JPH09138840A - Character recognition device - Google Patents

Abstract

PROBLEM TO BE SOLVED: To correct and complement a recognition result which is missed due to erroneous recognition, etc., by referring to an index table, comparing an appearing probability on plural candidate characters in a character recognition result and deciding an order of certainty. SOLUTION: The index table generation part 15 of a processor 14 reads a text inputted from an input part 3, processes the kind of the character which is possibly connected to a prescribed character (string) into a numerical value and generates the index table 8 of a tree structure to be displayed. A knowledge processing part 4 refers to the index table 8 and judges the propriety of a candidate character string being the character recognition result. Information including the appearance probability that the previously selected character appears next to the prescribed character is stored in the index table 8. The knowledge processing part 4 refers to the index table 8 and compares the appearance probability on the plural candidate characters stored in the character recognition result at each character and gives the order to the certainty of the candidate characters.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a morphological analysis device, a natural language processing device such as a character recognition device and a voice recognition device, and a knowledge processing for complementing a correct character missing due to erroneous segmentation or recognition. A character recognition device suitable for.

[0002]

2. Description of the Related Art When a document or the like described in natural language is input to an information processing device and made into a database,
Characters are optically read and recognition processing is performed. At the time of this character recognition processing, the recognition result is evaluated,
It is necessary to eliminate misreading. The following technologies have been developed to automatically perform such post-processing (Information Processing Society of Japan, Research Report Vol.95, No, 68 (95-NL-107)). This is O
It describes a method for detecting / correcting an erroneously read character in a character recognition result by a CR (optical character reader). This is based on the list of words that can appear in the recognition target character string and the information about whether or not the words can be connected, for each character, the ID (identification code) of the word in which the character appears, the appearance position of the character, and the word Create a character index that describes the length information, match the recognition results with multiple possibilities (candidate characters) with the character index, extract word candidates, and then validate the candidate characters as a sequence of words. The first candidate is determined by determining the sex.

[0003]

By the way, the above-mentioned conventional character recognition apparatus has the following problems to be solved. In the above method, since recognition is performed on a word-by-word basis, it is very effective for recognizing a document such as an address character string in which only a limited character string appears. However, if the word boundaries are unclear, or if the candidate character string does not exist in the word dictionary, it cannot be expected to perform adequately appropriate processing. Moreover, although the word dictionary used for recognition is prepared in advance, the user can freely operate the application condition of the dictionary, such as adding a new candidate word or changing the priority of the candidate word depending on the document. It is not easy to have such a function.

[0004]

The present invention employs the following structure to solve the above problems. <Structure 1> An input unit for inputting a character recognition result, an index table storing information including an appearance probability that a certain character selected in advance appears next to a certain character, and a character recognition result for each character is listed in the character recognition result. A knowledge processing unit that refers to the index table for the plurality of generated candidate characters and compares the appearance probabilities to rank the probability of the candidate characters.

<Explanation> The character recognition result is input from, for example, an optical character recognition device. In the character recognition process, a plurality of candidate characters for one character are listed as recognition results for each character as much as possible. When the appearance probability that a character appears next to a certain character is known, the probability can be ranked for each candidate character based on the relationship with the character immediately before it. As a result, it is possible to correct and supplement a part of the recognition result that has been lost due to erroneous recognition or erroneous cutting. If the index table has a sequence of characters that have a high probability of appearing in the text, the reliability of automatic post-processing of the character recognition result increases. The appearance probability does not have to be a numerical value that directly represents the probability, but may be any information that indirectly indicates the possibility that the character will appear next to a character using a numerical value. In addition, the reason why the characters are arranged is that the appearance probabilities are uniformly obtained regardless of whether or not a series of characters constitutes a word.

<Structure 2> The index table extracts a sequence of characters included in a text similar to the text to be recognized, and an appearance probability that a certain character included in the text appears next to a certain character. Is configured by digitizing.

<Explanation> It is highly probable that similar expressions and words are used in the text similar to the recognition target text. Therefore, the accuracy of character recognition can be improved by creating an index table that utilizes the character arrangement of the text and performing knowledge processing. Since this method is processed by paying attention to the arrangement of characters, it can be used without being aware of word divisions and whether or not the used word is an unknown word.

