1.3 (steps S1424 to S1427 / FIG. 9G).

  12. Since no more projection is possible (yes in step S1431 of the 10 recursive process), the 10 recursive process ends and the process returns to the 2 recursive process. Here, α = a, β = {b, c, d}, but b and c of β = {b, c, d} have been processed (FIG. 10B).

In the recursive processing of 13.2, d is selected as item out of β = {b, c, d}, and α = ad is set (step 1415, S1416 / FIG. 10B). Not | L | <K (k = 4) (step S1417), neither the index Χ ² (ad) = 0.2 nor the limit value Χ ² _max (ad) = 0.5 for α = ad exceeds the threshold τ _K = 1.3 ( Since all elements of β = {b, c, d} have been processed (steps S1423 and S1429) (step S1431), the recursive process of 2 is completed and the process returns to the first loop process. In the first loop processing, α = [], β = {a, b, c, d}, but a has been processed among β = {a, b, c, d}.

14. In the first loop processing, b is selected as item among β = {a, b, c, d}, and α = b is set (step 1415, S1416 / FIG. 10C). | L | <K (k = 4) is not satisfied (step S1417), and the index Χ ² (b) = 0 for α = b does not exceed the threshold τ _K = 1.3 (step S1423), but the limit value Χ ² _max ( Since b) = 2.7 exceeds the threshold τ _K = 1.3 (step S1429), projection is performed (step S1430), and β = {a, c, d} is set in the recursive processing (FIG. 10 ( C)).

15. Among 14 recursive processes, a is selected as item out of β = {a, c, d}, and α = ba is set (step 1415, S1416 / FIG. 10C). Not | L | <K (k = 4) (step S1417), and the index Χ ² (ba) = 0.5 for α = ba does not exceed the threshold τ _K = 1.3 (step S1423), but the limit value Χ ² _max ( Since ba) = 1.6 exceeds the threshold τ _K = 1.3 (step S1429), projection is performed (step S1430), and β = {c} is set in the recursive processing (FIG. 10C).

16. In 15 recursive processing, c is selected as item from β = {c}, and α = bac is set (step 1415, S1416 / FIG. 10C). Not | L | <K (k = 4) (step S1417), neither the index Χ ² (bac) = 0.3 nor the limit value Χ ² _max (bac) = 0.5 for α = bac exceeds the threshold τ _K = 1.3 ( Since all elements of β = {c} have been processed (steps S1423 and S1429) (step S1431), the recursive process of 15 ends and returns to the recursive process of 14. In the recursive processing of 14, β = {a, c, d}, but among β = {a, c, d}, a has been processed.

17. Among 14 recursive processes, c is selected as item among β = {a, c, d}, and α = bc is set (step 1415, S1416 / FIG. 10C). Not | L | <K (k = 4) (step S1417), and the index Χ ² (bc) = 0.2 for α = bc does not exceed the threshold τ _K = 1.3 (step S1423), but the limit value Χ ² _max ( Since bc) = 2.3 exceeds the threshold τ _K = 1.3 (step S1429), projection is performed (step S1430), and β = {a, c} is set in the recursive processing (FIG. 10C). ).

18. In 17 recursive processing, a of β = {a, c} is selected as item, and α = bca is set (step 1415, S1416 / FIG. 10C). Since | L | <K (k = 4) is not satisfied (step S1417), and the index Χ ² (bca) = 1.5 for α = bca exceeds the threshold τ _K = 1.3 (step S1423), [acc, 1.3] is deleted, a pair [bca, 1.5] of bca and index Χ ² (bca) = 1.5 is added to the elements of output list L, the elements of output list L are rearranged, and the threshold is τ _K = 1.3 (steps S1424 to S1427 / FIG. 9H).

