JPH0614375B2 - Character input device - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION Field of Application of the Invention The present invention relates to character recognition, so-called OCR (Optical Charact).
er Recognition) post-processing method or voice recognition post-processing method, and particularly to a character input device that evaluates and determines a so-called unreadable character that cannot be uniquely recognized from the viewpoint of language.

[Background of the Invention]

When recognizing handwritten characters or printed characters by checking their patterns, even if the characters are not readable and cannot be recognized correctly, some candidates often include correct characters. . Conventionally, as a post-processing method of character recognition for selecting a correct candidate in such a case, a method of determining correctness by comparing a recognized character string with a word dictionary is already known. Has been reported to be effective,
For general text input, the number of words is enormous, the processing is complicated and time-consuming, and it is difficult to put into practical use. Regarding the conventional character recognition method, for example, Japanese Patent Laid-Open No. 58-48
181, 58-166490.

[Object of the Invention]

An object of the present invention is to provide a character input device by character recognition that can reduce the number of correction steps for input characters by a post-recognition processing function by language analysis that can be applied to general sentences.

In the following description, the case of the Japanese language will be described as an example, but the present invention can also be implemented in other languages such as English, Korean, and Chinese. Further, although the output of character recognition will be described below in particular, recognition output by voice recognition can also be implemented.

[Outline of Invention]

The present invention recognizes a character or a voice and outputs a recognition result as a character candidate composed of at least a set of a character code, a similarity and a rank of the similarity, and the character candidate as a processing unit. A character string delimiter determining unit that delimits a character string, a character string candidate creating unit that creates a plurality of character string candidates by combining each character candidate in the character string that is the processing unit, and a language for the plurality of character string candidates In a character input device provided with a language analysis unit that outputs a language likelihood indicating the validity of, the maximum language likelihood among the language likelihoods obtained for the plurality of character string candidates is a predetermined language likelihood threshold. When the maximum linguistic likelihood is less than or equal to a predetermined linguistic likelihood threshold, the similarity rank is ranked first. Character candidates Characterized by comprising a string candidate selection means for outputting a character string candidates made.

Example of Invention

FIG. 1 shows that the character manuscript 1 is character-recognized by the character recognition unit 2 and the result is evaluated by the Japanese sentence analysis unit 3 to select the optimum recognized character candidate and code it as the text data 4. 2 is an outline of a processing procedure of. The present invention relates to the configuration and control of the Japanese sentence analysis unit 3 among them, performs the minimum necessary Japanese sentence analysis processing, and determines the optimum recognized character candidate based on both the evaluation value of the Japanese sentence analysis and the recognition similarity. It is for selective output.

An embodiment of the present invention will be described below with reference to the configuration diagram of FIG. 2, the flow chart of FIG. 3, and the data of FIGS. 4 to 6.

The character recognition unit 11 shown in FIG. 2 recognizes characters written on a character original 21 of a handwritten or printed document, and recognizes the recognition result for each character as a recognized character candidate composed of a plurality of character codes, similarity and rank. 22 is output. For example, the recognition character candidate 2 when the sentence "It is a big dog." Is recognized.
2 is output in the format shown in FIG. That is, the number of candidates 31 for each character and the character candidate 32 consisting of the character code a and the similarity degree b are output, and the rank is 1 in descending order of similarity degree.
The second place, the third place ...

The character string delimiter determination unit 12 divides the recognized character candidates 22 into analysis processing units so as to simplify the subsequent Japanese sentence analysis processing, and separates the first character candidates at punctuation marks or spaces. . In the example of FIG. 4, “large dog” is one analysis processing unit.

