JP2006004283A - Method and system for extracting/narrowing keyword from text information source - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To appropriately extract keywords from text documents so as to perform retrieval even in an environment that has a small recognition vocabulary. <P>SOLUTION: A keyword extracting/narrowing system has a noun extracting part 33 to extract nouns from text documents that morpheme analysis is performed; a compound noun extracting part 34 to extract compound nouns based on the extracted nouns; a katakana characters extracting part 35 to extract katakana characters based on the extracted nouns; a compound noun dividing part 36 to divide a compound noun into a plurality of nouns; a noun connection part 37, which connects noun elements using a noun-connection-postpositional-particle "no" if the noun-connection-postpositional-particle "no" is inserted between noun elements; a keyword narrowing part 38 to remove keywords, which are acquired through processings until the noun connection part 37 and satisfy specified conditions; and a keyword correlation narrowing part 39 to further narrow keywords, which are narrowed in the keyword narrowing part 38, considering appearance frequency. <P>COPYRIGHT: (C)2006,JPO&NCIPI

Description

  The present invention relates to a method and system for extracting and narrowing keywords from a text information source, and more particularly, keyword extraction and narrowing that are effective for performing information access from an environment where the number of recognized vocabulary words in speech recognition is limited. It relates to a method and a system.

  It is necessary to input a keyword for information access and search using the Internet or the like, or for a destination search in a car navigation system.

  By the way, compared with a fixed environment such as an office or home, in a mobile environment, restrictions on input / output interfaces due to hardware limitations are severe. Especially in the in-car environment, the driving task has the highest priority, and the car navigation device operation such as destination search and setting, and the telematics service such as information search and information browsing should not hinder driving. Therefore, a hands-free / eye-free voice interface is used in many car navigation systems. When a voice interface is used for inputting keywords or the like, a voice recognition technique is applied.

  Speech recognition includes word recognition that performs matching between dictation, which is continuous speech recognition, and words in a recognition dictionary. However, in a high noise environment such as in an automobile, it is difficult to ensure a sufficient recognition rate by dictation with a high degree of freedom. Furthermore, even in word recognition, the size of the recognition dictionary and the recognition rate are in a trade-off relationship, and it is necessary to narrow down to a few hundred recognition words in order to ensure a realistic recognition rate. More specifically, a trade-off between the recognition vocabulary scale and the recognition rate occurs at a position where the recognition vocabulary is small in a high noise environment in a car. Even in a quiet environment, even a system that can recognize tens of thousands to hundreds of thousands of vocabulary, even in a noisy environment, a few hundred vocabulary is the upper limit for obtaining a reasonable recognition rate. .

  Currently, search engines widely used on the Internet, etc., register tens of thousands to hundreds of thousands of keywords, and further search by a thesaurus is possible for many vocabularies. It is very difficult to apply to the search in the environment where the number of recognized vocabulary is limited. That is, a search engine widely used on the Internet assumes that a user interface that allows a user to input a search term is used so that a user can input a large number of vocabularies. A large number of keywords are set for a document to be searched in order to perform a search using one of the search terms. Specifically, when keywords that characterize the document are automatically extracted from each document in the search target document group and associated with the document, as many keywords as possible are assigned to the document and input. It ensures that word hits are made. However, when speech recognition in a noisy environment is used as a user interface for inputting search terms, the number of vocabularies that can be used as search terms is limited because the number of recognizable words is limited. Therefore, the detection efficiency is not improved even if keywords are added to the search target document. Rather, it is preferable that a relatively small number of accurate keywords be assigned. Of course, there is also a method of assigning keywords with human intervention, but in order to assign keywords to a large number of documents with various contents, a huge amount of people and time are required, which is not realistic.

  In addition, for now, it is possible to recognize words that are static words such as menu names. In other words, dynamic word speech recognition that should support the provision of dynamic information, which is a feature of telematics services, has not been done so far. For example, when performing a news search, it is required to set a dynamic keyword to enable voice recognition. In order to extract, set and recognize a dynamic recognition vocabulary, it is considered that (1) a recognition vocabulary extraction method and (2) a recognition vocabulary narrowing-down method are required.

