JPH02118785A - Method for correcting erroneous recognition - Google Patents

Abstract

PURPOSE:To improve a recognition rate by executing an unknown word processing to correct the part of the unknown word to be unsuccessful to the Japanese language analysis to KATAKANA (the square form of Japanese syllabary), an alphabet, figures and symbols. CONSTITUTION:Concerning the part of the character string to be unsuccessful finally to the Japanese language analysis, this is made into the unknown word, and in an unknown word checking part 15, a pointer Ps of the head character position and a pointer Pe of a final character position is written into an unknown word position saving memory 10. An unknown word processing part 9 refers to a candidate lattice memory 4 by a candidate character retrieving part 16, the candidate character is retrieved in the sequence from the first rank concerning the part of the unknown word and finally, a confusion character is retrieved. Concerning the retrieved character, a character species deciding part 17 decides the character species, and the character of KATAKANA, alphabet, figures and symbols found first and the first rank candidate are replaced. Thus, the KATAKANA word, etc., in which the correction is difficult for the erroneous recognition correcting processing by a Japanese language analysis, can be corrected and the recognition rate can be improved.

Description

[Detailed Description of the Invention] [Industrial Application Field] The present invention provides recognition results for Japanese characters in pattern recognition devices such as character recognition devices and speech recognition devices.
This paper relates to a method for correcting misrecognition using Japanese language analysis.

[Conventional technology]

Pattern recognition devices such as character recognition devices and voice recognition devices perform recognition based on the characteristics of input image and voice patterns, but recognition processing based on the characteristics of such individual patterns completely eliminates misrecognition. That is almost impossible. Therefore, it is necessary to improve the recognition rate by correcting misrecognition through Japanese language analysis (checking the grammar of each word using a word dictionary and a grammar dictionary).

Regarding correction of misrecognition through Japanese analysis, there is a character input processing method (Japanese Patent Application Laid-Open No. 62-111) that corrects misrecognition by performing Japanese analysis on all combinations of recognition candidate characters.
No. 19190), a misread character correction processing device that adds similar characters prepared in advance to recognition candidate characters and performs Japanese analysis on all character strings of candidate characters and additional characters to correct misrecognition (Japanese Patent Laid-Open No. 62 -251986) are known.

[Problem to be solved by the invention]

However, with this type of conventional recognition error correction method, there was a problem in that it was impossible to correct words that were not registered in the word dictionary (naturally, Japanese language analysis would fail).

To reduce the frequency of such uncorrectable situations, rF
It is possible to increase the number of words registered in the word dictionary, but
It is practically difficult to cover every word. In particular, there are an extremely large number of katakana words, and they are often coined, so it is impossible to register them all in a dictionary. The same applies to combinations of alphabets, numbers, and symbols. Furthermore, as the number of registered words increases, other problems arise, such as a decrease in recognition rate and an increase in dictionary search time due to an increase in similar words.

The purpose of the present invention is to correct words such as katakana, alphabets, numbers, or symbols that are difficult to correct depending on the Japanese language analysis, in the recognition results of pattern recognition devices such as character recognition devices and voice recognition devices for Japanese. The purpose of the present invention is to provide a method for correcting misrecognition that is possible.

[Means to solve the problem]

The present invention corrects erroneous recognition by Japanese language analysis on the recognition results of a character recognition device, etc., and also corrects unknown word parts for which Japanese language analysis has failed to be included in candidate characters or candidates. Performs unknown word processing that corrects characters and their similar characters to katakana, alphabetical h, numbers, or symbols.

This unknown word processing may be performed on all parts of the unknown word for which Japanese analysis has failed, but the part of speech of the part of the unknown word for which Japanese analysis has failed can be treated as a common noun or a strange noun and used as a word for the next successful Japanese analysis. The connection may be checked and unknown word processing may be performed only on the part of the unknown word where this connection is possible.

