JP5182845B2 - Information extraction apparatus, information extraction method, and information extraction program - Google Patents

Description

  The present invention relates to an information extraction technique, and more particularly, to an information extraction apparatus, an information extraction method, and an information extraction program that extract item expressions and unique expressions from an article group as information pairs.

As a conventional information extraction technique, for example, the following Non-Patent Document 1 relates to a technique for determining whether an inter-sentence relationship is a transition or an update using an inter-sentence relationship and extracting trend information using the information. It is described.
Namba, Kokusei, Fukushima, Aizawa, Okumura: “Extraction and visualization of trend information considering inter-document relations”, IPSJ SIG, 2005-NL-168, pp.67-74 (2005) .

  However, the above prior art does not automatically extract a pair of item expression and specific expression included in an article group. Further, the prior art does not display the pair of the extracted item expression and specific expression in a graph. Therefore, depending on the prior art, it is not possible to display at a glance the pairs of item expressions and unique expressions included in the article group, and it is difficult for the user to quickly understand the information of the article group.

  An object of the present invention is to solve the above-described problems of the prior art and provide an information extraction apparatus, an information extraction method, and an information extraction program for automatically extracting a pair of item expression and specific expression from an article group.

  The information extraction device of the present invention is an information extraction device, wherein main expression extraction means for extracting item expressions and types of specific expressions from article groups as main expressions, and the main expressions appear simultaneously from the article groups. And an information pair extracting unit that extracts a pair of an item expression and a specific expression described in the specified place as an information pair.

  Preferably, in the information extraction device of the present invention, the main expression extraction means extracts item expressions, types of unique expressions, and unit expressions from the article group as main expressions, and the information pair extraction means includes item expressions and A pair of a specific expression and a numerical expression related to the unit expression is extracted as an information pair.

  Preferably, in the information extracting apparatus of the present invention, the main expression extracting unit further includes the specific expression based on whether or not a predetermined word is attached before or after the type of the extracted specific expression. The types are classified, and the type of each classified unique expression is set as the main expression.

  The information extraction device of the present invention is an information extraction device, wherein main expression extraction means for extracting item expressions and unit expressions as main expressions from an article group, and the main expressions appear from the article group simultaneously. An information pair extracting unit that extracts a pair of an item expression described in the specified part and a numerical expression related to the unit expression as an information pair, and the main expression extracting unit includes: , Classifying the numerical expressions based on whether or not a predetermined word is attached before or after the numerical expressions related to the extracted unit expressions, and unit expressions related to each of the classified numerical expressions Is the main expression.

  Preferably, in the information extracting apparatus of the present invention, the main expression extracting means further extracts a word accompanying before or after the type of the extracted specific expression, and a word selected from the extracted words The accompanying specific expression type is the main expression.

  The information extraction device of the present invention is an information extraction device, wherein main expression extraction means for extracting item expressions and unit expressions as main expressions from an article group, and the main expressions appear from the article group simultaneously. An information pair extracting unit that extracts a pair of an item expression described in the specified part and a numerical expression related to the unit expression as an information pair, and the main expression extracting unit includes: Further, a word associated with the extracted unit expression before or after the numerical expression is extracted, and a unit expression associated with the numerical expression associated with the selected word from the extracted word is defined as the main expression. To do.

  Preferably, in the information extraction device of the present invention, the main expression extraction unit further determines a word that increases the degree of classification of the main expression when the main expression is attached to and not attached to the main expression. A main expression with a word and a main expression without the word are extracted as main expressions.

  The information extraction device of the present invention is an information extraction device, wherein main expression extraction means for extracting item expressions and unit expressions as main expressions from an article group, and the main expressions appear from the article group simultaneously. An information pair extracting unit that extracts a pair of an item expression described in the specified part and a numerical expression related to the unit expression as an information pair, and the main expression extracting unit includes: Further, a word having a high degree of distinction between the main expression and the case where it does not accompany the main expression is determined, and the main expression to which the determined word is attached and the main expression to which the word is not attached are mainly used. Extract as an expression.

  Preferably, in the information extracting apparatus of the present invention, the main expression extracting means further extracts a word accompanying or before the numerical expression related to the extracted unit expression, and the extracted word is attached. Obtaining a normal distribution of the numerical expression and a normal distribution of the numerical expression not accompanied by the extracted word, and a unit expression related to the numerical expression accompanied by the word when the ratio of the obtained normal distributions overlapping is smallest; A unit expression related to a numerical expression not accompanied by the word is defined as the main expression.

  Preferably, in the information extracting apparatus of the present invention, the main expression extracting unit further extracts a word attached before or after the extracted specific expression type, and the unique expression belonging to the extracted specific expression type. Based on the score value determined based on the similarity between the obtained unique expressions and information indicating whether or not the extracted word is attached to each unique expression. A word having a high degree of classification of the specific expression in the case where the word is attached and the case where the word is not attached, and a kind of the unique expression accompanied by the determined word and a kind of the specific expression not accompanied by the word The main expression.

  Preferably, in the information extracting apparatus of the present invention, the main expression extracting unit further extracts a word attached before or after the extracted unit expression, and between numerical expressions related to the extracted unit expression. The similarity is obtained and attached to the unit expression based on the score value determined based on the similarity between the obtained numerical expressions and information indicating whether or not the extracted word is attached to each numerical expression. A word having a high degree of division of the unit expression in a case and a case not accompanied is determined, and a unit expression accompanied by the determined word and a unit expression not accompanied by the word are set as the main expression.

  Preferably, in the information extraction device of the present invention, the main expression extraction unit further performs final determination based on a score value calculated according to a predetermined calculation formula based on the frequency of the extracted main expression in the article group. The main expression to be extracted is determined.

  Preferably, in the information extracting device of the present invention, the main expression extracting unit further extracts a cluster to which a word in the article group belongs from the article group, and each extracted cluster is represented by the specific expression. Kind.

  Preferably, in the information extraction device of the present invention, the information pair extraction unit further selects a specific main expression from the main expressions extracted by the main expression extraction unit based on a user's designated input, An information pair is extracted from the group of articles based on the selected specific main expression.

  Preferably, in the information extraction device of the present invention, the information pair extraction unit further extracts the information pair using a machine learning technique.

  Preferably, in the information extraction apparatus of the present invention, the information pair extraction unit further selects an information pair from the extracted information pairs based on an evaluation value calculated based on a predetermined evaluation value calculation formula .

  Preferably, the information extraction apparatus of the present invention further includes means for performing correlation analysis on the information pair extracted by the information pair extraction means.

  Preferably, the information extraction apparatus according to the present invention further includes display means for displaying the information pairs extracted by the information pair extraction means in a graph.

  Further, the information extraction method of the present invention is an information extraction method, wherein an item expression and a kind of specific expression are extracted from an article group as a main expression, and a place where the main expression appears simultaneously from the article group. A pair of the item expression and the specific expression described in the specified part is identified and extracted as an information pair.

  Also, the information extraction program of the present invention is an information extraction program, which extracts a main expression from a group of items as an item expression and a kind of specific expression, and the main expression simultaneously appears from the article group. And a process of extracting a pair of an item expression and a specific expression described in the specified part as an information pair.

  According to the information extraction device, the information extraction method, and the information extraction program of the present invention, for example, a pair of an item expression and a specific expression included in an article group can be extracted. Moreover, it can display so that the pair of the extracted item expression and specific expression can be grasped at a glance. As a result, the user can quickly understand the information of the article group.

  Hereinafter, embodiments of the present invention will be described with reference to the drawings. FIG. 1 is a diagram showing an example of a system configuration of the present invention. The information extraction device 1 is a processing device that extracts a plurality of information pairs as information pairs from an article group. For example, the information extraction apparatus 1 extracts a pair of one or more item expressions and one or more unique expressions as an information pair from an article group stored in a related article database (DB) 14 described later. Further, the information extraction apparatus 1 extracts a pair of one or more item expressions, one or more unique expressions, and one or more numerical expressions as an information pair from the article group stored in the related article DB 14.

  The information extraction apparatus 1 includes a main expression extraction unit 11, an information pair extraction unit 12, a display unit 13, and a related article database (DB) 14. The main expression extraction unit 11 extracts a main expression from an article group stored in a related article DB 14 described later. The main expression extraction unit 11 extracts, for example, one or more item expressions and one or more types of unique expressions as main expressions. For example, the main expression extraction unit 11 extracts one or more item expressions, one or more types of unique expressions, and one or more unit expressions as main expressions. The main expression is used when an information pair is extracted by the information pair extraction unit 12 described later. When extracting a main expression, for example, a corresponding expression that appears uniformly and frequently in the entire target article group is extracted.

  The main expression extraction unit 11 includes a main unit expression extraction unit 111, a main item expression extraction unit 112, and a main specific expression extraction unit 113. The main unit expression extraction unit 111 extracts a unit expression (main unit expression) necessary for extracting and organizing information pairs. For example, “yen” in the box office revenue “500 million yen” and “person” in the audience mobilization number “300,000” are extracted from the group of articles about the movie as main unit expressions.

  The main item expression extraction unit 112 extracts an item expression (main item expression) necessary for extracting and organizing information pairs. For example, “entertainment income”, “number of audience mobilization”, and the like are extracted as main item expressions from articles related to movies.

  The main specific expression extraction unit 113 extracts the types of specific expressions (types of main specific expressions) necessary for extracting and organizing information pairs. For example, from the group of articles related to movies, the type of specific expression “PERSON” indicating a person and the type of specific expression “LOCATION” indicating a place are extracted as types of main specific expressions.

  Based on the main expression extracted by the main expression extraction unit 11, the information pair extraction unit 12 sets a plurality of information pairs (for example, one or a plurality of item expressions and 1) from the articles constituting the article group in the related article DB 14. Alternatively, a pair with a plurality of specific expressions or a pair of one or a plurality of item expressions, one or a plurality of numerical expressions and one or a plurality of specific expressions) is extracted as an information pair. The specific expression is a specific expression belonging to the type of the specific expression, for example, “Japan”, “America”, etc., belonging to the specific expression type “LOCATION”. A specific expression belonging to the type of specific expression is extracted using a specific expression extraction technique described later.

  For example, in the article group stored in the related article DB 14, the information pair extraction unit 12 detects a location where the main expression (for example, item expression and unique expression type) extracted by the main expression extraction unit 11 appears at the same time. A pair of the unique expression and the item expression described in the place is identified as an information pair. In addition, for example, the information pair extraction unit 12 identifies a place where main expressions (for example, item expression and specific expression type and unit expression) appear simultaneously in the article group stored in the related article DB 14, and A pair of an item expression, a specific expression (a specific expression belonging to the type of specific expression) and a numerical expression described in a place is an information pair. The numerical expression is a numerical expression related to the unit expression as the main expression.

  That is, for the unit expression of the main expressions, the information pair extraction unit 12 simultaneously extracts a numerical value related to the unit expression (for example, a numerical value appearing in the article adjacent to the unit expression), The numerical value and the unit expression are combined and extracted as a numerical expression.

  For example, in the case of a movie article, the information pair extraction unit 12 extracts information pairs of “item expression: typhoon”, “numerical expression: No. 4”, and “LOCATION: Minami Daitojima”.

  The display unit 13 organizes and displays the information pairs extracted by the information pair extraction unit 12 (for example, displays them in a graph).

  Article groups are accumulated in the related article DB 14.

  According to an embodiment of the present invention, the main expression extraction unit 11 further determines the type of the specific expression based on whether a predetermined word is attached before or after the type of the extracted specific expression. And the type of each of the classified proper expressions may be the main expression.

  According to one embodiment of the present invention, the main expression extraction unit 11 extracts item expressions and unit expressions as main expressions from the article group stored in the related article DB 14, and the information pair extraction unit 12 extracts the article group. The location where the main expression appears at the same time is identified, the pair of the item representation described in the identified location and the numerical representation related to the unit representation is extracted as an information pair, and the main representation extraction The unit 11 classifies the numerical expression based on whether or not a predetermined word is attached before or after the numerical expression related to the extracted unit expression, and relates to each of the classified numerical expressions. The unit expression to be used may be the main expression.

  Here, the phrase “with a word” does not necessarily mean that the word appears concatenated with the type of specific expression or numerical expression. For example, the word includes the type of specific expression or numerical expression. It also means that it appears in the same sentence as the sentence that appears. In addition, words associated with the types of specific expressions and numerical expressions are also included in the words associated with the types of specific expressions and numerical expressions.

