JPS62285189A - Character recognition post processing system - Google Patents

Abstract

PURPOSE:To execute the correction of an erroneous recognizing character with high accuracy by executing a morpheme analysis to all character strings to be prepared from a candidate character. CONSTITUTION:The input character read by an OCR part 1, after the similarity with a standard character for respective input characters by a character recognizing part 2 and a candidate character is obtained, is sent to a processing part 3. A post processing part 3 segments a word for a candidate character by referring to a word dictionary 10, a speech part classification table 11, a term application table 12 and a connecting weight matrix table 13, and executes the morpheme analysis by the connecting checking with a preceding word. The morpheme analysis is executed concerning all character strings of the unknown word which can be prepared from the candidate character of respective characters of the unknown word part, and executed by the processing to select the character string with the highest estimation value. Thus, when the deciding capacity of the erroneous recognizing character position is not high, the correcting capacity of the high erroneous recognizing character can be expected.

Description

Detailed Description of the Invention 3. Detailed Description of the Invention [Technical Field] The present invention relates to a post-processing method in an OCR character recognition device or the like.

[Prior Art] In OCR character recognition devices and the like, recognition accuracy is generally improved by further performing post-processing on candidate characters that have been determined to be misrecognized. In recent years, it has been considered to utilize language grammatical knowledge as post-processing for this type of character recognition. A typical example is a method using morphological analysis. This involves performing morphological analysis on a character string that includes a character determined to be a misrecognized character, taking into account candidate characters, and checking connections between parts of speech to determine the correct character. However, in the conventional technology, if the ability to determine the position of an erroneously recognized character is low, the processing speed becomes slow and erroneous candidate characters are often determined to be correct.

[Purpose] The purpose of the present invention is to provide a character recognition post-processing method using morphological analysis that can be expected to have a high ability to correct erroneously recognized characters even when the ability to determine the position of erroneously recognized characters is not high. There is a particular thing.

[Configuration] The present invention performs character recognition on each input character to detect candidate characters, creates a character string only from the first candidate character, performs morphological analysis, and then performs morphological analysis. The input character is finally estimated by extracting the part and executing the following process 1 or 2 on the character string part (referred to as the unknown word part).

Process 1: Morphological analysis is performed on all character strings of the unknown word part that can be created from candidate characters of each character of the unknown word part, and the character string with the highest evaluation value is selected.

Process 2: For each character in the unknown word part, a process is performed to determine the position of a misrecognized character, and candidates for the unknown word part are created by sequentially substituting candidate characters only for character positions determined to be misrecognized characters. Perform morphological analysis on character strings and select the character string with the highest evaluation value.

An embodiment of the present invention will be described below with reference to the drawings.

FIG. 1 shows a block diagram of one embodiment of the invention.

The input sentence read by the OCR reading unit 1 is subjected to a character recognition unit 2 in which the degree of similarity with standard characters is calculated for each input character to obtain candidate characters. The candidate characters found by the character recognition section 2 are sent to the post-processing section 3. The present invention replaces this processing by the post-processing section 3. The post-processing unit 3 includes a word dictionary 101, a part-of-speech classification table 11, a conjugation table 12, and a connection weight matrix table 13 in order to perform morphological analysis.

As shown in FIG. 2, the word dictionary 10 includes, for each word,
It includes reading (the pronunciation of the word converted into hiragana), notation (output kana, kanji notation), part of speech, frequency rank, and other information.

The part-of-speech classification table 11 and the conjugation table 12 are based on the connection weight matrix table 1
This is a table showing row and column numbers when searching for 3. Part of speech classification table 11 corresponds to parts of speech that do not have a conjugated ending,
The record structure is as shown in Figure 3. The conjugation table 12 corresponds to parts of speech having conjugated endings, and has a record structure as shown in FIG. 3 (b). In 2N, the endings following the stems of verbs, adjectives, etc. are written in the conjugated ending column, and only when this ending matches the input character will it be evaluated.

As shown in FIG. 4, the connection weight matrix table 13 is a matrix with codes in the row direction and N records in the column direction, and each intersection position represents the connection weight. The processing means until the connection weight matrix table 13 is searched is to search the word dictionary 1o, search the part of speech of the word in the part of speech classification table 11 or the pragmatics division table 12, and find the meaning (the pragmatics). In this case, the conjugated ending and the subsequent character string are matched), and the connection is checked using the connection weight matrix table 13.

FIG. 5 shows a processing flowchart of the post-processing section 3 which forms the center of the present invention. Hereinafter, the processing of the post-processing section 3 will be described in detail based on FIG.