<Structure 3> The index table is a tree-structured table in which a sequence of characters included in a text is merged with a common character in a trunk node and different characters following the character are arranged in branch nodes. Then, by tracing the node, it is configured to indicate a continuous character string that may follow a certain character.

<Explanation> When a sequence of various characters is numerically expressed in association with other characters, if this is expressed by a tree structure, a search process for each character can be performed at high speed. All characters used in the text are displayed on the trunk, and if there is a common character immediately before, it has a common trunk, and if the following characters are different, it will gradually branch. If there are no subsequent characters, the branch will not extend any further.

[0010]

BEST MODE FOR CARRYING OUT THE INVENTION Embodiments of the present invention will be described below with reference to specific examples. <Specific Example 1> FIG. 1 is a functional block diagram of a character recognition device of the present invention. The illustrated apparatus includes an input unit 3 that receives the character recognition result 1, a knowledge processing unit 4, and a storage device 5. The character recognition result 1 read and processed by an OCR (optical character reading device) or the like (not shown) is input to this device from the input unit 3. The knowledge processing unit 4 is a unit for ranking the candidate characters in the character recognition result and obtaining the output 9 as described later. The index table 8 is referred to for this ranking process. This index table is generated in advance by the text 6 similar to the text to be recognized and stored in the storage device 5, as described later.

FIG. 2 is a block diagram showing a concrete device configuration of the present invention. The illustrated system is an input / output device 1
1, a processing device 14, and a storage device 5. The input / output device 11 has a function of inputting a text, displaying a result, and the like. The input / output device 11 includes an input unit 3 and an output unit 13.
Having. Here, the input unit 3 has a function of inputting a text for creating an index table and a character recognition result. The input unit 3 may be composed of, for example, a keyboard, or may be a device that accesses a text file stored in the storage device of the computer. The output unit 13 has a function of displaying processing results and the like. The output unit 13 is composed of, for example, a display and a printer.

The storage device 5 has a function of storing texts, processing results at each stage, and the like. The storage device 5 includes an input file 18 for storing input text, a recognition result file 19 for temporarily storing character recognition results, an index table 8, and an output file 21 for storing knowledge processing results. There is. The processing unit 14 has a general configuration such as an arithmetic unit, a memory, and a control unit, and has a function of creating the index table 8 and executing knowledge processing for a character recognition result according to a processing procedure described later. The processing device 14 includes an index table creating unit 15 and a knowledge processing unit 4. The index table creation unit 15 has a function of creating the index table 8 from the input text.

The knowledge processing unit 4 includes an index table 8
With the function of determining the validity of the candidate character string that is the result of character recognition. The index table 8 digitizes and displays the types of characters that may be connected to a certain character (column). In addition to the following characters, the number of appearances and the appearance probability may be described. The index table creation unit 15 reads the text and creates a tree-structured index. The input text may be described in any format such as a sentence, a section, a list of phrases, or an output of morphological analysis. The unit of the character string is also arbitrary. It may be separated by the number of characters or special characters, or the entire text may be considered as one character string. However, regarding the content, when knowledge processing is performed on the character recognition result, the more the same words and expressions as the processing target are included, the better the index quality becomes. It is desirable to use terms and the like.

FIG. 3 shows an example of a tree structure index. An index is a tree structure of arbitrary depth, which is created by merging a common front part into one node with an arbitrary character as a top node, and the succeeding characters are described in each node. Therefore, all the types of characters appearing in the input text are described in the node immediately after the top node, and the part following the node from the top node represents a continuous character string in the text. For example,
The index when the text "online" is input is as shown in FIG. In each node, characters and information on the number of appearances are described side by side. The "*" in the top node indicates an arbitrary character, that is, all characters in the text. The number of appearances of "*" is the total number of characters that form the text. Therefore, it is 5 in this example. Further, "@" indicates that this node is the end of the continuous character string pattern.

After any character, that is, in the first node,
There are letters "o", "n", "la", and "i", and "o" and "la"
The number of appearances of "a" is once and the number of appearances of "n" is twice. This is "o""la" in the text
It means that "a" appears once and "n" appears twice. The second node following "o" has one "n" and the number of appearances is one. This is "O"
It means that "n" appears once after "," that is, "on" appears once in succession. "Onla,""online," and "online" will only appear once in a row. The appearance probability can be calculated by "the number of appearances of subsequent characters / the number of appearances of preceding characters". This can be done by dividing the number of appearances of the child node by the number of appearances of the parent node.

