JPH11328317A - Method and device for correcting japanese character recognition error and recording medium with error correcting program recorded - Google Patents

Abstract

PROBLEM TO BE SOLVED: To precisely present a correction candidate for a short word without using the context by retrieving a word in a dictionary which is similar to a partial character string included in a character matrix according to a character mixing probability and a word appearance frequency and presenting a word string in the increasing order of the product of the concurrency probability of the word string and the character mixing probability of the respective characters. SOLUTION: An unknown word candidate generating means 2 generates a pair of the notation and appearance probability of a word in order to identify an unknown word included in the character matrix. Further, a similar word collating means 3 retrieves the word in the dictionary which is similar to the partial character string included in the character matrix in order to generate the correction candidate for a word whose correct characters are not included in the candidate characters according to the character mixing probability and word appearance probability without using the contexts of the precedent and following parts. A morpheme analyzing means 1 outputs only an arbitrary number of word strings among combinations of words in the dictionary included in the character matrix, unknown word candidates, and similar collatated words in the decreasing order of the probability according to a word division model 7.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a technique for correcting a character recognition error of a Japanese character recognition device for recognizing printed characters or handwritten characters, and more particularly, to an optical character recognition device (O).
For the Japanese text including the character recognition error output by CR), using a statistical language model, a statistical character recognizer model, and a probabilistic morphological analysis algorithm, the error included in the input sentence can be detected with high accuracy. Related to the technology that can be corrected.

[0002]

2. Description of the Related Art The conventional error correction method for character recognition is as follows.
The method can be broadly classified into two methods: a method using a character connection probability (character ngram probability), and (2) a method using a word connection probability (word ngram probability). Methods using character connection probabilities (for example, Sugimura and Saito, "Unreadable Character Judgment Processing Using Character Concatenation Information-Application to Character Recognition-" IEICE Transactions Vol.J68-D No.1, pp.64-71, 1985)
Detects or corrects an error by determining the possibility (or likelihood) of connection of adjacent character candidates using a character connection table or a character connection probability.

[0003] This method of using the concatenated information of characters is easy to realize and can achieve relatively large improvement in recognition performance. However, since the input sentence is treated as a simple character string, a character string that does not form a word or a character string that is not grammatically correct is often tolerated. In general, the correction accuracy is lower than the method using the connection probability of the word. Low. On the other hand, methods using word connection probabilities (for example, Takao and Nishino, "Implementation and Evaluation of Post-processing for Japanese Document Readers", Transactions of Information Processing Society of Japan Vol.33 No.5, pp.664-670, 1992 , Ito Maruyama "O
Error Detection and Automatic Correction of Japanese Sentences Inputted by CR "Transactions of Information Processing Society of Japan Vol.33 No.5, pp.664-670, 1992 etc.)
Examines the connection between words while searching words composed of candidate characters for character recognition from the word dictionary, and determines the best word string that reaches from the beginning to the end of the sentence as the correct answer string .

The method of using the word dictionary and the word connection information is as follows: (a) When a word not registered in the dictionary exists in the input sentence, or (b) when the correct character is not included in the candidate character, There is a problem that performance is greatly reduced. As a method for dealing with words that are not registered in the dictionary, there is a method in which unregistered words are found using an unregistered word template and the character string is limited using the connection probability of the character.
Nishino's dissertation). However, this unregistered word template
It must be created heuristically by hand, and there are problems with completeness and portability.

[0005] As a countermeasure when the correct character is not included in the candidate characters, there are a candidate character complement and a similar word collation. Candidate character complement is a method in which a set of similar characters for which character recognition is likely to be erroneous is defined in advance, and this similar character is unconditionally added as a candidate to the candidate character of the character recognition result (Sugimura "Candidate" Error Correction Processing of Kanji Recognition by Character Completion and Language Processing, Transactions of the Institute of Electronics, Information and Communication Engineers, Vol.J8
2-D-II No. 7, pp. 993-1000, 1989). Candidate character complementation has a problem in that although the complemented character does not always include the correct character, the amount of calculation required for error correction increases substantially uniformly.

[0006] Similar word matching is a method of searching not only words that are completely matched but also words that are similar (partially matched) in matching with a word dictionary (for example, Wa).
gner and Fishcer “The String-to-String Correction Pr
oblem "Journal of ACM Vol.21 No.1, pp.168-173,
1974). As a distance measure used for similar word matching,
It is common to use an edit distance that indicates the number of insertions, deletions, and replacements required to convert one character string to the other.

[0007] Similar word matching is effective in spell checking of a language having many long words (about 5 characters on average) such as English. However, it is not very effective in a language such as Japanese, which has many short words (about 2 characters on average). For example,
When measuring the similarity of words by the number of matched characters, there is a problem that there are very many sets of two-letter words in which one character matches, and all sets of one-letter words have the same similarity.

As a similar word matching method for Japanese in character recognition error correction, a method has been proposed in which a set of corrected word candidates is first created based on the edit distance, and the corrected word candidates are ranked based on contexts before and after. (Nagata, "Japanese Character Recognition Method and Apparatus", Japanese Patent Application No. 7-115926, 1995).
However, the method of using the context is that if the word length is short like a name and the surrounding context is not available,
There is a problem that it does not work effectively.

[0009]

SUMMARY OF THE INVENTION The present invention has been made in consideration of the above-mentioned problems of the conventional method. 1. when the length of the word to be corrected is short; An object of the present invention is to provide a Japanese character recognition error correction method and apparatus capable of correctly presenting a corrected word candidate when context information before and after a word to be corrected is not available.

[0010]

According to a first aspect of the present invention, there is provided a method for correcting a Japanese character recognition error, the method comprising the steps of: detecting a word occurrence frequency as a word candidate for correcting a character recognition error contained in an input Japanese sentence; The present invention is characterized in that an arbitrary number of words are presented in descending order of probability based on the character and the degree of graphic similarity between characters.

FIG. 1 is a diagram showing the principle of the present invention. The Japanese character recognition error correction device of the present invention corrects a Japanese text error including a character recognition error. As shown in the figure, the Japanese recognition error correction apparatus of the present invention receives a character matrix, that is, a list in which character candidates are arranged in the order of a character recognition score at each character position of an input sentence, and receives two consecutive characters. A word segmentation model 7 that gives a word pair probability that is a conditional appearance probability of a second word when a first word is given to a word, and a word dictionary and a character matrix that the word segmentation model 7 has Word matching means 4 for matching the included character strings. In addition, the Japanese recognition error correction device uses the first word in the average word length and the two consecutive characters.
And a word model 5 that gives the probability of appearance of a word in which an arbitrary character string is written, from the character pair probability that is the conditional appearance probability of the second character when the character is given, Unknown word candidate generating means 2 for generating an unknown word candidate based on the word model 5 from a partial character string included in the character matrix in order to identify a word that is not present. The Japanese recognition error correction device includes a character recognition device model 6 for giving a character confusion probability between any two characters, and a character confusion for generating a corrected word candidate when a correct character is not included in a candidate character. There is further provided a similar word matching unit 3 for searching for a word similar to the character string included in the character matrix based on the probability and the word appearance probability. In addition, the Japanese error correction device may select an arbitrary number of words based on the word segmentation model 7, the word model 5, and the character recognition device model 6 in descending order of the product of the joint probability of the word string and the character confusion probability of each character. It has a morphological analysis means 1 for obtaining a word string.

FIG. 2 is a flowchart for explaining the principle of the present invention. The Japanese recognition error correction method of the present invention takes a character matrix output by the character recognition device 8 as an input, and first identifies words in a word dictionary that completely match the character strings included in the character matrix. The word matching means 4 generates word candidates (step 1). Next, in order to identify an unknown word included in the character matrix, that is, a word that is not registered in the dictionary, an unknown word candidate generation unit 2
Generates an unknown word candidate, that is, a set of a word notation and an appearance probability (step 2). Further, in order to generate a correction candidate for a word whose correct character is not included in the candidate character, the similar word matching means 3 determines a word in the dictionary similar to the partial character string included in the character matrix by using the character confusion probability and the word A search is performed based on the appearance probability without using the context before and after (step 3). Finally, the morphological analysis unit 1 selects a morphological analysis candidate, that is, a word string, from a combination of words in the dictionary, unknown word candidates, and similar matching words included in the character matrix based on the word division model 7. Are output in arbitrary order in the order of higher probability (step 4).

Therefore, according to the present invention, even when the input sentence includes a word that is not registered, or when the correct character is not included in the character candidate, the context is determined based on the character confusion probability and the word appearance probability. It is possible to realize a Japanese character recognition error correction method and apparatus that generates correction candidates for short words without using them and presents the word strings in descending order of the product of the joint probability of the word strings and the character confusion probability of each character.

Hereinafter, various configurations of the Japanese character recognition method and the Japanese character recognition device of the present invention which achieve the above object of the present invention will be described. FIG. 3 is an explanatory diagram of the Japanese character recognition error correction method of the present invention. As shown in the figure, the Japanese character recognition error correction method of the present invention includes a word segmentation model 91 which inputs a Japanese sentence containing a character recognition error (step 91) and gives a joint probability of a word string constituting the sentence; A character recognizer model 92 that gives the probability of character confusion between any two characters
, A word in the dictionary 93 that is similar to the character string of the word to be corrected is searched (step 92).
Regardless of the length of the word to be corrected and the availability of context information before and after the word to be corrected, an arbitrary number of words are selected as correction word candidates for the word to be corrected in the order of higher probability. (Step 93).

In the above description and FIG.
Is provided separately from the word division model 91, but the dictionary 93 may be included in the word division model 91.
FIG. 4 is a block diagram of the Japanese character recognition error correction apparatus 10 that realizes the method of the present invention. The Japanese character recognition error correction device 10 includes a storage unit 12 for storing the frequency of occurrence of a word.
Calculating means 13 for calculating the graphic similarity of the character; and obtaining the appearance frequency of the words constituting the Japanese sentence given to the input means 11 from the storage means 12.
The character similarity of the characters constituting the Japanese sentence is calculated, and the character recognition error included in the Japanese sentence is determined based on the appearance frequency of the acquired word and the character similarity of the character. Means 14 for presenting an arbitrary number of words in order of increasing probability as word candidates to be corrected.

