JPH03156589A - Method for detecting and correcting erroneously read character - Google Patents

Abstract

PURPOSE:To enable the detection and correction of erroneous reading with high accuracy by determining whether a character extracted by comparison with the linkage table of two characters before and behind the postpositional particle of one character is a postpositional particle or not, dividing a character string and collating it with a word dictionary. CONSTITUTION:An input character feature pattern passing through a scanner 1 and a character recognizing device 2 is compared with a standard pattern by a personal computer 4 and the character is recognized. Then, the character string is partitioned by punctuation marks. From this partitioned character string, the postpositional particle of one Japanese character (ka, ga, shi, te, de, to, ni, no, ha, ba, he, mo, ya, wo) is extracted and the two characters before and behind the postpositional particle are compared with the stored table of linkage between two characters. Then, it is determined that the character is the postpositional particle. The character string is divided by the determined postpositional particle and when the erroneous read character is detected and corrected while collating it with the word dictionary by a longest coincidence method within the divided character string, the detection and correction can be executed with high accuracy.

Description

[Detailed description of the invention] [Industrial application field] The present invention relates to a method for detecting and correcting misread characters.

[Conventional technology]

In order to increase the credibility of the results recognized by a character recognition device, it is common practice to perform morphological analysis of the results as Japanese sentences, perform word matching at the word level, and detect and correct misread characters. Been name.

Conventionally, as a means of morphological analysis, punctuation marks are used as delimiters that can be used interchangeably, and character strings within the range separated by punctuation marks are further divided into character types (kanji).

A common method is to divide the resulting character strings at changing points (hiragana, katakana), and perform morphological analysis using the longest match method while checking the resulting character strings against a word dictionary (
However, the change in character type from kanji to hiragana is allowed due to the relationship between okurigana).

[Problem to be solved by the invention]

However, when morphological analysis is performed using the method described above, first, words that have a character type change other than the kanji-hiragana character type change, such as "Soviet Union", are detected.

``There is a problem that words such as Wandering J are not extracted as words, so even if they contain misread characters, they cannot be detected and corrected.Also, some kanji and katakana have very similar glyph shapes. (There are characters with the following characters (Yu and Yu, Chikara and Chikara, Ko and Ko, etc.), and if these characters are misread as characters of another similar character type, it is impossible to detect and correct the misreading, which is an inconvenience. is occurring.

Means for Solving Problem C] After recognizing the characters by comparing the human-generated characteristic patterns of the characters with the standard patterns of each character stored in advance,
Separate the recognized character strings with punctuation marks, and select one-character hiragana particles (ka, ga, shi, te, de, to, ni) from the character strings separated by these punctuation marks.

, wa, ba, he, mo, ya, wo) are extracted, and the characters before and after these extracted characters are extracted from a pre-memorized word containing a one-letter hiragana particle, and the two characters before and after are linked together. Determine whether the extracted character is a particle by comparison,
The character string is further divided into characters determined to be particles, and misread characters are detected and corrected while checking the divided character strings against a word dictionary using the longest match method.

[Effect]

Frequently occurring words with one character in hiragana (ka, ga, shi, de.

Extract characters that can come before or after a one-letter hiragana particle from a dictionary of words that include words such as de, to, ni, no, wa, ba, he, mo, ya, wo) and create a preceding and following character concatenation table. After separating the sentences with punctuation marks, detect the places where the above 14 types of hiragana appear, compare the character strings before and after that with the preceding and following character concatenation table above, and find out which part of the hiragana appears before or after it. If there is a hiragana in the word, it is marked to indicate that it may be a hiragana, but if there is no match, it is determined that this is a Kaida one-character particle, and morphological analysis is performed using that as a break. By detecting and correcting misread characters, it is possible to detect and correct words consisting of multiple character types, and to reduce misreading between character types (katakana-kanji, kanji-katakana).

〔Example〕

FIG. 1 is a flowchart showing an embodiment of the invention, and FIG.
The figure is a diagram showing a system to which the present invention is applied.

First, explanation will be given starting from FIG. In the figure, 1 is a scanner that inputs a document image into an image memory by optical scanning;
2 is a character recognition device (OC) consisting of a recognition unit that extracts and recognizes each character from a document image, and a correction unit that corrects misread characters using a word dictionary and a connection table (MAP).
R), 3 is a display that displays recognition results and 5 confirms the results;
4 is a data processing device such as a personal computer, and 5 is a keyboard for inputting data.

Next, FIG. 1 will be explained.

First, it is determined whether each character in the recognition result is a punctuation mark, and if it is a punctuation mark, the process ends without doing anything (see ■ to ■). If it is not a punctuation mark, determine whether it is a one-character hiragana particle (see ■), otherwise return to step ■.

If it is a one-character hiragana particle, check the munching with the preceding and following character concatenation table (see ■), and if the matching is successful, return to step ■.