The probability of "n" appearing after an arbitrary character * is 2/5, which is 0.4, but "o" and "rai".
In the example of the figure, the probability that "n" appears after "a" is 1 because both the parent node and the child node are 1. This is because the appearance probability of “n” in the entire text is 0.4, whereas “n” always appears after “o”, “rai”, and “a”. The index table 8 describes the information of such a tree structure index in a table format.

FIG. 4 shows an example of the contents of the index table when "online" is input. Here, the node level D1 of each node and the information for each character are the node number D2, the character D6, the number of appearances D4,
The appearance probability D5 is described. For example, node level 1
“O” has the node number “1”, the number of appearances “1”, and the appearance probability “0.2”. Even if new text is input, this index table can be updated simply by counting up the number of occurrences of each character in the node string containing the characters that appear in the text and recalculating the appearance probability. .

Next, the knowledge processing using the index table will be described. FIG. 5 is an operational flowchart of the knowledge processing in the present invention. In the knowledge processing, the recognition result file 19 first input from the input unit 3 is read (step S1). This character recognition result has one or more candidate characters for each character position i (i = 1, 2, ... N). For example, if there is a recognition result of 5 characters, candidate characters of about 1 to 3 characters are included for each character. The state set S [i] is associated with this character position i. S [i] represents the position of the node to which the (i-1) th candidate character matched, and the initial state is 0. When performing the comparison and collation processing of the character position i, the child node following the node described in S [i] is the processing target (steps S2 to S5). Each character position i (i = 1,
2, ... n) candidate character j (j = 1,2, ... m) is compared and collated with the character of the target node of the index table (step S6), and the matching candidate character is given a point. Addition processing is performed (step S8). The character with the highest point among the candidate characters at each character position i is
The certainty factor as a recognition result is high (step S
6 to S8). Finally, the comparison and comparison results of each character position are sorted and rearranged in the order of points (step S9), and the processing result is written in the output file (step S10).

FIG. 6 is an operation flowchart of the comparison and collation processing shown in step S6 of FIG. Here, all candidate characters j at the character position i are compared and collated with the comparison target node of the index table (step S
1, S2). The comparison target node is the node state set S.
It is indicated by [i]. If the i-1th candidate character does not match any node, the state set S [i] is 0.
Therefore, the comparison and collation processing is performed on the child node of the node of node number 0, that is, the characters of the node of node level 1.
If there is a match, S [i] contains 0 and the node number of the node with which the candidate character matched, so the node at the node level 1 and the child node of the node with the i-1th candidate character matched On the other hand, the comparison and collation processing is performed. If the candidate character j at the character position i matches the target node in the index table, point addition processing is executed for j (step S4), and the node position information is added to S [i + 1] (step S5). ). If j is the last candidate character, the process ends. If not, count up j by 1 (step S7),
The same process is repeated for the next candidate character (steps S2 to S7).

The point addition processing in step S4 may be a method of adding a certain score to the corresponding candidate character or a method of calculating and adding points according to the number of appearances of matched nodes and the depth of the node level. To be Here, the appearance probability is used as the score, the appearance probability of the node that matches the candidate character is added as the point, and the nodes are arranged in descending order of the score. When the scores are the same, the candidate character having a higher rank in the original recognition result is ranked higher.

