JP2005322120A - Extraction system - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To quickly and surely collect words in the field same to that of a small number of words input by a user. <P>SOLUTION: This extraction system is provided with input positive examples that are the plurality of small number of input words, a database D storing a fixed volume of document data, a pattern extraction part 1 for performing full text retrieval of the input positive examples that have been input in the database, and for extracting a pattern appeared in the periphery of the plurality of input positive examples, and a word extraction part 2 for performing full text retrieval of the pattern extracted by the pattern extraction part 1 in the database D, and for extracting an expression extracted by the pattern, and the word extraction part 2 is precluded from extracting an expression including a character type not existing in the input positive examples. <P>COPYRIGHT: (C)2006,JPO&NCIPI

Description

  The present invention relates to an extraction system that collects and outputs terms in the same field as a small number of words input by a user.

  In recent years, in the question answering processing system (see reference (1) below), the technique of extracting proper expressions, such as extracting proper nouns, has become a necessary technique, and the extraction of protein expressions is important in the biological information processing field. As seen and research related to it is active (see reference (2) below), research on specific expression extraction (extraction of proper nouns, numbers, etc.) and technical term extraction has become very important. ing. In addition, technologies and ideas related to specific expressions and technical terms are being applied to new fields and will continue to be applied, and the types of terms are becoming extremely diverse and diverse.

  Reference (1): Maki Murata “Current Status and Prospect of Question Answering System” Journal of IEICE, Vol.86, No.12 (2003), pp.959-963.

  Reference (2): Tomohiro Mitsumori, Sevrani Fation, Masaki Murata, Kouichi Doi, and Hirohumi Doi, “Boundary correction of protein names adapting heuristic rules” Fifth International Conference on Intelligent Text Processing and Computational Linguistics (CICLing 2004), (2004).

  Therefore, in the present invention, evaluation data for term extraction relating to various terms can be created. In this data, the term data is also created with as little leakage as possible, and can be used for performance evaluation such as calculating recall and precision in term extraction experiments. In the present invention, explanation of this data and a simple term extraction evaluation experiment using this data will be described. Furthermore, an application system for term extraction having a visualization function is also described. This application system is practical and useful for a system that collects, visualizes and presents terms in the same field as several words input by a user in about 20 seconds.

  Conventionally, extracting a term related to the input term from a small number of input terms (positive input example) takes time to process, and the related term cannot be extracted accurately.

Also, in recent years, research on question answering is gaining importance, but in the question answering system, for example, if binomial data (see Table 2) such as data of pairs of country names and capital names is provided in advance, It has been possible to process a question response with binary data of such a relationship with high performance (for example, see Non-Patent Document 1). Binary data can be used in such a system, and there is a binary data extraction system created for such a system (for example, see Non-Patent Documents 2 and 3).
Michael Fleischman, Eduard Hovy, and Abdessamad Echihabi "Offline strategies for online question answering: Answering questions before they are asked" Proceedings of the 41st Annual Meeting of the Association for Computational Linguistics (eds. Erhard Hinrichs and Dan Roth), (2003), pp.1-7. Sergey Brin "Extracting patterns and relations from the world wide web" WebDB Workshop at 6th International Conference on Extending Database Technology, EDBT'98, (1998). Maya Ando, Satoshi Sekine, Shun Ishizaki "Automatic Extraction of Subordinate Concept Words from Newspaper Articles Using Regular Expressions" IPSJ Natural Language Processing Study Group 2003-NL-157, (2003).

  The above conventional ones cannot collect words in the same field as a plurality of input terms at high speed and accurately.

  SUMMARY OF THE INVENTION An object of the present invention is to solve the above-described problems, and to quickly and accurately collect terms in the same field as a plurality of words input by a user.

  FIG. 1 shows an extraction system of the present invention. In FIG. 1, 1 is a pattern extraction unit, 2 is a term extraction unit, and D is a database.

  The present invention has the following means in order to solve the above conventional problems.

  (1): Input positive examples which are a plurality of small terms to be input, a database D storing a certain amount of document data, and a full-text search of the input input input examples in the database. A pattern extraction unit 1 for extracting patterns appearing around an example, and a term extraction unit 2 for performing a full text search on the database D for the pattern extracted by the pattern extraction unit 1 and extracting an expression extracted by the pattern The term extraction unit 2 does not extract an expression including a character type that was not included in the input example. For this reason, terms in the same field as the input positive example can be collected at high speed and reliably.

(2): Full-text search is performed in the database D by inputting the input positive examples which are a plurality of small terms to be input, the database D storing a certain amount of document data, and the input positive examples. A pattern extraction unit 1 that extracts patterns appearing in the vicinity of the positive example, and a full-text search of the pattern extracted by the pattern extraction unit 1 in the database D, and an expression extracted by the pattern is extracted, and at the same time, the score Score is large A term extracting unit 2 that sorts in order, and the term extracting unit 2 includes the input in which the pattern appears in the ratio ( _pi ) of input positive examples in the expression extracted by the pattern for the calculation of the score Score. the number of positive cases of (f _i) using said input positive example of the number obtained by multiplying the value obtained by dividing by _{(n i) (p i ×} f i / n i). Therefore, it is possible to reliably collect terms in the same field as the input positive example.

(3): Input positive examples which are a plurality of small terms to be input, database D storing a certain amount of document data, and full-text search in the database D for the input correct input examples, and a plurality of the input A pattern extraction unit 1 that extracts patterns appearing in the vicinity of the positive example, and a full-text search of the pattern extracted by the pattern extraction unit 1 in the database D, and an expression extracted by the pattern is extracted, and at the same time, the score Score is large A term extraction unit 2 that sorts in order, the term extraction unit 2 is configured not to extract an expression including a character type that was not in the input example, and is extracted in the pattern for calculating the score Score multiplied by the value input the pattern to the rate of positive cases (p _i) is divided by the emerging the input positive cases the number of (f _i) the number of the input positive sample (n _i) of the representation Value of _{(p i × f i / n} i) is used. For this reason, terms in the same field as the positive input example can be collected at higher speed and more reliably.

