JP2011070529A - Document processing apparatus - Google Patents

Abstract

A document structure extracting apparatus capable of efficiently extracting bibliographic data is provided.
An information storage unit 16 includes an extraction target metadata definition dictionary 161 describing an identifier of bibliographic information to be extracted and a metadata-specific feature definition dictionary 162 describing a feature amount of each metadata. Based on the document column structure determination result, character lines and their order are extracted, and for each character line, a metadata score is calculated based on the feature amount by metadata stored in the information holding unit 16, and based on the character line position. , Calculate the weighted moving average of metadata likeness from each metadata score, find the maximum point of any metadata score of the metadata to be extracted from the beginning position of the character line, and find two continuous maximums Among the points, the boundary position of the bibliographic information label is determined by the valley portion of the score value.
[Selection] Figure 1

Description

  In addition to electronic documents, the present invention labels various academic literature documents digitized by multifunctional machines, etc., with titles, author names, abstracts, English creator names, English abstracts, etc., providing academic information provision services, The present invention relates to a document processing apparatus that improves the efficiency of document reuse.

  In government offices and private companies, in addition to atypical operations such as pre-sales activities and planning, a large number of documents are generated every day in routine operations such as examination and application. These documents are usually only digitized, the contents are not labeled (metadata extraction), and are asleep in various document management systems without being organized. However, in recent years, there has been a growing demand for efficient access to stored documents from various viewpoints, such as compliance, labor productivity improvement, and customer service improvement, and for document reuse, knowledge sharing, response to inquiries, etc. . To do so, it recognizes the bibliographic information (title, subtitle, creator, creation date, etc.) described in the document and characteristic information (quotation deadline, order number, etc.) for each document, Corresponding management techniques are important.

In addition, academic information service institutions, etc., as a back-end system for literature search service sites, Japanese titles, Japanese creators, English titles, English creators, abstracts, English abstracts from a wide variety of academic documents (images or electronic documents) There is a need to efficiently extract bibliographic information such as the text area and cited reference information. Currently, bibliographic information is manually extracted and registered in a database for various academic documents and published on a literature search site. Bibliographic information is expressed in different ways (order and position) for each academic document / issued paper, making automation difficult.
Several techniques for extracting metadata of bibliographic information from such an atypical document have been disclosed. Patent Document 1 discloses a technique for acquiring document structure information such as a table of contents and chapters based on information on the character size, character type, line spacing, and head character position of a document. Non-Patent Document 1 discloses a technique for extracting a document structure based on a document model describing a correspondence relationship between a layout structure and a logical structure.

JP 2001-265762 A

Kise, “Document structure analysis based on layout model”, IEICE theory, 72-D2, 7, pp1029-1039

  However, the rule definition type document structure extraction apparatus as found in Non-Patent Document 1 can only process a document under a specific layout condition. For this reason, layout analysis rules can be used for several types of documents, but if you intend to support hundreds to thousands of types of academic documents that are provided in academic information services, a great deal of rule creation and maintenance is required. costly.

  Further, Patent Document 1 can be applied only to a relatively limited layout such as a table of contents or a chapter chapter. It is not possible to deal with cases in which the order or position in which bibliographic information such as columns (columns), author names, organization names, etc. are written is different or omitted for each academic document type.

  In view of such circumstances, the object of the present invention is to extract bibliographic information such as titles, author names, abstracts, etc. for various academic literature documents, and provide a scientific information providing service and a document processing apparatus that makes document reuse efficient. Is to provide.

  In order to achieve the above object, in the present invention, the document processing includes an information holding unit and a processing unit, and extracts metadata of bibliographic information by the processing unit based on the appearance position and the character string in the document of the character line. The information holding unit holds a feature amount of each metadata included in each metadata, and the processing unit extracts a character line and a character line position based on a result of the column structure determination of the document. Provided is a document processing apparatus that calculates a metadata score for each character line based on another feature amount and determines a boundary position of bibliographic information based on the metadata score.

  In the present invention, the processing unit further executes a propagation process of the metadata score for each metadata based on the character line position, and any one of the metadata to be extracted from the head position of the character line. Find the maximum point of the metadata score after the propagation processing of, determine the boundary position based on the change of the metadata score between the two consecutive maximum points searched, the character string before the determined boundary position Sets the metadata for the first local maximum point, and sets the metadata for the second local maximum point in the subsequent text line.

  Further, in the present invention, the processing unit executes the metadata score propagation process by performing a smoothing process by weighted moving average calculation of the metadata score for each piece of metadata based on the character line position.

  Furthermore, in the present invention, the information holding unit holds at least one of the number of characters constituting the character line, the English name pattern length, the average distance between words, and the average distance between characters as a feature amount for each metadata.

  In the present invention, preferably, in order to achieve the above object, the information holding unit preferably has an extraction target metadata definition dictionary describing an identifier of metadata to be extracted, and a metadata describing feature values of each metadata. A data-specific feature definition dictionary is prepared, and the processing unit (1) extracts character lines and their order based on the column structure determination result of the document. (2) Each character line is associated with the information holding unit. The metadata score is calculated based on the feature amount of each metadata stored, and (3) the metadata score propagation process is executed by calculating the weighted moving average of each metadata based on the character line position. , (4) Find the maximum point of any metadata score among the metadata to be extracted from the beginning position of the character line, and (5) Score value valley between two consecutive maximum points Bibliographic information boundary by part To provide a document processing apparatus for determining a location.

  As a more preferred aspect, the present invention includes an input unit capable of inputting document information, an information holding unit, a processing unit, and an output unit, and an appearance position in a document of a character line including a set of characters, and A document processing apparatus that extracts bibliographic information based on a character string, wherein the processing unit extracts character elements from the input document information, and generates a projection component histogram in the horizontal direction and the vertical direction of the extracted character elements. Generate and extract column information from the projection histogram, extract character lines based on the order of characters determined by the coordinate values of character elements and column information, and column information that includes the coordinates and character lines of the extracted character lines The character lines are rearranged as series data, and a metadata score is calculated for each rearranged character line based on the feature amount of each metadata stored in the information holding unit in advance. Based on the line position, execute the metadata score propagation process for each piece of metadata, search for the maximum point of the metadata score after the propagation process from the first character line, and between the two consecutive maximum points searched The boundary point is set based on the change in the metadata score of the first, the first maximum point metadata is set in the character line before the set boundary point, and the second maximum point is set in the subsequent character line. Provided is a document processing apparatus configured to set point metadata.

