JPH08221440A - Method and device for extracting keyword from network news article - Google Patents

Abstract

PURPOSE: To provide a method with which a retrieval key code suitable for retrieval can be extracted from a network news article. CONSTITUTION: Processing for excluding a header and a signature from the news article is performed (101-104). Any part to be quoted in the news article is specified and processing for raising a flag at that part is performed. A line feeding part in the news article is analyzed and an original sentence end is decided. Afterwards, a character string is extracted from the news article by executing morpheme analysis 107 regarding the partition of the types of characters important. The extracted character string is defined as an index keyword.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a method for extracting index keywords from news articles handled on a network and a device suitable for implementing the method.

[0002]

2. Description of the Related Art In order to be able to search for a desired article from a news article flowing through a network such as the Internet or a personal computer communication network by a keyword, an index keyword is extracted from the article in advance to index the article. There is a need. As one method for this, a morphological analysis developed by natural language processing, and a morphological analysis using a dictionary is used to extract index keyword candidates, and then an optimal index keyword is extracted from this candidate by the minimum matching method or maximum likelihood evaluation method. A possible method is to decide.

[0003]

However, when the morphological analysis method as described above is simply applied, the following problems occur.

(1). First, a news article flowing through the network has a unique structure such that it is composed of a header, a signature, a quotation part and a text. Here, the header is the group name of the article, the part that indicates who is the sender of the article, the signature is the part that describes the sender, and the citation section is when quoting a third-party article. The part where it is shown, the text, is the information that the sender wants to pass. When extracting the index key code from such a news article, it is better to exclude the header and the signature from the target for extracting the index key code of the text, and the quoting part can also be divided so that it can be recognized. This is because the index key code suitable for the text search can be extracted from the news article in this way. However, there has been no concrete method for achieving this in the past.

(2) Further, at the time of creating a news article, it is currently recommended that line breaks be made up of 30 full-width characters even if one sentence is not completed due to restrictions on character communication. ing. Therefore, the index keyword that should be originally extracted may be divided by a line feed, and as a result, a meaningless index keyword may be extracted. In order to avoid this, a method of completely ignoring the line feed code and setting the punctuation mark as one sentence and extracting the index keyword from there can be considered. However, in this case, for example, in the case of an article without punctuation marks such as a bulleted article, the keywords that should be divided this time are connected and the meaningless keyword is extracted. Therefore, it is desired to take measures against line feed symbols in network news articles.

(3) Further, in the case of the extraction method using the above-mentioned morphological analysis method, first the keyword candidates in the sentence are divided from the sentence by referring to the word dictionary, and then the optimum keyword is extracted by the minimum matching method or the like. However, in the case of a compound word formed by connecting kanji, for example, "electronic dictionary", it is preferable to divide it into the meaning and the retrieval efficiency as much as possible. Because, in the case of this "electronic dictionary", if "electronic" is not registered in the dictionary and only "electronic" is registered, "electronic" and "chemical dictionary" are index keywords. This is because the candidate is generated as a candidate, which causes a problem that a meaningless index keyword “chemical dictionary” is generated, and thus a meaningless search work is performed on the index keyword. In order to avoid this, compound words should be registered in the dictionary at appropriate times, but it is not realistic to register coined words generated every day in the dictionary.

[0007]

According to the method of extracting keywords from a network news article of the first invention of this application, when extracting an index keyword from a network news article, first, a header is extracted from the news article. At least the processing, the processing of extracting the signature from the news article, the processing of interpreting the quoted part in the news article, and the processing of interpreting the line break in the news article are performed at least based on predetermined rules. Next, a morpheme analysis that emphasizes character type delimitation is performed on the news article on which each of these processes has been performed, and a character string (including the case of one character; the same applies below) is extracted from the news article. Then, the character string (referred to as a primary keyword) is used as an index keyword.

