Form data search system, form data search method, and form data search program

Abstract

To provide a form data search system, a form data search method, and a form data search program capable of easily and exactly extracting necessary information from form data being electronic data.SOLUTION: In table item name model creation means 14, a word phrase relevancy model 26 composed of a plurality of word phrases assumed to be contained in form data CD and information indicating a degree of relevancy of the word phrases with each other. The table item name model creation means 14: acquires a designated item name from an item name dictionary 24; extracts, as an item name candidate, word phrases having a degree of relevancy of a fixed degree or higher with the designated item name by using the word phrase relevancy model 26; calculates a probability of matching of the extracted item name candidate and the designated item name; and creates a table item name model 28. Item name block extracting means 16 and contents block extracting means 18 extract an item name block BR1 in which specific data TD are stored and a contents block BR2 by using the table item name model 28 and the item name dictionary 24.SELECTED DRAWING: Figure 1

Description

[Technical field]

[0001] The present invention relates to a form data search system that automatically extracts and outputs specific data desired by a user from form data in which the positional relationships of rows and columns are defined and which can be expressed in a matrix, as well as a form data search method and form data search program executed by the form data search system. [Background technology]

[0002] Conventionally, technology that reads information written on paper forms using an OCR (Optical Character Reader) and converts it into electronic data has been put to practical use and is used in various fields as an effective way to improve business efficiency. In particular, in the business field, technology that converts information into electronic data (hereinafter referred to as form data) that can be expressed in a matrix format with the positional relationship between rows and columns defined is conveniently used.

[0003] The raw form data output by OCR contains almost all of the information contained in the paper form. Therefore, when a user wants to extract the desired information from the form data, they must visually look at the form data to identify the relevant parts, extract the necessary data, and transcribe it. This is not only a heavy burden for users, but also means that the desired information may not be extracted accurately, so there is a demand for technology that can extract the desired information easily and accurately.

[0004] In the past, as disclosed in Patent Document 1, for example, there was a document search device that searches for data in a specified matrix-format search table area within a matrix-format electronic document (document data), and that had a search condition receiving means, a label setting means, a search execution means, and a display means, and by operating these means, data (specific data) that meets the user's search conditions is automatically extracted.

[0005] The search condition receiving means is means for receiving search conditions desired by a user, and the label setting means is means for setting at least a part of the top row and column of a specified area as a label. The search execution means executes a search for data included in an area specified by a label that satisfies the user's search conditions among the labels set by the label setting means, and the display means operates to display a label on the display unit corresponding to the data obtained by the search execution means. [Prior art documents] [Patent documents]

[0006] [Patent Document 1] JP 2012-128606 A Summary of the Invention [Problem to be solved by the invention]

[0007] In the document search device of Patent Document 1, the user must select a specific search table area from the electronic document (document data) when searching. In other words, the user must visually identify the search table area that matches the purpose, which is a heavy burden for the user, especially when the electronic document is large in scale, and there is a possibility that the user may miss the search table area that matches the purpose.

[0008] The present invention has been made in consideration of the above-mentioned background technology, and has as its object to provide a form data search system, a form data search method, and a form data search program that can easily and accurately extract necessary information from form data, which is electronic data that can be expressed in a matrix. [Means for solving the problem]

[0009] The invention described in claim 1 is a computer system that automatically extracts and outputs specific data desired by a user from form data, which is electronic data that can be expressed in a matrix format with defined row and column positional relationships, and The specific data is composed of data of an item name block, which is a cell area in which item names that match the user's desire are arranged in one direction, and data of a content block, which is a cell area in which data corresponding to the item name block is arranged, when a map in which the form data is expressed in a matrix is ​​assumed. The method includes an item name dictionary, a table item name model creation means, an item name block extraction means, a content block extraction means, and a specific data output means, the item name dictionary is a dictionary that stores designated item names input by a user and data formats corresponding to the designated item names in association with each other, the table item name model creation means has a word association model configured of a plurality of words assumed to be included in the form data and information indicating the degree of association between the words, and the table item name model creation means acquires the designated item names by referring to the item name dictionary, and extracts the word association using an association model, extracting words having a certain degree of association with the specified item name or higher as item name candidates, calculating a probability value that the extracted item name candidates match the specified item name, and creating a table item name model in which the item name candidates are linked to the probability value; the item name block extraction means analyzes data in each cell in which the form data is stored, and extracts, as the item name block, a cell region of k rows and 1 column or 1 row and k columns (k is a natural number) in which words that are the same as or have a high degree of similarity with the item name candidates of the table item name model are arranged in one direction; the content block extraction means refers to the item name dictionary, recognizes the type of data associated with each of the words in the cells of the item name block that match the specified item name, and analyzes the data of cells in the cell area adjacent to the cell in the item name block in which the word matching the specified item name is stored, thereby identifying a cell area of ​​k rows and n columns or a cell area of ​​n rows and k columns (n ​​is a natural number) adjacent to the item name block, which includes all cells in which data of the type recognized in the item name dictionary is stored, and extracts the cell area as the content block; The specific data output means is a form data search system that performs a process of outputting the data stored in each cell of the extracted item name block and content block as the specific data.