If it is not in the preceding and following character conjunction table, it is definitely a one-character particle■
), and performs morphological analysis of the character string delimited by this one-character particle (see ■). As a result, if misreadings are detected, they will be corrected (see ■).

FIG. 3 shows a specific example of a character string, and FIG. 4 shows a part of a preceding and following character concatenation table.

In other words, the input character string in the case of Figure 3 is [last year - 1 among Soviet people in the year, character recognition result (first candidate)
is ``Last year - among the Soviet universities in the year''. If such a character string is divided by character type as in the conventional method (Kanji-Hiragana may be okurigana, it is excluded), "Last year - in the year / So / master If you match and correct the words within each separated character string, you can correct "annual" = "annual", but you can correct the mispronunciation of "master". It's impossible. On the other hand, in the method of this invention, "ni, no, ka, de" are picked up as one-character particle candidates from the above character string, so the character strings before and after them are
If you compare it with the character conjunction table for character particles, the preceding and following character strings of "ni, no, de" do not match the character combination table for single character particles, so these cannot be determined as particles. can. Also, for ``ka'', there is a character string ``naka'' in the preceding and following character concatenation table, so we mark ``ka'' as a possible hiragana character in the word, and also mark it as a character that has been determined as a particle. If you separate the string, it becomes "last year - year / in / so / master / inside / in /",
More accurate morphological analysis is possible than dividing by character change points. In other words, the word "Soviet person" can be divided into two types: "Soviet person/master" and "Soviet person/person," but since "Soviet person" is an abbreviation for "Soviet Union,""Sovietperson" ” as the correct answer, and it becomes possible to correct misreadings.

FIG. 5 shows another example of a character string. In this case, the character string is ``The troika came running'', and the character recognition result (first place) is ``The troika came running''. If we divide it by character type, we get "Troy/power/ran/came." Therefore, the result of word matching within the range divided by character types does not correct the mispronunciation of "power" - "power."

On the other hand, in the method according to the present invention, "ga, te" is raised as a candidate for a one-letter particle, but since "te" exists in the preceding and following character concatenation table in Figure 4, it is not considered a one-letter particle. is not confirmed. Therefore, the division result is ``Troika/ came running.'' Using the longest match method, ``Troika'' can be corrected to ``Troika.''

By the way, OCR recognition results include misread characters with a high probability, and not all words in a sentence are in the dictionary, so words containing misread characters and unregistered words (
(hereinafter collectively referred to as unknown words) must be extracted by morphological analysis. At this time, if an unknown word appears within a range separated by punctuation marks, the beginning of the unknown word can be determined by morphological analysis from the beginning, but it is difficult to extract the end of the unknown word. Even if you shift one character from the beginning of an unknown word and perform morphological analysis, the word often contains other words, so it may not be possible to extract a reliable range. Therefore, it is possible to do the following.

FIG. 6 is a flowchart showing an embodiment of the present invention based on this viewpoint.

First, in steps ■ to ■ in Figure 6, the above-mentioned hiragana 1
In addition to character particles, the final character of the verb "ru", the one-character auxiliary verb "
Extract "ta", "da" and frequently occurring two-letter hiragana words. Figure 7 shows the types of frequently occurring two-letter hiragana words, and here 23 types are targeted. Furthermore, as in the case of the embodiment shown in FIG. 1, single-character particles may be determined using a conjunctive table as shown in FIG. 4, but this method is not necessarily required.

In steps ■ to ■ in Figure 6, it is checked whether the relationship between the characters extracted as above and the character types before and after it forms a pattern like IFF, and if it matches this pattern, Makes the range delimited by punctuation into smaller parts.

a) Before and after the one-letter particle "wo".../wo/... b) Between other one-letter particles and characters other than hiragana...in the case of...,...ga/start... ...C)
Between the verb ending “ru” or the one-letter auxiliary verb “ta” and a letter other than hiragana...7 hours of waking up...,...7 hours of waking up...
d) Between a two-letter particle and a non-hiragana character... /starts from/ e) Between a non-hiragana character and a two-letter particle...reading/
etc. /... ") Between a character other than hiragana and a two-letter auxiliary verb... truth/desu/... g) Before or after a two-letter form noun and a one-letter auxiliary verb "da" or a one-letter auxiliary verb.../ thing / is / ..., ... / for / for / ...
h) Before and after two-letter verbs and two-letter formal nouns.../are/
for/...,.../to be/thing/...i) Before and after two-letter verbs and two-letter formal nouns.../to/goto/
... j) Before the two-letter verb at the end of the sentence.../. ,···/Masu.

k) Before the two-letter auxiliary verb at the end of the sentence ···/be. ,···/Become.

Note that the names of two-letter particles, two-letter formal nouns, etc. are based on the classification shown in FIG.