In implementing the first aspect of the present invention, the extracted character string is further divided using a dictionary and a predetermined statistical division rule, and the character string obtained by the division is used as an index keyword. (It is called a secondary keyword.)
Is preferred. Of course, further dividing means dividing when the above-mentioned statistical dividing rule is satisfied (the same applies to the second invention below). Further, the primary keywords that have not been divided can be used as they are as index keywords (the same applies to the second invention below).

Further, according to the keyword extracting apparatus for network news articles of the second invention of the present application (hereinafter, sometimes abbreviated as an extracting apparatus), a header extracting section for extracting a header from a news article, and A signature extracting unit that extracts a signature from a news article, a quoting unit interpreting unit that interprets the quoting unit in the news article, a line break interpreting unit that interprets a line break in the news article, the header extracting unit, and the signature extracting unit. , A morphological analysis unit for performing a morphological analysis that emphasizes character type delimitation for a news article that has been at least processed by the quotation unit interpretation unit and the line feed interpretation unit.

[0010]

According to the first and second inventions of this application, the header and the signature are separated from the network news article. In addition, it is possible to grasp the state of quotes and line breaks in network news articles. Then, from the news article in which the header and the signature are separated and the state of the quoted portion and the line feed is grasped, the character string is extracted in the unit of the character type.

[0011]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS Embodiments of each invention of a keyword extracting method and apparatus from a network news article of this application will be described below with reference to the drawings. However, the drawings used in the description are merely schematic representations so that the invention can be understood.

FIG. 1 is a diagram showing the structure of an extracting apparatus 100 of the embodiment. 2 is a diagram showing an example of a news article. 3 to 9 show header extraction processing,
It is explanatory drawing of each processing, such as a signature extraction process, a quotation part interpretation process, a line feed interpretation process, a morpheme analysis, and a secondary keyword extraction process.

The extraction device 100 of the embodiment comprises a primary keyword extraction unit 100a and a secondary keyword extraction unit 100b. In this embodiment, the primary keyword extracting unit 100a includes a header extracting unit 101, a header item index base 102, a signature extracting unit 103, a signature base 104, a quoting unit interpreting unit 105, and a line break interpreting unit 1.
06, the morphological analysis unit 107, and the automaton 108
And a primary keyword base 109. The secondary keyword extraction unit 100b, in this embodiment, includes a keyword division unit 110, a dictionary 111, a division rule storage unit 112, and a secondary keyword base 113.

1. Extraction of primary keyword Header extraction unit 1 provided in the primary keyword extraction unit 100a
01 is for extracting the header 120a from the news article 120. In this example, news article 1
20 (see FIG. 2), the header section 120a is extracted and removed from the news article 120. Further, the content itself such as the sender, the subject, and the transmission time of the news article is extracted from the header 120a, and these are extracted as the header item index. It is stored in the base 102. This series of processes in the header extraction unit 101 can be performed by the procedure described below with reference to FIG. 3, for example. In the news article 120, the header 120a and the body 120d are usually separated by a blank line, and the header item in the header 120a and the content shown in the header item are usually separated by a ":" symbol. Therefore, the data for one line of the article for which the keyword is to be extracted is read from a file (for example, an arbitrary file such as a floppy disk. The same applies hereinafter) in which news articles are stored (step S1 in FIG. 3). Next, it is determined whether or not the beginning of a line is composed of an English word and the following ":" symbol (step S2 in FIG. 3). If so, it is the header 120a, so this data is registered in the header item index base (also called the header item dictionary) (step S in FIG. 3).
3). Then, the process returns to step S1 of reading one line of data from the article, and the processes after step S2 are repeated. In addition, if the answer is NO in step S2, specifically, if the line is a blank line, the header extraction process ends (step S4 in FIG. 3). Such a header processing unit 10
1 can be configured by known computer technology or circuit technology, for example. Further, the header item index base 102 can also be configured by a known storage means. In this embodiment, the header 12
The example in which 0a is removed from the news article 120 has been described, but a header processing unit may be used in which the flag corresponding to the header of the data indicating the news article 120 is flagged and left in the news article. In such a case, this flag may be used in the subsequent morphological analysis to perform processing such as removing the header from the analysis target.