[0010] The phrase relevance model is preferably a learned model created by machine learning based on a plurality of the form data [the invention according to claim 2].

[0011] For the data of each cell in which the form data is stored, the item name block extraction means calculates the similarity to each of the item name candidates included in the table item name model, and for a cell in which the calculated similarity exceeds a first threshold value, assigns the probability value associated with the corresponding item name candidate, and for a cell in which the calculated similarity is equal to or less than the first threshold value, assigns a probability value of zero. Further, a correction value obtained by multiplying the probability value assigned to each cell by a coefficient a (0 < a < 1) of the probability value assigned to an adjacent cell in the row direction or column direction is added to correct the probability value of each cell. It is preferable to adopt a configuration in which a cell region of k rows and 1 column or 1 row and k columns (k is a natural number) in which the corrected probability value exceeds a second threshold value and is continuous in one direction is extracted as the item name block [the invention according to claim 3].

[0012] The specific data is composed of the data of the item name block, the data of the content block, and the data of an auxiliary item name block which is a cell region where the data corresponding to the content block is arranged. It is provided with auxiliary item name extraction means, and the auxiliary item name extraction means performs a process of extracting, as the auxiliary item name block, a cell region adjacent to a position intersecting with the item name block in the content block. The specific data output means may be configured to perform a process of outputting the data stored in each cell of the extracted item name block, content block, and auxiliary item name block as the specific data [the invention according to claim 4].

[0013] The invention described in claim 5 is an information processing method executed by a computer system for automatically extracting and outputting specific data desired by a user from form data, which is electronic data that can be expressed in a matrix format by defining the positional relationship between rows and columns, The specific data is composed of data of an item name block, which is a cell area in which item names that match the user's desire are arranged in one direction, and data of a content block, which is a cell area in which data corresponding to the item name block is arranged, when a map in which the form data is expressed in a matrix is ​​assumed. The computer system is provided with a word association model configured with a plurality of words expected to be included in the form data and information indicating the degree of association between the words, and further includes an item name dictionary that stores designated item names input by a user and data formats corresponding to the designated item names in association with each other; The method includes a table item name model creating step, an item name block extracting step, a content block extracting step, and a specific data output step, and in the table item name model creating step, the specified item name is acquired by referring to the item name dictionary, and a word or phrase having a certain degree of relevance or higher with the specified item name is extracted as an item name candidate using the word relevance model, and a probability value that the extracted item name candidate matches the specified item name is calculated, and a table item name model in which the item name candidate and the probability value are linked is created. In the item name block extracting step, data of each cell in which the form data is stored is analyzed, and a cell region of k rows and 1 column or 1 row and k columns (k is a natural number) in which words or phrases that are the same as or have a high similarity to the item name candidates of the table item name model are arranged in one direction is extracted as the item name block, and In the content block extraction step, the item name dictionary is referred to to recognize the type of data linked to each word in the cells of the item name block that matches the specified item name, and data in cells of a cell area adjacent to the cell in the item name block in which the word matching the specified item name is stored is analyzed to identify a cell area containing all cells storing data of the type recognized in the item name dictionary, the cell area of ​​k rows and n columns adjacent to the item name block, or a cell area of ​​n rows and k columns (n ​​is a natural number), and extract the cell area as the content block. In the specific data output step, the data stored in the extracted item name block and each cell of the content block is output as the specific data. This is a document data search method.

[0014] It is preferable that the phrase relevance model is a trained model created by machine learning based on multiple pieces of the document data [the invention described in claim 6].