  In addition, according to an embodiment of the present invention, the main expression extraction unit 11 may further perform before or after the extracted specific expression type, or before or after the numerical expression related to the extracted unit expression. A word associated with a numerical expression accompanied by a type of a unique expression accompanied by a word selected from the extracted word or a word selected from the extracted word; An expression may be used.

  Moreover, according to one Embodiment of this invention, the main expression extraction part 11 extracts item expression and unit expression as main expressions from the article group stored in related article DB14, and the information pair extraction part 12 is the said. A location where the main expression appears simultaneously from an article group is identified, and a pair of an item expression described in the specified location and a numerical expression related to the unit expression is extracted as an information pair, and the main expression is extracted. The expression extraction unit 11 further extracts a word associated with the numerical expression related to the extracted unit expression before or after, and a unit expression related to the numerical expression associated with the word selected from the extracted words May be the main expression.

  In addition, according to an embodiment of the present invention, the main expression extraction unit 11 further determines a word having a high degree of classification of the main expression when it is attached to the main expression and when it is not attached. A main expression accompanied by a word and a main expression not accompanied by the word may be extracted as main expressions.

  According to one embodiment of the present invention, the main expression extraction unit 11 extracts item expressions and unit expressions from the article group as main expressions, and the information pair extraction unit 12 extracts the main expression from the article group. A location that appears at the same time is identified, and a pair of an item representation described in the identified location and a numerical representation related to the unit representation is extracted as an information pair. Determining a word having a high degree of classification of the main expression between the case where it is attached to the main expression and the case where it is not attached, and the main expression including the determined word and the main expression not including the word as the main expression You may make it extract.

  In addition, according to an embodiment of the present invention, the main expression extraction unit 11 further extracts a word accompanying or before the numerical expression related to the extracted unit expression, and the extracted word is A unit related to a numerical expression accompanied by a word when a normal distribution of the accompanying numerical expression and a normal distribution of the numerical expression not accompanied by the extracted word are obtained and the ratio of the obtained normal distributions overlapping is the smallest The main expression may be an expression and a unit expression related to a numerical expression not accompanied by the word.

  In addition, according to an embodiment of the present invention, the main expression extraction unit 11 further extracts a word attached before or after the extracted specific expression type, and sets the extracted specific expression type as the extracted specific expression type. The degree of similarity between specific expressions belonging to each other is obtained, and based on a score value determined based on the degree of similarity between the obtained specific expressions and information indicating whether or not the extracted word is attached to each specific expression. A word having a high degree of classification of the unique expression in the case of accompanying with the expression and the case of not accompanying the expression is determined, and the type of the specific expression to which the determined word is attached and the kind of the unique expression to which the word is not attached are determined. The main expression may be used.

  In addition, according to an embodiment of the present invention, the main expression extraction unit 11 further extracts a word attached before or after the extracted unit expression, and a numerical expression related to the extracted unit expression Based on the score value determined based on the similarity between the obtained numerical expressions and information indicating whether or not the extracted word is attached to each numerical expression, the unit expression is obtained. A word that increases the degree of classification of the unit expression in the case of accompanying and not accompanying is determined, and the unit expression to which the determined word is attached and the unit expression to which the word is not attached are used as the main expression. It may be.

  Further, according to an embodiment of the present invention, the main expression extraction unit 11 is further calculated according to a predetermined calculation formula based on the frequency in the article group stored in the related article DB 14 of the extracted main expression. The main expression to be finally extracted may be determined based on the score value.

  Further, according to an embodiment of the present invention, the main expression extraction unit 11 further extracts a cluster to which a word in an article group of the related article DB 14 belongs from the related article DB 14 and extracts the cluster. Each cluster may be a type of the unique expression.

  In addition, according to an embodiment of the present invention, a dictionary of specific expressions created manually (for example, correspondence information between a station name, a movie name, a space shuttle name, and a word) is stored in a storage unit (not shown) in advance. The main expression extraction unit 11 determines the specific expression corresponding to the word in the article group of the related article DB 14 with reference to the specific expression dictionary, and the specific expression to which the determined specific expression belongs is stored. The type of expression may be extracted as the main expression.

  According to an embodiment of the present invention, the information pair extraction unit 12 further selects a specific main expression from the main expressions extracted by the main expression extraction unit 11 based on a user's designated input, Information pairs may be extracted from the articles constituting the article group based on the selected specific main expression.

  Further, according to the embodiment of the present invention, the information pair extraction unit 12 further extracts a specific expression type specified in advance (according to a user's specification input) as a specific expression type, and the main information is stored in the related article DB 14. A pair of the item expression and the specific expression described in the place where the item expression extracted by the expression extraction unit 11 and the type of the specific expression that appears at the same time may be extracted as an information pair. Good.

  Further, according to an embodiment of the present invention, the information pair extraction unit 12 may further extract the information pair using a machine learning technique.

Below, the detailed example of each component of the information extraction apparatus 1 which concerns on embodiment of this invention is demonstrated.
(Main Expression Extraction Unit 11)
The main expression extraction unit 11 extracts a main expression necessary for extracting and organizing information pairs. The main expression extraction unit 11 extracts, for example, item expressions and types of specific expressions as main expressions. The main expression extraction unit 11 extracts, for example, item expressions, types of specific expressions, and unit expressions as main expressions.

  The main expression extraction unit 11 extracts an item expression and a unit expression using, for example, ChaSen (see the following reference (1)).

Reference (1): Y. Matsumoto, A. Kitauchi, T. Yamashita, Y. Hirano, H. Matsuda and M. Asahara: Japanese morphological analysis system ChaSen version 2.0 manual 2nd edition ”(1999).
In the output of ChaSen, each expression is extracted using part of speech information. For unit expressions, take out the noun series connected to the front or back of the numerical value. For the item expression, for example, a noun series is taken out. Also, for example, expressions including expressions related to time (eg, “year”, “month”, “day”) may be removed from expressions obtained as unit expressions.

  Also, the main expression extraction unit 11 extracts the types of specific expressions by using the specific expression extraction technique described below, for example. At the time of extracting the type of the specific expression, the specific expression belonging to the type of the specific expression is extracted.

  A proper expression is a linguistic expression that means a specific thing / quantity, such as a proper noun such as a person name, a place name, or an organization name, or a numerical expression such as a monetary amount. For example, as types of specific expressions, “ORGANIZATION” indicating an organization, “PERSON” indicating a person, “LOCATION” indicating a place, “ARTIFACT” indicating an artifact, “DATE” indicating a date, and “TIME” indicating a time. , “MONEY” indicating the amount of money, and “PERCENT” indicating the ratio.

  The specific expression extraction technique is a technique for automatically extracting the types of specific expressions as described above and specific expressions belonging to the specific expressions from a sentence by a computer. For example, if a specific expression is extracted for a sentence “The Japanese prime minister is Junichiro Koizumi”, the types of specific expressions (for example, “PERSON” and “LOCATION”) and specific expressions belonging to the types of the specific expressions ( For example, a specific expression “Joiichiro Koizumi” belonging to “PERSON” and a specific expression “Japan” belonging to “LOCATION”) are extracted.

Hereinafter, an example of a general technique for extracting a specific expression will be described.
(1) A method using machine learning There is a method of extracting a specific expression using machine learning (for example, see the following reference (2)).

Reference (2): Masayuki Asahara, Yuji Matsumoto, Use of Redundant Morphological Analysis in Japanese Named Expression Extraction Information Processing Society of Japan Natural Language Processing Study Group NL153-7 2002
First, for example, the sentence “Japan's prime minister is Mr. Koizumi” is divided into each character, and the correct characters such as B-LOCATION and I-LOCATION are assigned to the divided characters as follows. Set the correct answer. The first column below is each divided character, and the correct tag of each character is the second column.
Sun B-LOCATION
I-LOCATION
O
Neck O
Phase O
Is O
Small B-PERSON
Izumi I-PERSON
O
N
At O
O
. O
In the above, B-? ? ? Is a tag that signifies the start of the type of proper expression below the hyphen. For example, B-LOCATION means the beginning of a specific expression indicating a place, and B-PERSON means the start of a specific expression indicating a person name. I-? ? ? Is a tag that means something other than the beginning of the type of proper expression below the hyphen, and O is something else. Therefore, for example, the character “day” is a character corresponding to the beginning of the specific expression indicating the place, and the character “book” is the specific expression indicating the place.

  In this way, the correct answer of each character is set, learned from such data, this correct answer is estimated with new data, and from this correct answer tag, the beginning of each proper expression and how far it is. Is recognized and the proper expression is estimated.

When learning from the correct data set for each character, the system uses various information in the form of features. For example,
Sun B-LOCATION
Part of
Japan-B Noun-B
Such information is used. Japan-B means the beginning of the word Japan, and noun-B means the beginning of the noun. For recognition of words and parts of speech, for example, morphological analysis by ChaSen described above is used. If ChaSen is used, the input Japanese can be divided into words. For example, ChaSen divides a Japanese sentence and estimates the part of speech of each word as described above. For example, if “go to school” is entered, the following results can be obtained.

School Gacco School Noun-General To He To particle-Case particle-General Go Iku Go Verb-independence
In this way, each line is divided so that one word is included, and reading and part-of-speech information are given to each word.

  For example, in the above reference (2), as features, characters constituting the input sentence itself (for example, “small” character), character type (for example, hiragana, katakana, etc.), part of speech information, Tag information (for example, “B-PERSON” or the like) is used.

  Learning using these features. Investigate what features appear in the characters that estimate the tag and the surrounding characters, learn what features are likely to appear when the features appear, and use the learning results to create new data Perform tag estimation. For machine learning, for example, a support vector machine is used.

  In addition to the above-described method, there are various methods for extracting the proper expression. For example, there is a technique for extracting a specific expression using a maximum entropy model and a rewrite rule (see reference (3)).

Reference (3): Kiyotaka Uchimoto, Maoi, Maki Murata, Hiromi Osaku, Masao Uchiyama, Hitoshi Isahara, Named Expression Extraction Based on Maximum Entropy Model and Rewriting Rules, Journal of the Language Processing Society, Vol.7, No.2 , 2000
Further, for example, the following reference (4) describes a technique for extracting Japanese proper expressions using a support vector machine.

Reference (4): Hiroyasu Yamada, Taku Kudo, Yuji Matsumoto, Extracting Japanese Named Expressions Using Support Vector Machine, Journal of Information Processing Society of Japan, Vol.43, No.1 ", 2002
(2) A method using a created rule There is also a method of manually creating a rule to extract a specific expression.

For example,
A noun + “san” means a person's name + “a prime minister” means a person's name + a “town” means a place + a “city” means a place.

  According to one embodiment of the present invention, the main expression extraction unit 11 may select a specific main expression from the extracted main expressions according to a user's designated input.

  The main expression extraction unit 11 extracts, as main expressions, main item expressions, types of unique expressions, and unit expressions that play a main role in the article group currently being handled. For example, a corresponding expression that appears uniformly and frequently in the entire target article group is extracted as a main expression.

Specifically, for the extraction of the main expression, Score (score) values as shown in the following formulas (1) to (3) are used, and the one with a large score value is extracted as the main expression.
(1) Okapi's TF term equation

(2) Total frequency

(3) Total number of appearing articles

Where i is the article number, Docs is the set of article numbers, TF _i is the number of appearances of the expression in article i, l _i is the length of article i, and Δ is the average length of articles in article group Docs. means. The expression of the TF term of Okapi has the effect of increasing the score of an expression that appears uniformly in a plurality of articles and has a high frequency. The length of the article is, for example, the number of words or characters included in the article. As for the type of specific expression, TF _i is the number of appearances of the specific expression belonging to the type of specific expression in article i.

For item expressions, for example, TF _i is not the number of appearances of an expression in article i so that a long character string can be preferentially fetched. A method of product of column length is also used.

  In the embodiment of the present invention, the value obtained by multiplying the value of equation (1) by IDF, that is, log N / DF, the value of equation (2) by the IDF, and the value of equation (3) by the above A value obtained by multiplying IDF may be used as the value of each score. Here, N means the total number of articles in a large-scale corpus (not shown), and DF means the number of articles in which the expression appears in the large-scale corpus, for example.

  In the embodiment of the present invention, the main expression extraction unit 11 extracts, for example, an expression having the highest calculated score value as the main expression. For example, the main expression extraction unit 11 may extract an expression having a calculated score value equal to or greater than a predetermined threshold as the main expression. In addition, the main expression extraction unit 11 may extract a predetermined number of expressions as the main expression from the one with the high calculated score value, for example.