Further, the signature extraction unit 103 detects the signature 120d from the news article 120 (see FIG. 2).
Is to be extracted. In the case of this embodiment, the signature 120d is extracted from the news article 120 and is removed from the news article 120. Further, data such as the sender of this news article, that is, the subject, and the time at which the signature was sent are stored in the signature base 10.
It is supposed to be stored in 4. Whether or not there is a signature in an article can be determined by regarding the same part when the last part of each article from each sender in the network is the same as the signature. Specifically, signature extraction can be performed by the procedure described below with reference to FIG. First, the buffer for signature extraction processing is cleared (step S11 in FIG. 4). Next, it is determined whether or not there is any signature file in the signature base 104 (see FIG. 1) (step S12 in FIG. 4). This judgment is
For example, it can be performed by searching the signature base with the data indicating the sender extracted in the header extraction processing and determining the presence or absence of the signature file. If there is a signature file, read the data for one line at the end of the article from which keywords are to be extracted and set this as A, and read the data for one line from the end of the signature file as B. (Step S1 in FIG. 4)
3). Next, it is determined whether or not the two data A and B match (step S14 in FIG. 4). Both data A and B
If they match, the data A is added to the beginning of the buffer (step S15 in FIG. 4). Then, the process returns to step S13 and the processes of S14, S15, and S13 are repeated. If the data A and B do not match in step S14, the data added to the buffer is stored as the signature of the caller in the signature base 104 (step S18 in FIG. 4). . If it is determined in step S12 that the signature file does not exist, it means that there is only one article by a certain sender, so the entire news article is copied to the buffer (step S16 in FIG. 4). , And stores it in the signature base 104 (step S17 in FIG. 4). However, the article stored in the signature base when it is determined that the signature file does not exist in step S12 is treated as the text when the keyword is extracted. Such a signature extraction processing unit 101 can be configured by, for example, known computer technology or circuit technology. In this embodiment, the signature 120
As described in the example of removing b from the news article 120,
The signature extraction processing unit may set a flag indicating that the signature is present in the relevant portion of the data indicating the news article 120 and leave the flag in the news article.
In such a case, this flag may be used in the subsequent morphological analysis to perform processing such as removing the signature from the analysis target.

The quoting section interpreting section 105 interprets the quoting section 120c in the news article 120. In the case of this embodiment, it is assumed that the process of identifying the quoted portion 120c in the news article 120 is performed. This processing can be performed, for example, by the procedure described below with reference to FIG. In general, a quoted part has a symbol at the beginning of the line indicating the quoted part. For example, a symbol in which a space symbol and a predetermined symbol are continuous, or a “>” symbol. It is a ":" symbol or a ">>" symbol. In the case of the news article 120 in FIG. 2, an example in which the quotation mark is “>>” is shown. Therefore, one line of data is read from the article subject to keyword extraction (step S2 in FIG. 5).
1). Next, it is determined whether the line head is a quotation mark (step S22 in FIG. 5). If so, the quotation mark is deleted, a flag indicating that the data in this line is a quotation is set, and this data is stored as a quotation in, for example, a predetermined memory (not shown) (step S23 in FIG. 5). ). And
Step S2 of reading one line of data from the article
The process returns to the process 1 and the processes after step S22 are repeated.
If the answer is NO in step S22, specifically, if the line is a blank line, the quotation part interpretation process is terminated (step S24 in FIG. 5). Such a quotation section interpreting section 105 can be configured by known computer technology and circuit technology, for example. As for the quoted portion, the line feed interpretation processing and the morpheme analysis processing, which will be described later, are also performed in the same manner as the main text as the extraction target sentence of the index key code. Of course, when the index key code is not extracted from the quotation part, the line feed interpretation process and the morphological analysis process may not be performed.