A character string consisting only of the first candidate character among the candidate characters sent from the character recognition unit 2 is created (step 101), and units are cut out from the character string (step 1o2). 2N
A unit is a string of characters separated by punctuation marks, commas, parentheses, and other delimiters. For example, morphological analysis is performed on this unit according to the procedure shown in FIG. 6 (step 103), that is, a predetermined number of characters are read from the beginning of the unit (step 201), and words are cut out by searching in the word dictionary 10 ( Step 2゜2). The part-of-speech classification table 11 or the conjugation table 12 is searched for this extracted word, and for conjugation words, the input character string and the conjugation ending are matched (step 203).
The connection with the immediately preceding word is checked using the connection weight matrix table 13 (step 204), the pointer is advanced by the number of characters of the extracted word (step 205), and the same process is repeated. It is assumed that the word dictionary used in this morphological analysis is sufficiently large and stores all the words in the input sentence.

By performing the above morphological analysis, character strings containing erroneously recognized characters are extracted as unknown word parts (step 10).
4) The following process 1 or process 2 is executed for this extracted unknown word portion.

In process 1, all possible candidate character strings are created by sequentially replacing each character of the unknown word portion with a candidate character (step 105). Next, perform morphological analysis similar to step 103 for each candidate character string in this unknown word part (step 106), and calculate the evaluation value for each candidate character string (step 1o7). Among them, the candidate with the maximum likelihood evaluation value. A character string is selected (step 108). After that, it is determined whether or not the unit ends (Step 109),
If it is not the end, advance the pointer indicating the character position within the unit to the next character position of the unknown word portion just processed and return to step 103 (step 110); if it is the end of the unit, proceed to processing the next unit. (Step 111).

On the other hand, in process 2, after determining the misrecognized character position based on the similarity (distance) calculated by the single character recognition unit 2 for the extracted unknown word part, the processes from step 105 onwards are performed. (Step 110).

In this case, in step 105, only the characters determined to be misrecognized are replaced with candidate characters to create a character string for the unknown word portion.

A specific example is shown in FIG. Assume that the input sentence (correct unit string) is "To perform such a calculation",
Now, as a result of performing morphological analysis in step 103 on the first candidate character string "To perform word calculation like this",
Assume that "nagosan" is extracted as an unknown word part.

In process 1, 33=27 character strings are first created as candidate character strings for the unknown word portion, such as [na word calculation J, rta word calculation], ``田 calculation'', and so on. Next, morphological analysis is performed on each of these character strings, and an evaluation value is calculated. In this case, the evaluation value is set to zero for those that cannot be analyzed. In this way, the character string "Na calculation" with the highest evaluation value is selected as the correct answer6. In process 2, the position of the misrecognized character is determined for the unknown word part, and as a result, the "word" Suppose that is determined to be a misrecognized character. Therefore, there are only three candidate character strings that are created: ``na word calculation'', ``na fraud'', and ``na calculation''. Thereafter, ``na calculation'' is selected by the same process as process 1. That is, it can be seen that Process 2 is faster than Process 1 when the ability to determine the position of an erroneously recognized character is high, and the correction ability is not inferior.

[Effects] As is clear from the above explanation, according to the present invention, when character recognition post-processing is performed using morphological analysis, a high number of misrecognized characters can be achieved even when the ability to determine the position of misrecognized characters is not high. If the correction ability is expected to be high and the ability to determine the position of erroneously recognized characters is high, processing speed can be increased.

[Brief explanation of drawings]

FIG. 1 is an overall configuration diagram of an embodiment of the present invention, FIG. 2 is a diagram showing an example of a word dictionary, FIG. 3 is a diagram showing an example of a part-of-speech classification table, and a conjugation table. A diagram showing an example of a connection weight matrix table, FIG. 5 is a diagram showing the processing flow of the post-processing section in FIG. 1, FIG. 6 is a diagram showing the processing flow of morphological analysis, and FIG.
The figure is a diagram showing a specific example of processing. 1...OCR reading section, 2...Character recognition section, 3.
... Post-processing unit, 10... Word dictionary, 11... Part of speech classification table, 12... Word conjugation table, 13... Connection weight matrix table 6 Figure 1 Figure 3 (Tl) M giant Royo = Dingyoro Figure 2 Figure 4

Claims (2)

[Claims] (1) In a post-processing method for character recognition using morphological analysis, a character string is created from the first candidate character for the input character, and the part of the character string that cannot be analyzed is extracted by performing morphological analysis. A character recognition post-processing method that performs morphological analysis on all character strings that can be created from candidate characters for each character in a character string part, and selects the character string with the highest evaluation value. (2) In the post-processing method of character analysis using morphological analysis, a character string is created from the first candidate character for the input character, and the part of the character string that cannot be analyzed is extracted by performing morphological analysis. The position of the misrecognized character is determined for each character in the string part, and candidate characters are substituted only for the characters determined to be misrecognized. Morphological analysis is performed on the candidate string created by A character recognition post-processing method characterized by selecting character strings with high evaluation values.