Currently, as a voice recognition technology for searching for destinations and entering commands in a car navigation system for high noise environments such as in automobiles, automakers or car navigation system manufacturers have set up a website on the Internet. The thing described in the nonpatent literatures 1-5 which are established is known.
“Internavi Premium Club” homepage provided by Honda Motor Co., Ltd., [Online], May 25, 2004 search, Internet <URL: http://premium-club.jp/PR/> "G-BOOK" homepage provided by Toyota Motor Corporation, [Online], searched on May 26, 2004, Internet <URL: http://gazoo.com/g-book/index.html> "Carwings" homepage provided by Nissan Motor Co., Ltd. [Online], searched on May 26, 2004, Internet <URL: http://www.nissan-carwings.com/> “Air Navi” website provided by Pioneer Corporation, [Online], May 26, 2004 search, Internet <URL: http://www.air-navi.com/> “MyDriveNet.Com” homepage provided by Alpine, Inc. [Online], May 26, 2004 search, Internet <URL: http://www.mydrivenet.com/mdnc/index.cgi>

  In consideration of the current voice recognition technology and the like, it is important to narrow down information when accessing information from a mobile environment, particularly an in-car environment. That is, since the number of vocabularies that can be input as search terms is limited, it is necessary to narrow down the keywords that are finally extracted when keywords are automatically extracted from the document to be searched.

  Accordingly, an object of the present invention is to provide a keyword extraction / narrowing method and system capable of automatically extracting a narrowed keyword from each document of an information source when a text document such as a news article is used as an information source. is there.

  A first keyword extraction / narrowing method of the present invention is a method for extracting a keyword from a text information source, acquiring text information from the text information source, performing morphological analysis on the text information, and extracting the text information. A compound noun composed of a noun extracted from text information decomposed into parts of speech, a noun is extracted from the text information decomposed into parts of speech to generate a set of simple nouns, and a complex noun composed of continuous nouns from the text information decomposed into parts of speech Generating at least one of a compound noun set composed of nouns and a katakana noun set composed of katakana words composed of three or more katakana characters.

  The second keyword extraction / narrowing method of the present invention is a method for extracting keywords from a text information source, the step of acquiring text information from the text information source, performing morphological analysis on the text information, A compound noun composed of a noun extracted from text information decomposed into parts of speech, a noun is extracted from the text information decomposed into parts of speech to generate a set of simple nouns, and a complex noun composed of continuous nouns from the text information decomposed into parts of speech A step of generating a noun set and determining whether the compound nouns in the compound noun set are partially matched with elements included in at least one of a simple noun set, a compound noun set, and a katakana noun set, and based on the determination result Dividing the compound noun to generate an extended compound noun set.

  A third keyword extraction / narrowing method according to the present invention is a method for extracting keywords from a text information source, the step of acquiring text information from the text information source, performing morphological analysis on the text information, When there is a syntactic particle “no” sandwiched between noun elements in the text information stage and the text information, those noun elements are combined via the syntactical particle “no”, and the noun group Generating.

  In each keyword extraction / narrowing method described above, an element included in any of the noun sets is set as a keyword candidate, and when the keyword candidate satisfies a predetermined condition, the keyword candidate is deleted and the keyword is narrowed down. Preferably further steps are provided. Predetermined conditions include: (1) the keyword candidate is a 2-mora word or 1-mora word, (2) the proportion of numbers in the keyword candidate is greater than or equal to a predetermined value, (3) keyword extraction performed in the past At least one of (4) candidate keywords appearing in news articles at a certain ratio or more across a plurality of news categories can be used. That is, when one or more of the four conditions (1) to (4) are satisfied, the corresponding keyword candidate may be deleted assuming that the predetermined condition is satisfied. Furthermore, when the keyword candidate is a noun element ending with “Mr.”, a step of deleting the word “Mr.” at the end may be provided.

  The first keyword extraction / narrowing system of the present invention is a system for extracting keywords from a text information source, an external interface for acquiring text information from the text information source, performing morphological analysis on the text information, and text information. A morphological analysis unit that decomposes into parts of speech, a noun extraction unit that extracts a noun from text information decomposed into parts of speech and generates a simple noun set, and a continuous noun from text information decomposed into parts of speech Means for generating at least one of a compound noun set composed of compound nouns composed of compound nouns and a katakana noun set composed of katakana words composed of three or more characters.

  A second keyword extraction / narrowing system according to the present invention is a system for extracting keywords from a text information source, an external interface for acquiring text information from the text information source, performing morphological analysis on the text information, and text information. A morphological analysis unit that decomposes into parts of speech, a noun extraction unit that extracts a noun from text information decomposed into parts of speech and generates a simple noun set, and a continuous noun from text information decomposed into parts of speech A compound noun extraction unit that generates a compound noun set composed of compound nouns, and a partial match between at least one of a simple noun set, a compound noun set, and a katakana noun set for compound nouns in the compound noun set And a compound noun dividing unit that divides compound nouns based on the determination result and generates an expanded compound noun set. That.