FIG. 5 is a block diagram of another Japanese character recognition error correction device 20 according to the present invention. The Japanese character recognition error correction device 20 shown in FIG. 1 includes a unit 22 that generates a word that matches a character string included in a Japanese sentence provided to an input unit 21 as a word candidate, and an unknown part included in the sentence. A means 23 for generating a character string as an unknown word candidate, a means 24 for generating a word similar to the partial character string included in the sentence as a corrected word candidate regardless of the context of the sentence, Means 25 for outputting an arbitrary number of word candidate strings in the order of probability as a corrected word string of a character recognition error included in the Japanese sentence from the combination of the word candidate and the corrected word candidate.

FIG. 6 is a diagram for explaining the operation of the error correction method implemented by the Japanese character recognition error correction device 20 of the present invention. As shown in the figure, the Japanese character recognition error correction method of the present invention generates a word that matches a character string included in a Japanese sentence given as an input as a word candidate (step 2).
1) An unknown partial character string included in the sentence is generated as an unknown word candidate (step 22), and a word similar to the partial character string included in the sentence is determined as a corrected word candidate regardless of the context of the sentence. Generate (Step 23), and from the combinations of the word candidates, the unknown word candidates, and the corrected word candidates, an arbitrary number of word candidate strings are set as the corrected word strings of the character recognition errors included in the Japanese sentence. (Step 24).

FIG. 7 is a block diagram of another Japanese character recognition error correction device 30 according to the present invention. A Japanese character recognition error correction device 30 for correcting a character recognition error included in a Japanese sentence given as an input shown in FIG.
A character recognizing device model storing means 32 for giving a character confusion probability between any two characters, and a similar word matching means 33 for searching a word in a dictionary 35 for similar matching with the character string of the word to be corrected. And the length of the word to be corrected,
Means 34 for presenting an arbitrary number of words as correction word candidates for the word to be corrected in the order of high probability, regardless of whether context information before and after the word to be corrected is usable. .

In the Japanese character recognition error correction device 30 for correcting a character recognition error included in a Japanese sentence given as an input according to the present invention, the similar word matching means 33 determines a joint probability of a word string constituting a sentence. Giving word segmentation model 3
1 and a character recognition device model 32 that gives a character confusion probability between any two characters, a word in the dictionary 35 that is similar to the character string of the word to be corrected is searched for, and the corrected word is searched. The candidate presenting means 34 may determine any number of correction word candidates for the word to be corrected, regardless of the length of the word to be corrected and the availability of context information before and after the word to be corrected. Words are presented in order of probability.

FIG. 8 is a block diagram of another Japanese character recognition error correction device 40 according to the present invention. The Japanese character recognition error correction device 40 shown in FIG. 3 includes a means 41 for inputting a character matrix which is a list of character candidates arranged in the order of the character recognition score for each character position of the Japanese sentence, and an average word length. A word model storage means 42 for storing a character pair probability indicating a conditional appearance probability of a second character when a first character is given out of two consecutive characters, and the word model storage means Based on the average word length and the character double set probability stored in 42, from among the partial character strings included in the character matrix, a set of an arbitrary number of unknown word expressions and appearance probabilities as unknown word candidates is determined. From the combination of word candidates including the unknown word candidate, regardless of whether the Japanese sentence includes a word not registered in the dictionary. Comprises means 44 for presenting the word candidate sequence of any number of sequentially product is large character confusion probabilities for each character constituting the joint probability and the Japanese sentence word strings constituting the sentence.

The Japanese character recognition error correction device 40
A character matrix, which is a list of character candidates arranged in the descending order of the character recognition score for each character position in the Japanese sentence, is input, and the average word length and the first character in two consecutive characters are input.
Is given as an unknown word candidate from among the partial character strings included in the character matrix, based on the character double set probability indicating the conditional appearance probability of the second character when the character is given. The set of the notation of the unknown word and the appearance probability is obtained in the order of high probability, regardless of whether the Japanese sentence includes a word not registered in the dictionary, from among the combinations of the word candidates including the unknown word candidate An arbitrary number of word candidate strings are presented in descending order of the product of the joint probability of the word strings constituting the Japanese sentence and the character confusion probability of each character constituting the Japanese sentence.

FIG. 9 is a block diagram of another Japanese character recognition error correction device 50 according to the present invention. The Japanese character recognition error correction device 50 shown in FIG. 1 includes a character matrix 51 which is a list of character candidates arranged in the order of the character recognition score for each character position of a Japanese sentence, and a character matrix. Searches words in the dictionary for similarity matching with the included character strings, and estimates them from a character confusion matrix that indicates which characters are recognized by which characters and how often, and character feature vectors that digitize the graphic features of the characters. Character recognition device model storage means 52 for storing the character confusion probability between the two characters to be obtained, and obtained from the character recognition device model storage means irrespective of whether or not the correct character is included in the character candidate. Means 53 for presenting an arbitrary number of word candidate strings in the order in which the character confusion probability of each character constituting the Japanese sentence is large.

The Japanese character recognition error correction device 50 of the present invention
Enters a character matrix, which is a list of character candidates arranged in the order of the highest character recognition score for each character position in the Japanese sentence, and searches for words in the dictionary that are similar to the character strings contained in the character matrix. Estimate the probability of character confusion between two characters from a character confusion matrix that indicates which characters are recognized by which characters and how often, and a character feature vector that digitizes the graphic characteristics of the characters. Regardless of whether or not it is included in the character candidates, an arbitrary number of word candidate strings are presented in the order in which the character confusion probability of each character constituting the Japanese sentence is large.

FIG. 10 is a block diagram of another Japanese character recognition error correction device 60 according to the present invention. The Japanese character recognition error correction device 60 shown in FIG. 7 includes a character matrix 61 which is a list of character candidates arranged in the descending order of the character recognition score for each character position of a Japanese sentence, and a character matrix. Searches words in the dictionary for similarity matching with the included character strings, and estimates them from a character confusion matrix that indicates which characters are recognized by which characters and how often, and character feature vectors that digitize the graphic features of the characters. A character recognition device model storage means 62 for storing a character confusion probability between two characters to be performed, a word division model storage means 63 for storing a word appearance probability, and the character confusion probability and the word appearance probability based on the character confusion probability and the word appearance probability. A similar word matching unit 64 for ranking words; and a word including the ranked words regardless of whether the correct character is included in the character candidates. Means 65 for presenting an arbitrary number of word candidate strings in descending order of the product of the simultaneous probability of the word strings constituting the Japanese sentence and the character confusion probability of each character constituting the Japanese sentence from among the complement combinations Consists of

The Japanese character recognition error correction device 60 of the present invention inputs a character matrix which is a list of character candidates arranged in the order of the character recognition score for each character position of the Japanese sentence, and inputs the character matrix to the character matrix. Searches words in the dictionary for similarity matching with the included character strings, and estimates them from a character confusion matrix that indicates which characters are recognized by which characters and how often, and character feature vectors that digitize the graphic features of the characters. The words are ranked based on the character confusion probability and the word appearance probability between the two characters to be determined, regardless of whether the correct character is included in the character candidates or not. From among combinations of word candidates that contain
An arbitrary number of word candidate strings are presented in the descending order of the product of the joint probability of the word strings constituting the Japanese sentence and the character confusion probability of each character constituting the Japanese sentence.

FIG. 11 is a block diagram of another Japanese character recognition error correction device 70 according to the present invention. The Japanese character recognition error correction device 70 shown in FIG. 7 includes a character matrix 71 which is a list of character candidates arranged in the order of a character recognition score for each character position of a Japanese sentence; Dictionary 7 for collating with partial character strings included in
A word matching means 72 for searching for a word in 3; an average word length; and a character pair probability indicating a conditional appearance probability of a second character when a first character is given in two consecutive characters And a word model storage unit 74 that stores the word in the character matrix on the basis of the stored average word length and the character double probability when the Japanese sentence includes a word not registered in the dictionary. Unknown word candidate generating means 75 for obtaining a set of an arbitrary number of unknown word expressions and appearance probabilities in the order of probability as unknown word candidates from among the partial character strings, and correct characters are not included in the character candidates In this case, search for words in the dictionary that are similar to the character strings contained in the character matrix, and digitize the character confusion matrix that indicates which characters are recognized by which characters and how often. did A character recognition device model storage unit 76 for storing a character confusion probability between two characters estimated from a character feature vector, a word segmentation model storage unit 77 for storing a word appearance probability, A combination of similar word matching means 78 for ranking the words based on the word appearance probability, a word in the dictionary included in the character matrix, the unknown word candidate, and a word candidate including the ranked word And a morphological analysis unit 79 that presents an arbitrary number of word candidate strings in the descending order of the product of the simultaneous probability of the word strings constituting the Japanese sentence and the character confusion probability of each character constituting the Japanese sentence. Consists of

The Japanese character recognition error correction device 70 of the present invention inputs a character matrix which is a list of character candidates arranged in the order of the character recognition score for each character position of the Japanese sentence, Search for a word in the dictionary to be matched with the substring included in. If the Japanese sentence includes a word not registered in the dictionary, the average word length and the first character in two consecutive characters are given. Based on the two-character probability set indicating the conditional appearance probability of the second character at the time of being selected, any number of sub-strings included in the character matrix in the order of higher probability as unknown word candidates When the correct character is not included in the above character candidates, a search is performed for a word in the dictionary that is similar to the character string included in the character matrix. Based on the character confusion matrix indicating how frequently the character is recognized and the character confusion probability between the two characters estimated from the character feature vector obtained by digitizing the graphic features of the character, and the word appearance probability. The word sequence forming the Japanese sentence from the combinations of the words in the dictionary included in the character matrix, the unknown word candidates, and the word candidates including the ranked words included in the character matrix. An arbitrary number of word candidate strings are presented in the descending order of the product of the joint probability of the characters and the character confusion probability of each character constituting the Japanese sentence.