  According to an embodiment of the present invention, the main expression extraction unit 11 further determines the type of the specific expression based on whether a predetermined word is attached before or after the type of the extracted specific expression. Classification may be performed, and the type of each unique expression thus classified may be the main expression. For example, the main expression extraction unit 11 classifies the types of specific expressions based on whether or not the word “alarm” is appended to the extracted specific expression type “ORGANIZATION”. Each of the accompanying “ORGANIZATION” and “ORGANIZATION” without the word “warning” may be used as the main expression.

  In addition, according to an embodiment of the present invention, the main expression extraction unit 11 further extracts a word attached before or after the type of the unique expression extracted, and a word selected from the extracted words The type of specific expression accompanied by may be the main expression. For example, the main expression extraction unit 11 extracts all the words attached before or after the extracted unique expression type “ORGANIZATION”, and sorts each word based on the frequency of appearance in the related article DB 14. The kind of specific expression accompanied by the word selected according to the user's designated input from the displayed word is defined as the main expression. Instead of extracting a word that comes before or after the type of the extracted specific expression, a character string that comes before or after the type of the extracted specific expression may be extracted.

  Further, according to an embodiment of the present invention, the main expression extraction unit 11 further extracts a word accompanying or before the numerical expression related to the extracted unit expression, and the extracted word is attached. Unit expression related to the numerical expression with which the normal distribution of the numerical expression to be obtained and the normal distribution of the numerical expression without the extracted word attached are obtained, and the ratio in which the obtained normal distributions overlap is the smallest And the unit expression related to the numerical expression not accompanied by the word may be the main expression. Instead of extracting a word associated with the extracted unit expression before or after the numerical expression, a character string associated with the extracted numerical expression related to the unit expression may be extracted.

  That is, for example, the main expression extraction unit 11 extracts words or character strings that appear before and after the numerical expression from the related article DB 14. Here, for convenience of explanation, the word or character string is referred to as a “pattern”. Note that patterns appearing in the same sentence may be extracted instead of patterns appearing adjacently before and after the numerical expression. Then, the main expression extraction unit 11 selects a predetermined number of patterns based on the frequency of appearance in the related article DB 14 from the extracted patterns, and sets the selected patterns as pattern candidates.

  Next, the main expression extraction unit 11 performs the following calculation for each pattern candidate. First, the main expression extraction unit 11 obtains an average and variance of numerical expressions with patterns preceding and following. Then, a normal distribution (first normal distribution) is obtained from the obtained average and variance. In addition, the main expression extraction unit 11 obtains an average and variance of numerical expressions in which no pattern is attached before and after. Then, a normal distribution (second normal distribution) is obtained from the obtained average and variance. Then, the main expression extraction unit 11 obtains a ratio in which the obtained first normal distribution and second normal distribution overlap each other. A pattern with a smaller overlapping ratio of the normal distributions has a higher degree (decomposition capability) of separating numerical expressions from each other. The main expression extraction unit 11 determines a pattern candidate when the obtained ratio is the smallest as a final pattern, and the unit expression related to the numerical expression accompanied by the determined final pattern and the pattern are A unit expression related to a numerical expression that is not attached is a main expression. For example, the main expression extraction unit 11 determines the word “hourly speed” as a final pattern, and relates to a unit expression related to a numerical expression accompanied by the word “hourly speed” and a numerical expression not accompanied by “hourly speed”. The unit expression is the main expression.

  According to an embodiment of the present invention, the information pair extraction unit 12 to be described later may extract an information pair including a numerical expression accompanied by the determined final pattern.

  According to an embodiment of the present invention, the main expression extraction unit 11 selects a predetermined number of the above pattern candidates in order of increasing proportion of the normal distributions, and the selected pattern candidates. A unit expression related to a numerical expression accompanied by a symbol and a unit expression related to a numerical expression not accompanied by a candidate for the pattern may be used as a main expression.

  Further, according to an embodiment of the present invention, the main expression extraction unit 11 selects and displays a predetermined number for each of the pattern candidates in a descending order of the overlapping ratio of the normal distributions as a list. A unit expression related to a numerical expression accompanied by a pattern candidate designated according to a user's designated input from the pattern candidates displayed as the list and a unit expression related to a numerical expression not accompanied by the pattern candidate are main expressions. You may make it do.

  In the present invention, a final pattern determination method other than the method described above may be used.

  In addition, according to an embodiment of the present invention, the main expression extraction unit 11 further extracts a word attached before or after the extracted specific expression type, and sets the extracted specific expression type as the extracted specific expression type. The degree of similarity between specific expressions belonging to each other is obtained, and based on a score value determined based on the degree of similarity between the obtained specific expressions and information indicating whether or not the extracted word is attached to each specific expression. A word having a high degree of classification of the unique expression in the case of accompanying with the expression and the case of not accompanying the expression is determined, and the type of the specific expression to which the determined word is attached and the kind of the unique expression to which the word is not attached are determined. The main expression may be used.

  That is, for example, the main expression extraction unit 11 extracts, from the related article DB 14, words or character strings that appear before and after the type of the unique expression. As described above, for convenience of explanation, the word or character string is referred to as a “pattern”. Note that patterns appearing in the same sentence may be extracted instead of patterns appearing adjacently before and after the numerical expression. Then, the main expression extraction unit 11 selects a predetermined number of patterns based on the frequency of appearance in the related article DB 14 from the extracted patterns, and sets the selected patterns as pattern candidates.

  Next, the main expression extraction unit 11 performs the following calculation for each pattern candidate. For example, the main expression extraction unit 11 uses a dictionary that classifies words, such as a classification vocabulary table that is stored in advance in the storage unit, and sets a higher similarity for words that have a closer meaning in the dictionary description. By doing so, the similarity between words is determined in advance.

  Instead of using a dictionary, the similarity between words may be determined by the following method. That is, the main expression extraction unit 11 acquires a word and a word that frequently co-occurs with the word (for example, a word that frequently appears in the same sentence) from a large-scale language corpus stored in advance. To do. Then, a vector having the co-occurrence word as a vector dimension and the number of times of co-occurrence of the co-occurrence word (frequency) as a vector element is created for each word. The similarity between words is defined as an angle (or cos) between word vectors, and the smaller (or larger) this angle (or cos), the higher the similarity.

  After determining the similarity between words as described above, the main expression extraction unit 11 calculates a score (score value) according to the following calculation formula for each pattern candidate.

score = Σ similarity between two specific expressions × f (first specific expression, second specific expression)
However, in the above formula, Σ is a process of adding every two combinations of specific expressions belonging to the specific expression type appearing in the related article DB 14. Further, the similarity between two unique expressions is the similarity between the above-described specific expressions, which is determined according to the above-described method for determining the similarity between words. The first specific expression and the second specific expression are two specific expressions among the specific expressions included in the specific expression belonging to the type of the specific expression. In addition, f (first specific expression, second specific expression) is the case where pattern candidates appear (accompany) in both the first specific expression and the second specific expression, or in which both the pattern candidates are The function is 1 when it does not appear, and is -1 when the pattern candidate appears only in one of them.

  The score is calculated for each pattern candidate, and the score value is obtained. A pattern candidate with a higher score value is a candidate pattern that has a higher degree of distinction (decomposition ability) of the unique expression when it is associated with the unique expression and when it is not associated with the specific expression.

  The main expression extraction unit 11 uses, for example, the above-described specific expression type accompanied by a pattern candidate at the highest score value and the specific expression type not accompanied by the pattern candidate as the main expression.

  According to an embodiment of the present invention, the information pair extraction unit 12 to be described later may extract an information pair including a unique expression accompanied by a pattern candidate at the highest score value.

  In addition, according to an embodiment of the present invention, a predetermined number of pattern candidates are selected in descending order of the score value, and the types of the unique expressions and the pattern candidates to which the selected pattern candidates are attached respectively. You may make it make it the main expression the kind of said specific expression which each does not accompany.

  Further, according to an embodiment of the present invention, the main expression extraction unit 11 selects a predetermined number of pattern candidates in descending order of the score value, displays the selected pattern candidates as a list, and The types of specific expressions accompanied by the pattern candidates designated according to the user's designated input from the pattern candidates displayed as a list and the types of unique expressions not accompanied by the pattern candidates may be used as the main expressions.

  In the present invention, a method for determining a main expression other than the method described above may be used.

  According to an embodiment of the present invention, a final pattern associated with the numerical expression described above may be determined by a method using the score value. In this case, for example, in the method using the score value described above, the type of specific expression is unit expression, the specific expression is a numerical expression (numerical value, numerical data), and the similarity between the numerical expressions is expressed as the closeness of the numerical value. It can be defined as shown. For example, a value obtained by dividing the difference between the numerical expressions by the maximum value of the difference between the numerical expressions is obtained, and a value obtained by subtracting the obtained value from 1 is set as the similarity between the numerical expressions. A score value (score = Σsimilarity between two numerical expressions) determined based on the similarity between the numerical expressions defined in this way and information indicating whether or not the extracted word is attached to each numerical expression. Based on degree × f (first numerical expression, second numerical expression), for example, a unit expression accompanied by a pattern candidate at the highest score value and a unit expression not accompanied by the pattern candidate. The main expression.

  However, in the formula indicating the score value, Σ is a process of adding every two combinations of numerical expressions that appear in the related article DB 14. f (first numerical expression, second numerical expression) is a case where a pattern candidate appears (attached) in both the first numerical expression and the second numerical expression, or a pattern candidate appears. The function is 1 when there is no pattern, and is -1 when a candidate for the pattern appears only in one of them.

  According to one embodiment of the present invention, the similarity between the numerical expressions may be defined as a value obtained by dividing the numerical expression having a larger value by the numerical expression having a smaller value.

  In addition, according to an embodiment of the present invention, the main expression extraction unit 11 further includes a word in an article group of the related article DB 14 from the related article DB 14 by using the following clustering method. Clusters may be extracted, and each extracted cluster may be a kind of unique expression as a main expression.

Hereinafter, an example of a clustering method will be described.
(Method by hierarchical clustering)
Among the members of a cluster, members who are closest to each other are joined together to form a cluster. Then, the clusters having the shortest distance are combined. There are various definitions of the distance between clusters. For example, the distance between cluster A and cluster B may be the smallest of the distances between members of cluster A (ie, words belonging to cluster A) and members of cluster B (ie, words belonging to cluster B). Good. Here, the distance between a certain member and another member is the distance between the position vector of a certain member and the position vector of another member. The position vector is a vector indicating the position of the member in the vector space. For example, the distance between cluster A and cluster B may be the largest of the distances between members of cluster A and cluster B. Further, for example, the distance between the cluster A and the cluster B may be an average of the distances between all the members of the cluster A and the members of the cluster B. Also, the average of the positions of all the members of cluster A is the position of cluster A, the average of the positions of all the members of cluster B is the position of cluster B, and the distance between the position of cluster A and the position of cluster B is It may be the distance between cluster A and cluster B.
(Method by Ward method)
The following W is defined.
W = ΣΣ (x (i, j) −ave_x (i)) ^ 2
^ Means exponent. For example, the first Σ in the above equation means the addition from i = 1 to i = g, and the second Σ means the addition from j = 1 to j = ni. Further, x (i, j) means the position of the j-th member of the i-th cluster, and ave_x (i) means the average of the positions of all the members of the i-th cluster. When the clusters are joined together, the value of W increases, but in the Ward method, the clusters are joined together so that the value of W does not become as large as possible.

(Ending condition for clustering)
The number of clusters may be determined in advance, and when the number of clusters reaches the predetermined number, it is possible to stop joining the clusters. Alternatively, a threshold value for the distance may be determined in advance, and the clusters that are separated by the threshold value or more may be stopped.

(Position of each member)
The position of each member (word) is obtained by using various information related to each member (for example, attribute information of each member). The attribute information of each member includes, for example, the type of characters (for example, whether there are hiragana, katakana, kanji, and others), the length of the word, the meaning of the word, etc. Use.

  According to one embodiment of the present invention, for example, the main expression extraction unit 11 uses a known keyword extraction technique from an article included in an article group in the related article DB 14 or the title of the article and the head sentence of the article. To extract words. Then, the position of each word (member) is expressed by a vector (position vector). The value of the element of the position vector indicating the position of the member includes, for example, the appearance frequency of each word, the Okapi expression of the word (for example, the value indicated by the above-described expression (1)), the tfidf of the word (the expression described above) (The value obtained by multiplying the value of (1) by logN / DF). For example, the position vector dimension is increased by the number of unit expressions and time expressions, and the numerical value appearing in the article adjacent to the unit expression and time expression in the article is used as the element value of the member position vector. Also good.