[0015] In the item name block extraction step, for the data of each cell in which the form data is stored, the similarity to each of the item name candidates included in the table item name model is calculated. For a cell in which the calculated similarity exceeds a first threshold value, the probability value associated with the corresponding item name candidate is assigned. For a cell in which the calculated similarity is equal to or less than the first threshold value, a probability value of zero is assigned. Further, a correction amount obtained by multiplying the probability value assigned to each cell by a coefficient a (0 < a < 1) of the probability value assigned to an adjacent cell in the row direction or the column direction is added to correct the probability value of each cell. It is preferable to adopt a configuration in which a cell region of k rows and 1 column or 1 row and k columns (k is a natural number) in which the cells in which the corrected probability value exceeds a second threshold value are continuous in one direction is extracted as the item name block [the invention according to claim 7].

[0016] The specific data is composed of the data of the item name block, the data of the content block, and the data of the auxiliary item name block which is a cell region in which the data corresponding to the content block is arranged. An auxiliary item name extraction step is provided. In the auxiliary item name extraction step, a process of extracting, as the auxiliary item name block, a cell region adjacent to a position intersecting with the item name block in the content block is performed. In the specific data output step, a process of outputting, as the specific data, the data stored in each cell of the extracted item name block, the content block, and the auxiliary item name block may be adopted [the invention according to claim 8].

[0017] The invention according to claim 9 is a program for each step execution for causing a computer system to execute the form data search method according to any one of claims 5 to 8. Equipped It is a form data search program.

Advantages of the Invention

[0018] According to the form data search system, form data search method, and form data search program of the present invention, when a user wishes to extract necessary information from form data, specific data that meets the user's desire can be automatically extracted and output simply by specifying related keywords (specified item names, etc.). Since there is no need to visually identify the area to be searched by looking at the form data, even when the form data is very large, searches can be performed accurately in a short time, and the burden on the user is greatly reduced. Furthermore, the large amount of useful information contained in the form data can be utilized to the maximum extent.

[0019] In addition, the process of extracting item name blocks (cell areas in which item names matching a specified item name are stored) from form data does not focus only on words (item names) that match the user's specified item name, but also on words (item names) inferred from the word relevance model and table item model, making it possible to more accurately extract item name blocks that match the user's intentions.The process of extracting content blocks (cell areas in which information corresponding to item name blocks is stored) from form data can also be performed easily and accurately by using a unique item name dictionary. [Brief description of the drawings]

[0020] [Figure 1] 1 is a system configuration diagram showing an embodiment of a form data search system of the present invention. [Diagram 2] 2 is a flowchart showing an embodiment of a form data search method of the present invention. [Diagram 3] This is a map in which the form data to be searched is expressed in a matrix. [Figure 4] 2 is a diagram showing a method for creating the phrase relevance model in FIG. 1. [Diagram 5] 2 is a diagram showing the contents of the item name dictionary in FIG. 1. [Figure 6] FIG. 3A is a diagram showing the flow of the table item name model creation step in FIG. 2, and FIG. 3B is a conceptual diagram showing the contents of the table item name model. [Figure 7]3A and 3B are diagrams showing the flow of processing performed in the item name block extraction step in FIG. 2. [Figure 8] 3A and 3B are diagrams showing the flow of processing performed in the content block extraction step in FIG. 2. [Figure 9] FIG. 3 is a diagram showing the processing performed in the auxiliary item name block in FIG. 2. [Figure 10] 3 is a map in which the specific data output by the specific data output block in FIG. 2 is expressed in a matrix. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0021] An embodiment of the form data search system, form data search method, and form data search program of the present invention will be described below with reference to the drawings. The form data search system 10 of this embodiment is a computer system that automatically extracts and outputs specific data TD desired by a user from form data CD, as shown in Fig. 1. The form data search method of this embodiment is a search method executed by the form data search system 10, and is represented as a flowchart in Fig. 2. The form data search program of this embodiment is a program for executing each step to cause the form data search system 10 to execute the form data search method shown in Fig. 2.

[0022] For ease of understanding, the following description will be given on the assumption that the document data CD to be searched contains various information about automobiles, and that the information the user wishes to extract is information about the body size of the automobile.

[0023] The document data CD is electronic data composed of text data in which the row and column positional relationships are defined and which can be expressed in a matrix, as shown in FIG. 3, and is, for example, CSV-format data or Excel-format data created by processing information written on a paper document with the OCR system 12.

[0024] In addition, as shown in Figure 10, the specific data TD is composed of data for an item name block BR1, which is a cell area in which item names matching the user's wishes are lined up in one direction in a map that represents the report data CD in a matrix, data for a content block BR2, which is a cell area in which data corresponding to the item name block BR1 is lined up, and data for an auxiliary item name block BR3, which is a cell area in which data corresponding to the content block BR2 is lined up.