The line break interpreter 106 interprets a line break in an article. In this embodiment, even if one sentence is not completed, a line is broken by a certain number of characters due to the current character communication restriction, so that the problem that the character string to be the index keyword is divided is compensated for. It is intended for the process of determining. This processing can be performed, for example, in the procedure described below with reference to FIG. Normally, the text is composed by inserting blank lines for each line. Therefore, first, the full text of the news article targeted for keyword extraction is read (step S3 in FIG. 6).
1). Note that header extraction processing, signature extraction processing,
If the quoting portion interpretation processing and the line feed interpretation processing have been completed, the reading of the text in step S31 is easy. Next, it is determined whether or not a line feed is performed for each line (that is, whether or not there is a blank line for each line) (step S32 in FIG. 6). If so, ignore line breaks at the end of each line of the text, and
The line feed of only one line is also ignored, and the place where the line feed continues is the end of the sentence (step S33 in FIG. 6). On the other hand, when the line feed is not performed for each line, the line feed at the end of the line is ignored for each line of the text, and the line containing only the line feed is set as the end of sentence (step S34 in FIG. 6). After that, the line feed interpretation process ends (step S35 in FIG. 6). Such a line break interpretation processing unit can be configured by known computer technology and circuit technology, for example.

It should be noted that the header extraction process, the signature extraction process, the quotation part interpretation process, and the line feed interpretation process may be performed in any order as long as each process is performed before the morphological analysis. Specifically, it is preferable to execute the header extraction process, the signature extraction process, the quotation part interpretation process, and the line feed interpretation process in this order.

The morpheme analysis unit 107 also performs morpheme analysis on the news article that has been subjected to the header extraction process, the signature extraction process, the quotation part interpretation process, and the line feed interpretation process, mainly using the character type delimiters. It is a thing. The processing in the morphological analysis unit 107 can be performed by a method described below using an automaton, for example, but not limited to the automaton shown in FIG. First, as pre-processing for this morphological analysis, full-width alphanumeric characters in the text of the news article are converted to half-width characters. Then, the body is divided into character strings by using an automaton that performs a state transition as shown in FIG. That is, when a character that is not in charge of the character type appears in each state (kanji, hiragana, etc. in FIG. 7) in the automaton, the state returns to State None, and the character string up to that point is divided (output). This allows the text to be divided at the character type. The character string thus divided is used as an index keyword (which is also referred to as a primary keyword in the following relationship with a secondary keyword). This primary keyword is stored in the primary keyword base 109. If a character string that is too long is divided, it is determined that there is an abnormality, for example, that bullets are connected, and post-processing is performed. As this post-processing, for example, a processing for re-performing the line feed interpretation processing and dividing this character string can be cited. For the character string divided by the post-processing, the morphological analysis is performed again using mainly character type delimiters.

The text 1 of the news article 120 shown in FIG.
The primary keywords extracted from 20d are as follows. Free software ftp ftp. oki. co. jp
/ Pub / mac eudora-1.2.2J Handed over when entering the Internet. Further, the primary keywords extracted by processing the citation portion 120c in the same manner as the text are as follows. Macintosh Internet mail reading E
udora software How to get it. When there are other news articles in the file, the respective processes are performed on these as well, and the primary keywords are extracted.

The primary keywords obtained by the processing up to this point are from the main body (which may also be a quoted part) in which the effects of the header, the signature, the quoted part, and the line feed symbol have been removed, and the delimiter of the character type is mainly used. Since it is extracted by the morphological analysis used for, it is not affected by the header, etc., and the character string is not divided more than necessary (for example, it becomes an index keyword as it is without dividing electronic dictionary etc.) .

2. Secondary Keywords As described above, primary keywords can be index keywords, but secondary keywords may be further extracted from the primary keywords in order to further improve search efficiency. This will be described below.