Further, the present invention provides a character recognition device model creation device applied to a Japanese character recognition error correction device. FIG. 12 is a configuration diagram of a character recognition device model creation device. As shown in the figure, a character recognition device model creation device 80 of the present invention includes a character confusion frequency storage unit 81 for storing a character confusion frequency which gives a character and how frequently a character is recognized, Character feature vector storage means 82 for storing a character feature vector obtained by digitizing a characteristic feature
And a character clustering unit 83 that classifies the characters into a plurality of character classes by combining character feature vectors that are close to each other into one character class, and a character of a certain character class based on the character confusion frequency and the character class. Means 84 for estimating the probability of character class confusion, which indicates how often a character is recognized by a character of which character class, and unobserved for estimating the sum of the probability that a character is recognized by a character which has not been observed before. Event probability estimating means 85, allocating the sum of the probabilities of the unobserved events in proportion to the character class confusion probability, and normalizing the sum to 1 to obtain the character confusion probability of the unobserved event, Means 86 for distributing the sum of the probabilities of the observed events in proportion to the probability of character confusion, and normalizing the sum of the probabilities to be 1; Characterized in providing the letter confusion probability between any two characters based on conformal similarity.

FIG. 13 is a diagram for explaining the operation of the character recognition device model creation device 80 of the present invention. As shown in the figure, the character recognition device model creation device 80 determines whether or not the character confusion frequency is 0 (step 81), and if it is 0, estimates the sum of the unobserved event probabilities (step 82). ), Find the class to which the input and output characters belong, estimate the character class confusion probability (step 83), distribute the sum of the unobserved event probabilities in proportion to the character class confusion probability, and find the sum of the overall probabilities. If the frequency of character confusion is not 0, the probability of the observed event is distributed in proportion to the probability of character confusion, and the sum of the probabilities is normalized to 1 (step 84). (Step 85).

The present invention also includes a recording medium on which software (program) for realizing the respective constituent requirements of the Japanese character recognition error correction device and the character recognition device model creation device of the present invention is recorded. Such a Japanese character recognition error correction program and a character recognition device model creation program of the present invention are stored in a disk device or the like, and installed on a computer of a character recognition error correction system as necessary, and each of them is provided with a character recognition error correction program. It is also possible to make corrections and create a character recognition device model.

A recording medium on which the Japanese recognition error correction program of the present invention is recorded has a process for storing the frequency of appearance of words, a process for calculating the degree of graphic similarity between characters, and a process for forming a Japanese sentence given as input. To obtain the appearance frequency of the stored word, and to obtain the calculated graphic similarity of the characters constituting the Japanese sentence, based on the obtained appearance frequency of the word and the graphic similarity of the character. Then, as a word candidate for correcting a character recognition error included in the Japanese sentence, a process of presenting an arbitrary number of words in order of high probability is recorded.

The present invention also provides a process for generating a word that matches a character string included in a Japanese sentence given as an input as a word candidate, and generating an unknown partial character string included in the sentence as an unknown word candidate. A process of generating a word similar to the partial character string included in the sentence as a corrected word candidate regardless of the context of the sentence, and a combination of the word candidate, the unknown word candidate, and the corrected word candidate. A process of outputting an arbitrary number of word candidate strings in the order of probability as a corrected word string of a character recognition error included in the Japanese sentence. It is a recording medium on which recording is performed.

Further, the present invention provides a process for giving a joint probability of a word string constituting a Japanese sentence given as an input, a process for giving a character confusion probability between any two characters, and a process for correction. Regardless of the process of searching for a word in the dictionary that performs similarity matching with the character string of the word, the length of the word to be corrected, and the availability of context information before and after the word to be corrected, And a process of presenting an arbitrary number of words as correction word candidates in the order of probability with high probability.

Further, the present invention provides a process of inputting a character matrix which is a list of character candidates arranged in the order of the character recognition score for each character position in a Japanese sentence, an average word length, and two consecutive words. A process of storing a character double probability indicating a conditional occurrence probability of a second character when a first character is given to a character, and a process of storing the average word length and the character double probability stored above. A process of obtaining a set of an arbitrary number of unknown words as unknown word candidates and appearance probabilities from the partial character strings included in the character matrix in the order of higher probability, and registering the Japanese sentence in the dictionary Regardless of whether or not it includes a word that is not performed, from among combinations of word candidates including the unknown word candidate,
A process of presenting an arbitrary number of word candidate strings in the descending order of the product of the simultaneous probability of the word strings constituting the Japanese sentence and the character confusion probability of each character constituting the Japanese sentence. This is a recording medium that stores a Japanese character recognition error correction program.

Further, according to the present invention, there is provided a process of inputting a character matrix which is a list of character candidates arranged in the descending order of the character recognition score for each character position of a Japanese sentence. A word in the dictionary to be collated is searched, and a character confusion matrix that indicates which character is recognized by which character and how often, and two character values estimated from the character feature vector that digitizes the graphic features of the character. The process of storing the character confusion probability between the characters and the character confusion of each character constituting the Japanese sentence obtained from the character recognition device model storage means regardless of whether the correct character is included in the character candidate or not. A process of presenting an arbitrary number of word candidate strings in the order of the highest probability. It is.

The present invention also relates to a process of inputting a character matrix which is a list of character candidates arranged in order of character recognition score for each character position of a Japanese sentence, and a process similar to the character string included in the character matrix. A word in the dictionary to be collated is searched, and a character confusion matrix that indicates which character is recognized by which character and how often, and two character values estimated from the character feature vector that digitizes the graphic features of the character. A process of storing the character confusion probability between the two, a process of storing the word appearance probability, a process of ranking the words based on the character confusion probability and the word appearance probability, and Regardless of whether they are included or not, the joint probabilities of the word strings constituting the Japanese sentence are selected from the combinations of the word candidates including the ranked words. A process of presenting an arbitrary number of word candidate strings in the order of the product of the character confusion probability of each of the characters constituting the Japanese sentence in the descending order. It is.

According to the present invention, there is also provided a process for inputting a character matrix which is a list of character candidates arranged in order of character recognition score for each character position of a Japanese sentence, and a process for inputting a partial character string included in the character matrix. A process of searching for a word in a dictionary to be matched with, a character-duplex probability indicating an average word length and a conditional appearance probability of a second character given a first character in two consecutive characters And when the Japanese sentence includes a word that is not registered in the dictionary, based on the stored average word length and the probability of double character combination, the partial character string included in the character matrix is A process of obtaining a set of notations and appearance probabilities of an arbitrary number of unknown words in the order of probability as unknown word candidates from the middle, and when the correct character is not included in the above character candidates , A character confusion matrix that indicates which words are recognized in a dictionary, and how often a character is recognized by a character in the dictionary that matches the character string included in the character matrix, and a character that digitizes the graphic characteristics of the character A process of storing a character confusion probability between two characters estimated from a feature vector, a process of storing a word appearance probability, and ranking the words based on the stored character confusion probability and word appearance probability. A combination of a word in the dictionary included in the character matrix, the unknown word candidate, and a word candidate including the ranked word, and a simultaneous probability of a word string constituting the Japanese sentence. And a process of presenting an arbitrary number of word candidate strings in the descending order of the product of the character confusion probability of each character constituting the Japanese sentence. Which is a recording medium for recording a Japanese character recognition error correction program that.

The present invention also relates to a process for storing a character confusion frequency which gives a character and how often a character is recognized, and a process for storing a character feature vector obtained by digitizing graphic features of a character. By combining character feature vectors that are close to each other,
A process of classifying the characters into a plurality of character classes;
A process of estimating a character class confusion probability indicating how often a character of a certain character class is recognized by a character of the character class based on the character confusion frequency and the character class, and observing a certain character so far. The process of estimating the sum of the probabilities of being recognized by the unobserved characters, and distributing the sum of the probabilities of the unobserved events in proportion to the character class confusion probability, and normalizing the sum to be 1 Determining the character confusion probability of the unobserved event, distributing the sum of the observed event probabilities in proportion to the character confusion probability, and normalizing the total probability to be 1; A recording medium storing a character recognition device model creation program characterized by giving a character confusion probability between any two characters based on a graphical similarity.

[0039]

FIG. 14 is a schematic block diagram of a Japanese character recognition error correction system according to one embodiment of the present invention. As shown in the figure, a Japanese character recognition error correction system includes a character recognition device 101 that receives a Japanese input sentence and outputs a character matrix, and a character recognition error that inputs a given character matrix and outputs a word string. And a correction device 100.

The character recognition error correction apparatus 100 includes a forward search section 102, a partial analysis table 103, a backward search section 1
04, average word length table 105, unknown word candidate generator 1
06, similar word matching section 107, word appearance probability table 1
08, word matching means 109, word double probability table 1
10, character double set probability table 111, character confusion probability calculation unit 112, character confusion frequency table 113, character class confusion frequency table 114, character clustering unit 11
5 and a character feature vector table 116.

The forward search unit 102 receives a character matrix output from the character recognition device 1 for an input sentence and uses a dynamic programming method (Dynamic Programming) that proceeds from the beginning of the input character matrix to the end of the sentence one character at a time. The word division of the input sentence is determined so that the product of the joint probability of the word string (ie, the product of the word pair probabilities) and the character confusion probability of each character is maximized. The forward search unit 102 calculates, for each last word, the probability of partial analysis (word sequence) that maximizes the product of the joint probability of the word string from the beginning of the sentence to a certain word and the character confusion probability of each character. , In the partial analysis score table 103.

In the forward search, word candidates are proposed by the word matching unit 109, the similar word matching unit 107, and the unknown word candidate generation unit 106. The product of the character confusion probability of each character constituting the word is given to the proposed word candidate by the character string confusion probability generation unit 112. Further, the word pair probability is given from the word pair probability table 110.

In the following, at each character position in the character matrix, a character string formed by selecting one character at a time from a list of character candidates at that character position is referred to as a "character string included in the character matrix". The word matching unit 109 matches the character string included in the character matrix with the word notation in the word appearance probability table 108, and proposes the matched word candidate.

The similar word matching means 107 performs similarity matching between the character strings included in the character matrix and the word expressions in the word appearance probability table 108 using the word appearance probabilities given from the word occurrence probability table 108. The unknown word candidate generation unit 106 regards a character string included in the character matrix that does not match with the word notation in the word appearance probability table 108 as an unknown word, and calculates the average word length given from the average word length table 105 and the character The word appearance probability is estimated using the character pair probabilities given from the pair probability table 111, and a predetermined number of unknown words are proposed as unknown word candidates in descending order of output probability.