[0025] 1, the form data search system 10 comprises a table item name model creation means 14, an item name block extraction means 16, a content block extraction means 18, an auxiliary item name block extraction means 20, and a specific data output means 22, and is further provided with an item name dictionary 24. In addition, a unique word association model 26 is set in advance in the table item model creation means 14.

[0026] The phrase relevance model 26 is composed of a plurality of phrases assumed to be included in the document data CD and information indicating the relevance between the phrases. FIG. 4 is a diagram showing an image of the data structure of the phrase relevance model 26, in which the distance between a plurality of phrases corresponds to the relevance, and the closer the distance, the higher the relevance. For example, "total length" in the upper left area of ​​the figure has a very high relevance to the nearby "total width" and "total height," and a low relevance to "wheelbase," "cabin dimensions," and "fuel economy," which are in a slightly distant area. This phrase relevance model 26 is a trained model created by machine learning based on a plurality of document data arbitrarily collected. This phrase relevance model 26 is set in advance by a system administrator before a user performs a search.

[0027] The item name dictionary 24 is a dictionary that stores information input by a user to specify data to be extracted, and as shown in FIG. 5, the specified item name is stored in association with the data type corresponding to the specified item. For example, when the specified item name is "total length," "wheel base," or "ground clearance," it generally corresponds to numerical data indicating length, so a "number string" is associated as the data type. In addition, when the specified item name is "drive system," it generally corresponds to character data indicating whether the vehicle is front-wheel drive or rear-wheel drive, so a "character string" is associated as the data type. The variations in notation in FIG. 5 will be described later.

[0028] The means 14, 16, 18, 20, and 22 shown in FIG. 1 execute the table item name model creation step S11, the item name block extraction step S12, the content block extraction step S13, the auxiliary item name block extraction step S14, and the specific data output step S15 shown in FIG.

[0029] Next, a form data detection method executed by the form data search system 10 will be described with reference to the flowchart of FIG.

[0030] First, a user inputs a desired designated item name and a data type corresponding to the designated item into the form data search system 10, and this information is stored in the item name dictionary 24. Then, the table item name model creation means 14 executes a table item name model creation step S11. In the table item name model creation step S11, the designated item name is acquired by referring to the item name dictionary 24, and a word relevance model 26 is used to extract words having a certain degree of relevance or higher with the designated item name as item name candidates, and a probability value that the extracted item name candidate matches the designated item name is calculated, and a table item name model 28 in which the item name candidate and the probability value are linked is created.

[0031] For example, as shown in Figures 6(a) and (b), when the item names specified by the user are "Overall Length", "Wheelbase", and "Ground Clearance", the table item name model creation means 14 uses the word association model 26 to extract nine words, "Overall Length", "Overall Width", "Overall Height", ..., and "Passenger Room Dimensions", which have a certain degree of association or higher with the specified item names "Overall Length", "Wheelbase", and "Ground Clearance", as item name candidates. In this example, having a certain degree of association or higher is equivalent to a probability value of approximately 0.70 or higher.

[0032] For example, the term "total length" in the phrase relevance model 26 matches the specified item name "total length," so it is inevitably determined to be an item name candidate and is associated with a probability value of 1.00. Also, the term "total width" in the phrase relevance model 26 does not match the specified item name, but is determined to have a certain degree of relevance or higher with "total length," so it becomes an item name candidate and is associated with a probability value of 0.88 calculated by a predetermined method. On the other hand, terms such as "fuel economy" and "horsepower" in the phrase relevance model 26 are determined not to have a certain degree of relevance or higher with the specified item name, so they are not item name candidates (even if a probability value were calculated, it would be determined to be less than 0.70).

[0033] After the table item name model creation step S11 is completed, the item name block extraction means 16 next executes the item name block extraction step S12. In the item name block extraction step S12, the data of each cell in which the form data CD is stored is analyzed, and a process is performed to extract, as an item name block BR1, a cell region of k rows and 1 column or 1 row and k columns (k is a natural number) in which words that are the same as or highly similar to the item name candidates in the table item name model 28 are arranged in one direction.

[0034] First, for the data of each cell in which the form data CD is stored, the similarity to each item name candidate included in the table item name model 28 is calculated, and cells whose calculated similarity exceeds a first threshold value are extracted. The method of calculating the similarity is not particularly limited, and can be freely selected from among known calculation methods.