Similarly, as a more preferable aspect, the present invention includes at least an input unit capable of inputting document information, an information holding unit, a processing unit, and an output unit. A document processing device that extracts bibliographic information based on an appearance position and a character string, wherein the processing unit extracts a character element from the input document information, and based on the coordinates of the extracted character element, Generate a projection component histogram in the vertical direction, extract column information extraction from the projection component histogram, extract character lines based on the order of characters determined by the coordinate values of the character elements and column information,
Based on the extracted character line coordinates and column information including the character line, the character line is rearranged as series data, and for each rearranged character line, based on the feature amount of each metadata stored in the information holding unit in advance. The metadata score is calculated, the cost when each character line is the metadata is set based on the metadata score of each character line, and the metadata stored in advance in association with the information holding unit Provided is a document processing apparatus configured to search for a metadata label sequence that minimizes the total cost of an entire character line based on identifier order candidates.

  According to the present invention, bibliographic information can be extracted from a wide variety of academic documents without defining the document structure for each document.

  Further, even when the number of academic literature types to be extracted increases, it is possible to extract bibliographic information metadata with high accuracy without newly creating a document structure definition.

  Furthermore, even when a plurality of documents are bundled and printed by a multifunction peripheral or the like, bibliographic information metadata can be extracted for each document, and mixed documents can be separated into individual documents.

It is a figure which shows schematic structure of the bibliographic information extraction system based on a 1st Example. It is a whole flowchart figure of the metadata extraction flow based on 1st Example. It is a figure which shows the example of the column structure extraction of a cover area | region based on 1st Example. It is a flowchart figure which shows the outline | summary of the column structure extraction process of a cover area based on 1st Example. It is a flowchart figure which shows the outline | summary of the vertical direction division | segmentation process based on 1st Example. It is a figure which shows an example of the projection component histogram of the X-axis and Y-axis direction used in the case of the column structure extraction of a cover area based on 1st Example. It is a flowchart figure showing the outline | summary of the character line extraction process based on 1st Example. It is a figure which shows the example of the character line extraction result based on 1st Example. It is a figure which shows the example which ordered and rearranged the extracted character line based on 1st Example. It is a figure which shows an example of the table of the feature definition dictionary classified by metadata which defined the feature-value used for the score calculation of each metadata based on a 1st Example. It is a figure explaining the word break determination method based on a 1st Example. It is a figure which shows an example of the table of the characteristic definition dictionary classified by metadata which defined the score calculation characteristic for every metadata based on 1st Example. It is a figure which shows an example of the table of the metadata feature definition dictionary which defined the restriction | limiting feature for every metadata based on a 1st Example. It is a flowchart figure showing the outline | summary of the calculation process of each metadata score of each character line based on a 1st Example. It is a figure which shows the example which drawn the metadata score based on 1st Example in order of a character line. It is a figure which shows the example which drawn the metadata score after a weighted moving average based on a 1st Example in order of a character line. It is a flowchart figure which shows the outline | summary of the metadata area | region determination process using the multiple metadata score graph based on 1st Example. It is the figure which illustrated the process which determines a boundary position from two local maximum points based on a 1st Example. It is a figure which shows an example of the metadata score graph with respect to a series of documents which copied the some literature continuously based on a 2nd Example. It is a figure which shows the example of the relationship between a character line position and a metadata score based on 3rd Example. It is a figure showing the example of a search of the minimum cost path | pass from the relationship matrix of a character line position and metadata score based on 3rd Example. It is a figure which shows the data structure which represented the possible metadata order by the tree structure based on 3rd Example. It is a flowchart figure which shows the outline | summary of the minimum cost label series search by a metadata score based on 3rd Example. It is a figure which shows the example of the metadata label cost between character lines based on a 3rd Example. It is a figure which shows an example of the table of the extraction object metadata definition dictionary which defined the metadata of extraction object in the bibliographic information extraction system shown in each Example.

  As described in detail below, the present invention, in addition to electronic documents, various academic literature documents digitized with a multifunction machine, etc., labeling the title, author name, abstract, English creator name, English abstract, etc. That is, it provides a document processing apparatus that extracts metadata that is bibliographic information, provides an academic information providing service, and makes document reuse more efficient.

  In this specification, “bibliographic information” refers to the Japanese title, Japanese creator, English title, English creator, abstract, English abstract, etc. in the academic document (image or electronic document). Bibliographic information in a narrow sense (title, subtitle, creator, creation date, etc.) and information characteristic of each document (estimation deadline, order number, etc.) are collectively referred to. Further, the “metadata score” means a value representing the metadata likeness of each metadata by a numerical value, and the larger the numerical value, the more likely it is that metadata. Further, the “character line position” means a character line position in the document or a relative value of the Y coordinate of the character line from the beginning of the document. Furthermore, “propagation processing” is a generic term for smoothing processing such as simple moving average and weighted moving average.

  <Configuration of metadata extraction device>

  FIG. 1 is a diagram illustrating a schematic configuration of a metadata adding apparatus, that is, a document processing apparatus constituting a bibliographic information extraction system according to the first embodiment. Here, in addition to an electronic document, a paper document digitized by a multifunction machine or the like is assumed.

  The bibliographic information extraction system 10 includes an input unit 11 including an input device 110 and an image input device 111, an output unit including a display device 12 and a printing device 14, a processing unit including a central processing unit (CPU) 13, and a work area 15. And a storage unit including the information holding unit 16.

  In addition to the operating system (OS) 151, the work area 15 already includes a communication program 152 and a document processing program 153, or loads them from the information holding unit 16 as necessary. The information holding unit 16 includes various dictionaries required by the document processing program 153.

  Examples of the input device 110 include a keyboard, a mouse, and a tablet for inputting data to be input, commands, and the like to the document processing program 153.

  Examples of the image input device 111 include devices such as a scanner for capturing a document as image data when a paper document is a processing target.

  The OS 151 has a function of controlling operations of the input unit 11, display device 12, CPU 13, printing device 14, communication program 152, document processing program 153, metadata editing program 154, other memory (not shown), and storage device. The communication program 152 has a communication function for acquiring a document to be treated via a network.

  The document processing program 153 according to the present exemplary embodiment has a function of extracting metadata (bibliographic information) using a document image input by the input device 111 or electronic document data acquired by the communication program 152 as a processing target. The data editing program 154 has a function of editing the obtained metadata.

  The information holding unit 16 in the memory (not shown) includes an extraction target metadata definition dictionary 161 and a metadata-specific feature definition dictionary 162. As will be described in detail later, these dictionaries function as a dictionary database that is referred to when the document processing program 153 extracts metadata.