19. Since no more projections are possible, in 17 recursive processes, c of β = {a, c} is selected as item and α = bcc is set (step 1415, S1416 / FIG. 10). (C)). Not | L | <K (k = 4) (step S1417), but the index Χ ² (bcc) = 1.3 for α = bcc is the threshold τ _K = 1.3 and the limit value Χ ² _max (bcc) = 1.3 is also the threshold τ _K = 1.3 is not exceeded (steps S1423 and S1429), and the determination in step S1431 is yes, so that the 17 recursive processing ends and returns to the 14 recursive processing. In the recursive processing of 14, β = {a, c, d}, but a and c of β = {a, c, d} have been processed.

20. In 14 recursive processes, d is selected as item out of β = {a, c, d}, and α = bd is set (step 1415, S1416 / FIG. 10C). Not | L | <K (k = 4) (step S1417), neither the index Χ ² (bd) = 0.2 nor the limit value Χ ² _max (bd) = 0.5 for α = bd exceeds the threshold τ _K = 1.3, (Steps S1423, S1429) and yes in the determination of step S1431, the 14 recursive processing ends, and the process returns to the first loop processing. In the first loop processing, β = {a, b, c, d}, but a and b of β = {a, b, c, d} have been processed.

21. In the first loop process, c is selected as item among β = {a, b, c, d}, and α = c is set (step 1415, S1416 / FIG. 11A). Not | L | <K (k = 4) (step S1417), and the index Χ ² (c) = 0.2 for α = c does not exceed the threshold τ _K = 1.3 (step S1423), but the limit value Χ ² _max ( b) = 3.1 exceeds the threshold τ _K = 1.3 (step S1429), projection is performed (step S1430), and β = {a, c} is set in the recursive processing (FIG. 11A). ).

22. In 21 recursive processing, a of β = {a, c} is selected as item, and α = ca is set (step 1415, S1416 / FIG. 11A). Since | L | <K (k = 4) is not satisfied (step S1417), and the index Χ ² (ca) = 1.5 for α = ca exceeds the threshold τ _K = 1.3 (step S1423), [abcc, 1.3] is deleted, a pair [ca, 1.5] of ca and the index Χ ² (ca) = 1.5 is added to the elements of the output list L, the elements of the output list L are rearranged, and the threshold is τ _K = 1.3 (steps S1424 to S1427 / FIG. 9I).

23. Since no more projections are possible, c is selected as item among β = {a, c} in 21 recursive processes, and α = cc is set (step 1415, S1416 / FIG. 11). (A)). Not | L | <K (k = 4) (step S1417), but the index Χ ² (cc) = 1.3 for α = cc is the threshold value τ _K = 1.3 and the threshold value Χ ² _max (cc) = 1.3 is also the threshold value τ _K = 1.3 is not exceeded (steps S1423 and S1429), and the determination in step S1431 is yes, so that the 21 recursive process ends and the process returns to the first loop process. In the first loop processing, β = {a, b, c, d}, but a, b, c of β = {a, b, c, d} have been processed.

24. In the first loop processing, d is selected as item out of β = {a, b, c, d}, and α = d is set (step 1415, S1416 / FIG. 11B). Since | L | <K (k = 4) is not satisfied (step S1417) and the index Χ ² (d) = 2.1 for α = d exceeds the threshold τ _K = 1.3 (step S1423), the output list L indicates [abc, 1.3] is deleted, a pair [d, 2.1] of d and index Χ ² (d) = 2.1 is added to the elements of the output list L, the elements of the output list L are rearranged, and the threshold is τ _K = 1.5 (steps S1424 to S1427 / FIG. 9J). Further, projection is performed (step S1428), and β = {a, b, c, d} is set in the recursive process (FIG. 11B).

25. In the recursive process of 24, a is selected as item out of β = {a, b, c, d}, and α = da is set (step 1415, S1416 / FIG. 11B). Since | L | <K (k = 4) is not satisfied (step S1417), and the index Χ ² (da) = 2.1 for α = da exceeds the threshold τ _K = 1.5 (step S1423), [ca, 1.5] is deleted, a pair [da, 2.1] of da and index Χ ² (da) = 2.1 is added to the elements of the output list L, the elements of the output list L are rearranged, and the threshold is τ _K = 1.5 (steps S1424 to S1427 / FIG. 9K). Further, projection is performed (step S1428), and β = {b, d} is set in the recursive process (FIG. 11B).