The keyword dividing unit 110 of the secondary keyword extracting unit 100b (see FIG. 1) stores the primary keywords stored in the primary keyword base 109 in the dictionary 111 and the dividing rule storage unit 112 in advance. Further processing is performed using the statistical division rule described above, and the character string obtained by the division is used as a secondary keyword as an index keyword. Here, in this embodiment, the predetermined statistical division rule is the rule of the following conditions (1) to (4). : Each character string stored as a primary keyword in the primary keyword base is sorted in the order of Kanji code, and a specific character string s of 2 or more characters is included in the sorted character string group (the head of the character string). If there is a portion (referred to as a continuous appearance portion) in which the primary keyword (character string) included in () is consecutively appearing N or more, this continuous appearance portion is removed from the character string group. Here, N is a predetermined positive integer. This N will be referred to as an appearance count threshold value. : How many character strings are not in the primary keyword base 109 or the dictionary 111 in each of the remaining parts when the respective character strings of the extracted character string group are divided into the s part and the remaining part. When the count value m is equal to or larger than a predetermined number M, the process for removing the s portion (division process) is not performed, and otherwise, the process for removing the character string (division process) is performed. This M will be referred to as an unregistered number threshold value. : If there are other consecutive occurrences, perform the same process. : This time, the character strings stored in the primary keyword base are rearranged in reverse order (for example, "character string" means "string character sentence"). The secondary keyword extraction processing using the division rules of the conditions 1 to 3 will be described more specifically with reference to FIGS. 8 and 9.

First, the appearance number threshold N and the non-registration number threshold M are set. Here, N is 1
00 and M are set to 1, for example. For convenience of explanation, the primary keyword group stored in the primary keyword base 109 is referred to as A, and the word group in the dictionary is referred to as B (step S41 in FIG. 8).

Next, the primary keyword group A is sorted in Kanji code order (step S42 in FIG. 8).

Next, the sorted primary keyword group A
If there is a portion in which N or more consecutive keywords each having a specific character string s of two or more characters appear in N, the portion is extracted (step S43 in FIG. 8). This will be described with a specific example. If there are 100 or more keywords having the character string “electronic” as the specific character string s at the beginning of the sorted primary keyword group, those are extracted. Some of the extracted keywords are shown below. Note that these keyword groups with the initials of "electronic" are referred to as C for convenience of explanation. Electronic Electronic Dictionary Electronic Conference :: Electronic Communication Next, the "electronic" part is removed from each of the keyword groups C starting with electronic. The character string group from which the character string of "electronic" is removed is called D (step S44 in FIG. 8). This character string group D is as follows. Next, the number of words that are not included in the primary keyword group A or the word group B of the dictionary among the character strings of the character string group D from which the “electronic” character string is removed is counted and m is calculated. Obtained (step S45 in FIG. 8). In this example, the character string group D that is not included in the primary keyword group A or the word group B of the dictionary is the "chemical dictionary", so that m = 1.

Next, this m is compared with the non-registration number threshold M set above (step S46 in FIG. 8). Here, since M = 1 and m = 1, the process jumps to step S48. That is, the primary keyword group A is left as it is. Therefore, since the “electronic dictionary” and the like are included in the primary keyword group as they are, unnecessary key codes such as the “chemical dictionary” need not be extracted. If "electronic dictionary" is not in the primary key code group, m = 0 and m ≧ M is not satisfied, and the process jumps to step S47. That is, "electronic"
A character string group D (“dictionary”, “meeting”, ..., “Communication”) from which electronic character strings are removed is put in the primary key code group A instead of the keyword group C starting with an electronic character, A
This is combined with the other primary keyword groups remaining in A to make it A, and then this A is sorted in Kanji code order (step S47 in FIG. 8). In the case of the latter example, the number of characters of each keyword is smaller than that of the original keyword group C (which includes the electronic conference), so that the memory capacity for storing the keyword can be reduced, for example. . Even in that case, the keyword "electronic conference" can be obtained from "electronic" and "meeting". Also, keywords with good search efficiency can be obtained.