  (4): input input examples which are a plurality of small terms to be input, database D storing a certain amount of document data, and full input search of the input input examples in the database D, and a plurality of the inputs A pattern extraction unit 1 for extracting patterns appearing in the vicinity of the positive example, a term extraction unit 2 for performing a full text search in the database D for the pattern extracted by the pattern extraction unit 1 and extracting an expression extracted by the pattern; The term extraction unit 2 does not extract an expression including a part of speech that was not in the last morpheme of the input positive example. For this reason, terms in the same field as the input positive example can be collected at high speed and reliably.

  (5): Full-text search is performed in the database D by inputting the input positive examples which are a plurality of small terms to be input, the database D storing a certain amount of document data, and the input positive examples. A pattern extraction unit 1 for extracting patterns appearing in the vicinity of the positive example, a term extraction unit 2 for performing a full text search in the database D for the pattern extracted by the pattern extraction unit 1 and extracting an expression extracted by the pattern; The term extraction unit 2 does not extract an expression that does not have a common partial character string of the input positive example. For this reason, terms in the same field as the input positive example can be collected at high speed and reliably.

  (6): In the extraction system of (1) to (5), the user can specify an expression that the term extraction unit 2 does not extract. For this reason, terms in the same field as the positive input example can be collected at high speed and more reliably.

  (7): In the extraction system of (1) to (6), a plurality of a few binary terms data is input as the positive input example, and the term extracting unit 2 expresses the terms of two or more terms. Extract. For this reason, it is possible to reliably collect two or more terms in the same field as the input positive example.

  (8): A full-text search is performed in the database D, a database D storing a certain amount of document data, and a database D in which a plurality of input two or more term data are input. A pattern extraction unit 1 for extracting patterns appearing around the plurality of input positive examples, and a full text search of the pattern extracted by the pattern extraction unit 1 in the database D, and two or more terms extracted by the patterns A term extracting unit 2 that extracts term expressions and sorts them in descending order of score, and the term extracting unit 2 uses a combination of the extracted patterns of data for calculating the score. Therefore, an accurate score can be calculated, and two or more terms in the same field as the input positive example can be reliably collected.

  (9): A procedure for inputting positive input examples which are a plurality of small terms, a procedure for storing a certain amount of document data in the database D, a full text search in the database, A pattern extraction procedure for extracting a pattern appearing in the vicinity of the input correct example, a term extraction procedure for full-text searching the extracted pattern in the database, and extracting an expression extracted by the pattern, and the input correct example A term extraction procedure for preventing an expression including a character type not found in the above from being extracted is a program for causing a computer to execute. Therefore, by installing this program in a computer, it is possible to easily provide an extraction system that can quickly and reliably collect terms in the same field as the input input example.

  The present invention has the following effects.

  (1): Patterns extracted by the pattern extraction unit 1 by a term extraction unit by performing full text search in a database and extracting patterns appearing around the plurality of input examples by the pattern extraction unit When searching for a full-text search in database D and extracting expressions extracted by the pattern, in order not to extract expressions containing character types that were not in the input positive examples, terms in the same field as the input positive examples are And can be collected reliably.

(2): The full pattern search is performed in the database D by the input correct example input by the pattern extracting unit, the patterns appearing around the input correct examples are extracted, and the pattern extracted by the pattern extracting unit by the term extracting unit When the full-text search is performed in the database D and the expressions extracted by the pattern are extracted and simultaneously sorted in descending order of the score Score, the ratio of the positive examples of the expression extracted by the pattern in calculating the score Score ( A value (p _i × f _i / n _i ) obtained by multiplying p _i ) by the number (f _i ) of the input positive examples in which the pattern appears divided by the number (n _i ) of the input positive examples is used. Therefore, terms in the same field as the input positive example can be reliably collected.

(3): Patterns extracted from the database D by full-text search in the database D, extracted patterns appearing around the plurality of input examples, and extracted by the pattern extraction unit by the term extraction unit Are retrieved from the database D, the expressions extracted by the pattern are extracted and simultaneously sorted in descending order of score Score, so that expressions including character types that are not in the input positive examples are not extracted, and In calculating the score Score, the number (f _i ) of the input positive examples in which the pattern appears in the ratio (p _i ) of the input positive examples in the expression extracted by the pattern is the number (n _i ) of the input positive examples. Since the value (p _i × f _i / n _i ) multiplied by the value divided by is used, terms in the same field as the input positive example can be collected at high speed and more reliably.

  (4): Patterns extracted by the full-text search in the database D by the pattern extraction unit, extracting patterns appearing around the plurality of input positive examples, and extracted by the pattern extraction unit by the term extraction unit In the same field as the input example, in order not to extract the expression including the part of speech that was not in the last morpheme of the input positive example when the database D is full-text searched and the expression extracted by the pattern is extracted Can be collected at high speed and reliably.

  (5): The full pattern search is performed in the database D for the input positive example input by the pattern extraction unit, the patterns appearing around the plurality of input positive examples are extracted, and the pattern extracted by the pattern extraction unit by the term extraction unit In order to avoid extracting expressions that do not have a common partial character string of the input positive example when extracting the expression extracted by the pattern. Can be collected at high speed and reliably.

  (6): Since the user designates an expression that the term extraction unit does not extract, terms in the same field as the input positive example can be collected at high speed and more reliably.

  (7): As a positive input example, a plurality of terms data of two or more terms are input, and the term extraction unit extracts expressions of two or more terms, so that two or more terms in the same field as the input positive example Can be collected reliably.