[For production]

Combinations of katakana, alphabets, numbers, or symbols that are not registered in the word dictionary or cannot be registered will fail to be corrected by Japanese analysis, but the unknown word processing described above will significantly improve the accuracy of such unknown words. Since correct words are obtained at a high rate, the recognition rate is greatly improved.

Additionally, kanji can also be unknown words, but if unknown word processing is performed only on unknown words that can be connected to the next word, it is possible to reduce the number of incorrect corrections of unknown words made up of kanji.

〔Example〕

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

FIG. 1 is a block diagram showing one embodiment of the present invention. 1
2 is a scanner for reading a document, and 2 is a character recognition device (OCR) that cuts out individual character patterns from the document image read by the scanner 1 and extracts candidate characters by comparing their characteristics with a pattern dictionary. The recognition result of the character recognition device 2 is corrected by Japanese language analysis (natural language processing), and parts that cannot be corrected by this Japanese processing are corrected by unknown word processing. .

Elements involved in correcting such misrecognition include a candidate character lattice generation section 3, a candidate lattice memory 4, a misrecognition correction section 5 based on Japanese language analysis, and a word dictionary 6 used for Japanese language analysis.
, a grammar dictionary 7, a candidate word memory 89, an unknown word processing section 9 that performs unknown word processing on the part of the unknown word for which Japanese language analysis has failed, and an unknown word for storing position information (pointer described later) of the unknown word part. There is a position save memory 1o.

11 is a recognition result output section for outputting a corrected recognition result; 12 is a control section that controls the misrecognition correction section 5, the unknown word processing section 9, and the recognition result output section 11.

The candidate character lattice generation unit 3 stores the candidate characters and distances output from the character recognition device 2 in the candidate character lattice memory 4.
Register. Furthermore, in this embodiment, since characters that are likely to be misrecognized can be predicted to some extent depending on the recognition method, such similar characters that are likely to be misrecognized (confusion characters) are prepared in advance, and the candidate character lattice generation unit 3 For example, a fusion character for the first candidate character is also written into the candidate lattice memory 4.

FIG. 2 shows an example of the contents of the candidate lattice memory 4. This example is the recognition result for the input sentence "The Japanese government's intention was rejected at the summit."

Only the first candidate and the confusion character are shown, but there are up to 10 candidates.

The misrecognition correction unit 5 detects unknown word portions when correcting the recognition results by Japanese language analysis. That is, the m-word matching unit 13 matches the combination string of candidate characters (not only the first candidate character but also all candidate characters) and fusion characters with the word dictionary 6 (word matching), and then checks all the matched characters. The word is written into the candidate word memory 8 as a candidate word. Next, the grammar check section 14 checks the grammar dictionary 7.
, and compare these candidate words with the word that was successfully parsed immediately before them (nouns, mundane nouns).

5-step conjugation verbs, etc.), and delete candidate words that cannot be connected from the candidate word memory 8. At this time,
When multiple word connections are successful.

Prioritizes long words (applies longest match method) and backtracks if parsing fails.

Next, perform a connection check on long candidate words.

Continue analysis.

The character of the word for which the Japanese language analysis was finally successful is the second character.
If it is a candidate character or a fusion character of the lower rank, the grammar check section], 4 replaces the first candidate character in the candidate lattice data memory 4 with the candidate character or fusion character of the second or lower rank. In other words, fix it.

In the example shown in Figure 2, Japanese language analysis is successful with the first candidate character up to "Japanese Government's", but Japanese language analysis is successful with "Intention" using the 6th candidate character for "Intention ga". However, the first candidate's ``Megumi'' and the sixth candidate's ``I'' are swapped. Next "Summit" (Note: #ToII is the kanji for "Boku")
Since the Japanese language analysis is not successful for , the Japanese language analysis is continued from the next "de". "So", "Rejected"
Japanese language analysis is successful for .

In addition, incorrect phrases are detected by performing Japanese analysis on the first candidate character string, and Japanese analysis is performed only for the detected incorrect phrases, including the second and lower candidate characters and confusion characters. You may also modify it by doing so.