26. In the recursive processing of 25, b is selected as item out of β = {b, d}, and α = dab is set (step 1415, S1416 / FIG. 11B). Not | L | <K (k = 4) (step S1417), neither the index Χ ² (dab) = 0.2 nor the limit value Χ ² _max (dab) = 0.5 for α = dab exceeds the threshold τ _K = 1.5 ( Steps S1423 and S1429).

27. Next, in 25 recursive processes, d is selected as item out of β = {b, d}, and α = dad is set (step 1415, S1416 / FIG. 11B). Not | L | <K (k = 4) (step S1417), neither the index Χ ² (dad) = 0.2 nor the limit value _max ² _max (dad) = 0.5 for α = dad exceeds the threshold τ _K = 1.5 ( Since all elements of β = {b, d} have been processed (steps S1433 and S1429) (step S1431), the recursive process of 25 ends and returns to the recursive process of 24. In the recursive process of 24, β = {a, b, c, d}, but a has been processed for β = {a, b, c, d}.

28. In the recursive processing of 24, b is selected as item among β = {a, b, c, d}, and α = db is set (step 1415, S1416 / FIG. 11B). Since | L | <K (k = 4) is not satisfied (step S1417) and the index Χ ² (db) = 2.1 for α = db exceeds the threshold τ _K = 1.5 (step S1423), the output list L indicates [bca, 1.5] is removed, the set [db between db and metrics ^{Χ 2 (db) = 2.1,} 2.1] is added to the elements of the output list L, and rearranged elements of the output list L is, the threshold tau _K = 2.1 (steps S1424 to S1427 / FIG. 9 (L)). Further, projection is performed (step S1428), and β = {a, c, d} is set in the recursive processing (FIG. 11B).

29. In the recursive processing of 28, a is selected as item among β = {a, c, d}, and α = dba is set (step 1415, S1416 / FIG. 11 (B)). Not | L | <K (k = 4) (step S1417), neither the index Χ ² (dba) = 1.5 nor the limit value Χ ² _max (dba) = 1.5 for α = dba exceeds the threshold τ _K = 2.1 ( Steps S1423 and S1429).

30. In the recursive processing of 28, c is selected as item among β = {a, c, d}, and α = dbc is set (step 1415, S1416 / FIG. 11B). Not | L | <K (k = 4) (step S1417), neither the index Χ ² (dbc) = 1.5 nor the limit value Χ ² _max (dba) = 1.5 for α = dbc exceeds the threshold τ _K = 2.1 ( Steps S1423 and S1429).

31. In the recursive processing of 28, d is selected as item out of β = {a, c, d}, and α = dbd is set (step 1415, S1416 / FIG. 11B). Not | L | <K (k = 4) (step S1417), neither the index Χ ² (dbd) = 0.5 nor the limit value Χ ² _max (dbd) = 0.2 for α = dbd exceeds the threshold τ _K = 2.1 ( Steps S1423 and S1429). Since the determination in step S1431 of 28 recursive processing is yes, the 28 recursive processing ends and returns to 24 recursive processing. In the recursive processing of 24, β = {a, b, c, d}, but a and b of β = {a, b, c, d} have been processed.

32. In the recursive processing of 24, c is selected as item out of β = {a, b, c, d}, and α = dc is set (step 1415, S1416 / FIG. 11B). Not | L | <K (k = 4) (step S1417), neither the index Χ ² (dc) = 1.5 nor the limit value Χ ² _max (dc) = 1.5 for α = dc exceeds the threshold τ _K = 2.1 ( Steps S1423 and S1429).

33. In the recursive process of 24, d is selected as item out of β = {a, b, c, d}, and α = dd is set (step 1415, S1416 / FIG. 11B). Not | L | <K (k = 4) (step S1417), neither the index Χ ² (dd) = 0.2 nor the limit value Χ ² _max (dd) = 0.5 for α = dd exceeds the threshold τ _K = 2.1 ( Steps S1423 and S1429). Since the determination in step S1431 of 24 recursive processing is yes, the 24 recursive processing ends, and the process returns to the first loop processing.