The backward search unit 104 inputs the partial analysis stored in the partial analysis table 103 in the forward search unit 2 and uses the A ^* algorithm that advances one word at a time from the end of the sentence to the beginning of the sentence to determine the joint probability of the word string and the Word string candidates are obtained one by one in order from the one with the largest product of the character and the character confusion probability. The character confusion probability calculation unit 112 obtains a character confusion probability from the character confusion frequency stored in the character confusion frequency table 113 and the character class confusion frequency stored in the character class confusion frequency table 114.

The character clustering section 115 classifies the character feature vector table 116 into a predetermined number of character classes using a vector quantization algorithm, and obtains the number of confusions between the character classes from the character confusion frequency table 113. In the following, first, the "information theoretical interpretation of character recognition error correction" which is the theoretical basis of the present invention will be described, followed by a character recognition device model, a word segmentation model, a word model, a forward search unit and a backward search unit. , An unknown word candidate generation unit, and a similar word collation unit.

Information Theoretic Interpretation of Character Recognition Error Correction In one embodiment of the present invention, the relationship between the input and output of the character recognition device 1 is based on the model of a noisy channel (noisy channel).
model). Assuming that the character recognition result for the input character string C is X, the character recognition error correction is performed on the character string that maximizes the posterior probability P (C | X).

[0048]

[Outside 1]

The problem is as follows. According to Bayes' theorem, the following relationship holds,

[0050]

(Equation 1)

P (X | C) Character string that maximizes P (C)

[0052]

[Outside 2]

It is sufficient to obtain

[0054]

(Equation 2)

Here, P (X | C) is called a character recognition device model, and P (C) is called a language model. Hereinafter, a character recognition device model and a language model used in one embodiment of the present invention will be described. · Character recognition device model character recognition device model P (X | C), each character c _i constituting the input character string C is can be calculated from the product of the probabilities that are recognized in the character x _i that make up the X.

[0056]

(Equation 3)

P (x _i | c _i ) is the character confusion probability (charact
It is called an "er confusion probability" and can be basically determined from a character confusion matrix which is frequency data of a set of input and output of the character recognition device. However, the character confusion matrix has low versatility because the character recognition method largely depends on the quality of the input image. Also, since Japanese has more than 3000 types of characters, it is not possible to collect a sufficiently large number of character recognition results for all characters. Since the character confusion matrix inevitably becomes sparse, it is inappropriate to estimate the character confusion probability P (x _i | c _i ) from the relative frequency with the maximum likelihood.

For example, assume that the word "environment" is erroneously recognized as "technical environment". FIG. 15 is a diagram illustrating an example of data of a character confusion matrix regarding characters “ring” and “border”.
A set of letters and numbers separated by a slash ('/')
Represents output characters and their frequencies for input characters. When 1,291 recognition results for the character "ring" are collected, the output characters are 1289 "ring", one "search",
There is only one "image", and there is no example that it is recognized as "skill".
Therefore, when the maximum likelihood is simply estimated, the character confusion probability P (technique |
Ring) = 0. This is called a “zero frequency problem” and is a problem that always occurs when a probability model is estimated from a small number of learning data.

Accordingly, in one embodiment of the present invention, a character feature vector, which is data obtained by digitizing the graphic characteristics of a character, is combined with a character confusion matrix, so that a character based on the graphic similarity of the character is used. Smooth the confusion probability (smoothi
ng). Therefore, the sum of the appearance probabilities of all unobserved events is first estimated, and then the probabilities are redistributed to each unobserved event based on the similarity of the characters. The method will be described below.

Estimation of Probability Occurring from Unobserved Event In one embodiment of the present invention, the method of Witten and Bell (Witten and Bill,
“The Zero-Frequency Problem: Estimating the Proba
bilities of Novel Events in Adaptive Text Compress
ion ", IEEE Transaction on information Theory, Vo
l.37, No.4, pp.1085-1094, 1991). Witten
According to Bell's method, the total number of events already observed, that is, the total calculated ignoring the difference in type is n, and the total number differently, that is, the total calculated as one for the same type is r. Then, the probability that a new event (unobserved event) will occur is estimated as r / (n + r). On the other hand, the probability of an event already observed c times is estimated to be c / (n + r). Witt
The method of en and Bell reflects the empirical rule that the probability of observing a new event increases as the total number of observed events increases and decreases as the total number of observed events increases. There is an advantage that accurate accuracy can be obtained. The method of Witten and Bell was devised in the field of text compression, and applying this method to character recognition error correction is a novel matter unique to one embodiment of the present invention.

The frequency of the event that the character c _i is recognized by the character c _j is C (c _i , c _j ), and the sum of the probability of the event that the character c _i is recognized by the unobserved character is β ( c _i ). From Witten and Bell's method, β (c _i ) is estimated as follows.

[0062]

(Equation 4)

Here, Θ (x) is a different step function introduced for obtaining the total number as follows.

[0064]

(Equation 5)

In the example shown in FIG. 15, the character “ring” is 1
It is input 291 times (= 1289 + 1 + 1) times, and the output includes three characters of "ring", "search", and "image". Therefore, the probability of observing a new character is 3 / (1291+
3) = 3/1294. Character clustering based on graphical similarity Witten and Bell's method can be used as a general method for calculating the sum of probabilities of unobserved events, but the method of redistributing probabilities to individual unobserved events is It is necessary to devise a new one according to the object to be modeled. The simplest probability redistribution method is to assume all unobserved events are of equal probability. However, in the case of the character confusion probability, recognition errors are more likely to occur for characters that are graphically similar, so the assumption of equal probability is inappropriate.

In one embodiment of the present invention, in order to redistribute the character confusion probability of an unobserved event based on the tendency of errors between characters that are graphically similar, first, based on the graphical similarity, The characters are classified into an appropriate number of classes, and the confusion frequency between the character classes is determined from the character confusion frequency. Then, the probability of the unobserved event is redistributed in proportion to the character class confusion frequency. The character class is obtained by clustering character feature vectors. Generally, character recognition is composed of feature extraction and classification. The feature extraction is an operation of extracting a plurality of feature amounts effective for character recognition from an image, and the classification is an operation of comparing a feature vector of an input character image with a representative vector of each character prepared in advance based on some distance scale. Therefore, if the characters are clustered based on the distance between the character feature vectors, a group of characters that are graphically similar and reflect the tendency of error in character recognition can be obtained.

In one embodiment of the present invention, the degree of outline direction contribution (Ogita / Naito / Masuda "Identification of handwritten kanji by outline direction contribution degree feature") is used as the character feature quantity, IEICE Transactions, Vo
l. J66-D, No. 10, pp. 1185-1192, 1983), and the LBG algorithm (Linde, Bu
zo and Gray “An algorithm for Vector QuantizerDes
ign ", IEEE Transactions on Communications, Vol.COM
-28, No. 1, pages 84-95, 1980). However, the character feature vector and the clustering method need only be a method that can finally obtain a class of characters that are graphically similar, and the contour direction contribution and LBG used in this embodiment are used.
It is not limited to algorithms. Although the contour direction contribution rate is originally a 1536-dimensional vector, a feature vector reduced to 256 dimensions is created using a two-stage feature selection method. Next, for 3021 Japanese characters, the feature vector of each character is calculated using the LBG algorithm.
One character was classified into 128 classes.

FIG. 16 is a diagram for explaining an example of a character class created by clustering. Class 29 is a character class including "ring", and class 119 is a character class including "skill". It can be seen that characters having graphically similar characteristics constitute one class. -Calculation method of character class confusion frequency and character confusion probability The procedure for obtaining the character class confusion frequency from the character class and the character confusion frequency will be described. The frequency at which a character belonging to the character class class _I is recognized as a character belonging to the character class class _J is called a character class confusion frequency, and C (class
s _I , class _J ). Character class confusion frequency C
(Class _I , class _J ) are class _I and c
It can be obtained from the sum of the elements of the character confusion matrix for the characters belonging to the class _J.

[0069]

(Equation 6)

Although less than the character confusion matrix, the character class confusion matrix also has zero elements. So, first,
We apply the Witten and Bell method to the character class confusion matrix to find the sum of the probabilities of the zero elements (unobserved events), and redistribute the probabilities assuming that the unobserved events are of equal probability. In this way, a class confusion probability can be defined between all character classes. FIG.
FIG. 7 is a diagram showing an example of a character class confusion matrix. In the figure, the character class confusion frequency of class 29 (character class including “ring”) is shown. Class 29 is 3 as input
It appears 1036 times and is recognized seven times by class 119 (a character class including "skill"). There are 36 different classes in the output. Therefore, the character class confusion probability is P (class ₁₁₉ | class ₂₉ ) = 7 / (31
036 + 36) = 7/31072.

The character confusion probability of an unobserved event is allocated so as to be proportional to the character class confusion probability. That is,

[0072]

(Equation 7)

Here, α (c _i ) is a normalization coefficient for making the sum of character confusion probabilities of unobserved events equal to β (c _i ).

[0074]

(Equation 8)

Class (c _i ) is a function for determining the class to which the character c _i belongs. Character class confusion probability P (cl
as (c _j ) | class (c _i )) and the normalization coefficient α (c _i ) can be calculated in advance, so that the calculation of the character confusion probability shown in equation (7) is realized very efficiently. it can. Language Model (Word Division Model) Subsequently, the language model (word division model), forward search and backward search will be described.

Character strings C = c ₁ c ₂ . . . If the input sentence composed of _cm is a word string W = w ₁ w ₂ . . . _Let it be divided into w _n . In one embodiment of the present invention, the probability P (C) of the character string C in Expression (2) is approximated by the probability P (W) of the most likely word string forming the character string C. Furthermore, in this embodiment, the joint probability P (W) of the word string is approximated by the product of the word pair probabilities as in the following equation.

[0077]

(Equation 9)

Here, "#" is a special symbol representing the beginning and end of a sentence. The word pair probability is learned in advance from a large amount of text data. Generally, P (W) is called a word segmentation model. When the word segmentation model is used as a language model as in Expression (9), the character recognition error correction performs posterior probability P (W
| X) word string that maximizes

[0079]

[Outside 3]

The result is the problem of finding Using Bayes' theorem as in equation (2), this results in P (X | W) and P
It is sufficient to maximize the product of (W).