  The main expression extraction unit 11 expresses the position of a word (member) in a plurality of articles as a position vector, and sets the distance between articles as the smallest of the distances between members of each article, and the distance is the longest. You may create a cluster by combining nearby articles.

Next, a method of top-down clustering (non-hierarchical clustering) will be described.
(Clustering with maximum distance algorithm)
Find a member and the member farthest away from the member, and make these members the center of each cluster. Next, the minimum value of the distance between the center of each cluster and each member is set as the distance of each member, and the member having the largest distance is set as the center of the new cluster. The process for obtaining the center of the cluster is repeated. For example, when the number of clusters reaches a predetermined number, the process for obtaining the center of the cluster is not repeated. Further, for example, when the distance between the clusters is equal to or less than a predetermined number, the process of obtaining the center of the cluster is stopped.

In addition, the goodness of the cluster is evaluated based on, for example, an AIC information amount standard, and the repetition of the process of obtaining the center of the cluster is stopped based on the comparison result between the value obtained by the evaluation and a predetermined threshold value. Good. According to the clustering by the above maximum distance algorithm, each member becomes a member of a cluster having a cluster center closest to each member.
(K-average method)
For example, clustering is performed to a predetermined number (k) by the following k-average method. First, k members are selected at random, and the selected k members are set as the center of the cluster. Each member is a member of a cluster having a cluster center closest to each member.

  Next, the member closest to the average position of each member in the cluster is set as the center of each cluster. Each member is a member of a cluster having a cluster center closest to each member. The member closest to the average position of each member in the cluster is set as the center of each cluster. The process for obtaining the center of the cluster is repeated, and when the center of the cluster stops moving, the process for obtaining the center of the cluster is stopped. According to an embodiment of the present invention, the process for obtaining the center of the cluster may be repeated for a predetermined number of times. Then, the cluster having the final cluster center is determined. Each member is a member of a cluster having the closest cluster center. Cluster members by the above method. The clustering method used in the present invention is not limited to the method described above.

  The information extraction apparatus 1 according to the present invention may perform clustering using various methods other than the above-described clustering method. For example, correspondence information between a word and a cluster to which the word belongs (for example, a sentence of an article including the word) is stored in advance in a storage unit (not shown) in the information extraction apparatus 1, and main expression extraction is performed. The unit 11 selects a specific word from the group of articles in the related article DB 14, and the selected word and a cluster to which the selected word belongs in the storage unit (for example, a sentence of an article including the word). Based on the correspondence information, the cluster to which the selected word belongs is determined, and the determined cluster may be the type of the unique expression as the main expression.

  Further, according to an embodiment of the present invention, the main expression extraction unit 11 inputs an item expression in accordance with a user's designation input, and displays a unit expression that co-occurs with the input item expression in the related article DB 14. May be extracted from the article group or bibliographic data previously stored in the storage means, and the item expression and the extracted unit expression may be used as the main expression. Then, the information pair extraction unit 12 extracts a pair of the numerical expression related to the unit expression and the item expression as an information pair from the articles constituting the article group based on the main expression, and the display unit 13 When the information pair extracted based on the main expression is displayed in a graph, the item expression may be displayed in a graph in association with a unit expression that appears in co-occurrence with the item expression.

  Further, according to an embodiment of the present invention, the main expression extraction unit 11 inputs an item expression in accordance with a user's designation input, and displays a unit expression that co-occurs with the input item expression in the related article DB 14. May be extracted from the article group or bibliographic data previously stored in the storage means, and the item expression and the extracted unit expression may be used as the main expression. Then, the information pair extraction unit 12 may extract a pair of the numerical expression related to the unit expression and the item expression as an information pair from the articles constituting the article group based on the main expression. .

  According to one embodiment of the present invention, the main expression extraction unit 11 inputs an item expression in accordance with a user's designation input, and the unique expression that appears along with the input item expression is included in the related article DB 14. The above-mentioned item expression and the kind of specific expression to which the extracted specific expression belongs may be used as the main expression. Then, the information pair extraction unit 12 may extract a pair of the unique expression and the item expression as an information pair from the articles constituting the article group based on the main expression.

  Here, a method of extracting a word A that appears co-occurring with the expression B in general (co-occurrence word extraction method) will be described. If the co-occurrence word extraction method is used, for example, the unit expression “person” can be obtained from the item expression “number of spectators mobilized”. Conversely, the item expression “number of spectators” can be obtained from the unit expression “people”. Further, if the co-occurrence word extraction method is used, for example, the unique expression “Kousuke Kitajima” can be obtained from the item expression “player”. Conversely, the item expression “player” or the like can be obtained from the unique expression “Kosuke Kitajima”.

  For example, when the unit expression “person” is obtained from the item expression “number of spectators”, candidates for the unit expression are extracted and the following calculation is performed by setting each as A. When the item expression “number of spectators” is obtained from the unit expression “people”, the item expression candidates are extracted, and the following calculation is performed with each of them as A. Also, for example, when the specific expression “Kousuke Kitajima” is obtained from the item expression “player”, the specific expression candidates are extracted, and the following calculation is performed with each set as A. When the item expression “player” or the like is obtained from the unique expression “Kousuke Kojima”, candidate item expressions are taken out and A is used for each of the following expressions.

The appearance rate of A in C and the appearance rate of A in B are obtained. here,
Appearance rate of A in C = Number of appearances of A in C / Total number of words in C. Appearance rate of A in B = Number of appearances of A in B / Total number of words in B. Then, the appearance rate of A in B / the appearance rate of A in C is obtained, and the larger this value, the word A becomes a word that appears co-occurring with a bias toward expression B.

  The appearance rate of A in B means the appearance rate of A when B and A co-occur, and the appearance rate of A in C is the bibliographic data stored in the storage means in advance. It means the appearance rate or the number of appearances of A.

  According to another embodiment of the present invention, the appearance rate of A in B means the appearance rate of A when B and A co-occur in the article group in the related article DB 14, The appearance rate of A in C may mean the appearance rate or the number of appearances of A in the article group in the related article DB 14.

  According to an embodiment of the present invention, for example, the item expression “player” may be obtained from a plurality of player names. For example, for each player name, an expression F that often appears biased and frequently appears together with the degree of bias, and for all players, the sum of the degree of bias for each expression F, or a product (multiply zero for zero). For example, the expression F having the largest score may be used as the item expression. As the degree of bias, for example, a value used when calculating a co-occurrence is used.

As a method for extracting the word A that appears co-occurring with the expression B, a method using a significant difference test may be used as follows.
(In case of binomial test)
The number of occurrences of A at C is N, the number of occurrences of A at B is N1, and N2 = N−N1. Probability that when A appears in C, the probability that it appears in B is 0.5, and out of N total occurrences, N appears less than N2 times and A appears in C and does not appear in B Ask for.

This probability is, P1 = ΣC (N1 + N2 , x) * 0.5 (x) * 0.5 (N1 + N2-x)
It is. However, in the above equation, Σ is the sum of x = 0 to x = N2, and C (N1 + N2, x) represents the number when x pieces are extracted from N1 + N2 different pieces.

  If the probability value expressed by the above equation is sufficiently small, it can be determined that N1 and N2 are not equivalent probabilities, that is, N1 is significantly larger than N2. If it is 5% test, P1 is smaller than 5%, and if it is 10% test, P1 is smaller than 10%.

For example, words in which N1 is determined to be significantly larger than N2 are words that appear well co-occurring with the expression B. In addition, a word having a smaller P1 is more likely to be biased toward the expression B and appear co-occurring.
(Chi-square test)
The number of occurrences of A in B is N1, the total number of occurrences of words in B is in F1, C and not in B, the number of occurrences of A is in N2, and the total number of occurrences in C is not in B Let it be F2. It is assumed that R1 = F1 / N1 and R2 = F2 / N2.

Here, N = N1 + N2
Chi-square value = (N * (F1 * (N2−F2) − (N1−F1) * F2) ² ) / ((F1 + F2) * (N− (F1 + F2)) * N1 * N2)
Ask for.

  Then, it can be said that the larger the obtained chi-square value is, the more significant difference between R1 and R2 is. If the chi-square value is larger than 3.84, it can be said that there is a significant difference of 5% of the risk rate, and the chi-square value is When it is larger than 6.63, it can be said that there is a significant difference of 1% of the risk rate.

For example, a word having N1> N2 and a larger chi-square value is more likely to be biased toward the expression B and appear co-occurring.
(Ratio test (ratio difference test))
p = (F1 + F2) / (N1 + N2), p1 = R1, p2 = R2,
Z = | p1-p2 | / sqrt (p * (1-p) * (1 / N1 + 1 / N2))
Ask for. sqrt means the root.

  And as Z is larger, it can be said that R1 and R2 are significantly different. When Z is larger than 1.96, it can be said that there is a significant difference of 5% of the risk rate. When Z is larger than 2.58, It can be said that there is a significant difference in the risk rate of 1%.

  For example, a word having N1> N2 and a larger Z is more likely to be biased toward the expression B and appear co-occurring.

The above-described three test methods may be combined with the above-described method of simply determining and determining the appearance rate of A in B / the appearance rate of A in C. For example, among those having a significant difference of 5% or more in risk rate, the larger the value of the appearance rate of A in B / the appearance rate of A in C, the more likely to appear in co-occurrence with a bias toward expression B .
(Information pair extraction unit 12)
The information pair extraction unit 12 specifies, for example, a part where the expressions (for example, item expression and specific expression type) extracted by the main expression extraction unit 11 appear simultaneously in the article group stored in the related article DB 14, For example, a pair of an item expression and a specific expression described in the place is an information pair. The specific expression included in the information pair is a specific expression belonging to the type of specific expression. The specific expression is extracted using the specific expression extraction technique described above.

  In addition, for example, the information pair extraction unit 12 identifies a place where the expressions extracted by the main expression extraction unit 11 (for example, item expression and specific expression type and unit expression) appear simultaneously, for example, A pair of an item expression, a unique expression, and a numerical expression related to the unit expression described above is an information pair. The numerical expression related to the unit expression is, for example, an expression obtained by combining the numerical value appearing in the article adjacent to the unit expression and the unit expression.

  In the embodiment of the present invention, for example, a special symbol indicating a punctuation mark, a line feed, or a document break is defined as a break, and a place where the main expression appears at the same time is defined as a place where it appears simultaneously.

  According to an embodiment of the present invention, the information pair extraction unit 12 further selects a specific main expression from the main expressions extracted by the main expression extraction unit 11, and sets the selected specific main expression as the selected specific main expression. Based on this, information pairs may be extracted from the article group.

  According to one embodiment of the present invention, the information pair extraction unit 12 uses the related article DB 14 (using the above-described specific expression extraction technique) based on the specific expression type (for example, LOCATION) specified in advance by the user. The unique expression belonging to the type of the specific expression is extracted from the item, and the item described in the location where the item expression extracted by the main expression extracting unit 11 and the extracted specific expression appear at the same time in the related article DB 14 A pair of an expression and a specific expression may be extracted as an information pair.

  According to an embodiment of the present invention, the information pair extraction unit 12 may further extract the information pair using a machine learning technique.

  Moreover, according to one Embodiment of this invention, the information pair extraction part 12 is based on the several types of information pair which the information pair extraction part 12 extracted based on the some main expression extracted in the main expression extraction part 11. Based on a predetermined evaluation value for each main expression, a main information pair may be selected (for example, an information pair having the largest evaluation value is selected).

  FIG. 2 is a diagram illustrating a configuration example of the information pair extraction unit 12 when a configuration for extracting information pairs using a machine learning technique is employed in the embodiment of the present invention. The information pair extraction unit 12 includes a teacher data storage unit 121, a solution-feature pair extraction unit 122, a machine learning unit 123, a learning result storage unit 124, an expression pair extraction unit 125, a feature extraction unit 126, a solution estimation unit 127, an information pair Extraction means 128 is provided.

  The teacher data storage unit 121 stores text data serving as teacher data used in the machine learning process. For example, the item expression is ai (i = 1, 2, 3,...), The specific expression belonging to the type of specific expression is bi (i = 1, 2, 3,...), And the numerical value related to the unit expression. Assuming that the expression is ci (i = 1, 2, 3,...), A pair of ai, bi, and ci (expression pair) appearing in the text data sentence is extracted as a problem and information pair as teacher data. A case where the information on whether or not the expression pair is to be made is a solution is stored. Specifically, for every ai, bi, ci pair appearing in text data, either an expression pair to be extracted as an information pair (positive example) or an expression pair that should not be extracted (negative example) A tag indicating the solution is manually attached.