Next, the similarity threshold determination unit 13 compares the similarity of each character candidate in the analysis processing unit with a predetermined similarity threshold, and extracts a combined character candidate 23 for creating a character string candidate to be subjected to Japanese sentence analysis. To do. Specifically, as shown in FIG. 5, for character candidates that exceed a predetermined similarity threshold, the number of candidates is calculated for each character, and the number of combined character candidates 41 corresponds to the corresponding combined character candidate 42. Record. However, if there is no character candidate that exceeds the predetermined similarity threshold, the number of combined character candidates is set to 1. FIG. 5 shows that the predetermined similarity threshold is 6
It is a combination character candidate when 0 is set.

The character string candidate creating unit 14 creates the character string candidate 24 by combining the combined character candidates 23. With respect to the combination character candidates of FIG. 5, a combination corresponding to the product of the combination character candidate number 41 is formed, and 12 kinds of character string candidates 24 are output as shown in FIG.

The Japanese sentence analysis unit 15 checks the validity of a given character string as a Japanese sentence and can be realized by using a known technique. A simple method is to compare the character string with a word dictionary, obtain the presence or absence of the corresponding word, the part of speech, or the frequency of use of the word, and evaluate the linguistic likelihood by evaluating the grammatical connection with the preceding and following character strings. You can ask. As a further advanced method, the technique found in the literature (“Kana-to-Kanji conversion of solid written sentences”, 1987 National Institute of Electronics and Communication Information Division National Convention 91) is also used for solid written sentences that are not separated by clauses. The likelihood of Japanese sentences can be calculated. A detailed description of the Japanese sentence analysis method is omitted, and the Japanese sentence analysis method is treated as having a function of outputting the Japanese likelihood J to a character string. The Japanese sentence analysis unit 15 in FIG. 2 obtains the Japanese likelihood 25 for each of the character string candidates 24. The processing by the Japanese sentence analysis unit 15 is continued until the Japanese likelihood 25 is obtained for all the character string candidates 24, as shown in the upper part of the flow chart in FIG. 3, and the result is as shown in FIG. become. In the example of FIG. 6, the Japanese likelihood is represented by a 5-step evaluation, and the likelihood of the third character string candidate is 5, which is the maximum.

In the character string candidate selection unit 16, comparing the Japanese likelihood 25,
Detect the maximum Japanese likelihood. Let this maximum Japanese likelihood be J _MAX . A predetermined Japanese likelihood threshold is J _TH . As shown in the lower half of the flow chart in Fig. 3, J _MAX and J _TH are compared, and when J _MAX > J _TH , J
_The character string corresponding to _MAX is selected and output as the optimum character string output 26. In FIG. 6, the Japanese likelihood of the third character string candidate “big dog” is J _MAX = 5, so J
_{If TH} = 4, "Big dog" is output. If J _MAX ≤ J _TH, the character string does not form the body of a Japanese sentence, and the correct character string candidate cannot be determined even by the Japanese sentence analysis. A character string consisting of is selected from the character string candidates 24 and output as the optimum character string output 26.

The result will be specifically described as follows. First, J _MAX ≦ J _TH is likely to occur when there is no correct answer in the character candidates. A specific example is shown in FIG. 7. This is an example in which the character "day" cannot be recognized when "Japanese nationals" is recognized, and it is not included in the candidates. In this case, the Japanese likelihood of the character string candidate combined with the erroneous candidate of the adjacent character is maximized. That is, the Japanese likelihood of the character string candidate "Shiraki Kuniwa" combined with the second wrong candidate is the maximum value J _MAX . However, the processing in the case of J _MAX ≦ J _TH can eliminate such a harm, and “Shiramoto nationals” consisting of the first character candidate is output.

In this example, if the similarity of the recognized character candidates is very low, there is a high possibility that a correct combination character candidate does not exist, and in that case, control is performed so as to stop wasted Japanese sentence analysis. It is desirable to do. Therefore, the flow chart of the embodiment in which control is added is shown in FIG. That is, in the similarity threshold determination unit 13, prior to the matching of the predetermined similarity threshold, the lower threshold of the foot cut similarity and the similarity of the first character candidate are compared, and the threshold of the foot cut similarity is equal to or less than the threshold. If the similarity of the first character candidate of No. 1 exists, the character string candidate creating unit 14 creates a character string composed of the character candidates having the first similarity, and directly outputs this as the optimum character string output 26. To do.