  The extraction target metadata definition dictionary 161 stores a dictionary describing a list or tree structure of identifiers representing bibliographic information to be extracted. The metadata-specific feature definition dictionary 162 stores a dictionary that describes the feature amounts of character lines labeled with each metadata.

  The display device 12 is a device such as a display for displaying the metadata extraction result by the document processing program 153 and the editing result by the metadata editing program 154. The CPU 13 loads various programs in the work area 15 on the memory (not shown), and executes the contents of the programs in cooperation with the OS 151. The printing device 14 is a device for outputting the conversion result of the input character string by the document processing program 153. The communication network 19 is a device for accessing data, a work area, and an information holding unit on another device such as the file server 21 connected via the network.

<Overall flow of metadata extraction>
FIG. 2 is an overall flowchart of the bibliographic information metadata extraction flow in the first embodiment.

  In the input P-A 101, the document data to be processed is loaded into the work area 15 through the input device 110 and the communication network 19 shown in FIG. In the column structure determination process P-A102, a column structure of an input document is determined in order to correctly extract a character line even for an academic document having a plurality of column structures. Details of the determination of the column structure of the input document will be described in FIG. In the character line extraction process P-A103, a character line serving as a labeling unit is extracted based on the column structure information (column information) extracted in the column structure determination P-A102. In the one-dimensional series processing PA-104, the character line extracted by the character line extraction process P-A103 is included in either the XY coordinates of the character line or the column (column) determined by the column structure determination P-A102. Sort by. In the character line feature setting P-A105, layout features such as line height, line position, average inter-character distance, and line character string pattern, which are stored in advance in the information holding unit 16 of FIG. 1 in association with each metadata. Based on the number of characters and the number of characters, the language and layout feature amount of each character line is set.

  In the score calculation process P-A106 for each metadata, the metadata likeness specified in the extraction target metadata definition dictionary is calculated and stored as a metadata score based on the feature amount extracted by the feature amount calculation P-A105. In the weighted moving average calculation P-A107, the weighted moving average is calculated according to the order extracted by the one-dimensional series processing P-A104 and the distance between the character lines for the metadata score calculated in the score calculation process P-A106. In the boundary position determination process P-A 108, the boundary area of each metadata and the metadata label for each area are determined based on the distribution pattern of each metadata score propagated to the adjacent area by the weighted moving average. In this specification, as described above, propagating a metadata score to an adjacent region by a simple moving average or a weighted moving average is collectively referred to as a score propagation process.

  According to the above processing flow, according to the present embodiment, even for hundreds to thousands of academic documents having an atypical layout, the metadata of the bibliographic information is appropriately obtained by the small dictionary of the extraction target metadata. Can be extracted. Further, the extraction can be performed at a higher speed than when the order is defined for each document and collation is performed for each document template. Furthermore, even when a plurality of documents are bundled and printed by a multifunction machine or the like, metadata can be extracted for each document, and mixed documents can be classified into individual documents.

  Hereinafter, details of each process of FIG. 2 in the present embodiment will be described with reference to FIGS.

<Column structure determination P-A102>
As shown in the example of FIG. 3, the bibliographic information description method for academic literature is not constant depending on each academic society and publisher. The cover layouts F-A001 to F-A003 shown in the upper part of the left part, the center part, and the right part of FIG. 3 are schematic representations of bibliographic data description patterns in a certain academic document. In addition, column structures F-A101 to F-A103 shown in the lower part of the figure represent column structures corresponding to them, and numbers such as circle 1, circle 2, and circle 3 represent region numbers. For example, on the cover F-A001, “Title”, “Author Name”, “English Title”, “English Author Name” are listed in one column, and “Summary” is double column. It is written at the top position in the left column. On the cover page F-A002, after bibliographic information, “Title”, “Author name”, and “Summary” are listed in one column, the organization information to which the author belongs appears in the lower left area of the double column. ing. In the cover F-A003, the entire page has a double column structure, and “title”, “author name”, and “abstract” are described from the top of the left column. In addition to the layout variations as described above, for example, in some academic literature, there are explicit keywords that start bibliographic data such as “paper”, “abstract”, “keyword”, but other academic literature. Then there can be situations where there are no such keywords. In the first embodiment of bibliographic metadata extraction, first, variations of these layouts are determined, and character lines are expressed as one-dimensional series data.

  Hereinafter, the detailed flow of the column structure determination P-A102 in FIG. 2 will be described with reference to FIGS. FIG. 4 is a flowchart showing an overview of the column structure determination processing of the cover area. Step P-B001 represents the vertical division process of the document area, and Step P-B002 represents the horizontal division process of the document area. As an example of vertical division, the division of circles 1 and 2 of the column structure F-A103 in FIG. 3, the division of circles 2 and 3 of the column structure F-A101, the circles 2 and 3 of the column structure F-A102 Can be divided.

  First, in the vertical division process P-B001, an attempt is made to execute vertical division based on the projection component histogram in the X-axis direction of the elements in the page. The division succeeds or fails depending on the pattern of the histogram. Details of this division processing will be described with reference to FIG. If it is successful, it is determined that the column is a double column, and the process ends. If the process fails, the horizontal division process P-B002 is executed. If the process is successful, the vertical division process P-B003 is executed again for the lower area after the division. When processing flow P-B101 is followed, a column structure such as column structure F-A103 in FIG. 3 is obtained, and when processing flow P-B102 is followed, column structure F-A101 and column structure F- in FIG. A column structure like A102 is obtained.

  FIG. 5 shows a detailed flow of the vertical region dividing process P-B001 of FIG. In process P-C001, a projection component to the X-axis is extracted for the text, ruled line, and image element in the cover page, and a projection component histogram is generated. The frequency of the histogram is the sum of the widths of each element in the Y-axis direction. Next, in process P-C002, a process of propagating the metadata score histogram extracted in process P-C001 to the adjacent region is performed. For the propagation process, a smoothing process such as a simple moving average or a weighted moving average can be used. The reason for performing the smoothing is to remove indentations that become noise in the next process P-C003. Next, in process P-C003, an X coordinate that is a candidate for division is searched. Specifically, when the histogram is searched in the ascending order of the X axis, a point where the shape of the histogram is concave is found (processing P-C004). Then, in process P-C005, when the histogram value of the concave point is equal to or less than a predetermined division threshold value with respect to the histogram value of the immediately previous convex point, it is determined that the division can be performed at the X coordinate. If an X coordinate that satisfies the division threshold is not found, the division fails. Satisfying the division threshold here means a case where the ratio of the histogram value of the concave point to the immediately preceding convex point is a certain value or less. Similarly, a projection component on the Y-axis can be extracted and a projection component histogram can be generated.