That is, in the embodiment of the present invention, for example,
(A1, b1, c1)-solution "positive example"
(A1, b2, c1)-solution "negative example"
・
・
・
(A2, b2, c2)-solution "negative example"
A pair of expression pair and solution is generated.

  The solution-feature pair extraction unit 122 extracts a set of a solution and a set of features from an example of text data stored in the teacher data storage unit 121. The feature is information used in the machine learning process. The solution-feature pair extraction unit 122 uses, as a feature, for example, the distance (character or word) between ai and bi, bi and ci, and ai and ci for each expression pair to which a solution is given in certain text data. A range including the expression pairs of ai, bi, and ci in text data, character strings before and after each of ai, bi, and ci, words, parts of speech information, and the like. In addition, the answer-feature pair extraction unit 122 may include information such as whether ai, bi, and ci are included in the title of text data, and parts of speech that appear between ai and bi, bi and ci, and ai and ci. Information or the like may be used as a feature. In the embodiment of the present invention, the position information of ai, bi, and ci in an article may be a feature. For example, in articles such as newspapers, the first main expression that appears first is often important.

  The machine learning means 123 uses a supervised machine learning method to determine what kind of solution is likely to be generated from a set of the solution extracted by the solution-feature pair extraction means 122 and the feature set. learn. The learning result is stored in the learning result storage unit 124.

  The expression pair extraction unit 125 uses a main expression (for example, item expression, type of specific expression, unit expression) extracted by the main expression extraction unit 11 to use ai (item expression) included in each article in the related article DB 14. ), Bi (specific expression belonging to the type of specific expression), and ci (numerical expression related to unit expression), all combinations (expression pairs) of three types of expressions are extracted. A combination of a numerical value appearing in an article connected to the unit expression and the unit expression is a numerical expression.

  The feature extraction unit 126 extracts a feature for each expression pair extracted by the expression pair extraction unit 125 by the same processing as the solution-feature pair extraction unit 122.

  The solution estimation unit 127 refers to the learning result of the learning result storage unit 124, and estimates the degree of the solution (classification destination) that is likely to be obtained in the case of a set of features for each expression pair.

  Based on the estimation result of the solution estimation unit 127, the information pair extraction unit 128 extracts an information pair that is estimated to have a high degree of expression pair (positive example) to be extracted as an information pair.

  Here, a machine learning method by the machine learning means 123 will be described. The machine learning method prepares many sets of problem-solution pairs, learns them, learns what kind of solution the problem becomes, and uses the learning result to create a new problem. This is a method that makes it possible to guess the solution (for example, see the following references (5) to (7)).

Reference (5): Masaki Murata, Language Processing Based on Machine Learning, Faculty of Science and Engineering, Ryukoku University. Invited lecture. 2004.http: //www2.nict.go.jp/jt/a132/members/murata/ps/rk1-siryou.pdf
Reference (6): Japanese-English translation of tense aspect modality using support vector machine, Maki Murata, Mao, Kiyotaka Uchimoto, Hitoshi Isahara, IEICE Language Understanding and Communication Study Group NLC2000-78, 2001 Year.
Reference (7): CRL's efforts in the SENSEVAL2J dictionary task, Masaki Murata, Masao Uchiyama, Kiyotaka Uchimoto, Ma Aoi, Hitoshi Isahara, IEICE Language Understanding and Communication Study Group NLC2001-40, 2001.
In order to convey the problem situation to the machine, what kind of problem is required, features (elements constituting the problem with information used for analysis) are required. The problem is expressed by the feature. For example, in the problem of estimating the tense of Japanese sentence ending expressions, the problem: “He speaks.” --- If the solution “present” is given, an example of a feature is “He speaks.” . "" Speaking. "" Su "". "

  In other words, the machine learning method prepares many sets of feature set-solution pairs, performs learning, learns what kind of solution the feature set becomes, and uses the learning result. This is a method of extracting a set of features from a new problem and inferring a solution in the case of the feature.

  The machine learning unit 123 uses a technique such as a k-nearest neighbor method, a simple Bayes method, a decision list method, a maximum entropy method, or a support vector machine method as a machine learning method.

  The k-nearest neighbor method is a method for obtaining a classification destination (solution) by using the k most similar cases instead of the most similar case, and by majority decision of the k cases. k is a predetermined integer number, and generally an odd number between 1 and 9 is used.

  The Simple Bayes method is a method of estimating the probability of each classification based on Bayes' theorem and determining the classification having the highest probability value as a classification destination.

  In the simple Bayes method, the probability of outputting the classification a in the context b is given by the following equation (4).

Here, the context b is a set of features f _j (εF, 1 ≦ j ≦ k) set in advance. p (b) is the appearance probability of the context b. Here, since it is independent of the classification a and is a constant, it is not calculated. P (a) (where P is a tilde at the top of p) and P (f _i | a) are the probabilities estimated from the teacher data, respectively, and the appearance probability of class a, and the feature f for class a means the probability of having _i . When a value obtained by performing maximum likelihood estimation as P (f _i | a) is used, the value often becomes zero, and it may be difficult to determine the classification destination because the value of Equation (5) is zero. Therefore, smoothing is performed. Here, what smoothed using the following formula | equation (6) is used.

Here, freq (f _i , a) means the number of cases having the feature f _i and the classification a, and freq (a) means the number of cases having the classification a.

  The decision list method uses features and combinations of classification destinations as rules, stores them in the list in a predetermined priority order, and when input to be detected is given, from the highest priority in the list This is a method in which input data is compared with the feature of the rule, and the classification destination of the rule having the same feature is set as the classification destination of the input.

In the decision list method, the probability value of each classification is obtained using only one of the features f _j (εF, 1 ≦ j ≦ k) set in advance as a context. The probability of outputting classification a in a context b is given by

p (a | b) = p (a | fmax) Equation (7)
However, fmax is given by the following equation.

P (a _i | f _j ) (where P is a tilde at the top of p) is the rate of appearance of the classification a _i when the feature f _j is in the context.

In the maximum entropy method, when a set of preset features f _j (1 ≦ j ≦ k) is F, an expression (entropy) that satisfies a predetermined conditional expression (equation (9)) below ( In this method, the probability distribution p (a, b) when 10) is maximized is obtained, and the classification having the largest probability value is obtained from the probabilities of the respective classifications obtained according to the probability distribution.

However, A and B mean a set of classifications and contexts, and g _j (a, b) is a function that is 1 if the context b has a feature f _j and the classification is a, and is 0 otherwise. means. Further, P (a _i | f _j ) (where P is a tilde at the top of p) means the rate of appearance of (a, b) in the known data.

  Formula (9) is to obtain the expected value of the frequency of the output and feature pair by multiplying the probability p and the function g meaning the appearance of the pair of output and feature. And the expected value calculated based on the probability distribution calculated on the left side is the constraint, entropy maximization (smoothing of the probability distribution) is performed to determine the probability distribution of the output and the context. Details of the maximum entropy method are described in the following references (8) and (9).

Reference (8): Eric Sven Ristad, Maximum Entropy Modeling for Natural Language, (ACL / EACL Tutorial Program, Madrid, 1997
(9): Eric Sven Ristad, Maximum Entropy Modeling Toolkit, Release 1.6beta, (http://www.mnemonic.com/software/memt,1998))
The support vector machine method is a method of classifying data composed of two classifications by dividing a space by a hyperplane.

  FIG. 3 shows the concept of margin maximization in the support vector machine method. In FIG. 3, a white circle means a positive example, a black circle means a negative example, a solid line means a hyperplane that divides the space, and a broken line means a surface that represents the boundary of the margin area. 3A is a conceptual diagram when the interval between the positive example and the negative example is small (small margin), and FIG. 3B is a conceptual diagram when the interval between the positive example and the negative example is wide (large margin). is there.

  At this time, if the two classifications consist of positive and negative examples, the larger the interval (margin) between the positive and negative examples in the learning data, the less likely it is to make an incorrect classification with open data. As shown in FIG. 3B, a hyperplane that maximizes this margin is obtained, and classification is performed using it.

  Basically, it is as described above. Usually, an extension of the method that the training data may contain a small number of cases in the inner area of the margin, or an extension that makes the linear part of the hyperplane nonlinear ( The one with the introduction of the kernel function is used.

  This extended method is equivalent to classification using the following discriminant function, and the two classes can be discriminated depending on whether the output value of the discriminant function is positive or negative.

Where x is the context (set of features) to be identified, and x _i and y _j (i = 1,..., L, y _j ∈ {1, -1}) mean the context and classification destination of the learning data. And the function sgn is
sgn (x) = 1 (x ≧ 0)
-1 (otherwise)
Each α _i is for maximizing the expression (12) under the constraints of the expressions (13) and (14).

  The function K is called a kernel function, and various functions are used. In this embodiment, the following polynomial is used.

K (x, y) = (x · y + 1) d Equation (15)
C and d are constants set experimentally. For example, C was fixed at 1 throughout all treatments. Moreover, two types of 1 and 2 are tried for d. Here, x _i where α _i > 0 is called a support vector, and the portion taking the sum of Expression (11) is usually calculated using only this case. That is, only actual cases called support vectors are used for actual analysis.

  The details of the extended support vector machine method are described in the following references (10) and (11).

Reference (10): Nello Cristianini and John Shawe-Taylor, An Introduction to Support Vector Machines and other kernel-based learning methods, (Cambridge University Press, 2000)
Reference (11): Taku Kudoh, Tinysvm: Support Vector machines, (http://cl.aist-nara.ac.jp/taku-ku//software/Tiny SVM / index.html, 2000)
The support vector machine method handles data with two classifications. Therefore, when handling cases with three or more classifications, a pair-wise method or a one-VS rest method is usually used in combination with this.

  In the pairwise method, in the case of data having n classifications, every pair (n (n-1) / 2) of two different classification destinations is generated, and a binary classifier indicates which is better for each pair. That is, it is obtained by the support vector machine method processing module and finally obtains the classification destination by majority decision of the classification destination by n (n−1) / 2 binary classification.

  For example, when there are three classification destinations a, b, and c, the one VS rest method generates three sets of classification destination a and other, classification destination b and other, classification destination c and other, The learning process is performed on the set of the support vector machine method. Then, in the estimation process based on the learning result, the learning results of the three sets of support vector machines are used. See how the three support vector machines are estimated as candidates to be estimated. Of the three support vector machines, it is the non-other classification target and the separation plane of the support vector machine. This is a method for obtaining a classification destination of a thing farthest from the object. For example, when a candidate is farthest from the separation plane in the support vector machine created by the learning process of “classification destination a and other”, the candidate classification destination is estimated as a.

  The method of determining the level of the solution (classification destination) that is likely to be the solution (classification destination) for each expression pair estimated by the solution estimation unit 127 differs depending on various methods used by the machine learning unit 123 as a machine learning method.

  For example, in the embodiment of the present invention, when the machine learning means 123 uses the k-nearest neighbor method as a machine learning technique, the machine learning means 123 sets the feature data extracted from the cases among the cases of the teacher data. The similarity between cases based on the ratio of overlapping features (the number of the same features) is defined, and the defined similarity and the case are stored in the learning result storage means 124 as learning result information. Keep it.

  Then, when a new expression pair (candidate) is extracted by the expression pair extraction unit 125, the solution estimation unit 127 refers to the similarity and the case defined in the learning result storage unit 124, and the expression pair extraction unit 127 For the candidates of the expression pairs extracted by 125, k cases are selected from the cases in the learning result storage means 124 in descending order of the similarity of the candidates, and the classification destinations determined by the majority vote in the selected k cases are selected. Estimated as the classification target (solution) of the expression pair candidate. That is, in the solution estimation means 127, the degree of what kind of solution (classification destination) is likely to be obtained for each expression pair is the number of votes of the majority vote in the selected k cases, here “to be extracted”. The number of votes obtained by classification.

  When the simple Bayes method is used as the machine learning method, the machine learning unit 123 stores, in the learning result storage unit 124, a set of a solution of the case and a set of features as learning result information for the example of the teacher data. Remember. Then, the solution estimation means 127, when a new expression pair (candidate) is extracted by the expression pair extraction means 125, based on the combination of the learning result information solution and the feature set in the learning result storage means 124. Based on Bayes' theorem, the probability of becoming each classification in the case of the feature pair of the expression pair candidates acquired by the feature extraction means 126 is calculated, and the classification having the largest probability value is selected as the candidate of the expression pair. It is estimated as the classification (solution) of the features of. That is, in the solution estimation means 127, the degree of the likelihood of becoming a certain solution in the case of a set of features of expression pair candidates is set as the probability of becoming each classification, here, the probability of becoming the classification “to be extracted”.