  (8): The full text search is performed in the database D for the input positive examples input by the pattern extraction unit, the patterns appearing around the input positive examples are extracted, and the patterns extracted by the pattern extraction unit are extracted from the database D When the full-text search is performed and the expression of two or more terms extracted by the pattern is extracted at the same time by the term extraction unit and simultaneously sorted in descending order of score, the combination of the extracted data of the plurality of patterns is used for the calculation of the score. Therefore, an accurate score can be calculated, and two or more terms in the same field as the input positive example can be reliably collected.

§1: Explanation of term extraction (1): Explanation of extraction system FIG. 1 is an explanatory diagram of the extraction system of the present invention. In FIG. 1, an extraction system for extracting terms includes a pattern extraction unit 1, a term extraction unit 2, and a database D.

Pattern extraction unit 1, extracts a pattern c _i and full-text search, which appeared around the positive examples of the at number (from the input means or the like) a few positive cases entered (terms) in the database D. The term extraction unit 2 extracts the pattern c _i and full-text search in the database D, at the same time extracts the representation exp extracted by the pattern c _i, extracted expression exp Score; in the order of decreasing (scoring value) It sorts and outputs (to an output means etc.). The database D stores data (a fixed amount of document data) extracted from, for example, newspapers, magazines, Web data (data on the network), and the like.

(Explanation based on flowchart)
FIG. 2 is a term extraction process flowchart. Hereinafter, description will be given according to the processing S1 to S3 of FIG.

  S1: A small number of positive examples are input to the pattern extraction unit 1 from the input means.

S2: the pattern extraction unit 1, enter a positive example full-text search database D, and extracts the appearance pattern around the plurality of input positive cases as c _i. (Definitions of patterns that appear in the vicinity are made as appropriate.)
S3: The term extraction unit 2, a pattern c _i extracted by the pattern extraction unit 1 and full-text search in the database D, at the same time extracts the representation exp extracted by the pattern c _i, extracted expression exp value Score Big Sort in order and output to output means.

(Description of pattern example)
1) Example of country name A and capital name B pair (binary data) Example of correct input example input to pattern extraction unit 1:
Japan Tokyo
China_Beijing
Korea_Seoul
Singapore_Singapore
Russia_Moscow ・ Examples of extraction patterns extracted by the pattern extraction unit 1:
A capital B
A capital B
In AB city
The [B branch office] A
B branch office] A Ministry of Foreign Affairs However, if there are A and B on the left and right, the left and right must be hiragana characters.

2) Example for country A ・ Input example:
Japan
China
Korea
Thailand
Korea ・ Example of extraction pattern (1): (Used at both ends, slow speed but good performance)
Sun, A army
A female
Japan is A
[A communication
Ministry. Ambassador to A ・ Example of extraction pattern (2): (Only one is used, one is Hiragana, and the speed is fast)
[..A country].

word. A
[..A country] side
[..A country]
A word input However, [..A ..] means that country name A itself matches. For example, [Country A] means that the end of the matched term is the country.

(2): Description of creation of term extraction evaluation data In the present invention, term extraction evaluation data (correct data) was created in order to determine the accuracy of the extraction system. Examples of the created data are shown in Tables 1 and 2. Table 1 shows data relating to country names, which are divided into rows for each country, and the head of the row is a representative form, and the rest is stored in the same line as a representation of the representative form. Table 2 is a pair of country name and capital name data. As in Table 1, the data is divided into lines for each country, and the head is represented as a representative form, and the rest is stored in the same line as a representation of the representative form.

  In the present invention, Table 1 and Table 2 are referred to as one-term data and two-term data, respectively. 58 types of each of these were made. Table 3 shows the scale of all data. The representative form number and the representative form + annotated number are the total number of them.

The 58 types of data created are "Solar system planets""Satellite""Zodiac""Holidays""Spaceshuttle""Taigadrama""Sumo-related""Flowername""Soccer defensive position""Professional baseball player name"" World heritage, village name and prefecture name pairs, holidays and date pairs, solar system planets and satellite pairs, composers and music pairs, etc.

Table 1: Example of unary data (country name data)
┌────────────────────────┐ │Iceland Iceland ISL │
│Ireland Republic of Ireland IRL │
│Azerbaijan Republic of Azerbaijan AZE│
│ Azores │
│Aduige Aduige Republic │
│Afghanistan Afghanistan │
│USA USA USA USA │
│ ... │
└────────────────────────┘

表３：データの規模
┌─────────┬─────┬─────┐
│ │一項データ│二項データ│
├─────────┼─────┼─────┤
│データの種類の数 │ 58│ 58│
│代表形数 │ 17696│ 19387│
│代表形＋異表記数 │ 26728│ 106850│
└─────────┴─────┴─────┘
（データの作成方法は以下の方法をとった）
ａ）単一の辞書・参考書などから手入力する。（例：太陽系惑星、衛星、十二支、祝日、スペースシャトル）
ｂ）Ｗｅｂのサイトから入手する。（例：大河ドラマの名称）
ｃ）複数の辞書・参考書またＷｅｂのサイトから得た情報を組み合わせる。（例：世界の山、花の名称、商品名）
ｄ）その分野の知識が豊富な人間が知識と記憶によって作成する。（例：サッカー守備位置）
（データの補充、異表記の作成には以下の方法をとった）
ａ）規則性を持った異表記を自動で生成する。（例：人名から姓を取り出す）
ｂ）規則性がない異表記をＷｅｂなどから取得する。（例：世界の山，花の名称，商品名）
ｃ）規則性がない異表記を思いつく範囲で入力する。（例：相撲決り手、星座）
ｄ）その分野の知識が豊富な人間が知識と記憶によって作成する。（例：サッカー守備位置）
データの作成の際には、それぞれのデータごとにその収集方法、異表記作成方法、代表形の基準の定義、異表記の基準の定義、その他のコメント、代表形の網羅度の情報を作成している。例えば、「国名データ」の代表形の基準の定義は「正式名称ではなく最も一般的に使用されるもの（例：フランス共和国→代表形「フランス」）」と記載されている。