  FIG. 6 shows an example of the projection component histogram in the X-axis and Y-axis directions described above. A histogram F-B001 shown in the upper part of FIG. 6 represents an example of a projection component histogram (after smoothing) in the X-axis direction. A histogram F-B002 shown in the lower part of the figure represents an example of a projection component histogram (after smoothing) in the Y-axis direction. Now, it is determined whether or not the X coordinate F-B101 of the concave point is a division point (processing P-C003 to P-C005 in FIG. 5). In this case, the immediately preceding convex point is the point F-B102, and the ratio for the determination in the process P-C005 is the ratio of the frequency β1 of the point F-B101 to the frequency α1 of the point F-B102, the value of β1 / α1 Is used. Also in the horizontal division process P-B002, the same process is performed using the projection component histogram in the Y-axis direction. For example, when it is determined whether or not the Y coordinate F-B201 of the concave point is a dividing point, if the value of the ratio β2 / α2 is equal to or less than a predetermined threshold value, the horizontal division is performed.

  By the method described in detail above, the column structure as shown by the column divisions F-A101, F-A102, and F-A103 in FIG. 3 can be determined.

<Character line extraction P-A103>
The character line extraction process P-A103 is performed using the result of the column structure determination process P-A102 of FIGS. 4 and 5 and the XY coordinates of each character element.

FIG. 7 is a flowchart showing an outline of the character line extraction P-A 103 in FIG. In process P-D001, the character set in the cover page is rearranged based on the XY coordinates and the column structure. For example, if the XY coordinates are expressed with the lower left origin,
Order of column areas that contain characters (upper left → lower right)
Sort the string according to the ascending order of the Y coordinate of the character and the descending order of the X coordinate of the character.

  Next, in process P-D002, character lines are extracted in order from the first character. The character is sequentially connected to the current character line until the character is traced from the beginning until it is determined as a new character line. When a line breaks, it starts as a new line.

  An example of a character line will be described with reference to FIG. FIG. 8 shows a character line extracted from the column structure F-A101 of FIG. An element F-C301 represents a character, and an internal number represents the order of characters rearranged by the process P-D001. The character line F-C201 represents the character line of the column region circle 1, and line breaks are made at the character number 9 and the character number 10. Therefore, the character line F-C201 constitutes the character line F-C201, and the character number 10 is broken. Therefore, it becomes the first character of a new line F-C202. In the column area circle 2, since a line break is made at the character number 23, the character numbers 19 to 22 constitute a character line F-C203.

<One-dimensional series P-A104>
Subsequently, the one-dimensional series processing P-A104 of the character lines in FIG. 2, which is the step of rearranging the character lines, uses the character lines extracted by the character line extraction processing P-A103 as the character rearrangement. Rearrange according to order. For example, if the XY coordinates are expressed with the lower left origin,
(1) Order of column area containing character lines (order from upper left to lower right)
(2) The character lines are extracted in the order of ascending order of the Y coordinate of the character line (3) descending order of the X coordinate of the character line.

  FIG. 9 shows an example in which character lines are made into a one-dimensional series. In FIG. 9, a cover page F-D001 on the left side represents a character line extraction result for the column structure F-A101 in FIG. Elements A to J in the cover page F-D001 represent character lines extracted by the above processing. The one-dimensional series data F-D101 shown on the right side of the figure represents character lines ordered according to the criteria (1) to (3).

<Characteristics calculation for each character line P-A105>
In the feature amount calculation P-A105 of each character line in FIG. 2, the layout feature amount and the character string feature amount of each character line are calculated. A table T-A100 in FIG. 10 shows an example of the feature-by-metadata definition dictionary 162 in FIG. 1 showing the feature amounts used in the metadata providing apparatus of the present embodiment. In the figure, item T-A001 represents the identification ID of each feature quantity. Item T-A002 represents the name of each feature. Item T-A003 represents the contents of each feature amount. Item T-A004 represents an example of weight calculation for each feature amount. Items T-A002, T-A003, and T-A004 constitute the definition of the feature amount specified by the item T-A001.

  The feature amount of identification ID 1 represents the number of words. The unit of the word is determined based on the space. FIG. 11 is a schematic diagram illustrating the explanation of word breaks based on spaces. A rectangle F-E101 represents a character. The rectangles b to h that follow the side also represent characters. Each character is separated by a space of a certain width. Here, based on the relative change of the space, it is determined whether or not it is a word break space. For example, the space F-E001 is relatively wider than the spaces F-E002, F-E003, and F-E004. This can be determined using a threshold relating to the space ratio or a relative threshold (such as page width) relating to the width difference. By counting the number of word break spaces by the above means, the number of words that is the feature amount of ID1 can be calculated.

  The feature amount of the identification ID No. 2 represents the number of characters constituting the character line. This can be used, for example, as one of the determination features of the gist and text area. Although the number of characters may be used as it is, the ratio of the character line extracted in the character line extraction process P-A103 to the maximum number of characters may be used.

  The feature amount of the identification ID No. 3 represents a pattern that appears characteristically in English names. In English names, as in “Taro HITACHI”, surnames are often expressed in uppercase letters and first names in uppercase and lowercase letters. Therefore, the ratio of the characters corresponding to this pattern to the number of character lines is used.

  The feature amount of the identification ID No. 4 represents a pattern that appears characteristically in an English word. In this case, the ratio of the alphabet + symbol character to the number of character lines is used.

  The feature amount of the identification ID No. 5 represents the number of characters of the personal name word. For the determination of a personal name word, morphological analysis technology or collation with a personal name dictionary can be used. As the feature value, the ratio of the number of pattern characters to the number of character characters is used.

  The feature amount of the identification ID No. 6 represents the number of characters of the organization name word. For the determination of the organization name word, morphological analysis technology or collation with the organization name dictionary can be used. As the feature value, the ratio of the number of pattern characters to the number of character characters is used.

  The feature amount of the identification ID No. 7 represents the number of characters of the place name word. For the determination of the place name word, morphological analysis technology or collation with the place name dictionary can be used. As the feature value, the ratio of the number of pattern characters to the number of character characters is used.

  The feature amount of the identification ID No. 8 represents the number of punctuation characters in the character line. By using the ratio of the number of punctuation characters to the number of words, it can be used as one of the features of the English name pattern.