[0035] For example, looking at column A of the form data CD shown in Fig. 7(a), the data in the four cells marked with "★", namely, the "total length" of cell cA, the "total width" of cell dA, the "foil base" of cell eA, and the "minimum ground height" of cell gA, are determined to have a similarity exceeding the first threshold with respect to the item name candidates. On the other hand, the data "tread" in cell fA is determined to have a similarity below the first threshold, and cells bA, hA, etc. where no data is stored are also determined to have a similarity below the first threshold for the data.

[0036] The data of cells other than column A (for example, cells in columns B to E) are also determined in the same way of thinking, and for all cells, it is determined that the similarity of the data is below the first threshold.

[0037] Then, as shown in Fig. 7(b), for cells cA, dA, eA, gA marked with "★" that are determined to have a similarity exceeding the first threshold, probability values 1.00, 0.88, 1.00, 0.96 associated with the corresponding item name candidates in the table item name model 28 are respectively assigned. On the other hand, for the other cells determined to have a similarity below the first threshold, a uniform probability value of 0 is assigned.

[0038] After assigning probability values to each cell, next, a process of correcting the probability values is performed. Specifically, a correction amount is calculated by multiplying the probability value assigned to an adjacent cell in the row direction or column direction (here, the row direction) by a predetermined coefficient a (0 < a < 1), and the calculated correction amount is added to the original probability value.

[0039] If the coefficient a is set to 1 / 2, for example, in the case of cell dA, the correction amount is 1.00, which is the sum of 0.50, which is obtained by multiplying the initial probability value 1.00 of the cell cA adjacent to the upper side by 1 / 2, and 0.50, which is obtained by multiplying the initial probability value 1.00 of the cell eA adjacent to the lower side by 1 / 2, and the probability value of cell dA is corrected to 1.88 by adding the correction amount 1.00 to the initial probability value 0.88. In the case of cell fA, the correction amount is 1.00, which is the sum of 0.50, which is obtained by multiplying the initial probability value 1.00 of the cell eA adjacent to the upper side by 1 / 2, and 0.50, which is obtained by multiplying the initial probability value 1.00 of the cell gA adjacent to the lower side by 1 / 2, and the probability value of cell fA is corrected to 1.00 by adding the correction amount 1.00 to the initial probability value 0. In the case of cell fB, the initial probability values ​​of the two adjacent cells above and below are 0, so the correction amount is 0, and the probability value of cell fB is corrected to 0 by adding the correction amount 0 to the initial probability value of 0 (in other words, the corrected probability value will be the same as the initial probability value). The probability values ​​of the other cells are corrected in a similar way.

[0040] When the correction of the probability value of each cell is completed, the item name block extraction means 16 identifies a cell area in which cells whose corrected probability value exceeds the second threshold are lined up in one direction, and extracts the identified cell area as the item name block BR1. In the example of Fig. 7(b), the second threshold is set to 0.80, and the cell area cA to gA (a cell area of ​​5 rows and 1 column) is extracted as the item name block BR1.

[0041] After the item name block extraction step S12 is completed, the content block extraction means 18 next executes the content block extraction step S13. The content block extraction step S13 executes steps S13-1 to S13-3 shown in Fig. 8(b) to identify a cell area in which data corresponding to the item name block BR1 is stored, and extracts this as a content block BR2.

[0042] In the example of Fig. 8(a), the cell area α adjacent to the item name block BR1 can become the content block BR2. First, the item name dictionary 24 is referenced to recognize the data type associated with each of the words in the item name block BR1 that match the specified item name (step S13-1). In other words, it is recognized that the data type "number string" is associated with each of the words "total length," "wheel base," and "ground clearance."

[0043] In this case, it would be good if there was an exact match with the word in the item name block BR1 in the specified item name dictionary 24, but there is a possibility that the corresponding data type cannot be recognized due to variations in the spelling of the word in the item name block BR1. Therefore, in order to avoid such a situation, when the user inputs the specified item name and data type, it is good to have them input a word that is synonymous with the specified item name (a word that is not considered to match the specified item name) as well. This makes it possible to obtain the item name dictionary 24 shown in Figure 5, which makes it possible to recognize the data type taking into account variations in the spelling of the word in the item name block BR1.

[0044] Next, the process proceeds to step S13-2, where the data type of cells (specific cells in cell area α) in the cell area adjacent to the cell in the item name block BR1 in which a word matching the specified item name is stored is analyzed, and cells in which data in the format recognized by the item name dictionary 24 is stored are extracted. For example, for "total length" in the item name block BR1, cells cB, cC, and cD with a data type of "numeric string" are extracted from the cells in row c in cell area α. For "wheel base", cells eB, eC, and eD with a data type of "numeric string" are extracted from the cells in row e in cell area α. Similarly, for "ground clearance", the corresponding cells gB, gC, and gD are extracted. Therefore, in the example of FIG. 8(a), a total of nine cells are extracted by performing step S13-2.