Table 3: Scale of data ┌─────────┬─────┬─────┐
│ │ One-term data │ Two-term data │
├─────────┼─────┼┼─────┤
│ Number of data types │ 58│ 58│
│Representative number │ 17696│ 19387│
│Representative type + number of different notations │ 26728│ 106850│
└─────────┴─────┴┴─────┘
(The following method was used to create the data)
a) Enter manually from a single dictionary or reference book. (Example: Solar system planets, satellites, zodiac, holidays, space shuttle)
b) Obtain from the Web site. (Example: Taiga drama name)
c) Combining information obtained from multiple dictionaries, reference books and websites. (Example: Mountains in the world, flower names, product names)
d) Created by knowledge and memory by a person with a wealth of knowledge in the field. (Example: soccer defensive position)
(The following methods were used to replenish data and create different notations)
a) Annotation with regularity is automatically generated. (Example: Extract last name from person's name)
b) Acquiring different notation with no regularity from the Web or the like. (Example: world mountains, flower names, product names)
c) Enter an illegitimate notation that is not regular. (Example: Sumo finalist, constellation)
d) Created by knowledge and memory by a person with a wealth of knowledge in the field. (Example: soccer defensive position)
At the time of data creation, for each data, the collection method, different notation creation method, definition of the standard for the representative form, definition of the standard for the different form, other comments, information on the coverage of the representative form are created. ing. For example, the definition of the criterion of the representative form of “country name data” is described as “not the official name but the most commonly used one (eg, France → representative form“ France ”)”.

  As the degree of coverage of the representative form, information such as almost 100% coverage or not coverage is given. Many of the terms used in the present invention are known to have a limited number of terms, such as country names, capital names, and solar system planets. If you do, you can create data that will be almost 100% covered.

  For this reason, the data of the present invention can be used for calculation of precision of precision and recall of term extraction. There are 58 types of data, which is very small compared to all types of terms in the world, but at the expense of the type of data, it can be easily collected and specified in terms within a specific field. They are creating data that can be used to calculate the precision of precision and recall with the completeness within the field.

  In addition, the term data of the present invention is universal, and is evaluation data that can be used regardless of what data is subject to automatic term extraction. For example, the result extracted from the daily newspaper may be evaluated with this data, and the result extracted from the Web may be evaluated with this data.

  Unary data is an ordinary term list, but binomial data is term pair data, which is more like knowledge than a term list. In recent years, research on question answering is gaining importance. If a question answering system is equipped with such binomial data in advance, the question answering with such relational binomial data can be processed with high performance. Binary data can be used for such a system, and can also be used for evaluation data of a binary data extraction system created for such a system.

(3): Specific Explanation of Term Extraction Since the evaluation data (correct data) has been created according to (2) above, a simple term extraction experiment using this was performed. In this experiment, the data whose coverage is “almost 100% covered” has 10 or more representative forms so that the experiment can be performed (40 types for 1-term data, 44 types for 2-term data) Was used. In this experiment, we used learning by using only positive examples that were learned using a small number of positive examples and picked up many positive examples. Experiment to see if the country can be extracted).

  In this experiment, different notation is not a positive example, but only a representative form is a positive example. As a few positive examples of input, we selected five terms that are famous from the most frequently used daily newspapers as representative examples of evaluation data. CD Mainichi Newspaper (Mainichi Newspaper recorded on a compact disc), 1999-2000 version, was designated as data (database) D for acquisition. The extraction procedure is as follows.

(1) a small number of positive cases full-text search database D, if the appearance pattern around the plurality of primary example is extracted as c _i (a pattern appearing on the periphery of the positive examples only that positive cases (one) Is not extracted). (Definitions of patterns that appear in the vicinity are made as appropriate). For example, when using three character strings before and after (left and right) of the positive example as patterns appearing in the periphery, there are cases where there are one, two, and three characters respectively before and after, so there are nine patterns in one positive example. A pattern can be made. Moreover, it can also be set as the pattern including a positive example (self).

(2) Next, full-text search is performed on the extracted pattern c _i in the database D, and the expression exp extracted based on the pattern c _i is extracted.

  (3) The extracted expression exp is sorted and output in descending order of the Score value.

  There are the following as Score.

・ Method 1 (decision list method)
Method 1 as Score of the extracted expression exp, those using p _i of the largest was pattern p _i in the pattern c _i.

・手法２（ベイズ法）
手法２は、抽出した表現ｅｘｐのＳｃｏｒｅとして、全てのパターンｃ_i のｐ_i を掛け合わせたものを使用する。

・ Method 2 (Bayes method)
Method 2 uses a score obtained by multiplying all the patterns c _i by p _i as the score of the extracted expression exp.

・手法３（類似度に基づく方法）
手法３は、抽出した表現ｅｘｐのＳｃｏｒｅとして、抽出されたパターンの個数（総数）を用いる。つまり、多くのパターンで抽出されたものほどＳｃｏｒｅを大きくする。

・ Method 3 (method based on similarity)
Method 3 uses the number (total number) of extracted patterns as the score of the extracted expression exp. That is, the score is increased as the number of patterns extracted is increased.

・手法４（下記研究(3) 参照）
手法４は、抽出した表現ｅｘｐのＳｃｏｒｅとして、ｐ_i の重みを加えた抽出されたパターンの個数を用いるものである。

・ Method 4 (Refer to Research (3) below)
Method 4 as Score of the extracted expression exp, is to use a number of the extracted pattern plus the weight of p _i.