  The feature amount of the identification ID No. 9 represents the average distance between words. For the word determination, the same means as that used for the feature amount of identification ID 1 is used. Since Japanese personal names tend to have more space, they can be used as one of the characteristics of Japanese personal names.

  The feature amount of the identification ID No. 10 represents the average distance between characters. The average distance between characters in Japanese names tends to be large because 1) there are fewer characters than titles and abstracts, and 2) there is a large space for word breaks. It can be used as one of the characteristics of Japanese names.

  The feature amount of the identification ID No. 11 is a feature amount based on the number of the character line that is one-dimensionally arranged in the process P-A104. For example, the following equation can be used for a title or the like that has a strong tendency to appear at the top.

  By using Equation 1, the weight of the top row is the highest at 1, and the remaining rows take values between 1 and 0 in descending order.

  The feature amount of the identification ID No. 12 represents the number of Japanese characters. This feature amount can be used as a constraint condition for a Japanese region, an English region, etc. in academic literature.

  The feature amount of the identification ID No. 13 represents the item dictionary lookup result. This feature quantity defines what can define an item word for each metadata, and uses an edit distance from each character line. The reason for using the edit distance is that there is a misread in the OCR result document. In addition, it is considered that the item word originally matches the head of the character line, but it is also possible that a symbol or the like is inserted and does not come to the head. By using the edit distance, it is possible to handle a case where a match is made in the middle of a line as a score, and it is possible to prevent a metadata determination error due to an item reading failure. By using the reciprocal of “edit distance + 1”, the item dictionary matches from the beginning of the character line, and the score is 1 in the case of complete match. In addition, the score takes a value of 1 to 0 each time the character is shifted from the beginning or inserted or missing.

  The feature amount of the identification ID No. 14 represents the relative positional relationship with the item word. When an item word exists, a character line that appears before the item word does not become the metadata, so that the feature amount can be used as a constraint condition.

  The feature quantities exemplified above are examples of what is considered to be effective for metadata determination of bibliographic information, and are not limited to those listed here. By using these feature amounts in combination, bibliographic information labeling, that is, metadata determination, can be performed on the character line extracted by processing P-A103.

<Metadata score calculation P-A106>
Next, in the metadata score calculation of the metadata score calculation P-A106 in FIG. 2, a metadata-like score is calculated according to the metadata-specific feature definition dictionary 162 recorded in association with each metadata. For example, the feature amount shown in FIG.

  FIG. 25 shows a table T-D100 showing the types of metadata to be extracted in the extraction target metadata definition dictionary 161 of FIG. An item T-D001 represents a meta ID that is an identifier representing a type of metadata that is bibliographic information. An item T-D002 represents a metadata name that is bibliographic information. An item T-D003 indicates whether or not it is an extraction target. When the value of the item T-D003 is “◯”, it indicates that it is an extraction target, and when it is “○”, it indicates that it is not an extraction target.

  FIG. 12 shows a table T-B100 in which feature amounts used for score calculation of metadata that is bibliographic information in the feature definition dictionary by metadata 162 of FIG. 1 are defined. An item T-B001 in the table represents an identifier meta ID indicating the type of metadata that is bibliographic information. Item T-B002 represents a metadata name which is the content of the bibliographic information. An item T-B003 represents an ID of a feature amount used for score calculation. Each ID of the feature amount corresponds to the item number T-A001 of the table T-A100 of the extraction target metadata definition dictionary 161 in FIG. For the score of each metadata, refer to the feature item T-B003 used for the score calculation of the table T-B100 in FIG. 12, and the item number defined therein is the weight calculation item T in the table T-A100 in FIG. Obtained by calculating the linear sum of the values of -A004. The calculation method is not limited to the linear sum, and for example, a value learned from a large sample by discriminant analysis may be used.

  FIG. 13 shows a table T-C100 that defines the constraint conditions (feature items to be constrained) that each metadata should satisfy in the metadata feature definition dictionary 162 of FIG. In the table T-C100, an item T-C001 represents an identifier indicating the type of metadata. Item T-C002 represents a metadata name. An item T-C003 represents an ID list of features used as constraints, and corresponds to the item number T-A001 in FIG. Item T-C004 represents a threshold value (lower limit value) that each constraint feature value should satisfy. Based on the table T-C100, first, whether or not the bibliographic metadata constraint condition is satisfied is calculated, and the metadata score is calculated only for those satisfying the constraint.

  The flow of the above metadata score calculation P-A106 can be represented by a processing flow as shown in FIG. In the metadata score calculation processing, first, it is determined by processing P-E001 whether or not the constraints for the metadata defined in the extraction target metadata definition dictionary 161 in FIG. 1 are satisfied (P-E001). The determination is based on whether or not the feature amount threshold illustrated in FIG. 13 is satisfied. A score of 0 is assigned to metadata that does not satisfy the condition. For metadata that satisfies the condition, in process P-E002, a metadata score is calculated (P-E002). For the score calculation, the feature value illustrated in FIG. 12 is used. The above processing is performed for all character lines or character lines up to a certain number of lines from the beginning (P-E003).

<Weighted moving average calculation P-A107>
FIG. 15 illustrates the score of each metadata calculated in process P-A106 in FIG. 2 in association with the position of the character line. An axis F-F200 is an axis indicating the value of the metadata score. The value is larger to the left. The axis F-F300 represents the position of the character line extracted by the process P-A104. Indicates the line number from the first line, or the Y coordinate on the series data up to the character line. Graph F-F001 illustrates the metadata score of the title among the bibliographic information. Graph F-F002 displays the metadata score of the author. Graph F-F003 displays the metadata score of the English title. For example, in the graph F-F001, the score of the first line is high, and the score is lowered thereafter. In addition, the graph F-F003 showing the English title has the highest score in the several lines from the top, and the score is low thereafter, but a portion with a suddenly high score appears halfway.

  As described above, by calculating the metadata score for each character line, it is possible to determine the metadata region and its boundary to some extent. However, in reality, there is not much difference between the scores of a plurality of metadata, and the metadata boundary may not be correctly determined only by the metadata score value. Therefore, in the present embodiment, in addition to the metadata score based on the feature amount, conversion processing to a metadata score graph reflecting layout information is performed.

  FIG. 16 shows a metadata score graph obtained by converting the metadata score based on the feature amount exemplified in FIG. 15 by the weighted moving average with respect to the Y coordinate on the character line position or the series up to the character line. Graph F-G001 shows the metadata score of the title among the bibliographic information. Graph F-G002 displays the author's metadata score. Graph F-G003 displays the metadata score of the English title.

