JPH03268065A - Article extracting system - Google Patents

Abstract

PURPOSE:To exactly segment plural articles into individual article by deriving a keyword containing in a document constituted of plural articles at every block of a character-string, and deciding to which keyword a block belongs, based on similarity of these keywords. CONSTITUTION:To a central processing unit 17, an image input part 18, storage device 19, input/output terminal 20, character-string block extracting mechanism 21, character segmenting mechanism 22, a character recognizing mechanism 23, keyword extracting mechanism 24, block deciding mechanism 25 and an article extracting mechanism 26 are connected. In such a state, a block being a unit for constituting an article is extracted from an input document image, and keywords are extracted from a character-string of each block thereof, respectively. Subsequently, similarity of each keyword is compared and based on a result of deciding whether each block constitutes the same article or not, a document is segmented into every article. In such a way, the document constituted of plural articles can be segmented exactly into individual article.

Description

DETAILED DESCRIPTION OF THE INVENTION [Object of the Invention] (Industrial Application Field) The present invention relates to an article extraction method for separating and extracting regions for each article from a document composed of a plurality of articles.

(Prior Art) In order to separate a document such as a newspaper consisting of a plurality of articles into individual articles, it is necessary to examine the connections among the blocks that make up the individual articles.

For example, as shown in Figure 2, if character string block 7 is connected to character string block 3, headline character string block 1 and body character string block 2°3.7 become article authors, and the headline 4 and the main text 5 and 6 become article B. Conversely, if character string block 7 has a connection relationship with character string block 6, character string block 7 belongs to article B.

Here, 8.9 is a ruled line used to clarify the boundaries of blocks.

As a method for investigating the connection relationship of such blocks, as described in Japanese Patent Publication No. 61-32712,
Blocks that may be connected, block 3 in Figure 2
Or check the last line of 6 and determine whether the end of that line ends with a small amount of black pixels equivalent to a period, and if it does not end with a period, there is always a continuation of the article in the next column. Based on this assumption, in the next row, the line with no spaces at the beginning of the rightmost line is extracted as the beginning of the continuation of the above article.

(Problem to be Solved by the Invention) However, if a block is judged only by the number of black pixels or the size of the last character of a block, a block ending with "," may be judged to be the last of the document. Conversely, there are cases where the person in charge of writing the document is written in parentheses () at the end of the document, and in such cases, there is a possibility that there is more to the document. be. Also,
In newspapers, where ruled lines play an extremely important role in defining the connections between text documents, various types of decorative ruled lines are used, and it is often difficult to extract the ruled lines accurately. If ruled lines are extracted incorrectly, it becomes extremely difficult to extract the connection relationships between blocks. For example, if a ruled line is made up of a collection of thin shapes, such as a broken line, one ruled line may be extracted as two ruled lines, or it may be extracted shorter than its actual length, or in some cases, it may not be extracted at all. There are things I can't do. As shown in Figure 3,
Ruled line 10 in FIG. 3(a) that should originally be extracted as a ruled line
is divided and extracted as two ruled lines 11 and 12, for example, as shown in FIG. 3(b), the character string block 13
It is not possible to determine whether the character string block 14 or the character string block 15 is connected to the character string block 15 only from the size and positional relationship of the figures.

The present invention has been made in view of the above, and its purpose is to eliminate the need to rely only on the position and size of the last character included in a block, and to eliminate the need to extract ruled lines etc. accurately. Another object of the present invention is to provide an article extraction method that can accurately determine the connections between blocks included in a document and accurately separate and divide a document composed of multiple articles into individual articles.

[Configuration of the invention J (Means for solving the problem) In order to achieve the above object, the article extraction method of the present invention is as follows:
A document reading device that divides document image data obtained by inputting a document containing multiple articles into regions for each article, and converts the character strings that make up each article into character code strings through character recognition. a central processing unit that controls processing; an image input unit that converts documents into images; and a storage device that stores input document images and the like;
An input/output terminal that displays the results of a series of processes and inputs commands, etc., a character string block extraction mechanism that extracts blocks that are the units that make up an article from the input document image, and the character string extraction block oil field A character extraction mechanism that extracts individual characters constituting a character string included in a block extracted by the mechanism; a character recognition mechanism that recognizes the individual characters extracted by the character extraction mechanism; A keyword extraction mechanism extracts keywords from the converted character string, and each block is extracted by comparing the similarity of the keywords extracted by the keyword extraction mechanism for each block extracted by the character string block extraction mechanism. A block determination mechanism that determines whether articles constitute the same article, and an individual article extraction mechanism that separates multiple articles arranged on the same paper page into articles based on the results obtained from the block determination mechanism. The gist is to have the following.

(Operation) In the article extraction method of the present invention, blocks, which are units that constitute an article, are extracted from an input document image, keywords are extracted from the character strings of each extracted block, and keywords are extracted from each block. The similarity of each extracted keyword is compared to determine whether each block constitutes the same article, and the document is divided into articles based on the determination result.

(Example) Hereinafter, the present invention will be explained in detail using the drawings.

FIG. 1 is a block diagram showing the configuration of an article extraction method according to an embodiment of the present invention. The article extraction method shown in the figure is
Central processing unit (CPU) that controls a series of processes, etc.
17, and the central processing unit 17 includes an image input section 1.
-8, storage device 19, input/output terminal 20, character string block extraction mechanism 21, character cutting mechanism 22, character recognition mechanism 2
3. A keyword extraction mechanism 24, a block determination mechanism 25, and an article extraction mechanism 26 are connected.