[0045] Then, proceed to step S13-3, identify a rectangular cell area that includes all the cells extracted in step S13-2 and is adjacent to the item name block BR1, and extract the cell area as the content block BR2. That is, in the example of Fig. 8(a), the cell area of ​​5 rows and 3 columns, (c~g) x (B~D), is extracted as the content block BR2.

[0046] After the content block extraction step S13 is completed, the auxiliary item name block extraction means 20 next executes the auxiliary item name block extraction step S14. In the auxiliary item name block extraction step S14, a process is performed to extract a cell area adjacent to the position where the content block BR2 intersects with the item name block BR1 as the auxiliary item name block. In the example of Fig. 9, data is stored in a cell area adjacent to the upper side of the content block BR2, so a cell area of ​​2 rows and 3 columns of (a~b) x (B~D) is extracted as the auxiliary item name block BR3. Alternatively, a cell area of ​​1 row and 3 columns of b x (B~D) may be set to be extracted.

[0047] When the auxiliary item name content block extraction step S14 is completed, the specific data output means 22 then executes the specific data output step S15. In the specific data output step S15, the data stored in each cell of the item name block BR1, content block BR2, and auxiliary item name block BR3 extracted in the previous step is output as specific data TD shown in Fig. 10. The characteristic data TD is data in CSV or Excel format in which the positional relationship between rows and columns is defined and matrix notation is possible.

[0048] In this way, by executing the five steps S11 to S15 shown in FIG. 2 in order, it is possible to accurately extract and output the specific data TD that meets the user's desire from the form data CD.

[0049] Here, we will provide additional information regarding the process contents and determination methods performed in the above-mentioned steps S11 to S15. For example, in the above explanation of the table item name model creation step S11, the words to be extracted as item name candidates are words with a probability value of approximately 0.70 or more, but the value of 0.70 is merely an example, and can be freely changed according to the type and characteristics of the form data CD to be searched, the item names specified by the user, etc.

[0050] Furthermore, in the item name block extraction step S12, it is determined whether the similarity is high or low based on a first threshold value, but the method of calculating the similarity and the first threshold value can be freely changed according to the type and characteristics of the report data CD to be searched, the item names specified by the user, etc.

[0051] In the above description of item name block extraction step S12, the probability value is corrected using coefficient α=1 / 2, and the second threshold value of 0.80 is used to determine whether the corrected probability value is high or low, but these values ​​can also be freely changed. However, looking at the example in FIG. 7(b), if α is set too small, the corrected probability value of cell fA may become too low, making it impossible to accurately extract item name block BR1, and if the second threshold value is set too low, there is a possibility that cells bA and hA that should not be included in item name block BR1 may be extracted as item name block BR1. Therefore, when setting coefficient α and the second threshold value, appropriate values ​​should be set according to the type and characteristics of the form data CD to be searched, the item names specified by the user, etc.

[0052] In the above description of content block extraction step S13, data is stored in all 15 cells that make up content block BR2, but there may be cases where no data is stored in a particular cell. In this case, the user can see the output specific data TD and know that "some data is missing from this form data CD."

[0053] In the specific example shown in Figures 7 to 10, a cell area of ​​k rows and 1 column is extracted as item name block BR1 to become the "row label," and correspondingly, a cell area of ​​k rows and n columns is extracted as content block BR2, and further a cell area of ​​2 rows and n columns is extracted as auxiliary item name block BR3 to become the "column label." However, depending on the form data CD to be searched, as a result of performing the same process as above, a cell area of ​​1 row and k columns is extracted as item name block BR1 to become the "column label," and correspondingly, a cell area of ​​n rows and k columns is extracted as content block BR2, and further n rows and 2 columns is extracted as auxiliary item name block BR3 to become the "row label."

[0054] As described above, according to the form data search system 10, form data search method, and form data search program of this embodiment, when a user wants to extract necessary information from the form data CD, the specific data TD that meets the user's desire can be automatically extracted and output simply by specifying related keywords, etc. (specified item names, etc.). Since there is no need to visually look at the form data to specify the area to be searched, even if the form data CD is very large, a search can be performed accurately in a short time, and the burden on the user is greatly reduced. Furthermore, the large amount of useful information contained in the form data CD can be utilized to the maximum extent.