In this way, if the finely divided range is the aforementioned one-letter or two-letter Hiragana word, that word is determined to have been determined. In other cases, morphological analysis is performed while checking the divided range against the word dictionary. Here, if there is an unknown word, it often becomes impossible to fully analyze it using morphological analysis. This is because, as mentioned above, although the beginning part of an unknown word can be determined, it is difficult to determine the last part.

Therefore, in this invention, analysis is performed in the reverse direction from the end of the divided range using the longest match method (see [Phase]).
, the end of the unknown word is determined, and the determined range is checked against a word dictionary. If the word contains mispronunciations, it is checked to see if it is a misread character, and corrections are made. If it is an unregistered word, the unregistered word list is created. Create it and use it later to check for unregistered words.

FIG. 8 is an explanatory diagram for explaining the above method.

In this case, the input character string is "Because I was arrested for the electrical work incident" as shown in the same figure (a), and the recognition result (first place)
As shown in the same figure (b), it is incorrectly read as ``rei'' - ``rei'' and ``re'' - ``rei.''

If this is divided using the conventional method, the division will be as shown in FIG. On the other hand, when the hiragana character string is separated using the method according to the present invention, as shown in FIG.
Compared to (c), it does not remain long and it is easy to perform word division using the longest match method. Furthermore, change the beginning and end of the unknown word (e),
After extracting as shown in (f) (use "?" to indicate the beginning or end), combine the candidate characters of the unknown word and if a word that matches the grammar can be extracted, correct it. If the word cannot be extracted, use it as an unregistered word. Restore it as an unregistered word, and later ask for confirmation as an unregistered word, and request corrections if any need to be corrected. In the case of (G) in the same figure, it is registered as an unregistered word, and in the case of (H) in the same figure, it is corrected as a case that can be corrected.

〔Effect of the invention〕

According to this invention, since the focus is on single-letter particles, it is possible to detect and correct misreadings in words consisting of multiple character types (other than kanji and hiragana), as well as across character types. . In addition to the one-letter particles mentioned above, by focusing on and extracting one-letter verb endings "ru", one-letter auxiliary verbs "ta", "da", and frequently occurring two-letter hiragana words, detection and correction can be made even more effective. The accuracy can be increased.

[Brief explanation of the drawing]

FIG. 1 is a flowchart showing an embodiment of the invention, and FIG.
The figure is a block diagram showing a system to which this invention is applied.
Figure 3 is an explanatory diagram for explaining a specific example of an input character string and recognition results, Figure 4 is an explanatory diagram for explaining an example of a preceding and following character concatenation table, and Figure 5 is an explanatory diagram for explaining an example of an input character string and recognition results. An explanatory diagram for explaining another example, FIG. 6 is a flowchart showing another embodiment of the invention, FIG. 7 is an explanatory diagram for explaining an example of frequently occurring two-letter hiragana words, and FIG. 8 is an input character FIG. 7 is an explanatory diagram for explaining still another example of columns and recognition results. Description of symbols 1...Scanner, 2...Character recognition device, 3...Display, 4...Data processing device (personal computer), 5...Keyboard.

Claims (1)

[Claims] 1) After recognizing characters by comparing the characteristic pattern of input characters with standard patterns of each character stored in advance, the recognized character string is separated by punctuation marks, and the punctuation marks are Hiragana one-letter particles from the string separated by (ka, ga, shi, te, de, to, ni, no, wa, ba, he, mo, ya, wo)
is extracted, and the characters before and after this extracted character are compared with a pre-memorized concatenation table of two characters before and after extracted from a word containing a one-character hiragana particle to determine whether the extracted character is a particle. A method for detecting and correcting misread characters, which further divides the character string by the characters determined to be particles and determines the particle, and detects and corrects the misread characters by comparing the divided character strings with a word dictionary using the longest match method. Detection and correction methods. 2) After recognizing the characters by comparing the characteristic pattern of the input characters with the standard pattern of each character stored in advance, the recognized character strings are separated by punctuation marks, and the character strings separated by these punctuation marks are created. Hiragana one-letter particles from among (ka, ga, shi, te, de, to, ni, no, wa, ba, he, mo, ya, wo)
, one-letter verb ending "ru", one-letter auxiliary verb "ta", "da"
and frequently occurring two-letter hiragana words (koto, thing, when, to, this, that, that, which, this, it, those, which, from, than, etc., until, not, desu, masu, aru, etc.)
Depending on whether these words and the character types before and after them satisfy a predetermined condition, the character string is further divided into a range of one to several words, and the longest within that range is extracted. Detect misreading by matching words using matching method.
A method for detecting and correcting misread characters, characterized by correcting them. 3) If a word that contains a mispronunciation or an unregistered word that is not in the word dictionary appears, use the longest match method to determine the word range from the front and back within the range delimited as above. 3. The method for detecting and correcting misread characters according to claim 2, further comprising extracting the remaining portion of the text and comparing it with a word dictionary to detect and correct words containing misread characters.