  34. Since the determination in step S1431 of the first loop processing is yes, the processing in step S14 ends, and the output list L stored in the storage unit 16e is output.

[Modifications, etc.]
The present invention is not limited to the embodiment described above. For example, in the above embodiment, when generating a database that combines the labeled data and the data with probability data in step S13, the probabilities Pr (P ₊ | x) and Pr (P ₋ | x) included in the data with probability data are generated. ) Is multiplied by the attenuation parameter λ. However, a configuration in which the attenuation parameter λ is not multiplied may be used. Conversely, the label value (1 or 2) of the labeled data may be multiplied by an amplification parameter ρ (ρ ≧ 1, preferably ρ> 1). Further, both the attenuation parameter λ and the amplification parameter ρ may be multiplied.

  Further, the following formulas (31) to (34) may be used instead of the formulas (25) to (28).

N = ρ ・ | D ^L | + | D ^U |… (31)

或いは、式(25)-(28)の代わりに以下の式(35)-(38)が用いられてもよい。

Alternatively, the following formulas (35) to (38) may be used instead of the formulas (25) to (28).

N = ρ ・ | D ^L | + λ ・ | D ^U |… (35)

すなわち、データベースが含むラベルありテキストの総数|D^L|から定まる値が、当該総数|D^L|と所定の増幅パラメータとの積であり、データベースが含むラベルありテキストのうち所定の集合に属することを表すラベルに対応付けられたものの総数から定まる値が、当該ラベルありテキストのうち所定の集合に属することを表すラベルに対応付けられたものの総数と増幅パラメータとの積であり、データベースが含むラベルありテキストであって所定の集合に属することを表すラベルに対応付けられたラベルありテキストのうち第１パターンαを含むものの総数から定まる値が、当該ラベルありテキストであって所定の集合に属することを表すラベルに対応付けられたラベルありテキストのうち第１パターンαを含むものの総数と増幅パラメータとの積であり、第１パターンαを含むラベルありテキストの総数から定まる値が、当該第１パターンαを含むラベルありテキストの総数と増幅パラメータとの積であってもよい。

That is, the database label has the total number of text containing the | is determined from the value, the total number | | D ^L D ^L | is the product of the predetermined amplification parameters, that belonging to a predetermined set of labels have text database contains The value determined from the total number of labels associated with the label representing the product is the product of the total number of labels associated with the label representing that they belong to a predetermined set of the labeled text and the amplification parameter, and the label included in the database A value determined from the total number of texts including the first pattern α among the texts with labels that are associated with the labels indicating that the texts belong to the predetermined set is the text with labels and belongs to the predetermined set. The total number of texts with the first pattern α among the labeled texts associated with the labels representing the The value determined from the total number of labeled texts including the first pattern α may be the product of the total number of labeled texts including the first pattern α and the amplification parameter.

In the above-described embodiment, Χ ² (α) itself is used as an index. However, other values corresponding to the broad monotonic function value (monotonic non-decreasing function value) of Χ ² (α) may be used as an index. The value corresponding to the weakly monotonically function value of chi ² (alpha) is a concept including a chi ² (alpha) itself. For example, if the value corresponding to the broad monotone increasing function value of Χ ² (α) is used as an index, it can be said that the pattern α having a larger index value contributes more to the label classification. Further, for example, if a value corresponding to a broad-sense monotone decreasing function value of Χ ² (α) is used as an index, it can be said that a pattern α having a smaller index value has a higher degree of contribution to label classification. In addition, a convex function value indicating the degree of relevance between the pattern α and the classification result when the text including the pattern α is classified into a set to which the text belongs may be used as an index.

Similarly, in the above-described embodiment, Χ ² _max (α) itself is used as an index. However, other values corresponding to broad monotone function values of Χ ² _max (α) may be used as an index. The value corresponding to the weakly monotonically function value of Χ ² _max (α) is a concept including a Χ ² _max (α) itself.