  A third keyword extraction / narrowing system according to the present invention is a system for extracting keywords from a text information source, an external interface for acquiring text information from the text information source, performing morphological analysis on the text information, and text information. Morphological analysis part that decomposes into part of speech, and in the text information, if there is a coalesced particle `` no '' sandwiched between noun elements, these noun elements are combined via the merging particle `` no '', A linking unit that generates a linking noun set.

  In the keyword extraction / narrowing system of the present invention, an element included in any of the noun sets is a keyword candidate, and when the keyword candidate satisfies a predetermined condition, the keyword candidate is deleted and the keyword is narrowed down It is preferable to provide a narrowing part. As the predetermined condition, it is possible to use at least one of (1) the keyword candidate is a 2-mora word or 1-mora word, and (2) the ratio of the number in the keyword candidate is a predetermined value or more. In addition, the keyword extraction / narrowing system of the present invention updates the appearance frequency database for storing the appearance frequency in the keyword extraction performed in the past and the application frequency database using the elements included in any of the noun sets as keyword candidates. And a keyword correlation narrowing unit that deletes keyword candidates whose appearance frequency stored in the appearance frequency database is equal to or greater than a predetermined value.

  In the present invention, as nouns are mainly extracted from text articles, not only simple nouns but also compound nouns and katakana nouns are extracted, and further, keywords are extracted by dividing compound nouns. Furthermore, an appropriate keyword can be obtained by using a method of narrowing down the extracted keyword candidates based on the number of characters, the ratio of numbers, the appearance frequency, and the like.

  The present invention has an effect that an appropriately narrowed keyword can be automatically obtained when keywords are extracted from text information. By using such keywords, an appropriate recognition vocabulary can be extracted while suppressing the recognition dictionary size when inputting a search word by voice recognition or the like. Therefore, the keyword extraction according to the present invention has a high affinity for a voice input interface in a high noise environment such as in an automobile. In addition, keywords can be obtained while processing text information. By using the obtained keywords as recognition vocabulary sequentially, dynamic word speech that supports the provision of dynamic information that is a feature of telematics services. Recognition can be realized.

  Next, a preferred embodiment of the present invention will be described with reference to the drawings. FIG. 1 is a block diagram showing a configuration of a keyword extraction / narrowing system according to an embodiment of the present invention. This keyword extraction / narrowing system 10 uses morphological analysis, and uses a “recognition vocabulary extraction method” that focuses on the extraction of nouns and compound nouns, and a “recognition vocabulary reduction method” that considers appearance frequency and the like. Extraction and refinement.

  The keyword extraction / narrowing system 10 shown in FIG. 1 acquires an external interface 31 that acquires a text document such as a news article as text information from an external information source 20 that is a text information source, and a text document acquired by the external interface 31. A morpheme analysis unit 32 that performs morpheme analysis, a noun extraction unit 33 that extracts a noun from a text document subjected to morpheme analysis, and generates a simple noun set, and extracts a compound noun from the extracted nouns, and a compound noun set A compound noun extraction unit 34 that generates a katakana word from the extracted nouns to generate a katakana noun set, a compound noun included in the compound noun set, a simple noun set, a compound noun A compound noun splitting unit 36 for generating an extended compound noun set by dividing the set and the compound noun set with reference to the set; When the particle “no” is sandwiched between noun elements, the noun elements are connected with the noun element “no” to generate a noun group, and a simple noun set, compound noun set, katakana Each element of the noun set, the extended compound noun set, and the combined noun set is used as a keyword candidate, and the keyword narrowing-down section 38 that removes those keyword candidates that satisfy a specific condition, and the keyword narrowing-down section 38 A keyword correlation narrowing unit 39 that further narrows down the keyword candidates in consideration of the appearance frequency, and an appearance frequency database (DB) 40 that is used when the keyword correlation narrowing unit 39 narrows down the keywords. ing. The keyword extracted and narrowed down from the text document is output from the keyword correlation narrowing unit 39 as a recognized vocabulary output.