[0055] The process of extracting item name blocks BR1 from form data CD focuses not only on words (item names) that match the item names specified by the user, but also on words (item names) estimated from the word relevance model 26 and table item model 28, making it possible to more accurately extract item name blocks BR1 that match the user's intentions. In addition, the process of extracting content blocks BR2 from form data CD can also be performed easily and accurately. Furthermore, by performing the process of extracting auxiliary item name blocks BR3, both row labels and column labels when displayed in a matrix are extracted, making it possible to output search results in a format that is easy for the user to see.

[0056] The form data search system, form data search method, and form data search program of the present invention are not limited to the above embodiment. For example, the phrase relevance model used in the table item name model creation step is preferably a trained model created by machine learning based on multiple form data, such as the phrase relevance model 26 shown in Fig. 4, but a model set by a method other than machine learning (for example, a model originally devised by a system administrator, etc.) may be used.

[0057] The above embodiment is configured to include an auxiliary item name block extraction means 20 and an auxiliary item name block extraction step S14, but depending on the content of the information desired by the user, there may be cases where it is not necessary to extract the auxiliary item name block BR3, in which case the auxiliary item name block extraction means and the auxiliary item name block extraction step can be omitted.

[0058] In addition, in the explanation of the above embodiment, it is assumed that "the report data CD to be searched is data containing various information related to automobiles." However, the report data CD may be data that contains a mixture of information related to automobiles and other miscellaneous information (for example, information that is completely unrelated to automobiles), and by setting a word relevance model 26 that corresponds to the content of the report data CD, information on the body size of automobiles can be easily extracted in the above manner, and even information that is unrelated to automobiles can be extracted without any problems in a similar manner. [Explanation of symbols]

[0059] 10 Search Data Search System 12 OCR System 14 Table entry name creation method 16 Item name block extraction method 18 Content block extraction methods 20. Extraction method for auxiliary item name block 22 Specific data output means 24 Item Name Dictionary 26 Phrase Association Model 28 Table Item Name Model BR1 Item Name Block BR2 Content Block BR3 Sub-item name block CD report data S11 Table item name model creation step S12 Item name block extraction step S13 Content block extraction step S14 Auxiliary item name block extraction step S15 Specific data output step TD specific data

Claims

1. A computer system that automatically extracts and outputs specific data desired by a user from form data, which is electronic data that can be expressed in a matrix format with defined row and column positional relationships, The specific data is composed of data of an item name block, which is a cell area in which item names that match the user's desire are arranged in one direction, and data of a content block, which is a cell area in which data corresponding to the item name block is arranged, when a map in which the form data is expressed in a matrix is ​​assumed. The system includes an item name dictionary, a table item name model creating means, an item name block extracting means, a content block extracting means, and a specific data outputting means, the item name dictionary is a dictionary that stores designated item names input by a user and data formats corresponding to the designated item names in association with each other; a word association model configured by a plurality of words assumed to be included in the form data and information indicating the degree of association between the words is set in the table item name model creation means; the table item name model creation means, when acquiring the designated item name by referring to the item name dictionary, uses the word relevance model to extract words having a degree of relevance or higher with the designated item name as item name candidates, calculates a probability value that the extracted item name candidates match the designated item name, and performs a process of creating a table item name model in which the item name candidates are linked to the probability value; the item name block extraction means performs a process of analyzing data of each cell in which the form data is stored, and extracting, as the item name block, a cell region of k rows and 1 column or 1 row and k columns (k is a natural number) in which words that are the same as or have a high similarity to the item name candidates of the table item name model are arranged in one direction; the content block extraction means refers to the item name dictionary, recognizes the type of data associated with each of the words in the cells of the item name block that match the specified item name, and analyzes the data of cells in the cell area adjacent to the cell in the item name block in which the word matching the specified item name is stored, thereby identifying a cell area of ​​k rows and n columns or a cell area of ​​n rows and k columns (n ​​is a natural number) adjacent to the item name block, which includes all cells in which data of the type recognized in the item name dictionary is stored, and extracts the cell area as the content block; the specific data output means performs processing for outputting the data stored in each cell of the extracted item name block and content block as the specific data. A form data search system comprising:

2. The form data search system according to claim 1 , wherein the word association model is a trained model created by machine learning based on a plurality of the form data.