  The weighted moving average of the width n can be calculated by the following formula, for example.

In Equation 2, P _Y represents the metadata score of the character line that is the Y coordinate (Y coordinate on the character line series from the first character line, that is, the character line position) of the weighted moving average calculation target. n represents the window size of the weighted moving average. With the metadata score propagation process according to this calculation formula, it is possible to smooth the noise score and create an essential metadata score graph.

<Metadata area determination P-A108>
Next, the process of determining the boundary position of the metadata area in FIG. 2 performed using the metadata score graph based on a plurality of feature amounts created by the above means will be described.

  FIG. 17 is a flowchart showing an outline of the metadata area boundary determination P-A 108 including the processing of the weighted moving average calculation P-A 107 in FIG. First, in process P-F001, the weighted moving average of the metadata score described in the process flow of FIG. 14 is calculated. This processing is as illustrated in Equation 2, and corresponds to the weighted moving average calculation P-A107 in FIG.

  Next, in process P-F002, the maximum point P1 of the metadata score to be extracted is searched from the beginning position of the character line. Next, in process P-F003, a metadata score maximum point P2 next to the score maximum point of the metadata M1 currently focused on is searched for. Put this metadata as M2. Next, in process P-F004, a region boundary point between the metadata M1 and the metadata M2 is obtained from between the local maximum point P1 and the local maximum point P2.

  FIG. 18 is a schematic diagram illustrating the process of determining the boundary position from the two maximum points in process P-F004. The axis F-H301 represents the score value. The axis F-H302 represents the character line position (Y coordinate on the character line series from the first character line). A graph F-H001 represents a score graph of the metadata M1. A graph F-H002 represents the metadata M2. A point F-H201 on the graph represents the maximum point P1 of the metadata M1 score, and a point F-H202 on the graph represents the maximum point P2 of the metadata M2 score. As a method for determining the boundary between P1 and P2, there is a method using the intersection F-H203 of the graph F-H001 and the graph F-H002 as a boundary. Further, the position where the inter-class variance is maximized as determined by the following equation may be used as a metadata boundary point.

  Equation 3 represents a variance value between two regions when the boundary position is y, and a position where this value is maximized is selected as the boundary position y.

<Modification of weighted moving average calculation P-A107>
Although the first embodiment has been described in detail with reference to the drawings, the Y coordinate is used as it is as the distance value when executing the weighted moving average calculation P-A107, or the distance can be converted by a pattern of continuous lines. Good. Then, a weighted moving average is performed based on the converted distance. For example, paying attention to the following features, the virtual position of each character line is changed.

(1) When the length of the lower line of the continuous line is α times the length of the upper line
=> Multiply line spacing by α ・ ε.

(2) The lower line of the continuous line is shifted to the left by X0 from the upper line.
=> Multiply line spacing by (X0 / width of lower line).

<Automatic segmentation method for multiple document data>
The metadata extraction method of the bibliographic information extraction system according to the first embodiment described above can be applied even when a plurality of documents are printed continuously.

  FIG. 19 is a diagram illustrating an example of a metadata score graph for a series of documents in which a plurality of documents are continuously copied. In the upper part of the figure, axis F-I100 represents the axis of the copied page position. The axis F-I200 represents the metadata score, and the higher the score is, the higher the score is. Graphs F-I001, F-I002, F-I003, and F-I004 represent graphs after the weighted moving average calculation of the metadata score shown in FIG. In this example, four documents are printed in succession, and a maximum point group of metadata scores appears repeatedly four times. The maximum point group of the metadata score can be found, for example, as an area where a certain number of local maximum points exist within a certain window size.

  Labels F-I301, F-I302, F-I303, and F-I304 shown at the bottom of the figure apply the means described using the flow chat in FIG. 17 to a part of a series of copied pages. Represents the metadata label assigned. Boundary lines F-I301, F-I302, and F-I303 represent document boundaries that are determined based on the assignment results of the metadata labels F-I301, F-I302, F-I303, and F-I304. The document boundary can be delimited in front of the page containing the maximum score group of metadata scores.

  By the above means, even when a plurality of documents are bundled and printed by a multifunction machine or the like, metadata can be extracted for each document, and mixed documents can be separated into individual documents.

<Calculation of 2-axis weighted moving average of metadata score>
In the first embodiment described above, the weighted moving average of the metadata score is calculated after the character lines are arranged one-dimensionally. This calculation of the weighted moving average is not limited to fixing in the Y-axis direction, and may be performed by selecting the X-axis direction. In addition, the weighted moving average calculation can be executed not only in one axis direction but also in two axis directions. In the second embodiment, a weighted moving average is performed on two-dimensional coordinates without performing a one-dimensional series of character lines. The calculation formula in this case is as follows.

  The boundary region can be determined by finding the boundary point of the metadata score for each of the X axis and the Y axis, as in the case of the one-dimensional case. According to the present embodiment, it is possible to easily determine a two-dimensional boundary region.

<Determination method of metadata area boundary by dynamic programming>
Next, as a third embodiment, a configuration in which dynamic programming is used as a method for determining a boundary based on a metadata score during the processing of the document processing program 153 executed by the CPU 13 will be described. By using dynamic programming, it becomes possible to label metadata reflecting the relative positional relationship of metadata.

  FIG. 20 is a diagram illustrating an example of the relationship between the character line position and the metadata type. The matrix F-J100 shown in the upper part of the figure is a matrix showing the relationship between the character line position and the normalized metadata score. The axis F-J110 represents the character line position, and the farther to the right, the farther from the beginning. An axis F-J120 represents an axis of the metadata type. The metadata identifier F-J121 represents a provisional metadata appearance order. The numbers in the matrix represent the cost based on the temporary metadata score of the character line. A matrix F-J200 shown in the lower part of the figure represents a metadata selection cost matrix based on each metadata score ranking for each character line. This can be calculated, for example, by setting “3” if the score ranking of the metadata T of a certain character line is third.

  Assuming one metadata order, find the path with the lowest total cost among the paths to the bottom line on matrix F-J100 or matrix F-J200. Data allocation can be determined. For the search of the shortest path, a dynamic programming method often used in calculating the edit distance between character strings and character strings can be used. In the case of the dynamic programming method using the metadata score in the present embodiment, the matrix moving cost is defined as follows.