研究(3):Ellen Riloff and Rosie Jones "Learning dictionaries for information extraction by multi-level bootstrapping" Proceedings of AAAI-99,(1999)。

Study (3): Ellen Riloff and Rosie Jones "Learning dictionaries for information extraction by multi-level bootstrapping" Proceedings of AAAI-99, (1999).

・ Method 5 (Refer to the following document (4))
Method 5 uses at least one value that is likely to be the score of the extracted expression exp.

上記式(5) は、確からしくない（１−ｐ_i ）を掛け合わせることで一つも確からしくないことになり、そして、これを１から引くと少なくとも一つは確からしくなる。

In the above formula (5), by multiplying (1− _pi ) which is not certain, one is not certain, and when subtracting this from 1, at least one is certain.

  Reference (4): Masaki Murata and Hitoshi Isahara "Automatic Acquisition of Knowledge about Paraphrases Based on Matching Synonymous Texts" IPSJ SIG 2001-NL-142, (2001).

However, f _i is the number of input positive examples in which the pattern c _i appears, and p _i is the ratio of input positive examples in the expression extracted by the pattern c _i (the probability, that is, the certainty level). In the methods 1, 2, 4, and 5, when the score is the same, the score is sorted by the score of the method 3, and in the method 3, the score is sorted by the score of the method 5.

  Using one-term data, an experiment was performed using a combination of 1 to 3 characters including either the left or the head of the positive example and that on the right as a pattern. The results are shown in Table 4.

表４において、ＡＰは、情報検索（下記文献(5) 参照）で用いるaverage precision の平均であり、正解記事を上位から取ったたびに求めた適合率の平均である。本願の内容の場合は、正解正例分を上位から取ったたびに求めた適合率の平均（ただし、入力正例は正解正例から除く）である。

In Table 4, AP is an average of average precision used in information retrieval (see the following document (5)), and is an average of relevance ratios obtained every time correct articles are taken from the top. In the case of the contents of the present application, it is the average of the relevance ratios obtained every time the correct correct example is taken from the top (however, the input correct example is excluded from the correct correct example).

  RP is an average of r-precision, and is a ratio of including correct articles when searching for the number of correct articles. In the case of the contents of the present application, it is a ratio in which correct correct examples are included when only correct correct examples are extracted. The relevance rate is the same as the correct answer rate, and is a ratio including correct correct examples. TP is the average accuracy of the top five.

(Explanation of extraction method based on constraints)
(A) Method of Using Character Type and KR For the extraction method in the example of Table 4, a method of further using character type and KR was used. Here, the character types are kanji, katakana, hiragana, symbols, numbers, and the like. For example, in English, alphabets, numbers, symbols, and whether the beginning of a word is capitalized or the like.

  The method using the character type is a method that does not extract an expression including a character type that was not included in a small number of input terms (five in this example). For example, when there are no hiragana characters in five input terms, an expression including hiragana characters is not extracted.

In the method using KR, p _i is replaced with p _i * f _i / n _i . The advantage of this method is that the certainty factor can be changed by the value of f _i / n _i even if p _i is the same. However, n _i is the number of positive input examples, and in the case of method 3, in the case of KR, 1 is replaced with f _i . In the evaluation, the positive results were excluded from the extracted results. In addition to the character type method, there are also the following methods.

(B) Method based on part of speech In the method based on part of speech, for example, if there is only a noun in the input expression, the expression other than the noun is omitted at the time of output, and if the input expression only has an adjective, the expression other than the adjective is omitted at the time of output. That's it. Furthermore, when the expression is composed of a plurality of words, the part of speech information of the last word (morpheme) can be used.

(Description 1 by example)
If the input example is:
“Fun” “sad” “happy” “very happy” “very sad”
If the following is obtained as an extract:
"Very""New""Beautiful""Verybeautiful""Verydifficult"
Estimate the part of speech of the last word in the expression of the above extract, and in the above input example, the part of speech of the last word is only “adjective”, so the part of speech of the last word in the extract is “adjective”. Not output adverb ("very").

(Description 2 by example)
If the input example is:
"Fun""joy""sadness""sad"
In the above input example, when there are multiple types such as “adjective” and “noun”, those parts of speech are output, and expressions other than those parts of speech are not output.

  In order to obtain part-of-speech information such as the estimation of the part-of-speech of the last word (morpheme) as described above, the following morpheme analysis system (morpheme analysis means) is required.

-Description of morphological analysis system In order to divide Japanese into words, it is necessary to use the morphological analysis system in the term extraction unit 2. Here, ChaSen will be explained (published on the morphological analysis system Chasen http://chasen.aist-nara.ac.jp/index.html.jp developed at Nara Institute of Technology).

  This splits the Japanese sentence and also estimates the part of speech of each word. 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.

(C) Method based on common partial character string For example, when input expressions all have the same common end expression “Shi”, an expression that does not have “Shi” is omitted at the time of output. This can be done not only at the end but also at the top character string.

(Description by example)
If the input example is:
"Sad""fun""happy"
If the following is extracted:
"Joy""Sorrow""Beautiful""New"
Since the common partial character string of the above input example is "Shi", "Joy" and "Sadness" without "Shi" are deleted and output.

(D) Specification of constraint by user Up to now, it has been a method of automatically obtaining a constraint from an input expression, but this constraint can be made to be made by the user. For example, when the user selects the option of “Kanji only”, the output using a character type other than Kanji can not be output. In addition, when the user selects the option “Shi” at the end, it is possible to prevent the output not having “Shi” at the end in the output. Furthermore, when the user selects the option that the part of speech is a noun, the output is made so that expressions other than the noun are not output.