[0081]

(Equation 10)

The word sequence that maximizes the expression (10) is obtained by a dynamic programming method that proceeds one character at a time from the beginning of the sentence to the end of the sentence. Furthermore, if an A ^* search that proceeds from the end of the sentence to the beginning of the sentence is used, an arbitrary number of word string candidates can be obtained in the order of increasing probability. In one embodiment of the present invention, a dynamic programming that proceeds one character at a time from the beginning of a sentence to the end of a sentence is referred to as a forward search, and an A ^* search that proceeds from the end of a sentence to the beginning of a sentence is referred to as a backward search.
Hereinafter, the forward search and the backward search will be described.

Forward Search A procedure in which the forward search unit 102 obtains a word string that maximizes the expression (10) will be described. And the joint probability of the word sequence from the beginning of a sentence to the i-th word _{P (w 1 ... w i)} , confusion each character of the characters that make up a word string probability P (x _j | c _j) and the maximum value of the product the defining and phi (w _i), the equation (9), the following relation is established.

[0084]

[Equation 11]

Here, q and r represent the start position and end position of the word w _i . That is, w _i = c _{q + 1} . . . a c _r, x _{q +} 1. . . x _r is a character recognition result corresponding to w _i. Equation (11) represents the following relationship. Maximum value of the product of the character confusion probabilities for each character constituting the joint probability and the word string from beginning of a sentence to the i-th word phi (w _i) are, i-1 from the beginning of a sentence
Th of the maximum value of the product of the character confusion probability of each of the characters that make up the joint probability and word string of up to word φ and (w _i-1), i-th word of the word duplicate probability P (w _i | w _i-1 ) is multiplied by the product of the character confusion probabilities of the characters constituting the i-th word. Based on this relationship, by obtaining from the beginning of a sentence in order phi (w _i), it is possible to obtain the maximum value of the probability from beginning of a sentence to the end of the sentence φ (w _n).

FIG. 18 is an operation flowchart of the forward search section 102 according to one embodiment of the present invention. The forward search implements the calculation of equation (11) using dynamic programming. In this case, referred to as the probability of φ (w _i) the partial analysis, φ (w _i)
Is referred to as a partial analysis table 103.
The forward search operation will be described below with reference to FIG. The forward search starts at the beginning of the input sentence and proceeds one character at a time toward the end of the sentence. In step 101, the search start position is set at the beginning of the input sentence.

At step 102, it is determined whether or not the search has reached the end of the sentence. If the end of the sentence has been reached, the forward search ends. Otherwise, the following processing is performed at each character position. In step 103, all partial analyzes that reach the current character position are searched from the partial analysis table 103, and one of them is selected as the current partial analysis.

In step 104, it is determined whether or not all the partial analyzes have been checked. If so, step 11
At 4, the search proceeds to the next character position. Otherwise, the following processing is performed for each partial analysis. Step 1
At 05, a set of all word candidates starting from the current character position is created, and one word is selected as the current word from the set. A set of all the word candidates is a word in the dictionary (obtained from the word matching unit 109) to be matched with the character string included in the character matrix, and a word (similarity) in the dictionary to be similarly matched with the character string included in the character string matrix. Word collating unit 10
7) and an unknown word candidate (obtained from the unknown word candidate generating unit 106) generated from a character string included in the character matrix that is not matched with the dictionary.

In step 106, it is determined whether all the words have been checked. If so, in step 113 the next partial analysis is selected. Otherwise, the following process is performed for each word. In step 107, it is determined whether or not the current word (partial analysis with the last word) is registered in the partial analysis table 103. If so, go to step 109. Otherwise, in step 108, this word is stored in the partial analysis table 103.
After the probability of the partial analysis is initialized to 0, the process proceeds to step 109.

In step 109, the probability of a new partial analysis based on the combination of the current partial analysis and the current word is obtained. The probability of the new partial analysis is

[0091]

[Outside 4]

Is as follows. Step 110 checks whether the probability of the new partial analysis is greater than the probability of the previous partial analysis where the last word is the same. If so, in step 111, the new partial analysis probability is stored in the partial analysis table 103, and the process proceeds to step 112. If not, the process proceeds directly to step 112.

At step 112, the next word is selected and the routine returns to step 106. In step 113, the next partial analysis is selected, and the process returns to step 104. In step 114,
The search proceeds to the next character position, and the process returns to step 102. -Backward search A procedure in which the backward search unit 104 obtains word segmentation candidates one by one in the descending order of the probability of Expression (10) will be described.

Some word division candidates W = w ₁ . . . w _n , the joint probability P of the word string from the end of the sentence to the i-th word
If the product of (w ₁ ... W _n ) and the character confusion probability P (x _j | c _j ) of each character constituting the word string is defined as ψ (w _i ), it is the same as in the case of the forward partial analysis. From equation (9),
The following relationship is established.

[0095]

(Equation 12)

Here, s and t represent the start position and the end position of the word wi _{+ 1} . That is, w _{i + 1} = _{cs + 1} . . .
a c _t, x _{s +} 1. . . _xt is the character recognition result corresponding to _{wi + 1} . φ (w _i) the probability of a positive partial analysis,
ψ the (w _i) will be referred to as the probability of the backward partial analysis.
Probability of word segmentation candidate, for any w _i, the probability of the probability φ (w _i) and the backward partial analysis of the prospective partial analysis ψ (w
_i ).

[0097] P (X | W) P ( W) = φ (w i) ψ (w i) (13) in the backward search of the present invention, in order to obtain one by one word segmentation candidates in descending order of probability A ^* Use an algorithm.
The A ^* algorithm is an algorithm for obtaining a minimum cost path of a state space graph (for details of the A ^* algorithm, see a textbook on artificial intelligence, for example, Avron
See Barr, Edward A. Feigenbaum, edited by Kokichi Tanaka and Kazuhiro Fuchi, "Artificial Intelligence Handbook, Volume I, Chapter II Search, Kyoritsu Shuppan, 1983).

Here, the A ^* algorithm will be briefly described. When an arbitrary node in the graph is n, the cost of the optimal path from the initial state to n is g (n),
Let h (n) be the cost of the optimal path from n to the final state. The cost f (n) of the optimal path through n is given by: f (n) = g (n) + h (n) (14) Consider the problem of finding the optimal path from the initial state to the final state. If f (n) is accurately known, a solution can be obtained by tracing the node where f (n) is minimum from the initial state. In practice, neither g (n) nor h (n) is exactly known, so a search is needed.

Let g (n) be the one with the least cost on the road to n known so far. The estimate of h (n) is

[0100]

[Outside 5]

Then, if the estimated cost

[0102]

[Outside 6]

Is less than the true cost h (n), ie

[0104]

[Outside 7]

Then,

[0106]

[Outside 8]

It can be proved that the optimal solution can be obtained by tracing the node where is minimized. A graph search strategy utilizing this property is called an A ^* algorithm. Also,
When a search algorithm can always find the optimal solution,
The search algorithm is said to be admissible. If estimated cost

[0108]

[Outside 9]

If the true cost h (n) matches,
The A ^* algorithm can immediately find the optimal route (without passing through nodes other than the optimal route). In general, the closer the estimated cost is to the true cost, the smaller the amount of search will be. If the estimated cost is always 0, the acceptable condition that the estimated cost is smaller than the true cost is satisfied, but it is practically the same as performing a breadth-first search using g (n), which is very efficient. Is bad.

In one embodiment of the present invention, the backward partial analysis is considered as a node of the graph, and the absolute value of the logarithm of the probability of the backward partial analysis is defined as a function g (n), and the probability of the forward partial analysis is defined as a function h (n). Use the absolute value of the logarithm of As a result, the solution with the highest probability of the word segmentation candidate is the cost function f
(N) corresponds to the minimum solution. H by forward search
Since the true value of (n) is known, the backward search is acceptable and the optimum solution can always be found immediately.
When the optimal solution is obtained, the node is removed and the search is continued to obtain the next optimal solution. In this way, in the backward search, an arbitrary number of word division candidates can be obtained in descending order of probability.

FIG. 19 is an operation flowchart of the backward search unit 104 according to one embodiment of the present invention. In general, A ^*
The algorithm uses two lists, open and closed. The list open has already been created,
This is a set of nodes for which the cost function h (n) of the remaining routes has been calculated, but has not been expanded (not checked). This list is a priority queue based on the value of the function h (n). The list closed is a set of already expanded (examined) nodes.

In the A ^* algorithm, one node is expanded in each step until a node corresponding to the final state is generated. In each step, the node having the lowest cost, which has been generated but has not been expanded yet, is expanded. That is, create a subsequent node of the selected node,
Calculate the cost function h (n) of the remaining routes, check whether they have already been generated, and add them to the list open.
This check ensures that each node appears only once in the graph. When two or more routes generate the same node, only the node with the smaller cost is recorded.

In the following, according to FIG.
The operation of No. 04 will be described. In step 201, a backward partial analysis representing the end of a sentence is substituted into the list open.
Also, an empty list is substituted for the list closed. In step 202, it is checked whether or not the list open is an empty list. If so, the search is terminated after notifying that the search failed because no solution was found. Otherwise, the following processing is performed.

In step 203, the first element of the list open is extracted and the current backward partial analysis is performed. Then, in step 204, it is checked whether or not the search has reached the beginning of the sentence. If the search has reached the beginning of the sentence, the current backward partial analysis is the optimal solution, and the fact that the search was successful is notified and the backward search is terminated. Otherwise, the following processing is performed. In addition, the following processing is performed when the search is successful but the next lowest cost (highest probability) solution is to be obtained.

In step 205, the current backward partial analysis is inserted into the list closed, and the list closed
Is the cost f of the entire path from the initial state to the final state
Sort in the order of (n). In step 206, all the words that can be connected to the left side of the current backward partial analysis are searched, and one of them is set as the current word.

In step 207, it is determined whether all the words have been checked. If so, step 202
Proceed to. Otherwise, the following processing is performed. In step 208, a new backward partial analysis representing a path from the current backward partial analysis to the current word is created, and this is substituted into a variable newpath. In step 209, it is checked whether a backward partial analysis for transitioning to the current word is already included in the list open. If not included, the process proceeds to step 213. If it is included, the following processing is performed.