Cost due to matrix movement pattern:
(1) When moving to the right in the matrix: This indicates that consecutive character lines are given the same metadata label. If the space between character lines is more than a certain amount, add a cost.
(2) When moving down the matrix: This indicates that multiple metadata is assigned to the same character line (or metadata is omitted). Add a deletion cost.
(3) When moving diagonally through the matrix: Indicates transition to the next metadata. If the space between character lines is sufficiently narrow, a transition cost is added.

  In the present embodiment, the movement pattern cost on the matrix and the path with the shortest total value of the matrix cost are obtained by dynamic programming.

  FIG. 21 is a diagram illustrating a search example of the minimum cost path using the matrix F-J100 or F-J200 of FIG. The axis F-K110 is an axis representing the position of the character line. The axis F-K120 is an axis representing the metadata type. The metadata series F-K121 represents a temporary order of metadata (an example). A cell F-K201 represents a cell having the lowest cost among the costs when the matrix is searched up to “B” of the metadata series. When searching for the minimum cost path, it is possible to obtain the minimum cost path by comparing the values of the diagonally upper left or left cells and tracing the cell with the smaller value. If the cell values are the same, priority is given to the upper left corner. In the example of FIG. 21, the hatched portion of the cells in the matrix F-J100 corresponds to the minimum cost path. The metadata label of the character line is obtained from the metadata of the row position of the cell by tracing the target character line vertically (in the direction of the axis F-K120) and finding the cell of the minimum cost path.

  As the space cost between character lines, (A) a relative value with respect to the character line height, (B) a relative value with respect to the page height, (C) a ranking when the distances between the character lines are arranged in descending order, and the like can be used. FIG. 24 is a diagram illustrating an example of a metadata label cost between character lines. Elements F-M100 to F-M107 represent character lines. The width F-M001 represents the page height. A width F-M002 represents a distance between character lines between the character line F-M100 and the character line F-M101. A width F-M003 represents a distance between character lines between the character line F-M103 and the character line F-M104. A width F-M004 represents a distance between character lines between the character line F-M006 and the character line F-M007. The width F-M005 represents the height of the character line F-M100. The width F-M006 represents the height of the character line F-M103. A width F-M007 represents the height of the character line F-M106. Using the above values, the above (A), (B), and (C) are expressed as follows.

Example (A):
-Cost when the metadata is the same for character lines F-M100 and F-M101 ((1) of matrix cost) = α1 / β1
-Cost when the metadata is the same for the character lines F-M103 and F-M104 ((1) of matrix cost) = α2 / β2
-Cost when the metadata is the same for character lines F-M106 and F-M107 ((1) of matrix cost) = α3 / β3
Example (B):
-Cost when the metadata is the same for character lines F-M100 and F-M101 ((1) of matrix cost) = α1 / γ1
-Cost when the metadata is the same for the character lines F-M103 and F-M104 ((1) of the matrix cost) = α2 / γ2
・ Cost when the metadata is the same for character lines F-M106 and F-M107 ((1) of matrix cost) = α3 / γ3
Example of (C): Use the reciprocal of the rank of the distance between character lines (α2>α3> α1)-Cost when the metadata is the same for character lines F-M100 and F-M101 ((1) of matrix cost) = 1/3
-Cost when the metadata is the same for character lines F-M103 and F-M104 ((1) of matrix cost) = 1/1
-Cost when the metadata is the same for character lines F-M106 and F-M107 ((1) of matrix cost) = 1/2
By the above means, when a temporary metadata order is given, optimal metadata labeling can be performed in consideration of the metadata likeness, metadata order and layout of each character line.

  The method introduced above is a process when a metadata order is given. Finally, a method for executing a search for a label sequence that minimizes the total metadata cost when the metadata order is indefinite will be described. Simply, the above dynamic programming is executed for all metadata orders, and the result of the metadata column having the minimum cost may be selected. However, the same partial series is duplicated in all metadata orders, and the same processing is repeatedly executed, which is inefficient. Thus, by representing the order of metadata that can appear in a tree structure, it is possible to partially avoid duplication processing.

  FIG. 22 shows a data structure that represents a possible metadata order in a tree structure. By performing an all-path search of this tree structure, it is possible to avoid duplication of processing for the same order from the beginning. Tree node F-L001 represents a root node. The identifier of the root node is T (title). Tree nodes F-L002, F-L003, and F-L004 represent child nodes of the node F-L001. In the example, NE (English name), TE (English title), and N (Japanese author name) can appear after T. As can be seen from the above example, each path of the tree corresponds to the appearance order of the metadata. The links F-L201 and F-L202 are links in which the beginnings of paths that share a partial character string on the tree path are linked. In this link, a link is made from the subsequent partial character string to the previous partial character string based on the order of performing a full tree search. With the above configuration, for example, when the tree node F-L005 is reached during the search, the link can be traced and the result of the already processed tree node F-L006 can be used. By this processing, it is possible to avoid duplication of processing for a portion having a common partial sequence at the end.

  In this embodiment, even when the number of academic literature types to be extracted increases by using the above-described tree structure search and the dynamic programming method based on the metadata score described with reference to FIGS. Bibliographic metadata information can be extracted with high accuracy without creating a new structure definition.

  FIG. 23 is a flowchart showing an outline of the minimum cost label sequence search based on the metadata score in the third embodiment described above. It goes without saying that this processing flow can be realized by the CPU 13 executing the document processing program 152 stored in the work area 15 of the storage unit of the bibliographic information extraction system 10 shown in FIG. .

  First, in process P-G001, a tree composed of possible metadata orders is constructed. At this time, a link for the common suffix series is also created.

  As in the first embodiment, the input P-G 101 loads document data to be processed into the work area 15 through the input device 110 and the communication network 19 shown in FIG. In the column structure determination process P-G102, the column structure of the input document is determined in order to correctly extract the character lines even for academic documents having a plurality of column structures. In the character line extraction process P-G103, a character line serving as a labeling unit is extracted based on the column structure information extracted by the column structure determination P-G102. In the one-dimensional series processing PG-104, the character lines extracted by the character line extraction process P-G103 are rearranged according to whether they are included in the XY coordinates of the character line or the column determined by the column structure determination P-G102. . In the character line feature setting P-G105, layout features such as line height, line position, average inter-character distance, and character string pattern of the line, which are stored in advance in the information holding unit 16 of FIG. 1 in association with each metadata. Based on the number of characters and the number of characters, the language and layout feature amount of each character line is set.