FIG. 7 shows an example explanatory diagram of the recognition result. Furthermore, a specific flow of knowledge processing will be described using the index table of FIG. 4 and the example of FIG. 7. First, the recognition result is read in step S1 of FIG. 5, and the length “5” of the input character string is set to n (step S2 of FIG. 5). Next, i is set to "1" and S [i] is set to 0 (steps S3 and S4 in FIG. 5). Further, the number of candidate characters of the i (= 1) th character “3” is set to m (step S in FIG. 5).
5) Proceed to the comparison and collation processing. In the comparison and collation processing in step S6 of FIG. 5, first, j = 1 is set (step S1 of FIG. 6). j indicates the position of the i-th candidate character to be collated. First, i (= 1)
The index table is searched by using the jth (= 1) candidate character "age" of the character as a key (step S2 in FIG. 6).
At this time, since the state set S [i] is {0}, the nodes following the “0” in the index table, that is, the nodes with node numbers “1” to “4” are the search targets. Since there is no node that matches "age" in these nodes (step S3 in FIG. 6), j = 1 + 1 = 2 is set (step S7 in FIG. 6), and the jth (= 2) th candidate character "o" is keyed. Then, the nodes "1" to "4" in the index table are searched (step S2 in FIG. 6). Here, node number "1"
"O" matches (step S6, step S6
3). The appearance probability “0.2” is added as a point to the candidate character “o” (step S4 in FIG. 6), and the node number {1} that matches the state set S [i + 1], that is, S [2] is added. Add (step S5 of FIG. 6). Then j = 2
+ 1 = 3 (step S7 in FIG. 6), the j-th (= 3)
Index table “1” with candidate character “thousand” as key
~ Search for "4" node. These nodes include
There is no node that matches "thousand" (step S in FIG. 6).
3). Here, since j = m, the processing is terminated, and
Return to step S7.

Next, i = 1 + 1 = 2 is set (step S8 in FIG. 5), and the comparison and collation processing of the candidate character of the i (= 2) th character is performed. At this time, since S [2] is {0, 1}, the nodes to be compared are child nodes following “0” and “1”, that is, node numbers “1” to “4” and "1
1 "node. Hereinafter, the same process is performed up to the fifth candidate character" n ".

FIG. 8 shows the result of the comparison and collation processing. In this way, the probabilities are ranked for each candidate character. Finally, the candidate characters are rearranged in order from the one having the most points (step S9 in FIG. 5), and the result is output to the output file (step S10 in FIG. 5). FIG. 9 shows the contents of the output file after the knowledge processing is completed. This result is used for post-processing of the character recognition result.

As described above, according to the present invention,
It is possible to create an index table that describes the nature of the arrangement of character strings that appear from the text specified by the user, and perform character recognition processing based on that table. This index table uses an arbitrary character string as a unit, and a character from a pattern in which a character appears before and after to a pattern of a continuous character string following a character is 1
It is described in one format. Also, by changing the input text or adding a character string to be prioritized to the index table, it is possible to easily create the index table according to the conditions. In the knowledge processing using this table, the probability of candidate characters is ranked by preferentially adopting candidate characters that match the sequence of characters that are likely to appear consecutively. Therefore, even if the candidate character string does not exist in the index table as it is, the character recognition process can be performed by partially using the index table. Moreover,
By changing the index table to be applied according to the content or weighting the score of the character arrangement of the important words, it is possible to easily change the application condition according to the recognition target.

[Brief description of the drawings]

FIG. 1 is a functional block diagram of a character recognition device of the present invention.

FIG. 2 is a block diagram showing a specific device configuration.

FIG. 3 is an explanatory diagram of an example of a tree structure index.

FIG. 4 is a diagram illustrating an example of an index table.

FIG. 5 is an operation flowchart of knowledge processing.

FIG. 6 is an operation flowchart of comparison processing.

FIG. 7 is a diagram illustrating an example of a recognition result.

FIG. 8 is an explanatory diagram of points of comparison and collation processing.

FIG. 9 is an explanatory diagram of contents of an output file after the knowledge process is completed.

[Explanation of symbols]

 1 Character recognition result 3 Input unit 4 Knowledge processing unit 5 Storage device 8 Index table

Claims (3)

[Claims] 1. An input unit for inputting a character recognition result, an index table storing information including an appearance probability that a character selected beforehand appears next to a certain character, and a list in the character recognition result for each character. A character recognition device, comprising: a knowledge processing unit that compares the appearance probabilities with respect to a plurality of updated candidate characters to rank the probabilities of the candidate characters. 2. The index table extracts a sequence of characters included in a text similar to the text to be recognized, and numerically expresses the appearance probability that a certain character included in the text appears next to a certain character. The character recognition device according to claim 1, characterized in that 3. The index table is a tree-structured table in which a sequence of characters included in a text is merged with a common character in a trunk node and different characters following the character are arranged in branch nodes, respectively. By following the node,
The character recognition device according to claim 1 or 2, wherein the character recognition device is configured to indicate a continuous character string that may follow a certain character.