  In addition, the various processes described above are not only executed in time series according to the description, but may be executed in parallel or individually according to the processing capability of the apparatus that executes the processes or as necessary. Needless to say, other modifications are possible without departing from the spirit of the present invention.

  Further, when the above-described configuration is realized by a computer, processing contents of functions that each device should have are described by a program. The processing functions are realized on the computer by executing the program on the computer.

  The program describing the processing contents can be recorded on a computer-readable recording medium. As the computer-readable recording medium, any recording medium such as a magnetic recording device, an optical disk, a magneto-optical recording medium, and a semiconductor memory may be used.

  The program is distributed by selling, transferring, or lending a portable recording medium such as a DVD or CD-ROM in which the program is recorded. Furthermore, the program may be distributed by storing the program in a storage device of the server computer and transferring the program from the server computer to another computer via a network.

  A computer that executes such a program first stores, for example, a program recorded on a portable recording medium or a program transferred from a server computer in its own storage device. When executing the process, the computer reads a program stored in its own recording medium and executes a process according to the read program. As another execution form of the program, the computer may directly read the program from a portable recording medium and execute processing according to the program, and the program is transferred from the server computer to the computer. Each time, the processing according to the received program may be executed sequentially. Also, the program is not transferred from the server computer to the computer, and the above-described processing is executed by a so-called ASP (Application Service Provider) type service that realizes the processing function only by the execution instruction and result acquisition. It is good. Note that the program in this embodiment includes information that is used for processing by an electronic computer and that conforms to the program (data that is not a direct command to the computer but has a property that defines the processing of the computer).

  In this embodiment, the present apparatus is configured by executing a predetermined program on a computer. However, at least a part of these processing contents may be realized by hardware.

DESCRIPTION OF SYMBOLS 1 Pattern extraction apparatus 11 Training part 12 Classification part 13 Database composition part 14 Extraction part 15 Control part 16a-16e Storage part

Claims (10)

Data that represents a set to which a text belongs, with an element consisting of one or more symbols as an item, a series of one or more items as text, a series of one or more items included in the text as a pattern, A storage unit that stores the unlabeled text when the text associated with a certain label is labeled text and the text that is not associated with the label is unlabeled text;
A statistical model generated by using the labeled text as training data, and a classification model configured to output probability data representing a probability that an applied arbitrary text belongs to a predetermined set. Applying to the unlabeled text and generating probability data representing the probability that the unlabeled text belongs to the predetermined set;
Using a value determined from the probability data generated by the classification unit, a first pattern that is an arbitrary pattern, and a classification result when the text including the first pattern is classified into a set to which the text belongs. An extractor for generating an index representing the degree of relevance;
A pattern extraction apparatus having: The pattern extraction device according to claim 1,
The extraction unit includes:
When the index satisfies a predetermined first output condition, the first pattern is output as an element of an output list, and a second pattern is a series in which one or more items are added to the first pattern; A second index indicating the level of relevance between the second text including the second pattern and the classification result when the second text is classified into the set to which the second text belongs is generated, and the second index is a predetermined second When the output condition is satisfied, the second pattern is configured to be output as an element of the output list.
A pattern extraction apparatus characterized by that. The pattern extraction device according to claim 2,
The extraction unit includes:
The relationship between an arbitrary pattern that is an arbitrary sequence in which one or more items are added to the first pattern, and a classification result when the arbitrary text including the arbitrary pattern is classified into a set to which the text belongs. Generating a limit value of an index representing height, and outputting the first pattern as an element of an output list when the index does not satisfy the first output condition but the limit value satisfies a predetermined search condition A third pattern that is a series in which one or more items are added to the first pattern, and a classification result when the third text including the third pattern is classified into a set to which the third text belongs, Generating a third index representing the degree of relevance of the third pattern, and outputting the third pattern as an element of the output list when the third index satisfies a predetermined third output condition;
A pattern extraction apparatus characterized by that. The pattern extraction device according to any one of claims 1 to 3,
The value determined from the probability data is a product of the probability data and a predetermined attenuation parameter.
A pattern extraction apparatus characterized by that. The pattern extraction device according to any one of claims 1 to 4,
The indicator is
| D ^L | (| D ^L | ≧ 0) of the labeled text and | D ^U | (| D ^U |> 0) of the unlabeled text from which the probability data is generated, respectively. Generated against the database,
The first pattern is α,
The sum of the value determined from the total number of labeled texts included in the database | D ^L | and the value determined from the total number of unlabeled texts included in the database | D ^U | is N,
Sum of a value determined from the total number of texts with labels included in the database and associated with a label indicating that the text belongs to the predetermined set, and a total number of values determined from the probability data generated by the classification unit Is M,
A value determined from the total number of texts with labels that are included in the database and that include the first pattern α among the texts with labels that are associated with labels that belong to the predetermined set, and the first pattern α. Y (α) is the sum of the total number of values determined from the probability data representing the probability that the unlabeled text to be included belongs to the predetermined set,
When the sum of the value determined from the total number of labeled texts including the first pattern α and the value determined from the total number of unlabeled texts including the first pattern α is x (α),