Now, regarding the part of the character string (erroneous phrase) that ultimately failed in the above Japanese analysis, this is treated as an unknown word, and the unknown word check section 15 uses the pointer Ps of the first character position.
and the pointer Pe of the final character position are written into the unknown word position save memory 10.

In the example in Figure 2, Japanese parsing fails at the position of "su" in "summit", so pointer P
Save as s, continue Japanese parsing from the next character,
Japanese parsing is successful with "de". The character position before this successful word "de", that is, the character position of "to (kanji)" is saved as a pointer Pa.

The unknown word processing unit 9 is a part that performs unknown word processing on the portion between the pointer Ps and the pointer Pe. That is, the candidate character search unit 16 refers to the candidate lattice memory 4, searches for candidate characters in the unknown word portion in order from the first position, and finally searches for the fusion character. Then, the character type determination unit 17 determines the character type for the searched characters, and the first found katakana, alphabet,
Swap (correct) the number or symbol character and the first candidate. If you cannot find any katakana, alphabet, number, or symbol, use the first candidate as is.

In the example in Figure 2, “summit” is the last letter “to” (kanji 1).
``Boku'')'' is rewritten as ``1-'' in katakana, and the word ``summit'' is written in katakana.

FIG. 3 shows a flowchart of the misrecognition correction process and unknown word process based on the Japanese language analysis described above.

FIG. 4 is a block diagram showing another embodiment of the present invention.

In this embodiment, a grammar check unit 19 is added to the unknown word processing unit 9a, and it is possible to convert the unknown word into a common noun or a strange noun and connect it with the word that succeeded in the next Japanese analysis. The candidate character search unit 16 and character type determination unit 17 perform unknown word processing only on unknown words determined to be connectable. By performing such a grammar check before processing unknown words, it is possible to reduce the number of erroneous corrections in which unknown words consisting of kanji are rewritten into katakana, alphabets, numbers, or symbols.

The processing of this embodiment is shown in FIG. 5 as a flowchart.

Although the embodiments applied to character recognition devices have been described above, the present invention can be similarly applied to voice recognition devices, etc.
Furthermore, each function may be realized by either hardware or software.

〔Effect of the invention〕

As explained in detail above, according to the present invention, it is possible to avoid the increase in the capacity of word dictionaries and the disadvantages caused by them, and to avoid the disadvantages of increasing the capacity of word dictionaries. It can be corrected and the recognition rate can be improved.

[Brief explanation of the drawing]

FIG. 1 is a block diagram showing an embodiment of the present invention, FIG. 2 is a diagram showing an example of the contents of a candidate lattice memory, FIG. 3 is a processing flowchart of FIG. FIG. 5, a block diagram showing the embodiment, is the processing flowchart i- of FIG. 4. 2... Character recognition device, 3... Candidate character lattice generation section, 4... Candidate lattice memory, 5... Erroneous recognition correction section, 6... Word dictionary, 7... Grammar dictionary, 8 ...
Candidate word memory, 9, 9a...Unknown word processing unit, 10.
...Unknown word position save memory, 13...Word matching unit, 14.19...Grammar check unit, 15...Unknown word check unit, 16...Candidate character search unit, 17...Character type determination Department.

Claims (2)

[Claims] (1) Correcting misrecognition through Japanese language analysis of the recognition results, and replacing parts of unknown words that failed Japanese language analysis with those included in the candidate characters or with the candidate characters and their similar characters. Includes katakana, alphabet,
A misrecognition correction method characterized by performing unknown word processing to correct it to numbers or symbols. (2) The feature is that the part of speech of the unknown word part is used as a common noun or a paranormal noun, and the connection with the next word that was successfully analyzed in Japanese is checked, and unknown word processing is performed only on the part of the unknown word where this connection is possible. The misrecognition correction method according to claim (1).