In step 210, the element of the list open that transits to the current word is substituted for the variable oldpath. In step 211, the total path cost of the variable newpath is compared with the total path cost of the variable oldpath. If the total path cost of the variable newpath is larger, the process directly proceeds to step 218. If the cost of the variable newpath is smaller, step 2
In step 12, the element corresponding to the variable oldpath is deleted from the list open, the element corresponding to the variable newpath is inserted into the list open, and then sorted in order of cost. Then, the process proceeds to step 218.

In step 213, it is checked whether or not the backward partial analysis representing the transition to the current word is included in the list closed. If not included, the process proceeds to step 217. If it is included, the following processing is performed. In step 214, the element in the closed list that transitions to the current word is assigned to the variable oldpath.

In step 215, the variable newpath
Is compared with the total path cost of the variable oldpath. If the total path cost of the variable newpath is larger, the variable ol
Delete the element corresponding to dpath and set the variable newpat
After inserting the element corresponding to h into the list closed,
Sort by cost. Then, the process proceeds to step 218.

In step 217, the variable newpath
Are inserted into the list open, and then sorted in order of cost. Then, the process proceeds to step 218. In step 218, the next word is selected. Next, the word model and the unknown word candidate generation unit 106 will be described.

The word model is a calculation model for obtaining the appearance probability of a word not registered in the dictionary. This is, string c ₁ that make up the unknown word w _i. . . It is defined as the joint probability distribution of _ck . This can be represented by the product of the word length probability and the word notation probability without loss of generality.

[0122]

(Equation 13)

[0123] Here, k is the length of the string that make up the unknown words, <UNK> is a special symbol that represents the unknown word.
The word length probability P (k) is approximated by a Poisson distribution using the average word length as a λ parameter. This means that word segmentation is approximated by a stochastic process in which a word boundary symbol having a length of 0 appears randomly between characters at an average interval λ.

[0124]

[Equation 14]

The word notation probabilities P (c ₁ ... C _k ) are approximated by the product of the character pair probabilities as follows.

[0126]

(Equation 15)

Here, "#" is a special symbol representing the beginning and end of a word. Incidentally, the average word length and the character pair probability are learned in advance from a large amount of text data, and
It is stored in the average word length table 105 and the character double set probability table 111. -Generation of unknown word candidates The generation of unknown word candidates is performed to cope with the case where the input sentence includes a word that is not registered in the dictionary. In the forward search, the unknown word candidate generation unit 106 determines, at each character position of the input sentence, a partial character string included in a character matrix starting from the character position and not matching the dictionary, as an unknown word candidate, using an expression ( 15) A predetermined number is generated in the descending order of the appearance probability based on the word model. When it is desired to reduce the calculation amount, the unknown word candidate may be generated only from the partial character string of the character string composed of the first candidates for character recognition.

This method is also effective as a method of estimating a word break when a word registered in the dictionary is apparently a word not registered in the dictionary due to a character recognition error. FIG. 20 is a flowchart illustrating the operation of unknown word candidate generation in the unknown word candidate generation unit 106 according to an embodiment of the present invention. Here, for the sake of simplicity, an unknown word candidate is generated using only a character string composed of the first candidates for character recognition.

The unknown word candidate generation unit 106 is given as parameters a character string composed of the first candidate of the character matrix for the input sentence and a character position at which the unknown word candidate is to be generated. In step 301, the current character position is assigned to a variable i representing the start point of the unknown word candidate, and i + 1 is assigned to a variable j representing the end point of the unknown word candidate. An empty list is substituted into the unknown word candidate list as an initial value.

At step 302, it is checked whether j is smaller than the length of the input sentence. If not, go to step 309. If so, the following processing is performed. In step 303, it is checked whether or not the length j-1 of the unknown word candidate is longer than a predetermined maximum length of the unknown word. If so, go to step 309. Otherwise,
The following processing is performed.

In step 304, it is checked whether or not the j-th character of the input sentence is a delimiter. Delimiters are symbols that cannot be part of a word, such as punctuation marks (.) And punctuation marks (,). If so, go to step 309. Otherwise, the following processing is performed. Step 3
In 05, a partial character string from the character position i to the character position j of the input character string is used as an unknown word candidate notation.

At step 306, it is checked whether or not the notation of the current unknown word candidate matches the dictionary. If so, go to step 308. Otherwise, the following processing is performed. In step 307, first, the appearance probability of the current unknown word candidate is obtained based on the word model. Next, this unknown word candidate is added to the unknown word candidate list, and step 30
Proceed to 8.

At step 308, a variable j representing the character position at the end of the unknown word candidate is incremented by one, and at step 30
Return to 2. In step 309, the obtained unknown word candidate list is sorted in descending order of probability, and a predetermined number of unknown word candidates are selected. Word similarity (ngram distance and edit distance) Subsequently, a method for calculating the word similarity and the similar word matching unit 107
Will be described. Here, in order to simplify the description, a case will be described where similar word matching is performed on a character string consisting of only one word. Actually, it is necessary to perform similar word matching on an arbitrary partial character string of the input sentence.

The character string output by the character recognition apparatus 1 is X, and the corrected word candidate for this character string is W. Equation (10)
If the following relationship is satisfied, X should be corrected by W. P (X) P (X | X) <P (W) P (X | W) (18) The left side of Expression (18) represents the probability that an unknown word X occurs and X is correctly recognized. . The right side of the expression (18) represents a probability that the word W occurs and W is erroneously recognized as X. P
(X) is obtained from the unknown word model, and P (W) is obtained from the word segmentation model. P (X | X) and P (X |
W) is obtained from the product of the character confusion probabilities of the characters constituting X and W.

Word appearance probability P (W) and word confusion probability P
A word W having a larger product of (X | W) is more likely as a correction candidate of the character string X. Therefore, in the similar word matching unit 107 according to the embodiment of the present invention, the character string X is determined in advance in the descending order of the value of P (W) P (X | W) from all the words in the dictionary. Suggested number of correction candidates.
However, when the size of the word dictionary is tens of thousands of words or more, for any combination of the partial character string X in the input sentence and all the words W in the word dictionary, P (W) P (X | W) A simple realization method such as finding a value and selecting a word having a large product of probabilities requires an enormous amount of calculations.

Thus, in one embodiment of the present invention, first, ng
Randomized word candidates are generated based on the ram distance, then the corrected word candidates are narrowed down based on the edit distance, and finally the corrected word candidates are determined based on the product of the word appearance probability and the character confusion probability. By using a procedure consisting of steps, efficient similar word matching is realized.
The details will be described below.

The ngram distance is the number of characters ngram that two character strings commonly include as a distance measure (similarity) between the character strings. In the case of character recognition, since the number of insertion errors and deletion errors is small, for simplicity of description, the character string lengths of the input and output of character recognition are the same here. Strings of the same length each contain m characters ngram, of which c
If the numbers match, the ngram distance between the two character strings is (mc) / m.

As an example, consider the bigram distance between “Internet” and “inkjet”. When a character bigram is created from each character string, it is as follows. Internet =># i, in, nta, tar, ne, net, net, g # Inkjet =># i, in, nku, lottery, j, g, et, g # where # is a word Is a special symbol representing the beginning or end of Since four of the eight character bigrams match, the big of "Internet" and "inkjet"
The ram distance is 4/8. It can be said that the smaller the value is, the more similar the two character strings are.

However, the character ngram distance has a disadvantage that the distance between words (anagrams) whose character positions are interchanged is small (similarity is large). For example, if there is a word “net inter”, “internet” and “net inter” are common characters bi.
Since there are five gram, the bigram distance becomes 3/8.

Internet =># i, in, nta, tar, ne, net, n, net # net internet =># net, net, n, toy, in, nanta, tar, n # Since the degree of similarity of “net inter” to “internet” is greater than that of “inkjet”, the ngram distance is not suitable as a distance measure for error correction of character recognition.

However, a certain character n is located at the ngram distance.
If a list of words including a gram is created in advance from a word dictionary, there is an advantage that a word having a small gram distance can be searched very quickly. Therefore, in one embodiment of the present invention, first, ngra is applied to a character string to be corrected.
After searching for a word whose m distance is equal to or less than a threshold, the editing distance between the searched word set and the correction target character string is calculated, and words whose editing distance is equal to or less than the threshold are corrected word candidates.

The edit distance uses the minimum value of the number of insertions / deletions / replacements required to convert one character string into the other character string for two character strings as a distance measure. In the case of character recognition, since the number of insertion errors and deletion errors is small, only replacement errors are considered here for the sake of simplicity. In the case of only a substitution error, two character strings having the same length n are c
When only characters match, the edit distance between the two strings is (n
-C) / n.

For example, since “Internet” and “inkjet” have the same four characters, the editing distance is 4/7.
It is. On the other hand, since there is no matching character between "Internet" and "NetInter", the editing distance is 7/7. Generally, the smaller the editing distance is, the more similar the two character strings are. In general, for a long word, specifically, a word having a length of 3 or more, a corrected word candidate can be accurately proposed based on the edit distance. For example, if the output of character recognition is “ink-out”, if the edit distances are arranged in ascending order, the following words are proposed as correction candidates.
Here, the numerical value on the right side is the edit distance.

Inkjet 2/7 Internet 2/7 Intersoft 3/7 Intourist 3/7 However, for a short word, specifically, a word having a length of 2, if similar word matching is performed based on the edit distance, it is very difficult. , Many corrected word candidates are obtained. In Japanese, most of the words are two letters and there are more than 3000 kinds of letters, so this problem cannot be avoided.

For example, if "environment" is erroneously read as "technical environment", if a word having a short editing distance is searched using the character string "technical environment" as a key, at least the following 2
0 words are obtained. Technology Skills Technicians Skills Skills Skills Skills Environment Borders Predicament Mentality Crossing Borders Lower Borders Older Borders Frontiers Prefectural Borders Lookouts Frontiers Adversity These are all 1/2 edit distances and cannot be further narrowed down by the edit distance alone. If 10 to several tens of corrected word candidates are proposed for a character string having an arbitrary length 2 in the input sentence, the amount of calculation of morphological analysis, which is a process of examining all combinations of word candidates, is extremely large. Problem occurs.