  When the decision list method is used as the machine learning method, the machine learning unit 123 stores, in the learning result storage unit 124, a list in which rules of features and classification destinations are arranged in a predetermined priority order with respect to examples of teacher data. Remember. Then, when a new expression pair (candidate) is extracted by the expression pair extraction unit 125, the solution estimation unit 127 selects candidate expression pairs extracted in descending order of priority in the list of the learning result storage unit 124. The feature is compared with the feature of the rule, and the classification destination of the rule having the identical feature is estimated as the candidate classification destination (solution). That is, the solution estimation means 127 assigns the degree of the likelihood of becoming a solution in the case of a set of candidate features of the expression pair to a predetermined priority or a numerical value or scale corresponding thereto, in this case, “to be extracted”. Priority in the list of probabilities.

  When the maximum entropy method is used as the machine learning method, the machine learning means 123 specifies a class that can be a solution from the example of the teacher data, and maximizes an expression that satisfies a predetermined conditional expression and shows entropy. A probability distribution consisting of a set of features and a class that can be a solution is obtained and stored in the learning result storage means 124. When a new expression pair (candidate) is extracted by the expression pair extraction unit 125, the solution estimation unit 127 uses the probability distribution of the learning result storage unit 124 to identify the features of the extracted expression pair candidate. The probability of the classification that can be the solution for the set of is determined, the classification that can be the solution having the largest probability value is identified, and the identified classification is estimated as the candidate solution. That is, in the solution estimation means 127, the degree of the likelihood of becoming a certain solution in the case of a set of features of expression pair candidates is set as the probability of becoming each classification, here, the probability of becoming the classification “to be extracted”.

  When the support vector machine method is used as the machine learning method, the machine learning unit 123 specifies a class that can be a solution from the example of the teacher data, divides the class into a positive example and a negative example, A hyperplane that maximizes the interval between the positive and negative examples of a case and divides the positive and negative examples by a hyperplane in a space whose dimension is a set of case features according to a predetermined execution function using Is stored in the learning result storage means 124. Then, when a new expression pair (candidate) is extracted by the expression pair extraction unit 125, the solution estimation unit 127 uses the hyperplane of the learning result storage unit 124 to identify the feature of the extracted expression pair candidate. Whether the set is on the positive example side or the negative example side in the space divided by the hyperplane is specified, and the classification determined based on the specified result is estimated as the candidate solution. That is, in the solution estimation means 127, the degree of the likelihood of being a solution in the case of a set of candidate expression pairs is the distance from the separation plane to the space of the positive example (expression pair to be extracted). . More specifically, when the expression pair to be extracted is a positive example and the expression pair that should not be extracted is a negative example, the case located in the space on the positive example side with respect to the separation plane is referred to as “example to be extracted”. The distance from the separation plane of the case is determined as the degree of the case. In the above description, an example in which the information pair extraction unit 12 extracts an information pair by the machine learning method using the item expression as the main expression, the type of the unique expression, and the unit expression has been described. Using the machine learning technique, the information pair extraction unit 12 may extract the information pair using the item expression as the main expression and the kind of specific expression.

  According to an embodiment of the present invention, the information pair extraction unit 12 further selects an information pair that is finally targeted for extraction from the extracted information pairs, and an information pair that is not finally targeted for extraction. Information that is determined as a negative example, machine-learned about the extracted information pair using the above-described machine learning technique, using the determined positive example and negative example as teacher data, and finally extracted. You may make it determine a pair.

For example, after extracting 100 information pairs, the information pair extracting unit 12 finally extracts 3 of 6 information pairs included in the 100 information pairs in accordance with a user's designated input. The target information pair (positive example) is determined, and the remaining three are finally determined as information pairs not to be extracted (negative example). Then, the above-described machine learning is performed using the six information pairs determined as positive examples or negative examples as teacher data, and the remaining 94 information pairs are converted into positive examples or negative examples using the learning results. Classify. Then, the three information pairs determined as positive examples according to the user's designated input and the information pairs classified as positive examples are finally determined as information pairs to be extracted.
(Display unit 13)
The display unit 13 organizes the information pairs extracted by the information pair extraction unit 12 and displays them in a graph, for example.

  According to an embodiment of the present invention, for example, the display unit 13 may select each main information from a plurality of types of information pairs extracted by the information pair extraction unit 12 based on a plurality of main expressions extracted by the main expression extraction unit 11. A configuration may be adopted in which a main information pair is selected (for example, an information pair having the largest evaluation value is selected) based on a predetermined evaluation value for expression, and then the selected main information pair is graphed. .

As the evaluation value calculation method, for example, any one of the following four evaluation value calculation formulas is used. Here, a case where the main expressions extracted by the main expression extraction unit 11 are one item expression, two types of unique expressions, and one unit expression will be described as an example.
(Method 1): Frequency of numerical expression and score of main expression are used.

Evaluation value M = Freq × S1 × S2 × S2 ′ × S3
(Method 2): The numerical expression frequency and the main expression score are used.

Evaluation value M = Freq × (S1 × S2 × S2 ′ × S3)
(Method 3): The frequency of numerical expression is used.

Evaluation value M = Freq
(Method 4): The score of the main expression is used.

Evaluation value M = S1 × S2 × S2 ′ × S3
Here, Freq is the number of numerical expressions extracted by the information pair extraction unit 12 based on the main expression, and S1 is a score as shown in the above-described expressions (1) to (3) for the item expression. The values S2, S2 ′ are the score values as shown in the above-described formulas (1) to (3) for each of the two types of proper expressions, and S3 is the above-described formula (1) for the unit representation. It is a score value as shown in Formula (3).

  According to one embodiment of the present invention, for example, the main expression extraction unit 11 selects a predetermined number of item expressions, types of specific expressions, and unit expressions from the above-described high score values. Then, the display unit 13 selects, for example, one item expression, two types of specific expressions, and one unit expression from the selected item expressions, types of specific expressions, and unit expressions, and all of them. As the evaluation value M obtained by calculating the evaluation value M with respect to the combination of is larger, it is determined that the graph is more useful, and among the information pairs extracted by the information pair extraction unit 12, for example, the evaluation value M An information pair extracted based on one item expression, one unique expression type, and one unit expression having the largest is displayed in a graph.

  According to one embodiment of the present invention, the information extracting apparatus 1 further includes means (not shown in FIG. 1) for performing a correlation analysis on the information pair extracted by the information pair extracting unit 12. Also good. Further, the display unit 13 may perform the correlation analysis and display the correlation analysis result.

  FIG. 4 is a diagram showing an example of an information extraction processing flow in the embodiment of the present invention. First, the information extraction apparatus 1 extracts a main expression from an article group in the related article DB 14 (step S1). Next, the information extraction device 1 extracts an information pair using the extracted main expression (step S2). Then, the information extraction device 1 displays the extracted information pair (step S3).

  5 to 14 are diagrams showing examples of display by the display unit. The display example shown in FIG. 5 is an information pair in the case where the item expression as the main expression is “end price”, the types of specific expressions are “LOCATION” and “ORGANIZATION”, and the unit expressions are “kilo” and “yen”. A display example is shown. The item “sentence” in the table shown in FIG. 5 indicates a sentence in which each information pair appears simultaneously.

  The display example shown in FIG. 6 shows a display example of an information pair when the item expression as the main expression is “ballistic missile” and the types of specific expressions are “ARTIFACT” and “LOCATION”. The item “sentence” in the table shown in FIG. 6 indicates a sentence in which each information pair appears simultaneously.

  In the display example shown in FIG. 7, the item expression as the main expression is “sponsored by Mainichi Shimbun”, and the types of unique expressions are “DATE”, “LOCATION”, “ORGANIZATION”, “PERSON”, and “TIME”. The example of a display of an information pair is shown. The item “sentence” in the table shown in FIG. 7 indicates a sentence in which each information pair appears simultaneously.

  The display example shown in FIG. 8 shows a display example of an information pair when the item expression as the main expression is “typhoon”, the type of specific expression is “LOCATION”, and the unit expressions are “No.” and “Kilo”. Yes. The item “sentence” in the table shown in FIG. 8 indicates a sentence in which each information pair appears simultaneously.

  The display example shown in FIG. 9 is an information pair in which the item expression as the main expression is “middle advance”, the types of specific expressions are “ORGANIZATION” and “PERSON”, and the unit expressions are “times” and “points”. A display example is shown. The item “sentence” in the table shown in FIG. 9 indicates a sentence in which each information pair appears simultaneously.

  The display example shown in FIG. 10 is an information pair in which the item expression as the main expression is “unemployed”, the types of specific expressions are “PERSON” and “TIME”, and the unit expressions are “floor” and “floor”. A display example is shown. The item “sentence” in the table shown in FIG. 10 indicates a sentence in which each information pair appears simultaneously.

  In the display example shown in FIG. 11, the item expression as the main expression is “male”, the types of specific expressions are “LOCATION”, “ORGANIZATION”, “PERSON”, and the unit expressions are “rank” and “meter”. The example of a display of an information pair is shown. The item “sentence” in the table shown in FIG. 11 indicates a sentence in which each information pair appears simultaneously.

  In the display example shown in FIG. 12, the item expression as the main expression is “bribery crime”, the types of specific expressions are “DATE”, “LOCATION”, “MONEY” and “PERSON”, and the unit expressions are “people” and “yen” ”Is a display example of information pairs. The item “sentence” in the table shown in FIG. 12 indicates a sentence in which each information pair appears simultaneously.

  The display example shown in FIG. 13 is a graph display example of an information pair when the item expression as the main expression is “typhoon”, the specific expression type is “LOCATION”, and the unit expressions are “No.” and “Kilo”. Show. FIG. 13 shows a graph display example of the information pair (No. 4-210 km-Minamidaitojima) in the first row in the display example shown in FIG. Referring to the display example shown in FIG. 13, it can be seen that Typhoon No. 4 is 210 km from Minami Daitojima.

  FIG. 14 is a graph display example of information pairs when the item expression as the main expression is “end price”, the types of specific expressions are “LOCATION” and “ORGANIZATION”, and the unit expressions are “kilo” and “yen”. Show. The display example shown in FIG. 14 indicates that the Osaka Customs Fushiki Customs Branch seized 9.3 kilos of stimulant with a terminal price of 740,000 yen from a Russian flag ship.

  A modification of the present invention will be described. In the modification of the present invention, the information pair extraction unit 12 clusters each article constituting the article group from which the information pair is extracted into a cluster to which each article belongs. Then, the display unit 13 performs correlation analysis for each cluster on the information pairs extracted from the articles belonging to each cluster, and the information pairs extracted from the articles belonging to each cluster are clustered based on the correlation analysis result. Each graph is displayed. Correlation analysis refers to, for example, analyzing the correlation between two data. For example, in the present invention, when information pairs are plotted on a two-axis graph of x-axis and y-axis, data corresponding to the x-axis (for example, numerical expression) and data corresponding to the y-axis (for example, numerical values) Correlation analysis is performed to see if there is a correlation with the expression. Further, for example, in the present invention, when information pair plots are arranged on a two-axis graph of x-axis and y-axis, data corresponding to the x-axis (for example, a specific expression belonging to the specific expression type DATE) And a correlation analysis is performed to determine whether there is a correlation between the data corresponding to the y axis (for example, a numerical expression).

  It can be said that there is a correlation if the plot of the graph is close to a straight line. For example, when the correlation analysis is performed on the plot of the graph, the display unit 13 may calculate a correlation coefficient indicating how linear the two data are.

  The display unit 13 determines whether or not there is a correlation between the numerical expression corresponding to the x axis and the numerical expression corresponding to the y axis in the graph data for each cluster generated based on the information pairs extracted from the articles belonging to each cluster. Analysis may be performed, and only graph data that is determined to be correlated as a result of the analysis or graph data having a correlation coefficient equal to or greater than a predetermined value may be displayed as a graph.

  Further, the display unit 13 has a correlation between the specific expression corresponding to the x axis and the numerical expression corresponding to the y axis in the graph data for each cluster generated based on the information pairs extracted from the articles belonging to each cluster. As a result of the analysis, only graph data determined to be correlated or graph data having a correlation coefficient equal to or greater than a predetermined value may be displayed in a graph.

  The display unit 13 displays the above-described evaluation value of the information pair corresponding to the graph data, the graph data determined to be correlated as a result of the correlation analysis, or the graph data having a correlation coefficient equal to or greater than a predetermined value. Sorting may be performed in descending order of M, and each graph data may be displayed as a graph.