Next, the arrangement order of the characters of each key code of the primary key code group A is reversed (step S48 in FIG. 9). This is to prepare for extracting highly probable keywords that come to the end of the character string. Next, the primary key code groups in which the respective key codes are in the reverse order are sorted in the Kanji code order (step S49). Next, a sequence of occurrences is detected in the same procedure as in the case of the above word to obtain a keyword group C, the specific character s is removed from each keyword of each keyword group to obtain a character string group D, and this character string group D is obtained. Then, m is obtained from this, and this m is compared with M, and processing is performed according to the result of this comparison (step S5 in FIG. 9).
0). In the series of processes of S48 to S50, highly probable keywords that come to the end of the character string can be extracted.

Next, the character arrangement of each keyword, which has been reversed in step S48, is reversed again to restore the original character arrangement of each keyword (step S51 in FIG. 9).

Next, it is determined whether or not there is a division target (step S52 in FIG. 9). Depending on the determination result, the process jumps to either step S42 or step S53.

According to this secondary keyword extraction processing, it is possible to divide a part-of-speech that is easy to start at the primary keyword and can be completely divided, and a part-of-speech that is easy to end at the primary keyword and can be completely divided. It can be divided and these divided ones can be stored in the secondary keyword base 109 as secondary keywords. Of course,
Secondary keywords can be extracted from compound words such as technical terms without being divided.

The secondary keyword obtained after performing the above-described secondary keyword extraction processing on the primary keyword obtained from the text of the news article 120 in FIG. 2 is “case”,
It will be "handed". Therefore, news article 12 in FIG.
The index key code extracted from the body of 0 is as follows. In other words, from "case handing" to "case",
"Handing" will be added. Free software ftp ftp. oki. co. jp
/ Pub / mac eudora-1.2.2J Internet handed over.

Although the embodiments of the method and apparatus for extracting keywords from the network news article of this application have been described above, these inventions are not limited to the detailed embodiments. For example, the procedures of the header extraction processing, the signature extraction processing, the quotation portion interpretation processing, and the line feed interpretation processing are not limited to the above-mentioned examples, and other procedures according to the object of the present invention may be used. Further, the present invention can be applied to a program synthesis, a knowledge acquisition system for an expert system, and the like.

[0034]

As is apparent from the above description, according to the keyword extracting method from the network news article of the first invention, the header extracting process, the signature extracting process, the quoting part interpreting process and the line feed interpreting process are performed on the news article. After at least performing each process, a morphological analysis that emphasizes character type separation is performed to extract a character string from the news article, and the character string is used as an index keyword. Therefore, the news article in which the header and the signature are separated is used. And from the news article in which the state of the quotation part and the line feed was grasped,
A character string is extracted for each unit of character type. For this reason, it is possible to prevent the influence of the header, signature, quotation, and line feed, and to prevent the search keywords that are meaninglessly divided from being extracted. Therefore, it is possible to extract index keywords suitable for news articles including compound words. be able to.

Further, the keyword extracting device from the network news article of the second invention facilitates the implementation of the first invention.

[Brief description of drawings]

FIG. 1 is a diagram showing a configuration of an extraction device according to an embodiment.

FIG. 2 is an explanatory diagram of a news article.

FIG. 3 is an explanatory diagram of header extraction processing.

FIG. 4 is an explanatory diagram of signature extraction processing.

FIG. 5 is an explanatory diagram of a quotation portion interpretation process.

FIG. 6 is an explanatory diagram of line feed interpretation processing.

FIG. 7 is an explanatory diagram of a morpheme analysis, and is an explanatory diagram of an example of an automaton used for it.

FIG. 8 is an explanatory diagram of secondary keyword extraction processing.