  In addition, there is already research (3) in which a small number of positive examples are entered.In this research (3), a bootstrap algorithm is added to the extraction method to increase the number of positive examples and increase the accuracy. It uses a method of extracting terms. The accuracy of the method of the present invention will also be improved by using bootstrap. There is method 4 as research (3), but the result of the method which changed the formula a little and the method which made various ideas can be compared because there is the data for evaluation. In the comparison of the present invention, it is understood that the method 5 using the character type / KR is the best.

  Reference (5): Masaki Murata, Ma Aoi, Kiyotaka Uchimoto, Hiromi Osaku, Masao Uchiyama, Hitoshi Isahara "Information Retrieval Using Location Information and Field Information" Journal of Language Processing Society, Vol.7, No.2, ( 2000).

  Next, an experiment using binomial data was performed. As a pattern, the order of the positive example indicating which of the binomial data first appears, and between the first and second positive examples including one of the left and top of the first positive example The combination of expressions, the order of positive examples, and the combination of expressions between 1 to 3 characters including either the right or the rear of the second positive example and the two positive examples were used.

  When the character string of the pattern is not extracted on the left or right side of the positive example so that the boundary of the word of the positive example is determined, the pattern condition is that they are hiragana characters. In this experiment, the accuracy of the method 3 using the character type and KR was the best, AP was 0.026, RP was 0.040, and TP was 0.141.

In order to speed up the extraction system, the pattern ci is _defined as “one to three characters including either the left or the beginning of the input example and the right of the input example is a hiragana character, or the input example 1 to 3 characters including either the right or the rear of the character and the left of the input example is a hiragana character.

§2: Description of two-dimensional display (1): Description of display system FIG. 3 is an explanatory diagram of the display system of the present invention. In FIG. 3, the display system includes a pattern extraction unit 1, a term extraction unit 2, a term display unit 3, a visualization device 4, and a database D.

Pattern extraction unit 1, extracts a pattern c _i and full-text search, which appeared around the positive examples of the at number (from the input means or the like) a few positive cases entered (terms) in the database D. The term extraction unit 2 extracts the pattern c _i and full-text search in the database D, at the same time extracts the representation exp extracted by the pattern c _i, extracted expression exp sorted in the order of decreasing Score (Score) (To output means etc.) For each term, the term display unit 3 uses the pattern used for extracting the term as a context, obtains the similarity between the terms from the information of the context, and means the semantic distance between the terms. Create a matrix. The visualization device 4 displays (visualizes) a two-dimensional figure using a self-organizing map or the like (two-dimensional display means). The database D stores data (a certain amount of document data) extracted from, for example, newspapers, magazines, Web data, and the like.

(Explanation based on flowchart)
FIG. 4 is a flowchart of the display process. Hereinafter, description will be given according to processing S11 to S14 of FIG.

  S11: A small number of positive input examples are input to the pattern extraction unit 1 from the input means.

S12: In pattern extraction unit 1, enter a positive example full-text search database D, and extracts the appearance pattern around the plurality of input positive cases as c _i. (Definitions of patterns that appear in the vicinity are made as appropriate.)
S13: The term extraction unit 2, a pattern c _i extracted by the pattern extraction unit 1 and full-text search in the database D, at the same time extracts the representation (term) exp extracted by the pattern c _i, extracted expression exp of Score Sort the items in descending order and pass them to the term display unit 3.

  At this time, data indicating which pattern is used for each extracted term is also attached.

  S14: In the term display unit 3, for each term, the pattern used for extracting the term is used as a context, and the similarity between the terms is obtained from the information of the context (terms having the same pattern are similarities) And a matrix representing the semantic distance between the terms is created, and this matrix is input to the visualization device 4 (self-organizing map SOM_PAK tool) to determine the coordinates of each term in the two-dimensional map. And the figure which displayed the term in the calculated coordinate is outputted.

(2): Explanation of Visualization Device (Self-Organizing Map SOM_PAK Tool) As an automatic construction machine for semantic maps, Self-Organizing Map (SOM for short) is Kohonen's self-organizing neural network model. (Kohonen, T .: Self-organizing maps, 2nd edition, Springer, 1997.). The SOM is composed of nodes of a two-dimensional array having a high-dimensional input, and high-dimensional data can be mapped to a two-dimensional space so as to reflect the characteristics by self-organization described below.

但し、参照ベクトルの要素μ_ijはノードｉと入力要素ξ_j の間の重みであり、自己組織過程において少しずつ修正される。入力ベクトルｘが与えられたとき、まず、その入力をすべてのノードの参照ベクトルと比較し、ユークリッド距離の一番短いノードを活性化する。マッピング処理段階ではこのノードのみ活性化される。このノードを勝者ノードと呼ぶ。即ち、勝者ノードｃは以下の式(6) のように選ばれる。

However, the element μ _{ij of the} reference vector is a weight between the node i and the input element ξ _j and is corrected little by little in the self-organization process. When an input vector x is given, the input is first compared with the reference vectors of all nodes, and the node with the shortest Euclidean distance is activated. Only this node is activated in the mapping process stage. This node is called the winner node. That is, the winner node c is selected as shown in the following formula (6).

一方、自己組織化過程では、グローバルに自己組織化が行われるように、勝者ノードだけでなくその近傍のノードも活性化させ、リラックス処理を行う。即ち、活性化されたすべてのノードに対し、それらの参照ベクトルを入力ベクトルに近づくように修正を行う。

On the other hand, in the self-organization process, not only the winner node but also the nearby nodes are activated and the relaxation process is performed so that the self-organization is performed globally. That is, all the activated nodes are modified so that their reference vectors are close to the input vector.

ここで、ｔは学習回数で、ｈ_ci(t) は、例えば以下の式(8) のように定義された近傍関数である。

Here, t is the number of learning times, and h _ci (t) is a neighborhood function defined as, for example, the following equation (8).