  In addition, according to an embodiment of the present invention, the display unit 13 performs correlation analysis on the information pair extracted by the information pair extraction unit 12, and graphs and displays the information pair based on the result of the correlation analysis. You may make it do. For example, the display unit 13 performs the above-described correlation analysis with respect to the information pairs of the number of combinations (25) of the five types of numerical expressions and the five types of item expressions, and there is a correlation as a result of the correlation analysis. Only the information pairs that are determined to be or the information pair whose correlation coefficient is equal to or greater than a predetermined value may be displayed in a graph. In addition, for example, the display unit 13 performs the above-described correlation analysis for the number of types of information pairs (25) of the combinations of the five types of specific expressions and the five types of item expressions. Only information pairs that are determined to be present or only information pairs having a correlation coefficient equal to or greater than a predetermined value may be displayed in a graph.

When displaying the specific expression on the x-axis of the graph, the display unit 13 determines the display order of the specific expression and displays it on the graph. There are five ways of determining the display order: (1) to (5) below.
(1) The display order is determined manually in advance, and the display order is stored in the storage means as tabular data. The order is determined with reference to tabular data stored in the storage means.
(2) A specific expression and a word that co-occurs with it (for example, a word in the same sentence as the specific expression in a large corpus) are extracted, and the number of co-occurrence is obtained. To summarize.

Word 1 word 2 word 3
Proper expression 1 2 0 1
Proper expression 2 3 2 1
Proper expression 3 1 0 0
A numerical analysis such as principal component analysis or dual scaling is performed on the tabular data. (See, for example, Reference (12): “Multivariate analysis understood by illustration” (Nippon Jitsugyo Publishing Co., Ltd.), Reference (13): “Practical Workshop Excel Thorough Use Multivariate Analysis” (Hidewa System)).

The unique expressions are rearranged in the order of the values corresponding to the first eigenvalues, and this is used as the order of the unique expressions.
(3) A specific expression and a word that co-occurs with it (for example, a word in the same sentence as the specific expression in a large corpus) are extracted, and the number of co-occurrence is obtained. The results are summarized in the following table format.

Word 1 word 2 word 3
Proper expression 1 2 0 1
Proper expression 2 3 2 1
Proper expression 3 1 0 0
For each unique expression, a vector is created with the type of co-occurring words as the vector dimension and the number of co-occurring words as the vector elements. The inner product of two vectors (for example, v1 and v2) or Cos (v1, v2) is set as the similarity between the proper expressions. The specific expressions with large similarity are connected in order. However, at most one unique expression should be two unique expressions. When all the proper expressions are connected, they are stretched in a straight line, and the order of proper expressions is set as the order of proper expressions from the end.
(4) The display order of the unique expressions is set to the order of the Japanese syllabary.
(5) The unique expressions are arranged in the order of the longest character string length, and the arranged order is set as the display order of the unique expressions.

  FIGS. 15 to 24 show evaluation examples of information pairs extracted using the present invention. FIG. 15 is a diagram showing the number of evaluations of information pairs. Only the number of information pairs described in the table shown in FIG. 15 is manually evaluated. NEx (x = 1 to 8) indicates x number of unique expressions and one item expression. The number of articles X indicates an article group having X articles. That is, the number described in the cell where the column in which the number of articles X is described and the row in which NEx is described intersects with the x number of unique expressions extracted from the group of articles whose number of articles is X. It shows the number of information pairs consisting of one item expression.

  FIGS. 16A, 16B, 17A, and 17B are diagrams showing an example of evaluation results for the number of information pairs described in the table shown in FIG. FIG. 16A shows the number of information pairs when information pairs extracted using the present invention are more than 75% correct (for example, information pairs appear simultaneously in an article). ) Shows the result of dividing the number of information pairs shown in the evaluation result shown in FIG. 16A by the evaluation number of the corresponding information pair shown in FIG.

  FIG. 17A shows the number of information pairs when the information pairs extracted using the present invention are 50% or more correct (for example, information pairs appear simultaneously in an article). (B) shows the result of dividing the number of information pairs shown in the evaluation result shown in FIG. 17A by the evaluation number of the corresponding information pair shown in FIG. 18 is a graph showing the evaluation results shown in FIG. 16B, and FIG. 19 is a graph showing the evaluation results shown in FIG. 17B.

  FIG. 20 is a diagram illustrating the number of evaluations of information pairs. Only the number of information pairs described in the table shown in FIG. 20 is manually evaluated. NEx (x = 1 to 8) indicates x unique expressions, one item expression, and two numerical expressions (numerical expressions related to the unit expression). The number of articles X indicates an article group having X articles. That is, the number described in the cell where the column in which the number of articles X is described and the row in which NEx is described intersects with the x number of unique expressions extracted from the group of articles whose number of articles is X. The number of information pairs consisting of one item expression and two numerical expressions is shown.

  FIGS. 21A, 21B, 22A, and 22B are diagrams showing an example of evaluation results for the number of information pairs described in the table shown in FIG. FIG. 21A shows the number of information pairs when the information pairs extracted by using the present invention are 75% or more correct (for example, information pairs appear simultaneously in an article). ) Shows the result of dividing the number of information pairs shown in the evaluation result shown in FIG. 21A by the evaluation number of the corresponding information pair shown in FIG.

  22A shows the number of information pairs when the information pairs extracted using the present invention are 50% or more correct (for example, information pairs appear simultaneously in an article). (B) shows the result of dividing the number of information pairs shown in the evaluation result shown in FIG. 22A by the evaluation number of the corresponding information pair shown in FIG. FIG. 23 is a graph showing the evaluation results shown in FIG. 21B, and FIG. 24 is a graph showing the evaluation results shown in FIG. 22B.

  The experiment in the present invention will be further described. With the information extraction apparatus of the present invention, an information pair extraction experiment was performed when the types of item expression and specific expression were used as the main expression. Here, we used the articles (220,078 articles) of the Mainichi Shimbun for two years of 1998 and 1999. In this experiment, the entire extracted information pair was the subject of the experiment, and some of them were selected and evaluated manually. The evaluation results are shown in FIGS. 25 and 26 show the evaluation results of the information pair when only the specific expression is used as the main expression (FIG. 25 shows the accuracy, FIG. 26 shows the total number of extractions), and FIGS. FIG. 27 shows the evaluation results of information pairs when ˜6 unique expressions and one unit (numerical value) expression are used as main expressions (FIG. 27 shows accuracy, and FIG. 28 shows the total number of extractions). NEx in FIGS. 25 to 28 means a case where x number of unique expressions are used. Evaluation is based on the number of extracted articles (the number of articles with one sentence in which the main expression appears at the same time) taken at random from each of the data that was exactly 10, 30, 50, 70, 90. Was examined manually. “Evaluation A” indicates that the data is correct when the correct information is shown for one topic having 75% of the information pairs of the numerical value / specific expression extracted from the number of extracted articles, and the correct data Mean percentage. “Evaluation B” is the data that is determined to be correct when it shows correct information for one topic with 50% of the information pairs of numerical values and specific expressions extracted from several extracted articles. Means the percentage of However, if multiple equivalent expressions of the same type appear in the same sentence, any of them can be interpreted as a correct answer. In FIG. 27, only the result of evaluation B is shown. In the evaluation A, the total data was 0.084 in the case of only unique expression information, and 0.018 in the case of numerical value / specific expression information. 26 and 27 show the total number of extracted data.

  The number of data that can be extracted in evaluation A was 24 when both the numerical expression and the specific expression were used when only the specific expression was used. The accuracy of the entire evaluation B data was 0.28 when only the specific expression was used, and 0.26 when the numerical value / specific expression information was used. We also estimated how much useful data can be extracted by multiplying the total number of extractions and accuracy. For example, the product of the total number of extractions of NE2 with the number of articles of 21 to 40 and the accuracy of the number of articles 30 is obtained by a procedure to obtain useful data that can be extracted in the case of NE2 with the number of articles of 1 to 40. In this estimation, the data of evaluation A that can be extracted is about 20,000 when only the specific expression is used and when both the numerical value and the specific expression are used.

  Data (information pairs) extracted by the information extraction apparatus 1 of the present invention are shown in FIGS. FIG. 29 shows information pairs obtained when the specific expression and the item expression are the main expressions. FIG. 29A shows an information pair in the case where the item expression “slider” and the type of the unique expression of the person name and the organization name are the main expression set. From FIG. 29 (A), the player who was throwing the slider at that time and the team name can be seen. FIG. 29B shows an information pair in the case where the main expression is the item expression “ballistic missile” and the type of the unique expression of the artifact name and the place name. From FIG. 29 (B), the missile name related to the ballistic missile at that time and the country of possession of the missile are known. In addition, there are various data such as the date, place, sponsoring organization, name of the name of the person sponsored by the Mainichi Shimbun, such as Go Shogi, the data of the organization, date, place, person, amount of money, and related laws that received the house search. can get.

  FIG. 30 is a display example of information pairs obtained when the specific expression and the item expression are the main expressions. This is the case where the main expression is the item expression “bribery crime”, the unit expressions “people”, “yen”, and the unique expressions of person names and place names. The horizontal axis in FIG. 30 indicates the number of people who committed bribery, and the vertical axis indicates the amount of bribery. Each plot lists the location associated with the person's name. However, although the name of the person is taken in the system, it is displayed anonymously here. In addition, we obtained a variety of graphs showing how many floors and how many fires occurred, the names and times of the residents, the ranking of sports competitions, the number of athletes, the players, the organization, and the location.

  Another modification of the present invention will be described. In this example, the display unit 13 extracts a sentence including the information pair extracted by the information pair extraction unit 12 from the article group in the related article DB 14, and highlights the information pair in the extracted sentence.

  For example, if the information pair (number expression, specific expression, item expression pair) extracted by the information pair extraction unit 12 is “◯ No.”, “○ day”, “typhoon”, the display unit 13 A sentence in which the three expressions appear simultaneously is extracted, and the three expressions are highlighted in the extracted sentence. When there are a plurality of expressions in the same sentence, for example, the first appearing is highlighted with a double line and the others are highlighted with a single line. The result is shown in FIG. A configuration may be adopted in which the above three expressions are appropriately color-coded and displayed.

  The extracted sentence clearly shows the state of the typhoon at that time, and it seems that the sentence having the same effect as the important sentence extraction in the summary research can be extracted. In other words, the name of the place through which the typhoon passed and the damage situation are also described, and an important description about the typhoon is included in the extracted sentence.

  In addition, the seventh data in the figure includes a plurality of data of typhoon No. 7 and typhoon No. 8. In addition to the extracted information, if there is a main expression currently focused on, it is also a single underline. By highlighting with, you can immediately see that the data is in the data. In addition, the system may mistake the set of information pairs to be retrieved, but this highlighting helps to find the error early. Here, only the extracted sentence is highlighted, but there is a possibility that the sentence to be extracted remains in the article. If the same highlighting is applied to the entire article, such a leak may be extracted. Thus, a configuration may be adopted in which the entire original article is highlighted.

  Next, in one embodiment of the present invention, as described above, the main expression extraction unit 11 performs before or after the type of the extracted unique expression or before the numerical expression related to the extracted unit expression. Or a word associated with a numerical expression associated with a type of a unique expression accompanied by a word selected from the extracted word or a word selected from the extracted word, or a unit expression associated with a numerical expression accompanied by a word selected from the extracted word A description will be given of how the word is selected in the case of performing processing as the main expression.

  FIG. 32 and FIG. 33 are diagrams showing a table in which appearance words are arranged with words attached before or after the type of the extracted unique expression and words attached before or after the numerical expression related to the extracted unit expression. is there. In the example shown in FIG. 32 and FIG. 33, the result of arranging words up to 3 characters before and after the kind of numerical expression or numerical expression is arranged in order of appearance frequency.

  For example, the word attached as the last three letters of the numerical expression “9.3 kg | 9.3 kg | 120 kg |...” Related to the unit expression “kilo” in FIG. , The word that accompanies the specific expression “Tamamaru | Citeirote ...” that belongs to the type of specific expression “ARTIFACT” is “ship.” For example, the main expression extraction unit 11 is shown in FIG. A table as shown is created and displayed, and the word “ship” is selected in accordance with the designation input by the user who saw the display.

  Referring to FIG. 33, for example, the word accompanying the first two letters of the numerical expression “No. 4 | No. 5 | No. 6 | ...” related to the unit expression “kilo” is “typhoon”. The word attached as the first two letters of the numerical expression “15 km | 25 km | 75 km” related to the unit expression “kilo” is “speed”. The main expression extracting unit 11 creates and displays a table as shown in FIG. 33, for example, and selects the word “speed” according to the designation input by the user who saw the display.