Therefore, in one embodiment of the present invention, the product P (W) P (X | W) of the word appearance probability and the character confusion probability is calculated for the corrected word candidate searched by the edit distance, and A predetermined number is proposed as a correction word candidate in descending order of the value. When the word set having an edit distance of に 対 す る with respect to “technique” in the above example is sorted in the order of the value of P (W) P (X | W), words including characters that are graphically similar as shown below are obtained. Become a high rank.

Environmental Technology Techniques Borders Predicaments. . . According to the experiment, since the correct word is included in most cases up to the top five, the number of proposed corrected word candidates can be significantly reduced. Thus, by narrowing down the corrected word candidates while sequentially applying the three similarity measures of the ngram distance, the editing distance, and the product of the word occurrence probability and the character confusion probability, one embodiment of the present invention is achieved. In the example, the correction word candidate generation process is efficiently realized, and at the same time, the number of correction word candidate proposals is reduced, thereby reducing the calculation amount of morphological analysis.

FIG. 21 shows a similar word matching unit 107 according to an embodiment of the present invention.
5 is a flowchart for explaining the operation of FIG. Hereinafter, the operation of similar word matching will be described with reference to FIG. The similar word matching unit 107 is provided with a character string of the first candidate of the character matrix for the input sentence and a character position where a similar word is to be searched.

In step 401, the current character position is substituted for a variable i representing the start point of a character string to be compared with a similar word, and a variable j representing the end point of the character string to be compared with a similar word is represented by i + 1. Is assigned. Also, an empty list is substituted into the corrected word candidate list as an initial value. Step 402
Then, it is checked whether j is smaller than the length of the input sentence.
If not, the process ends. in that case,
The following processing is performed.

In step 403, it is checked whether or not the length j-1 of the character string to be compared is greater than a predetermined maximum length of similar words. If so, the process ends.
Otherwise, the following processing is performed. In step 404, it is checked whether the j-th character of the input sentence is a delimiter. If so, the process ends. Otherwise, the following processing is performed.

At step 405, a partial character string from the character position i to the character position j of the input character string is set as a target character string for similar word comparison. In step 406, the dictionary is searched for a word whose ngram distance from the character string to be subjected to similarity matching is smaller than a predetermined threshold value. In step 407,
The edit distance between the character string to be subjected to similarity matching and the word searched in step 406 is obtained, and only words whose value is smaller than a predetermined threshold are left.

In step 408, the value of the product of the word occurrence probability and the character confusion probability from the word notation to the target character string between the character string to be subjected to similarity matching and the word left in step 407 is calculated. Then, the specified number of words are added to the corrected word candidate list in the order of increasing values. Step 409
Then, the variable j indicating the character position at the end of the character string to be subjected to similarity matching is increased by one, and the process returns to step 402.

[0153]

EXAMPLES Finally, examples of the present invention will be described. FIG.
It is a figure showing the example of a process of the character recognition error correction system by one example of the present invention. FIG. 2 shows a character matrix output by the character recognition device 1 for a character string “organization change”, and a word candidate sequence output by the character recognition error correction device 100 according to an embodiment of the present invention on this character matrix. ,
In addition, an example of a word candidate generated by the character recognition error correction device 100 according to the embodiment of the present invention at each character position of the character matrix is shown.

In this processing example, the character matrix uses up to the fifth candidate. For example, the first to fifth candidates for the character "Cup" are "net", "rope",
They are "shrink", "fine" and "summary". A character position is a number starting from 0 given between characters. The word candidate string output by the character recognition error correction device 100 indicates up to the third candidate. Each word is represented by a set of “notation / reading / part of speech”, and word boundaries are represented by blanks. In the row before the word string, the logarithm of the probability of the word string obtained from equation (9) is shown.
Therefore, the smaller the absolute value of this numerical value, the higher the probability.

In this processing example, the correct character is not included in the input character matrix for the two characters “set” and “fur”. However, corrected word candidates “organization” and “change” are generated by similar word matching, and the correct answer “organization change” is obtained as the first candidate output. In order to show the operation of similar word matching, in this processing example, a list of word candidates generated at each character position is shown. For example, at character position 0, one-letter words matching the character candidates included in the character matrix are “net / ami / noun”, “class / tuna /
Nouns ”and the like are generated, and“ thin line / seisen / noun ”is generated as a two-character word to be matched with a combination of character candidates included in the character matrix.

In the similar word collation at the character position 0, a word whose editing distance from the character string “Amiori” of length 2 composed of the first candidate of the character matrix is １ ／, that is, one of the two words Words such as "organization / soshiki / verb" and "retina / moumak / noun" are generated from words in which one character matches, based on the graphic similarity of the character and the frequency of appearance of the word. In this example, there is no similar matching word having a length of 3 or more.

As an unknown word candidate at character position 0,
Partial character strings of the character string "Amiori Higashi" composed of the first candidate of the character matrix "Ami", "Amiori", "Amiori",
Among "Amiori Higashi", three "Amiori", "Amiori Hen", and "Amiori Higashi" that do not match the word notation contained in the word dictionary are generated as unknown word candidates. . Where N
IL indicates that the word reading is not defined, and <U
NK> indicates an unknown word.

FIG. 23 is a diagram showing the accuracy of character recognition error correction according to one embodiment of the present invention. The horizontal axis represents the character recognition accuracy before applying error correction according to one embodiment of the present invention, and the horizontal axis represents the character recognition accuracy after applying error correction according to one embodiment of the present invention. For example, (0.70, 0.88) means that when the character recognition rate of the first candidate of the input text is 70%, the character recognition rate is improved to 88% by the error correction according to the embodiment of the present invention. To be performed. If the point representing the accuracy is on a 45-degree diagonal straight line, it indicates that there is no change in the recognition rate before and after the error correction. The higher the point is on this straight line, the higher the accuracy of the error correction is Represents

FIG. 23 shows the graphic similarity of a character in order to compare with the character recognition error correction according to one embodiment of the present invention in which the correction word candidates are ranked using the graphic similarity of the character. The accuracy of the method of appropriately estimating the character confusion probability from only the character matrix candidate rank without using it is also shown. For the sake of simplicity, here, the error correction method according to an embodiment of the present invention is called a character similarity method, and the method of estimating the character confusion probability from the candidate rank is called a candidate rank method.

In the candidate ranking method, when the i-th character of the input text is c _i and the j-th character recognition candidate for this c _i is x _ij , the character confusion probability P (x _ij | c _j ) is Approximation.

[0161]

(Equation 16)

Here, p is the average correct answer rate of the first candidate for character recognition. Equation (19) is obtained by ignoring the difference in recognition rate between characters and approximating, by a geometric distribution, how reliable the nth character candidate is on average. FIG.
As shown in FIG. 7, even when the recognition rate of the input text changes (70-95%), the character recognition error correction (character similarity method) according to an embodiment of the present invention using the graphic similarity of the character. Has a significantly higher error correction accuracy than the conventional method (candidate ranking method) that does not use graphic similarity.

Further, the configuration of the character recognition error correction system is not limited to the example described in the above embodiment. Each component of the character recognition error correction system is constructed by software (program), It can be stored in a device or the like, and installed on a computer of a character recognition error correction system as needed to perform character recognition error correction. Further, the constructed program can be stored in a portable recording medium such as a floppy disk or a CD-ROM, and can be used for general purposes in a case where such a system is used.

The present invention is not limited to the above-described embodiments, but can be variously modified and applied within the scope of the claims.

[0165]

As described above, according to the present invention, there is provided a word segmentation model for providing a word pair probability and a word appearance probability,
Word matching means for searching a word in a dictionary that matches completely with a partial character string included in a character matrix, a word model composed of an average word length and a character double set probability, and a partial character string included in a character matrix Unknown word generation means for generating an unknown word candidate from a character recognition device model for giving a character confusion probability between any two characters, and a partial character included in a character matrix based on the character confusion probability and word appearance frequency Using similar word matching means for searching words in a dictionary similar to a string and morphological analysis means for presenting the word strings in descending order of the product of the joint probability of the word strings and the character confusion probability of each character, the context is determined. A Japanese character recognition error correction method and apparatus capable of accurately presenting a short word correction candidate without using it is realized.

[Brief description of the drawings]

FIG. 1 is a principle configuration diagram of the present invention.

FIG. 2 is a flowchart illustrating the principle of the present invention.

FIG. 3 is an explanatory diagram of a Japanese character recognition error correction method of the present invention.

FIG. 4 is a configuration diagram of a Japanese character recognition error correction device of the present invention.

FIG. 5 is a configuration diagram of the Japanese character recognition error correction device of the present invention.

FIG. 6 is a diagram illustrating the operation of the Japanese character recognition error correction device of the present invention.

FIG. 7 is a configuration diagram of a Japanese character recognition error correction device of the present invention.

FIG. 8 is a configuration diagram of a Japanese character recognition error correction device of the present invention.

FIG. 9 is a configuration diagram of a Japanese character recognition error correction device of the present invention.

FIG. 10 is a configuration diagram of a Japanese character recognition error correction device of the present invention.

FIG. 11 is a configuration diagram of a Japanese character recognition error correction device of the present invention.

FIG. 12 is a configuration diagram of a character recognition device model creation device of the present invention.

FIG. 13 is an operation flowchart of the character recognition device model creation device of the present invention.

FIG. 14 is a schematic block diagram of a Japanese character recognition error correction system according to one embodiment of the present invention.

FIG. 15 is a diagram illustrating an example of a character confusion frequency.

FIG. 16 is a diagram illustrating an example of a character class.

FIG. 17 is a diagram illustrating an example of a character class confusion frequency.

FIG. 18 is an operation flowchart of a forward search according to an embodiment of the present invention.

FIG. 19 is an operation flowchart of a backward search according to an embodiment of the present invention.

FIG. 20 is an operation flowchart of unknown word candidate generation according to an embodiment of the present invention.

FIG. 21 is an operation flowchart of similar word matching according to an embodiment of the present invention.

FIG. 22 is a diagram illustrating a processing example of character recognition error correction according to an embodiment of the present invention.