  Next, the operation of the keyword extraction / narrowing system 10 will be described. Here, the case where a keyword is extracted from each news article for news search will be described. The system 10 of the present embodiment first extracts nouns from news articles that are text documents, and extracts keyword candidates using extraction methods such as compound nouns, katakana nouns, and compound noun division. After that, keywords are narrowed down from a set of keyword candidates by using a restriction method based on the number of characters, a ratio occupied by numbers, an appearance frequency, and the like. This makes it possible to extract appropriate keywords while reducing the number of keywords. The keyword extracted in this manner makes it possible to extract an appropriate recognition vocabulary (that is, a search word corresponding to the keyword) while suppressing the size of the recognition dictionary when inputting a search word by voice recognition. In this case, each extracted keyword is stored in the recognition dictionary.

  First, in the keyword extraction / narrowing system 10, the external interface 31 acquires a news article from the external information source 20, and the morpheme analysis unit 32 divides the news article into morphemes.

《Keyword extraction》
When the division into morphemes is completed, a keyword extraction process is executed next. FIG. 2 shows an example of keyword extraction. The noun extraction unit 33 extracts the morphemes by paying attention to the nouns. The extracted noun set is named “simple noun set”. Given a text article (original sentence) as shown in FIG. 2A, the simple noun set is as shown in FIG. 2B. Subsequently, the compound noun extraction unit 34 extracts a compound noun composed of two or more consecutive nouns. This set of compound nouns is named “compound noun set”. A compound noun set extracted from the original sentence shown in FIG. 2A is shown in FIG. Next, the katakana word extraction unit 35 extracts katakana words (words composed only of katakana) having three or more characters as nouns (such as proper nouns) of unknown words. This set of Katakana words is named “Katakana noun set”. FIG. 2D shows a katakana noun set extracted from the original sentence shown in FIG.

  Subsequently, the compound noun dividing unit 36 divides compound nouns in the compound noun set. This is done using simple noun sets, katakana noun sets and compound noun sets. Specifically, when there are simple nouns, katakana nouns or (other) compound nouns that partially match compound nouns, the previous and subsequent parts that partially match are newly added as an “extended compound noun set”. To register. In the example shown in FIG. 2, the nouns “Kyoto”, “Fu”, “Tamba”, and “Machi” exist consecutively in the original sentence, and these four nouns are stored in a simple noun set. “Tanba-cho, Kyoto” is stored in the compound noun set. The compound noun dividing unit 36 checks whether the vocabulary in the simple noun set partially matches the vocabulary in the compound noun set. Here, the four nouns “Kyoto”, “Fu”, “Tamba”, “Town” partially match “Tanba-cho, Kyoto”, so the first part after the part that matches “Kyoto” = “Fu” "Tamba-cho" is divided, then the front part of the part that matches "Fu" = "Kyoto" and the rear part = "Tamba-cho" is divided, and then the front part of the part that matches "Tamba" = “Kyoto Prefecture” and the rear part = “town” are divided, and finally, the front part of the part matching “town” = “Kyoto Prefecture Tamba” is divided. Here, "Kyoto" and "Town" are excluded because they are already in the simple noun set, and "Fu-Tamba-cho", "Tamba-cho", "Kyoto-fu", and "Kyoto-fu Tamba" are stored in the expanded compound noun set. Is done. FIG. 2E shows the extended compound noun set obtained in this way.

  Next, the linking unit 37 performs linking with “no”, which is a linking particle. If there are simple nouns, compound nouns, katakana nouns, or extended compound nouns (collectively called noun elements) across the coalesced particle “no”, these are connected and registered in the noun set. In this case, the noun elements before and after the union particle “no” are compound nouns, extended compound nouns, katakana nouns, and simple nouns in the order of priority. Is possible. FIG. 2 (f) shows a noun set obtained in this way.

<Keywords>
A simple noun set (FIG. 2B), a compound noun set (FIG. 2C), a katakana noun set (FIG. 2D), an extended compound noun set (FIG. 2) obtained by the above processing. (E)), each element of the noun group (FIG. 2 (f)) is a keyword candidate extracted from the original sentence (FIG. 2 (a)). However, some of these keyword candidates do not realize an efficient search when input as a search term. If all of these keyword candidates are keywords, the total number of keywords is too large when the number of recognized vocabulary is limited. Therefore, the keyword is narrowed down. In the present embodiment, as keyword narrowing, both simple narrowing by the keyword narrowing unit 38 and correlation narrowing by the keyword correlation narrowing unit 39 are performed in this order.

  First, single narrowing will be described. The keyword narrowing unit 38 narrows down according to the following rules.