Next, an embodiment in which the present invention is used as an input adapter of a word processor will be described with reference to FIG. In this embodiment, O
The CR 63 is used as an input means of the word processor 61. The document input control device 62 recognizes a handwritten document or a printed document by the OCR 63, stores the recognition result in the recognized character candidate memory 64, creates a character string candidate to be subjected to language analysis and stores it in the character string candidate memory 65, language. Each control such as control of the analysis device 66 and storage of the processing result in the character string output text memory 67 is performed.

The document input control device 62 is composed of a microprocessor, and has a recognized character candidate memory 64 and a character string candidate memory 6.
5. The character string output text memory 67 is a memory of the microprocessor, and the OCR 63 and the language analysis device 66 are each configured as an additional device of the microprocessor.

According to the present embodiment, among the characters read by OCR, it is possible to determine an unread character that cannot be uniquely determined by evaluating it from another viewpoint of likelihood as a language. It is possible to use OCR for initial input.

Here, it goes without saying that the same effect can be obtained even when a voice recognition device is used instead of the Kanji OCR, and a word processor can be realized by voice input.

〔The invention's effect〕

As described above, according to the present invention, the processing time can be reduced by first narrowing down the character string candidates to be subjected to language analysis to the necessary minimum. Next, in the case of a character string that is not well evaluated by linguistic analysis, the character candidate with the first highest degree of recognition similarity is selected and output. Since it is possible to eliminate the adverse effect of erroneous determination of, and to enhance the effect of character recognition post-processing by applying language analysis, OCR
And voice recognition can be used for general text input.

[Brief description of drawings]

FIG. 1 is a diagram showing an outline of the present invention, FIG. 2 is an embodiment of the present invention, FIG. 3 is its flow chart, FIG. 4 is a recognition character candidate, FIG. 5 is a combination character candidate, and FIG. Is a character string candidate and its Japanese likelihood, FIG. 7 is another example of a combination candidate, FIG. 8 is a flow chart when a control function is added in the configuration of FIG. 2, and FIG. 9 is another embodiment. FIG. 11 ... Character recognition unit, 12 ... Character string delimitation determination unit, 1
3 ... Similarity threshold determination unit, 14 ... Character string candidate creation unit,
15: Japanese sentence analysis unit, 16: Character string candidate selection unit.

Claims (2)

[Claims] 1. A recognition unit for recognizing a character or a voice and outputting a recognition result as a character candidate composed of at least a set of a character code, a similarity and a rank of the similarity, and the character candidate as a processing unit. A character string delimiter determining unit that delimits the character string into a character string, a character string candidate creating unit that creates a plurality of character string candidates by combining each character candidate in the character string serving as the processing unit, and a language for the plurality of character string candidates. In a character input device provided with a language analysis unit that outputs a language likelihood indicating the validity as, among the language likelihoods obtained for the plurality of character string candidates,
When the maximum linguistic likelihood exceeds a predetermined linguistic likelihood threshold, a character string candidate corresponding to the maximum linguistic likelihood is output, and when the maximum linguistic likelihood is less than or equal to the predetermined linguistic likelihood threshold. In the character input device, there is provided a character string candidate selecting means for outputting a character string candidate composed of the character candidates having the first rank of similarity. 2. The character string delimiter according to claim 1, wherein the character string which is the processing unit and is divided by the character string delimiter determination unit has a predetermined similarity among the character candidates having the first rank. When there is a character candidate that is equal to or less than the cutoff similarity threshold, the character string candidate creating unit determines whether the first rank of the similarity rank is the first.
A character input device characterized in that a character string composed of character candidates for positions is created and directly output.