Is a value equivalent to the monotonic function value in the broad sense of
A pattern extraction apparatus characterized by that. The pattern extraction device according to claim 5,
The extraction unit further includes:
Χ ² (α) when y = x is Χ ² (α, y = x), and Χ ² (α) when y = 0 is Χ ² (α, y = 0), max (ν, μ) when ν ≧ μ is ν, and max (ν, μ) when ν <μ is μ,
Χ ² _max (α) = max (Χ ² (α, y = x), Χ ² (α, y = 0))
Generates the limit value of the index corresponding to the broad monotonic function value of
A pattern extraction apparatus characterized by that. The pattern extraction device according to any one of claims 5 to 6,
The value determined from the total number of unlabeled texts | D ^U | included in the database is a product of the total number of unlabeled texts | D ^U | and a predetermined attenuation parameter, and the value determined from the probability data is the probability A product of the probability represented by the data and the attenuation parameter, and a value determined from the total number of unlabeled text including the first pattern α is the product of the total number of unlabeled text including the first pattern α and the attenuation parameter. Is,
And / or
The label has the total number of text wherein the database comprises | D ^L | determined by the value, the total number | D ^L | is the product of the predetermined amplification parameters, the predetermined set of labels have text wherein the database comprises A value determined from the total number of items associated with the label indicating belonging is a product of the total number of items associated with the label indicating belonging to the predetermined set of the labeled text and the amplification parameter, A value determined from the total number of texts with labels that are included in the database and that include the first pattern α among the texts with labels that are associated with labels that belong to the predetermined set is the text with labels. The first pattern of the labeled text associated with the label indicating belonging to the predetermined set. The value determined from the total number of labeled texts including the first pattern α is the product of the total number including the first pattern α and the amplification parameter. Is the product of
A pattern extraction apparatus characterized by that. The pattern extraction device according to any one of claims 1 to 7,
The extraction unit is configured to output a predetermined number or less of patterns selected in descending order of the size of the corresponding index as an element of an output list.
A pattern extraction apparatus characterized by that. A set in which the classification unit has an element including one or more symbols as an item, a series of one or more items as text, a series of one or more items included in the text as a pattern, and the text belongs to When the text associated with the label that is data representing the text is a text with a label and the text not associated with the label is a text without a label, the text with the label is used as training data and generated A classification model configured to output probability data representing a probability that an applied arbitrary text belongs to a predetermined set, to the unlabeled text stored in the storage unit. Applied to generate probability data representing the probability that the unlabeled text belongs to the predetermined set. The method comprising the steps of,
Classification result when the extraction unit classifies the first pattern which is an arbitrary pattern and the text including the first pattern into a set to which the text belongs, using a value determined from the probability data generated by the classification unit Generating an index indicating the degree of relevance of
A pattern extraction method comprising:   A program for causing a computer to function as the pattern extraction device according to claim 1.

Images