3. The item name block extraction means Calculating a similarity between data of each cell in which the form data is stored and each of the item name candidates included in the table item name model, and allocating the probability value associated with the corresponding item name candidate to a cell in which the calculated similarity exceeds a first threshold, and allocating a probability value of zero to a cell in which the calculated similarity is equal to or less than the first threshold; Furthermore, the probability value of each cell is corrected by adding a correction amount obtained by multiplying the probability value assigned to the adjacent cell in the row direction or column direction by a coefficient a (0<a<1) to the probability value assigned to each cell; A document data search system as described in claim 1 or 2, which performs a process of extracting a cell area of ​​k rows and 1 column or 1 row and k columns (k is a natural number) in which cells whose corrected probability value exceeds a second threshold value are consecutive in one direction as the item name block.

4. the specific data is composed of data of the item name block, data of the content block, and data of an auxiliary item name block which is a cell area in which data corresponding to the content block are arranged, a sub-item name extraction means for extracting a cell area adjacent to a position where the content block intersects with the item name block as the sub-item name block; 4. A form data search system as claimed in any one of claims 1 to 3, wherein the specific data output means performs processing to output the data stored in each cell of the extracted item name block, content block and auxiliary item name block as the specific data.

5. An information processing method executed by a computer system for automatically extracting and outputting specific data desired by a user from form data, which is electronic data that can be expressed in a matrix format by defining the positional relationship between rows and columns, comprising: The specific data is composed of data of an item name block, which is a cell area in which item names that match the user's desire are arranged in one direction, and data of a content block, which is a cell area in which data corresponding to the item name block is arranged, when a map in which the form data is expressed in a matrix is ​​assumed. The computer system is provided with a word association model configured with a plurality of words expected to be included in the form data and information indicating the degree of association between the words, and further includes an item name dictionary that stores designated item names input by a user and data formats corresponding to the designated item names in association with each other; The method comprises a table item name model creation step, an item name block extraction step, a content block extraction step, and a specific data output step, In the table item name model creation step, a designated item name designated by a user is acquired by referring to the item name dictionary, and a word or phrase having a degree of association with the designated item name or higher is extracted as an item name candidate using the word relevance model, and a probability value that the extracted item name candidate matches the designated item name is calculated, and a table item name model in which the item name candidate and the probability value are linked is created; In the item name block extraction step, data of each cell in which the form data is stored is analyzed, and a cell region of k rows and 1 column or 1 row and k columns (k is a natural number) in which words that are the same as or highly similar to the item name candidates of the table item name model are arranged in one direction is extracted as the item name block; In the content block extraction step, the item name dictionary is referenced to recognize the type of data associated with each of the words in the cells of the item name block that match the specified item name, and data in cells of the item name block that are adjacent to the cell in which the word matching the specified item name is stored is analyzed to identify a cell area that includes all cells in which data of the type recognized in the item name dictionary is stored, the cell area being k rows and n columns or the cell area being n rows and k columns (n ​​is a natural number) adjacent to the item name block, and a process of extracting the cell area as the content block is performed; In the specific data output step, a process is performed to output the data stored in each cell of the extracted item name block and content block as the specific data. A form data search method comprising:

6. The form data search method according to claim 5 , wherein the word relevance model is a trained model created by machine learning based on a plurality of the form data.

7. In the item name block extraction step, Calculating a similarity between data of each cell in which the form data is stored and each of the item name candidates included in the table item name model, and allocating the probability value associated with the corresponding item name candidate to a cell in which the calculated similarity exceeds a first threshold, and allocating a probability value of zero to a cell in which the calculated similarity is equal to or less than the first threshold; Furthermore, the probability value of each cell is corrected by adding a correction amount obtained by multiplying the probability value assigned to the adjacent cell in the row direction or column direction by a coefficient a (0<a<1) to the probability value assigned to each cell; A form data search method as described in claim 5 or 6, which performs a process of extracting a cell area of ​​k rows and 1 column or 1 row and k columns (k is a natural number) in which cells whose corrected probability value exceeds a second threshold value are consecutive in one direction as the item name block.

8. the specific data is composed of data of the item name block, data of the content block, and data of an auxiliary item name block which is a cell area in which data corresponding to the content block are arranged, An auxiliary item name extraction step is provided, in which a cell area adjacent to a position where the content block intersects with the item name block is extracted as the auxiliary item name block, A form data search method as claimed in any one of claims 5 to 7, wherein the specific data output step performs processing to output the data stored in each cell of the extracted item name block, content block and auxiliary item name block as the specific data.

9. 9. A form data retrieval program comprising programs for executing each step, for causing a computer system to execute the form data retrieval method according to claim 5.

Images