従って、項‖ｒ_c −ｒ_i ‖は近傍ノードｉが勝者ノードｃから離れて行くにつれ、ｈ_ciが小さくなりｍ_i (t) の修正量が小さくなることを意味する。また、α(t) は学習率で、σ(t) は近傍の大きさ（半径）である。これらは時間と共に単調に減少していく関数であればよい。

Thus, term ‖r _c -r _i ‖ means that neighboring node i as the moves away from the winning node c, the correction amount of h _ci decreases m _i (t) becomes smaller. Α (t) is the learning rate, and σ (t) is the size (radius) of the neighborhood. These may be functions that monotonously decrease with time.

  The learning process usually consists of an “alignment” phase and a “fine adjustment” phase. In the “alignment” phase, the initial values of α (t) and σ (t) are both large and decrease with time. The basic form of node placement is formed in this phase. On the other hand, in the remaining phases, α (t) and σ (t) take a long time while keeping small values, and finely adjust the basic shape formed in the initial phase.

  As an example of the visualization apparatus that the present inventors have proposed so far, words that are semantically close between terms (nodes) are close to each other, and words that are semantically far from each other are distant from each other. There is an automatic construction method of monolingual semantic maps (for example, reference (6) described below, and Ma, Q., Zhang, M., Murata, M., Zhou, M., Isahara, H .: (See Self-Organizing Chinese and Japanese Semantic Maps, The 19th International Conference on Computational Linguistics (COLING'2002), Taiwan, pp. 605-611, August, 2002.).

(3): Explanation by Specific Example The term extraction method has a feature that it can be calculated at high speed because it uses a full-text search of a character string. There is a fast search algorithm called suffix array for full text search of this string. Therefore, a simple application system using the above method was developed. First, the user inputs several favorite words. Using these few words as a small number of positive examples, the above-mentioned method 5 using the character type / KR is used to collect words in the same field as the input several words.

Pattern c _i in the algorithm of the time for speeding the §1 (3) (specific description of term extraction), to be more simple, also several times paw data threshold in the course of the calculation process Cut and reduce. To be exact, the experiment program described above is slightly cut off. Still, one of the above takes about 10 hours.

Visualize and output the words in the same field so that they are easy to see. For the visualization, the semantic map of the following document (6) was used. In the use of the semantic map, the number of learning is set to 1/10 for speeding up. Otherwise, the same method as in Ref. (6) was used. The context used was the pattern c _i excluding the feature that uses the internal representation of the positive example. In the present invention, the pattern c _i used for term extraction is used for the context of the semantic map. However, in addition to the pattern c _i , the context of each term can be extracted from large-scale data and the semantic map can be displayed. It is possible.

  With this implementation, operation in about 20 seconds is now realized. (There are currently implementation parts in Perl, and it is possible to increase the speed by improving the algorithm in the future.) Since the performance of this system is information that has been cut by speeding up, the accuracy of the one-term data from Table 4 As a result, AP was 0.111, RP was 0.164, and TP was 0.310.

  Reference (6): Ma Aoi, Kyoko Kanzaki, Maki Murata, Kiyotaka Uchimoto, Hitoshi Isahara “Self-Organization of Semantic Maps of Japanese Nouns” Information Processing Society Journal, Vol.42, No.10, (2001), pp.2379-2391.

(Specific example (1))
FIG. 5 is an explanatory diagram of term extraction related to “color” by the user. FIG. 5 shows the outputs (the top 20 outputs) when the user actually inputs “red”, “blue”, “yellow”, “purple”, and “brown” into this system. Here, the symbol “●” is attached to the input (input positive example). The number attached to each expression indicates the number obtained by the extraction method. In general, expressions having similar patterns are arranged close to each other by a semantic map so that it is easy to see. In this example, “red” and “red” appear near “red” in the upper left, and it can be seen that similar expressions may appear nearby. In addition, expressions that are not "tone", "different color", and "same color" are set in the middle from the top right, but in the visualization on the semantic map, such incompatible expressions are collected somewhere ( There is also an effect that it is easy to see that it should be excluded.

(Specific example (2))
In the above-described examples and the like, nouns are used as input positive examples, but evaluation expressions such as adjectives can also be handled. For example, if “● sad” “● fun” “● sad” “● happy” is input as a positive input to the pattern extraction unit, the following terms can be obtained as the output of the term extraction unit.

"New""Beautiful""Difficult""Strict""Intense""Friendly""Lonely""Unusual""Right""Right""Poor""Detailed""Poor""Cool""Friendly""Iwant""Brandnew""Busy""Remarkable""Equal""Heavily""Delicious""Ugly""Polite""Pleasant""Friendly""Tough"
In this way, the input term can handle not only nouns but also anything.

  In addition, although the method of the self-organization map was demonstrated as a visualization apparatus, other methods, such as the method of using a principal component analysis, can also be used. In addition, although the symbol “●” is attached to the input correct example, in addition to this, any other mark or color may be added so that the input correct example can be understood.

In the examples so far, a matrix indicating semantic distance is created from context information and this is input to the visualization device. However, the visualization device itself can be visualized even if other matrices are input. Can do. For example, a term can be visualized by inputting the following matrix in which context information for a word is arranged on the right.

Context 1 Context 2 Context 3 ・ ・ ・ ・ Context N
Terminology 1 2 0 1 1
Terminology 2 2 1 1 2
...
The term M 0 0 1 0

The number of each element of the matrix means the number of occurrences of the term in the context. Also, a method using principal component analysis can be visualized with such a format as an input.

(4): Summary As described above, in the present invention, data that can be used for evaluation at the time of term extraction is created. As this evaluation data, one-term data and two-term data such as “country name” and “country name and capital name pair” were created. The data of the present invention is highly comprehensive in terms, and can be used for performance evaluations such as calculating recall and relevance ratios when extracting terms from any data. The binary data created in the present application is closer to knowledge than a term list, and can be used for research on knowledge processing such as question answering.