  Next, in process P-G002, the score of each metadata is calculated for the character line to be processed. Next, in process P-G003, a metadata rank based on the metadata score for each character line is determined, and a matrix having the rank as a cost is constructed. The method of calculating the cost is not limited to this, and a value such as “1-normalized metadata score” may be used. Next, in process P-G004, one path of the tree to be searched is determined. In process P-G005, the minimum cost path search based on the metadata score is performed as shown in FIG. 21 using the metadata appearing in the path. On the way, it is determined whether or not there is a link in determination P-G101. If there is a link, the process proceeds to process P-G007, and if there is no link, the minimum cost path search is continued in process P-G006. In process P-G007, the link is traced, and the search is completed using the minimum cost path search result of the link destination. When the search of all tree paths is completed, the metadata labeling result extracted in the tree path that minimizes the cost in the process P-G008 is finally acquired as the final result.

  Through the above processing, according to the present embodiment, even when the number of academic literature types to be extracted increases, metadata that is bibliographic information can be extracted with high accuracy without newly creating a document structure definition. It becomes possible.

  In addition to electronic documents, the present invention described in detail above extracts metadata of bibliographic information such as titles, author names, abstracts, names of English creators, English abstracts, etc., for various academic literature documents that have been digitized by multifunction peripherals. It is useful as a document processing device that performs academic information providing services and reuses documents efficiently.

DESCRIPTION OF SYMBOLS 11 ... Input part 110 ... Input device 111 ... Image input device 12 ... Display device 13 ... CPU
14 ... Printer 15 ... Work area 151 ... OS
152 ... Communication program 153 ... Document processing program 154 ... Bibliographic information (metadata) editing program 16 ... Information holding unit 161 ... Extraction target metadata definition dictionary 162 ... Feature definition dictionary 19 for each metadata ... Communication network 21 ... File server.

Claims (14)

A document processing apparatus comprising an information holding unit and a processing unit, and extracting metadata of bibliographic information by the processing unit based on an appearance position and a character string in a document of a character line,
The information holding unit
Holds a feature definition dictionary for each metadata that describes the feature values for each metadata.
The processor is
Based on the document column structure determination result, extract the character line and character line position,
A metadata score is calculated for each of the character lines based on the feature amount for each metadata held in the information holding unit,
Determining a boundary position of the bibliographic information based on the metadata score;
A document processing apparatus characterized by that. The document processing apparatus according to claim 1,
The processor is
Based on the character line position, a propagation process of the metadata score for each metadata is executed, and the meta data after the propagation process of any one of the metadata to be extracted from the start position of the character line. The maximum point of the data score is searched, the boundary position is determined based on the change of the metadata score between the two consecutive maximum points searched, and the character line before the determined boundary position is set to one. Set the metadata for the first maximum point, and set the metadata for the second maximum point in the subsequent text line.
A document processing apparatus characterized by that. The document processing apparatus according to claim 2,
The processor is
Based on the character line position, performing the metadata score propagation process by performing a smoothing process by weighted moving average calculation of the metadata score for each of the metadata,
A document processing apparatus characterized by that. The document processing apparatus according to claim 1,
The information holding unit
The number of characters constituting a character line is held as the feature amount for each metadata.
A document processing apparatus characterized by that. The document processing apparatus according to claim 1,
The information holding unit
As a feature quantity for each metadata, an English name pattern length is retained.
A document processing apparatus characterized by that. The document processing apparatus according to claim 1,
The information holding unit
As the feature amount for each metadata, the average distance between words and the average distance between characters are retained.
A document processing apparatus characterized by that. The document processing apparatus according to claim 1,
The processor is
Character elements are extracted from the input document information, horizontal and vertical projection component histograms of the extracted character elements are generated, and column information is extracted from the extracted projection component histograms to extract the document information. Perform column structure determination,
A document processing apparatus characterized by that. Bibliographic information is extracted based on the appearance position and character string in the document of a character line consisting of a set of characters, at least having an input unit capable of inputting document information, an information holding unit, a processing unit, and an output unit. A document processing device,
The processor is
Extract character elements from the input document information,
Generating a horizontal and vertical projection component histogram of the extracted character elements, extracting column information from the projection histogram;
Based on the order of characters determined by the coordinate values of the character elements and the column information, character lines are extracted,
Based on the extracted coordinates of the character line and the column information including the character line, the character line is rearranged as series data,
For each of the sorted character lines, calculate a metadata score based on the feature amount of each metadata stored in advance in the information holding unit,
Based on the character line position, execute the propagation process of the metadata score for each of the metadata,
The maximum point of the metadata score after propagation processing is searched from the first character line, and the boundary point is set based on the change of the metadata score between the two consecutive maximum points searched. The first maximum point metadata is set in the character line before the boundary point, and the second maximum point metadata is set in the subsequent character line.
A document processing apparatus characterized by that. The document processing apparatus according to claim 8, wherein:
The processor is
Performing the propagation process by performing a weighted moving average calculation of the metadata score for each of the metadata;
A document processing apparatus characterized by that. The document processing apparatus according to claim 8, wherein:
The information holding unit
As the feature amount for each metadata, it holds at least one of the number of characters constituting the character line, the English name pattern length, the average distance between words, and the average distance between characters.
A document processing apparatus characterized by that. Bibliographic information is extracted based on the appearance position and character string in the document of a character line consisting of a set of characters, at least having an input unit capable of inputting document information, an information holding unit, a processing unit, and an output unit. A document processing device,
The processor is
Extract character elements from the input document information, generate horizontal and vertical projection component histograms based on the extracted coordinates of the character elements, and extract column information extraction from the projection component histograms ,
Based on the order of characters determined by the coordinate values of the character elements and the column information, character lines are extracted,
Based on the extracted coordinates of the character line and the column information including the character line, the character line is rearranged as series data,

For each of the sorted character lines, calculate a metadata score based on the feature amount of each metadata stored in advance in the information holding unit,
Based on the metadata score of each character line, set the cost when each character line is the metadata, and based on the metadata identifier order candidates stored in advance associated with the information holding unit , Search for a metadata label sequence that minimizes the total cost of the entire text line,
A document processing apparatus characterized by that. The document processing apparatus according to claim 11,
The processor is
Determining the rank based on the metadata score of each character line, and setting the rank as the cost;
A document processing apparatus characterized by that. The document processing apparatus according to claim 11,
The processor is
A value obtained by subtracting a value obtained by normalizing the metadata score of each character line from 1 is set as the cost.
A document processing apparatus characterized by that. The document processing apparatus according to claim 11,
The processor is
Set the distance between the character lines as the cost,
A document processing apparatus characterized by that.

Images