(1) Delete two-mora words and one-mora words:
Delete a two-letter or one-letter vocabulary. In addition, compound nouns and extended compound nouns that start with simple nouns of 2 Mora or 1 Mora are deleted. Mora is a unit of prosody in Japanese corresponding to syllable in English.

(2) Delete vocabulary with a fixed percentage of numbers:
Vocabulary in which the number occupies a certain percentage or more is deleted because it is less important as a recognized vocabulary.

(3) Delete “Mr” from words that end with “Mr”:
Since the voice name in the voice interface hardly includes “Mr” in the name of the person, that is, since “Mr” is rarely added as an input as a search term, “Mr” is deleted.

  FIGS. 3A to 3E show sets obtained as a result of performing the above-described simple narrowing down in the sets shown in FIGS. 2B to 2F.

  Next, correlation narrowing will be described. The keyword correlation narrowing unit 39 stores the keywords extracted in the keyword extraction in the appearance frequency database 40 in the form of vocabulary, number of appearances, and number of parameters each time the keywords are extracted from the news articles and narrowed down. The parameter is the total number of news articles processed by the keyword extraction / restriction system 10, that is, the number of news articles that have been extracted / restricted by key thirds. In the record, the value of the parameter is the same.

  The keyword correlation narrowing unit 39 refers to the appearance frequency database 40 with respect to the keywords narrowed down by the keyword narrowing unit 38, searches for keywords whose number of appearances exceeds a certain number, and deletes them. As a result, general keywords that appear frequently and do not represent the characteristics of the news article are removed.

  FIGS. 4A to 4E show sets obtained as a result of performing correlation narrowing in the sets shown in FIGS. 3A to 3E. For example, in the simple noun set shown in FIG. 4A, from the set after simple narrowing (FIG. 3A), “environment”, “occurrence”, “national”, “local”, “site” Keywords that are too general and unlikely to be useful for search are removed.

  FIG. 5 shows the result of merging the sets after eliminating the duplication of keywords in each set for each set after the correlation narrowing ((a) to (d) in FIG. 4). Shows the recognition vocabulary obtained as keywords. In order to search for news articles by voice recognition, keywords are extracted and narrowed down for each article in this way, and the finally obtained keywords are stored as recognition vocabulary in a recognition dictionary for voice recognition. .

  In the keyword correlation narrowing down, it is possible to expect a more precise narrowing down by removing keywords having a certain appearance frequency as general keywords across a plurality of news genres such as economy, politics, and international.

  The keyword extraction / narrowing described above can also be realized by causing a personal computer or the like to read a computer program for realizing the keyword and executing the program. A program for keyword extraction / narrowing is read into a computer by a recording medium such as a magnetic tape or a CD-ROM, or via a network. Such a computer generally includes a CPU (Central Processing Unit), a hard disk device for storing programs and data, a main memory, an input device such as a keyboard and a mouse, a display device such as a CRT, a CD- It comprises a reading device that reads a recording medium such as a ROM, and a communication interface that connects to a network and acquires text information from an external information source. The text information may be stored in advance in the hard disk device. The hard disk device, main memory, input device, display device, reading device, and communication interface are connected to the CPU. In this computer, a recording medium storing a program for extracting and narrowing keywords is loaded into a reading device, and the program is read from the recording medium and stored in a hard disk device, or such a program is stored on a hard disk via a network. When the CPU executes the program downloaded to the device and then stored in the hard disk device, it functions as the keyword extraction / narrowing system 10 described above. Such a program or a recording medium storing the program is also included in the category of the present invention.

It is a block diagram which shows the structure of the keyword extraction and narrowing-down system of one Embodiment of this invention. It is a figure which shows an example of the process of keyword extraction. It is a figure which shows the example of simple keyword narrowing down. It is a figure explaining the example of keyword narrowing down by correlation narrowing down. It is a figure explaining the example of a recognition vocabulary.