FIG. 9 is an explanatory diagram of the secondary keyword extraction processing continued from FIG. 8;

[Explanation of symbols]

 100: Extraction Device of Example 100a: Primary Keyword Extraction Unit 100b: Secondary Keyword Extraction Unit 120: News Article 120a: Header 120b: Signature 120c: Citation Section 120d: Body

 ─────────────────────────────────────────────────── ─── Continuation of the front page (72) Inventor Hiromi Hanyuda 1-7-12 Toranomon, Minato-ku, Tokyo Oki Electric Industry Co., Ltd. (72) Inventor Tetsuo Kinoshita 1-12-12 Toranomon, Minato-ku, Tokyo Oki Electric Industry Co., Ltd.

Claims (8)

[Claims] 1. When extracting an index keyword from a network news article, a process of extracting a header from a news article, a process of extracting a signature from the news article, a process of interpreting a quoted part in the news article, and the news article. At least each process of interpreting a line break in the above is performed, and a morpheme analysis that emphasizes the delimitation of the character type is performed on the news article on which each process is performed to extract a character string from the news article, A method for extracting a keyword from a network news article, which is characterized by using a character string as an index keyword. 2. The method of extracting keywords from a network news article according to claim 1, wherein the extracted character string is further divided using a dictionary and a predetermined statistical division rule, and the division is performed. A method for extracting a keyword from a network news article, characterized in that the obtained character string is used as an index keyword. 3. The method of extracting keywords from a network news article according to claim 2, wherein the division processing using a dictionary and a statistical division rule sorts the extracted character string groups in Kanji code order. Each character string of the character string group in which the character string having a specific character string s of 2 or more characters at the beginning of the sorted character string group continues for N or more is divided into the character string s and the rest. A method of extracting a keyword from a network news article, wherein the division is performed when the number that the remaining portion does not exist in the extracted character string group or dictionary is m. Only when m is smaller than a predetermined value M, where N and M are predetermined positive integers.). 4. The method of extracting keywords from a network news article according to claim 3, wherein a character string group in which the character arrangement order of the character strings of the character string group is reversed is generated, and the character string group is generated. A method of extracting a keyword from a network news article, further comprising: 5. An apparatus for extracting an index keyword from a network news article, a header extracting section for extracting a header from a news article, a signature extracting section for extracting a signature from the news article, and a quoting section in the news article. A quote section interpreting section that interprets a line break in the news article, a line break interpreting section that interprets a line break in the news article, and a news article that is at least processed by the header extraction section, the signature extraction section, the quote section interpretation section, and the line break interpretation section. A keyword extraction device from a network news article, comprising: a morphological analysis unit that performs morphological analysis with an emphasis on character type delimiters. 6. The keyword extraction device from a network news article according to claim 5, wherein the character string divided by the morpheme analysis unit from the news article is divided using a dictionary and a predetermined statistical division rule. A keyword extracting device from a network news article, further comprising a secondary keyword extracting section for obtaining a secondary keyword as an index key code. 7. The apparatus for extracting keywords from network news articles according to claim 6, wherein the secondary keyword extraction unit sorts the group of character strings divided by the morpheme analysis unit in Kanji code order. Each character string of the character string group in which a character string having a specific character string s of two or more characters at the beginning of the completed character string group continues for N or more is divided into the character string s and the remaining portion, A device for extracting keywords from network news articles, characterized in that these divided character strings are used as secondary keywords (however, in the division, the remaining part is a group of the extracted character strings or in a dictionary). It is limited to the case where m is a number that does not exist in M, and m is smaller than a predetermined value M. Here, N and M are predetermined positive integers.). 8. The apparatus for extracting keywords from a network news article according to claim 7, wherein the secondary keyword extracting unit is a character string group in which the character arrangement order of the character strings of the character string group is reversed. And a keyword extracting device from a network news article, characterized in that the dividing process is further performed on the character string group.