  In this application, a simple term extraction experiment was performed and various term extraction methods were compared. The method 5 (reference (4)) using the character type / KR is the best for the one-term data, and the method 3 (method based on the similarity) using the character type / KR is the best for the binary data. Furthermore, an application system is shown in which when a user inputs several favorite words, the words in the same field as those several words are collected and visualized in about 20 seconds. In this system, similar terms are gathered and unnecessary terms tend to be gathered, which is useful for displaying extracted terms.

  In the above example, two pairs of terms (binary data) have been described as positive input examples, but the present invention can be similarly implemented using three or more pairs of terms.

(5): Explanation of program installation The pattern extraction unit 1, the term extraction unit 2, the term display unit 3, the visualization device 4 and the like can be configured by programs, and are executed by the main control unit (CPU) and stored in the main memory. It is stored. This program is generally processed by a computer. This computer is composed of hardware such as an input device as input means such as a main control unit, main memory, file device, display device, and keyboard.

  The program of the present invention is installed on this computer. In this installation, these programs are stored in a portable recording (storage) medium such as a floppy disk or a magneto-optical disk, and a drive device for accessing the recording medium provided in the computer is used. Alternatively, it is installed in a file device provided in the computer via a network such as a LAN. Then, the program steps necessary for processing are read from the file device into the main memory and executed by the main control unit.

It is explanatory drawing of the extraction system of this invention. It is a term extraction processing flowchart of the present invention. It is explanatory drawing of the display system of this invention. It is a display processing flowchart of the present invention. It is explanatory drawing of the term extraction regarding the "color" by the user of this invention.

Explanation of symbols

1 Pattern extractor 2 Term extractor D Database

Claims (9)

A positive input, which is a small number of terms to enter,
A database that stores a certain amount of document data;
A full-text search in the database for the input positive examples that have been input, and a pattern extraction unit that extracts patterns that appear around the input positive examples, and
A full-text search in the database for the pattern extracted by the pattern extraction unit, and a term extraction unit for extracting an expression extracted by the pattern;
The term extraction unit is configured to prevent an expression including a character type not included in the input example from being extracted. A positive input, which is a small number of terms to enter,
A database that stores a certain amount of document data;
A full-text search in the database for the input positive examples that have been input, and a pattern extraction unit that extracts patterns that appear around the input positive examples, and
A full text search of the pattern extracted by the pattern extraction unit, and a term extraction unit that extracts expressions extracted by the pattern and simultaneously sorts them in descending order of scores,
The term extraction unit calculates the number (f _i ) of the input positive examples in which the pattern appears in the ratio (p _i ) of the input positive examples in the expression extracted by the pattern for the calculation of the score. An extraction system characterized by using a value (p _i × f _i / n _i ) multiplied by a value divided by the number of (n _i ). A positive input, which is a small number of terms to enter,
A database that stores a certain amount of document data;
A full-text search in the database for the input positive examples that have been input, and a pattern extraction unit that extracts patterns that appear around the input positive examples, and
A full text search of the pattern extracted by the pattern extraction unit, and a term extraction unit that extracts expressions extracted by the pattern and simultaneously sorts them in descending order of scores,
The term extraction unit is configured not to extract an expression including a character type that was not in the input positive example, and to calculate a ratio of the input positive example in the expression extracted by the pattern in the calculation of the score ( _pi ). A value (p _i × f _i / n _i ) obtained by multiplying the number (f _i ) of the input positive examples in which the pattern appears by the number (n _i ) of the input positive examples is used. Extraction system. A positive input, which is a small number of terms to enter,
A database that stores a certain amount of document data;
A full-text search in the database for the input positive examples that have been input, and a pattern extraction unit that extracts patterns that appear around the input positive examples, and
A full-text search in the database for the pattern extracted by the pattern extraction unit, and a term extraction unit for extracting an expression extracted by the pattern;
The extraction system according to claim 1, wherein the term extraction unit is configured not to extract an expression including a part of speech that was not included in the last morpheme of the input positive example. A positive input, which is a small number of terms to enter,
A database that stores a certain amount of document data;
A full-text search in the database for the input positive examples that have been input, and a pattern extraction unit that extracts patterns that appear around the input positive examples, and
A full-text search in the database for the pattern extracted by the pattern extraction unit, and a term extraction unit for extracting an expression extracted by the pattern;
The term extraction unit does not extract an expression having no common partial character string of the input positive example. The extraction system according to claim 1, wherein a user can specify an expression that the term extraction unit does not extract. 7. The input of a plurality of two or more terms as a positive example, and the term extraction unit extracts expressions of two or more terms according to any one of claims 1 to 6. Extraction system. A positive input example that is a plurality of a few binomial term data to be input,
A database that stores a certain amount of document data;
A full-text search in the database for the input positive examples that have been input, and a pattern extraction unit that extracts patterns that appear around the input positive examples, and
A full text search of the pattern extracted by the pattern extraction unit, and a term extraction unit that extracts expressions of two or more terms extracted by the pattern and sorts them in descending order of scores,
The term extraction unit uses a combination of the extracted data of the plurality of patterns for the calculation of the score. A procedure for entering positive input examples that are multiple minority terms;
A procedure for storing a certain amount of document data in a database;
A full-text search in the database for the input positive examples that have been input, and a pattern extraction procedure for extracting patterns that appear around the input positive examples,
A full-text search for the extracted pattern in the database, and a term extraction procedure for extracting an expression extracted by the pattern;
A term extraction procedure for preventing an expression including a character type that was not in the input example from being extracted,
A program that causes a computer to execute.

Images