Explanation of symbols

DESCRIPTION OF SYMBOLS 10 Keyword extraction / narrowing system 20 External information source 31 External interface 32 Morphological analysis part 33 Noun extraction part 34 Compound noun extraction part 35 Katakana word extraction part 36 Compound noun division part 37 Linkage part 38 Keyword refinement part 39 Keyword correlation narrowing part 40 Appearance frequency database (DB)

Claims (12)

A method for extracting keywords from a text information source,
Obtaining text information from the text information source;
Performing morphological analysis on the text information and decomposing the text information into parts of speech;
Extracting a noun from the text information decomposed into parts of speech to generate a simple noun set;
From the text information decomposed into parts of speech, a compound noun set composed of compound nouns composed of compound nouns composed of consecutive nouns, and a katakana noun set composed of katakana words composed of three or more characters, Generating at least one of the
A keyword extraction / narrowing method. A method for extracting keywords from a text information source,
Obtaining text information from the text information source;
Performing morphological analysis on the text information and decomposing the text information into parts of speech;
Extracting a noun from the text information decomposed into parts of speech to generate a simple noun set;
Generating a compound noun set composed of compound nouns composed of consecutive nouns from the text information decomposed into parts of speech;
The compound noun in the compound noun set is determined to partially match an element included in at least one of the simple noun set, the compound noun set, and the katakana noun set, and the compound noun is determined based on the determination result. Dividing and generating an extended compound noun set;
A keyword extraction / narrowing method. A method for extracting keywords from a text information source,
Obtaining text information from the text information source;
Performing morphological analysis on the text information and decomposing the text information into parts of speech;
In the text information, when there is a coalesced particle `` no '' sandwiched between noun elements, the noun elements are combined via the coalesced particle `` no '', and a noun group is generated,
A keyword extraction / narrowing method.   4. The method according to claim 1, further comprising a step of defining an element included in any of the noun sets as a keyword candidate, deleting the keyword candidate when the keyword candidate satisfies a predetermined condition, and narrowing down the keyword. The keyword extraction / narrowing method according to claim 1. The predetermined condition is:
(1) The keyword candidates are 2-mora or 1-mora.
(2) The ratio of numbers in the keyword candidates is equal to or greater than a predetermined value.
(3) It is a keyword candidate whose appearance frequency in keyword extraction performed in the past is a predetermined value or more.
(4) A keyword candidate that appears in news articles of a certain ratio or more across a plurality of news categories.
The keyword extraction / narrowing method according to claim 4, comprising at least one of the following.   The keyword extraction / narrowing method according to claim 4, further comprising a step of deleting the ending “Mr” when the keyword candidate is a noun element ending with “Mr”. A system for extracting keywords from a text information source,
An external interface for obtaining text information from the text information source;
Morphological analysis is performed on the text information, and the text information is decomposed into parts of speech.
A noun extraction unit that extracts a noun from the text information decomposed into parts of speech and generates a simple noun set;
From the text information decomposed into parts of speech, a compound noun set composed of compound nouns composed of compound nouns composed of consecutive nouns, and a katakana noun set composed of katakana words composed of three or more characters, Means for generating at least one of
A keyword extraction / narrowing system. A system for extracting keywords from a text information source,
An external interface for obtaining text information from the text information source;
Morphological analysis is performed on the text information, and the text information is decomposed into parts of speech.
A noun extraction unit that extracts a noun from the text information decomposed into parts of speech and generates a simple noun set;
A compound noun extractor that generates a compound noun set composed of compound nouns composed of consecutive nouns from the text information decomposed into parts of speech;
The compound noun in the compound noun set is determined to partially match an element included in at least one of the simple noun set, the compound noun set, and the katakana noun set, and the compound noun is determined based on the determination result. A compound noun divider that divides and generates an extended compound noun set; and
A keyword extraction / narrowing system. A system for extracting keywords from a text information source,
An external interface for obtaining text information from the text information source;
Morphological analysis is performed on the text information, and the text information is decomposed into parts of speech.
In the text information, when there is a linking particle “no” sandwiched between noun elements, a linking unit that combines these noun elements via the linking particle “no” to generate a linking noun set. When,
A keyword extraction / narrowing system.   The keyword refinement unit further includes a keyword refinement unit that refines a keyword by using an element included in any of the noun sets as a keyword candidate, and deleting the keyword candidate when the keyword candidate satisfies a predetermined condition. The keyword extraction / narrowing system according to any one of the above. The predetermined condition is:
(1) The keyword candidates are 2-mora or 1-mora.
(2) The ratio of numbers in the keyword candidates is equal to or greater than a predetermined value.
The keyword extraction / narrowing system according to claim 10, comprising at least one of the following. Elements included in any of the noun sets are keyword candidates,
An appearance frequency database for storing appearance frequencies in keyword extraction performed in the past,
While updating the application frequency database, a keyword correlation narrowing unit that deletes keyword candidates whose appearance frequency stored in the appearance frequency database is a predetermined value or more;
The keyword extraction / narrowing system according to claim 7, further comprising:

Images