The image input unit 18 uses an image scanner or the like to quantize and sample the document, and stores the results in a memory bag M19.

The input/output terminal 2o is used to display intermediate results of processing, input necessary commands, etc. A character string block extraction mechanism 21 extracts a character string block that is a set of character strings within a document.

This character string block extraction mechanism 21 is, for example, Journal of the Institute of Electronics and Communication Engineers J69-D, No. 8. pp, 1187-119
6 (Showa 6l-08) [marginal distribution, linear density,
This can be realized by the method described in [Document Image Region Segmentation Using Circumscribed Rectangle Features]. The above-mentioned FIG. 2 shows a state in which the heading character string 1.4 and the main text blocks 2.degree. 3.5 and 6.7 have been extracted from the document.

If complete blocks cannot be extracted using these methods, display the input image etc. on a terminal device such as a CRT, and use a mouse or similar tool to remove incorrectly extracted blocks, unextracted blocks, or in some cases all blocks. You may input using a device.

Further, the character cutting mechanism 22 cuts out individual characters from the character strings constituting the character string blocks of the document extracted by the character string block extraction mechanism 21. This character cutting mechanism 22 is, for example,
D.No.10. Rip,! 194-1201 (Sho 59
This can be realized by "character cutting out from printed matter by non-contact character priority cutting" described in 10).

Furthermore, the character recognition mechanism 23 recognizes the individual characters cut out by the character cutout mechanism 22,
This is, for example, Journal of the Institute of Electronics and Communication Engineers J66-D, No. 1.
0. pp, 1185-1192, (Sho 58-10)
This can be achieved using the method described in ``Identification of handwritten kanji using contour direction contribution characteristics'' published in . It should be noted that the method described here is also effective for printed characters according to Technical Direction PRU86 of the Institute of Electronics and Communication Engineers Pattern Recognition and Understanding Study Group.
-33. pr), 31-40 [Automatic reading of unknown format documents].

The keyword extraction mechanism 24 performs morphological analysis of the recognized character code string, and extracts keywords based on the result. The keyword extraction mechanism 24 is, for example, IEICE Journal J65-D, No.

10, pp. 1195-1202. (1981-10) [Keyword Extraction by Text Analysis of Patent Claims] or NTTi Research and Practical Application Report Vol. 1.36. N
o, 9. I) I), 1151-1158 (1987)
“Keyword automatic extraction system (I)” described in
This can be achieved by the method described in NDXER) J.

The block determination mechanism 25 uses keywords extracted from different blocks to determine whether these blocks belong to the same article. In the block judgment mechanism 25, in addition to the criterion of how many of the same keywords are included, the distance between the included keywords is calculated using concepts such as higher rank and lower rank in the thesaurus, and based on that distance. Good too. The above-mentioned figure 2 shows the character string block 2.3.
contains many education-related keywords such as school and subject, and character string blocks 5 and 6 contain words such as semiconductor, DRAM,
This shows that keywords in the electronic field such as memory are included. In this state, since the character string block 7 containing keywords such as semiconductor and computer contains the same keywords as the keywords contained in blocks 5 and 6, or keywords in the same field, the block determination mechanism 25 It is determined that column blocks 5.6 and 7 belong to the same article A. The article extraction mechanism 26 determines the area of each article based on the result determined by the block determination mechanism 25. In Figure 2, string block 1
．． 2.3 is output as a single article, and 4, 5, 6, and 7 are output as separate articles. The results of dividing each article may be stored in the storage device 19, or may be displayed using the input/output terminal 20.

[Effects of the Invention] As explained above, according to the present invention, keywords included in a document consisting of a plurality of articles are obtained for each block of character strings, and blocks are created based on the similarity of these keywords. Since it determines which keyword the article belongs to, it is possible to accurately convert these multiple articles into individual articles, even if the last period of the article is not extracted or the borders etc. are not extracted accurately. It can be separated.

[Brief explanation of drawings]

FIG. 1 is a block diagram showing the configuration of an article extraction method according to an embodiment of the present invention, FIG. 2 is a diagram showing a sequence of documents consisting of two articles, and FIG. 3 is a block diagram of character strings formed by ruled lines. FIG. 17... Central processing unit, 21... Character string block extraction mechanism, 22... Character extraction mechanism, 24... Keyword extraction mechanism, 25... Block determination mechanism, 26... Article extraction mechanism .

Claims (1)

[Claims] A document reading device that divides document image data obtained by inputting a document containing multiple articles into regions for each article, and converts the character strings that make up each article into character code strings through character recognition. a central processing unit that controls processing; an image input unit that converts documents into images; and a storage device that stores input document images and the like;
An input/output terminal that displays the results of a series of processes and inputs commands, etc., a character string block extraction mechanism that extracts blocks that are the units that make up an article from the input document image, and the character string extraction block extraction A character extraction mechanism that extracts individual characters constituting a character string included in a block extracted by the mechanism; a character recognition mechanism that recognizes the individual characters extracted by the character extraction mechanism; A keyword extraction mechanism extracts keywords from the converted character string, and each block is extracted by comparing the similarity of the keywords extracted by the keyword extraction mechanism for each block extracted by the character string block extraction mechanism. A block determination mechanism that determines whether articles constitute the same article, and an individual article extraction mechanism that separates multiple articles arranged on the same paper page into articles based on the results obtained from the block determination mechanism. An article extraction method characterized by having the following.