  Next, an experiment in which the information extraction apparatus 1 performs a correlation analysis on the information pair extraction result by the information extraction apparatus 1 of the present invention will be described. In general, a correlation coefficient is used as an index representing the correlation of data. Given data xi, yi (i = 1, 2, 3,... N), the correlation coefficient r between x and y is

It becomes.
here,

It is.

  r always takes a value between -1 and 1, and there is a strong correlation when close to 1 (or -1), and no correlation when close to 0. When the correlation coefficient is a positive value, there is a positive correlation, and as x increases, y also increases. When the correlation coefficient is a negative value, there is a negative correlation, and when x increases, y decreases. The determination of the presence or absence of correlation is performed using a test such as a t test.

  For example, the information extraction apparatus 1 extracts information pairs related to a plurality of sets (a set of various fields such as marathon, typhoon, bribery) from large-scale data such as newspapers for two years, and correlates the extracted information pairs. The information extraction apparatus 1 performs correlation analysis on the information pair extraction result.

  For example, a correlation analysis result as shown in FIG. 34 is obtained. The correlation coefficient in the data shown in FIG. 34 is a correlation coefficient between numerical expressions related to unit expressions. A flag “1” is set for data that has been correlated in the test.

  For example, the original data of the sixth data in FIG. 34 (data in which the item expression is “between”, the unit expression is “ku” and “kilo”, and the type of specific expression is “ORGANIZATION”) FIG. 35 shows the extraction result of the information pair that is the original data).

  FIG. 36 is a graph showing a correlation analysis result for the numerical expression related to the unit expression “ku” shown in FIG. 35 and the numerical expression related to the unit expression “kilo”. The correlation coefficient between the numerical expression related to the unit expression “ku” and the numerical expression related to the unit expression “kilo” is positive (about 0.783). In addition, from the graph shown in FIG. 36, for example, it can be seen that when the numerical value of “K” increases, the numerical value of “kilometers” (for example, the distance traveled by relay stations) increases.

  According to the correlation analysis process by the information extraction device 1, a series of data having correlation can be easily extracted.

  According to an embodiment of the present invention, the information extraction apparatus 1 extracts a set of a plurality of types of main expressions from data in one field (for example, a field called typhoon), and extracts the extracted data based on correlation analysis. The main expression to be extracted may be finally determined from the set of the main expressions.

  Next, another experiment using the information extraction apparatus 1 of the present invention will be described. FIG. 37 is a diagram illustrating a score for each unique expression accompanied by a pattern candidate (word or character string) before or after. Here, the pattern candidate means a word or a character string that appears before or after the type of the unique expression in the article group.

  That is, FIG. 37 shows information pair extraction by the information extraction apparatus 1 when the item expression is “end price”, the unit expression is “kilo” and “yen”, and the types of specific expressions are “LOCATION” and “ORGANIZATION”. Based on the result, the score for each unique expression accompanied by a pattern candidate before or after the specific expression type “LOCATION (location name)” is shown. In FIG. 37, data is displayed in descending order of score.

  In FIG. 37, the item “expression” is a specific expression accompanied by a pattern candidate before or after, the item “score” is a score for each specific expression, and the item “example” is a specific example of the specific expression. Show.

  For example, the first data in FIG. 37 indicates that the score of the specific expression accompanying the word “person” as the last character is 798, and a specific example of the specific expression is Bhutan | Bhutan | Colombia | Colombia | China |.

The above score is the above-described formula score = Σsimilarity between two specific expressions × f (first specific expression, second specific expression)
Calculate based on

  The above “similarity between two unique expressions” indicates two specific expressions (first specific expression and second specific expression) obtained from the specific expressions displayed in the item “example” shown in FIG. The degree of similarity. For example, if the unique expressions are both country names, or both are other than country names, the similarity between the specific expressions is 1, and if one of the unique expressions is a country name and the other is not a country name, the similarity between the specific expressions is -1. To do.

  According to an embodiment of the present invention, the similarity between specific expressions may be converted using a predetermined conversion formula, and the score may be calculated using the converted similarity. For example, when the similarity between two specific expressions is an angle (or cos) between vectors according to the specific expression determined by a predetermined vector generation method, since the similarity is 0 to 1, for example, the cos The value obtained by doubling the value of 1 and subtracting 1 may be used as the similarity between the two unique expressions. In the present invention, any other conversion method of similarity may be used.

  In addition, f (first specific expression, second specific expression) is the case where pattern candidates appear (accompany) in both the first specific expression and the second specific expression, or in which both the pattern candidates are The function is 1 when it does not appear, and is -1 when the pattern candidate appears only in one of them.

  In FIG. 37, the data for two lines corresponding to the same “expression” includes data for a unique expression in which a pattern candidate appears in the upper line, and data for a unique expression in which the pattern candidate appears in the lower line. It is.

  For example, the specific expression “Bhutan | Bhutan | Colombia | Colombia | China | ...” displayed in the item “example” in the first data in FIG. 37 is followed by the pattern candidate “people”. The specific expression “Russia | Fushiki Port | Russia | Narita Airport |...” Displayed in the “example” item in the second data is a pattern. A specific example of a specific expression in which the candidate “person” is not attached as the last character is shown.

  By looking at the data displayed in FIG. 37, it is possible to automatically acquire patterns (for example, “people”, “nationality”, etc.) useful for separating (grouping) specific expressions extracted as information pairs. I understand that.

  The present invention can also be implemented as a program that is read and executed by a computer. The program for realizing the present invention can be stored in an appropriate recording medium such as a portable medium memory, a semiconductor memory, or a hard disk, which can be read by a computer, provided by being recorded on these recording media, or communication. It is provided by transmission / reception using a network via an interface.

It is a figure which shows an example of the system configuration | structure of this invention. It is a figure which shows the structural example of an information pair extraction part. It is a figure which shows the concept of margin maximization of a support vector machine method. It is a figure which shows an example of an information extraction process flow. It is a figure which shows the example of a display by a display part. It is a figure which shows the example of a display by a display part. It is a figure which shows the example of a display by a display part. It is a figure which shows the example of a display by a display part. It is a figure which shows the example of a display by a display part. It is a figure which shows the example of a display by a display part. It is a figure which shows the example of a display by a display part. It is a figure which shows the example of a display by a display part. It is a figure which shows the example of a display by a display part. It is a figure which shows the example of a display by a display part. It is a figure which shows the evaluation number of an information pair. It is a figure which shows an example of the evaluation result about an information pair. It is a figure which shows an example of the evaluation result about an information pair. It is a graph which shows an example of the evaluation result about an information pair. It is a graph which shows an example of the evaluation result about an information pair. It is a figure which shows the evaluation number of an information pair. It is a figure which shows an example of the evaluation result about an information pair. It is a figure which shows an example of the evaluation result about an information pair. It is a graph which shows an example of the evaluation result about an information pair. It is a graph which shows an example of the evaluation result about an information pair. It is a figure which shows an example of the evaluation result about an information pair. It is a figure which shows an example of the evaluation result about an information pair. It is a figure which shows an example of the evaluation result about an information pair. It is a figure which shows an example of the evaluation result about an information pair. It is a figure which shows an information pair. It is a figure which shows an information pair. It is a figure which shows the example of the highlight display of an information pair. It is a figure which shows the table | surface which arranged the word which accompanies before or after the type of the extracted specific expression, and the word which accompanies before or after the numerical expression relevant to the extracted unit expression in appearance frequency. It is a figure which shows the table | surface which arranged the word which accompanies before or after the type of the extracted specific expression, and the word which accompanies before or after the numerical expression relevant to the extracted unit expression in appearance frequency. It is a figure which shows a correlation analysis result. It is a figure which shows the extraction result of an information pair. It is a graph which shows a correlation analysis result. It is a figure which shows score for every specific expression with which the pattern candidate precedes or follows.

Explanation of symbols

DESCRIPTION OF SYMBOLS 1 Information extraction apparatus 11 Main expression extraction part 12 Information pair extraction part 13 Display part 14 Related article DB
111 Main Unit Expression Extraction Unit 112 Main Item Expression Extraction Unit 113 Main Specific Expression Extraction Unit 121 Teacher Data Storage Unit 122 Solution-Feature Pair Extraction Unit 123 Machine Learning Unit 124 Learning Result Storage Unit 125 Expression Pair Extraction Unit 126 Feature Extraction Unit 127 Solution Estimating means 128 Information pair extracting means

Claims (9)

An information extraction device,
The main expression extraction means for extracting from the article group the item expression that is the main expression, the item expression that is the main expression, and the specific expression that belongs to the type of the specific expression that is the main expression,
The location where the item representation and the specific representation belonging to the type of the specific representation appear simultaneously from the article group, and the pair of the item representation and the specific representation described in the identified location is information An information pair extracting means for extracting as a pair,
The main expression extraction means further uses machine learning to make use of the correct data feature to estimate proper nouns from the article group as new data, and obtains proper expressions including proper nouns obtained by the estimation. The specific expression type is extracted as a specific expression belonging to the specific expression type, and the specific expression type is determined based on whether a predetermined word is attached before or after the specific expression belonging to the specific expression type. Classifying and defining at least one of the classified proper expression types as a main proper expression type,
The information pair extraction means further identifies a location where the item representation and the specific representation belonging to the type of the main specific representation appear simultaneously from the article group, and is described in the identified location A pair of the item expression and a specific expression belonging to the type of the main specific expression is extracted as an information pair. The information extraction device according to claim 1,
The main expression extraction means further extracts a word attached before or after the specific expression extracted as belonging to the specific expression type, and the specific expression extracted as belonging to the specific expression type Based on the score value determined based on the similarity between the obtained unique expressions and information indicating whether or not the extracted word is attached to each unique expression, the unique expressions are obtained. A word having a high degree of classification of the specific expression in the case where it is attached and a case where it is not attached is determined, and the type of the specific expression to which the determined word is attached and the kind of the specific expression to which the word is not attached are respectively determined. An information extraction apparatus, wherein the information is classified into different types of main expressions, and at least one of the classified types of specific expressions is set as a type of main specific expressions. In the information extraction device according to claim 1 or 2,
The main expression extracting means further extracts a cluster to which a word in the article group belongs from the article group, and sets each extracted cluster as a type of the unique expression. . In the information extraction device according to any one of claims 1 to 3,
The information pair extracting unit further extracts the information pair by using a machine learning technique. In the information extraction device according to any one of claims 1 to 4,
The information pair extraction unit further selects an information pair from the extracted information pairs based on an evaluation value calculated based on a predetermined evaluation value calculation formula. The information extraction device according to any one of claims 1 to 5, further comprising:
An information extraction apparatus comprising: means for performing correlation analysis on the information pair extracted by the information pair extraction means. In the information extraction device according to any one of claims 1 to 6,
An information extraction apparatus comprising: display means for displaying the information pairs extracted by the information pair extraction means in a graph. An information extraction method,
The main expression extraction means that extracts the item expression and the specific expression from the article group as the main expression, the item expression that is the main expression, and the specific expression belonging to the type of the specific expression that is the main expression. The proper noun is estimated from the article group which is new data by learning using the feature of the data, and the proper expression including the proper noun obtained by this estimation is extracted as the proper expression belonging to the type of the proper expression. And classifying the types of the specific expressions based on whether or not a predetermined word is attached before or after the specific expressions belonging to the types of the specific expressions, and the types of the classified specific expressions At least one of
The location where the item representation and the specific representation belonging to the type of the specific representation appear simultaneously from the article group, and the pair of the item representation and the specific representation described in the identified location is information The information pair extraction means for extracting as a pair further identifies a location where the item representation and the specific representation belonging to the type of the main specific representation appear simultaneously from the article group, and describes the identified location An information extraction method comprising: extracting, as an information pair, a pair of the item expression and the specific expression belonging to the main specific expression type. An information extraction program,
On the computer,
The type of item expression and specific expression is the main expression, the item expression that is the main expression, and the specific expression belonging to the type of specific expression that is the main expression from the article group,
Further, the proper nouns are estimated from the article group as new data by learning using the features of correct data by machine learning, and the proper expressions including the proper nouns obtained by this estimation are converted into the types of the proper expressions. The specific expression type is extracted based on whether or not a predetermined word is attached before or after the specific expression belonging to the specific expression type. A process in which at least one of the types of specific expressions is a type of main specific expressions;
Furthermore, the part where the item expression and the specific expression belonging to the type of the main specific expression appear simultaneously from the article group, and the item expression and the main expression described in the specified part are specified. An information extraction program for executing a process of extracting a pair with a specific expression belonging to a type of specific expression as an information pair.

Images