FIG. 23 is a diagram showing the accuracy of character recognition error correction according to one embodiment of the present invention.

[Explanation of symbols]

 DESCRIPTION OF SYMBOLS 1 Morphological analysis means 2 Unknown word candidate generation means 3 Similar word collation means 4 Word collation means 5 Word model storage means 6 Character recognition device model storage means 7 Word division model storage means 8 Character recognition device

Claims (23)

[Claims] 1. As a word candidate for correcting a character recognition error included in a Japanese sentence given as an input, an arbitrary number of words are sorted in descending order of probability based on the frequency of appearance of the word and the degree of graphic similarity of the character. Japanese character recognition error correction method characterized by being presented to a user. 2. A word matching a character string included in a Japanese sentence given as an input is generated as a word candidate, an unknown partial character string included in the sentence is generated as an unknown word candidate, and included in the sentence. A word similar to the partial character string to be generated is generated as a corrected word candidate irrespective of the context of the sentence, and a character included in the Japanese sentence is selected from a combination of the word candidate, the unknown word candidate, and the corrected word candidate. A Japanese character recognition error correction method, comprising outputting an arbitrary number of word candidate strings in the order of probability as a word string for correcting a recognition error. 3. A Japanese character recognition error correction method for correcting a character recognition error included in a Japanese sentence given as an input, comprising: a word segmentation model for giving a simultaneous probability of a word string constituting a sentence; Based on a character recognition device model that gives a character confusion probability between characters, a word in the dictionary that is similar to the character string of the word to be corrected is searched for in the dictionary, and the length of the word to be corrected is corrected. Japanese characters characterized by presenting an arbitrary number of words in the order of high probability as correction word candidates for the word to be corrected, regardless of whether context information before and after the word to be used can be used. Recognition error correction method. 4. A character matrix which is a list of character candidates arranged in order of character recognition score for each character position in a Japanese sentence is input, and the average word length and the first character in two consecutive characters are input. Is given as an unknown word candidate from among the partial character strings included in the character matrix, based on the character double set probability indicating the conditional appearance probability of the second character when given. Of the notation and the appearance probabilities in the descending order of the probabilities. Regardless of whether the Japanese sentence includes a word that is not registered in the dictionary, from among combinations of the word candidates including the unknown word candidates, Japanese character recognition error characterized by presenting an arbitrary number of word candidate strings in the descending order of the product of the joint probability of the word strings constituting the Japanese sentence and the character confusion probability of each character constituting the Japanese sentence Correction method. 5. A word in a dictionary for inputting a character matrix, which is a list of character candidates arranged in order of character recognition score for each character position of a Japanese sentence, and performing similarity matching with a character string included in the character matrix Is searched, and a character confusion matrix indicating which character is recognized by which character and how often is used, and a character confusion probability between the two characters is estimated from a character feature vector obtained by digitizing the graphic characteristics of the character, A Japanese language characterized by presenting an arbitrary number of word candidate strings in the descending order of the character confusion probability of each character constituting the Japanese sentence, irrespective of whether or not correct characters are included in the character candidates. Character recognition error correction method. 6. A word in a dictionary for inputting a character matrix which is a list of character candidates arranged in order of character recognition score for each character position of a Japanese sentence and performing similarity matching with a character string included in the character matrix And a character confusion matrix that indicates which character is recognized by which character and how often, and the character confusion probability between the two characters estimated from the character feature vector that digitizes the graphic features of the character and The above words are ranked based on the word appearance probabilities, and regardless of whether or not the correct character is included in the above character candidates, the combination of the above word candidates is selected from the combinations of word candidates including the above ranked words. Japanese characters characterized by presenting an arbitrary number of word candidate strings in the descending order of the product of the simultaneous probability of the word strings constituting the word sentence and the character confusion probability of each character constituting the Japanese sentence識誤 Ri correction method. 7. A dictionary in which a character matrix which is a list of character candidates arranged in an order of a character recognition score for each character position of a Japanese sentence is inputted, and a character string included in the character matrix is collated. When a word is searched for and the Japanese sentence includes a word not registered in the dictionary, the average word length and the conditional appearance of the second character when the first character is given in two consecutive characters Based on the character pair probabilities indicating the probabilities, a set of expressions of an arbitrary number of unknown words and appearance probabilities are determined from the partial character strings included in the character matrix in the order of high probability as unknown word candidates. If the correct character is not included in the above character candidates, search for words in the dictionary that are similar to the character string included in the character matrix, and indicate which character is recognized by which character and how often. Rank the words based on the character confusion matrix and the word appearance probability between the two characters estimated from the character confusion matrix and the character feature vector obtained by digitizing the graphic features of the characters. From the combinations of the words in the included dictionary, the unknown word candidates, and the word candidates including the ranked words, the joint probabilities of the word strings constituting the Japanese sentence and the respective characters constituting the Japanese sentence A character recognition error correction method for Japanese characters, characterized by presenting an arbitrary number of word candidate strings in the order of increasing the product of the character confusion probability and the character confusion probability. 8. A storage means for storing the frequency of appearance of a word, a calculation means for calculating a graphic similarity of a character, and an appearance frequency of a word constituting a Japanese sentence given as an input is obtained from the storage means. Acquiring the graphic similarity of the characters constituting the Japanese sentence calculated by the calculating means, and including the obtained word in the Japanese sentence based on the appearance frequency of the acquired word and the graphic similarity of the character. A Japanese character error correction device, characterized by comprising means for presenting an arbitrary number of words in order of probability as word candidates for correcting character recognition errors. 9. A means for generating a word that matches a character string included in a Japanese sentence given as an input as a word candidate, a means for generating an unknown partial character string included in the sentence as an unknown word candidate, Means for generating a word similar to the partial character string included in the sentence as a corrected word candidate irrespective of the context of the sentence; and a combination of the word candidate, the unknown word candidate, and the corrected word candidate. Means for outputting an arbitrary number of word candidate strings in order of probability as a corrected word string of a character recognition error included in a Japanese sentence. 10. A Japanese character recognition error correction device for correcting a character recognition error included in a Japanese sentence given as an input, comprising: a word segmentation model storage means for giving a joint probability of a word string constituting a sentence; A character recognition device model storing means for giving a character confusion probability between two characters, a similar word matching means for searching a word in a dictionary that performs a similar matching with the character string of the word to be corrected, a word to be corrected Means for presenting an arbitrary number of words in the order of high probability as correction word candidates for the word to be corrected, irrespective of the availability of context information before and after the word to be corrected, and A Japanese character recognition error correction device, comprising: 11. A means for inputting a character matrix which is a list of character candidates arranged in order of character recognition score for each character position of a Japanese sentence, an average word length, and a first character value in two consecutive characters. Word model storage means for storing a set of character pairs indicating the conditional appearance probability of the second character when the character is given, and an average word length and two characters stored in the word model storage means. Unknown word candidate generating means for obtaining a set of an arbitrary number of unknown word expressions and appearance probabilities as unknown word candidates from the partial character strings included in the character matrix based on the set probabilities in the order of higher probability, Regardless of whether or not the Japanese sentence includes a word that is not registered in the dictionary, from the combination of word candidates including the unknown word candidate, the simultaneous probability of the word string constituting the Japanese sentence and the Japanese sentence Means for presenting an arbitrary number of word candidate strings in the order of increasing the product of the character confusion probability of each character constituting. 12. A means for inputting a character matrix which is a list of character candidates arranged in the order of a character recognition score for each character position of a Japanese sentence, and a dictionary for performing similarity matching with a character string included in the character matrix. Character confusion matrix between two characters estimated from a character confusion matrix indicating which characters are recognized by which characters and how often Character recognition device model storage means for storing probabilities; and characters of each character constituting the Japanese sentence obtained from the character recognition device model storage means, regardless of whether or not the correct character is included in the character candidate. Means for presenting an arbitrary number of word candidate strings in the order of the greatest confusion probability. 13. A means for inputting a character matrix which is a list of character candidates arranged in an order of a character recognition score for each character position of a Japanese sentence, and a dictionary for performing similarity matching with a character string included in the character matrix. Character confusion matrix between two characters estimated from a character confusion matrix indicating which characters are recognized by which characters and how often Character recognition device model storage means for storing probabilities; word segmentation model storage means for storing word appearance probabilities; similar word matching means for ranking the words based on the character confusion probabilities and the word appearance probabilities; Regardless of whether or not the correct character is included in the character candidate, the Japanese sentence is constructed from a combination of word candidates including the ranked words. Means for presenting an arbitrary number of word candidate strings in the descending order of the product of the word string joint probability and the character confusion probability of each character constituting the Japanese sentence. apparatus. 14. A means for inputting a character matrix which is a list of character candidates arranged in order of character recognition score for each character position of a Japanese sentence, and a dictionary for collating with a partial character string included in the character matrix Word matching means for searching for a word inside, an average word length, and a character double set probability indicating a conditional appearance probability of a second character when a first character is given in two consecutive characters. A word model storing means for storing, a partial character string included in the character matrix based on the stored average word length and character double probability when the Japanese sentence includes a word not registered in the dictionary; Unknown word candidate generating means for obtaining a set of notation and appearance probabilities of an arbitrary number of unknown words in the order of probability as unknown word candidates from among the above, and when the correct character is not included in the above character candidates , A character confusion matrix that indicates which words are recognized in a dictionary, and how often a character is recognized by a character in the dictionary that matches the character string included in the character matrix, and a character that digitizes the graphic characteristics of the character A character recognition device model storing means for storing a character confusion probability between two characters estimated from a feature vector; a word division model storing means for storing a word appearance probability; and the stored character confusion probability and word appearance probability. And a similar word matching unit that ranks the words based on a word in the dictionary included in the character matrix, the unknown word candidate, and a combination of word candidates including the ranked word. An arbitrary number of word candidate strings are presented in the descending order of the product of the joint probability of the word strings constituting the Japanese sentence and the character confusion probability of each character constituting the Japanese sentence. A Japanese character recognition error correction device comprising morphological analysis means. 15. A character confusion frequency storing means for storing a character confusion frequency which gives a character and how often a character is recognized, and a character for storing a character feature vector obtained by digitizing graphic characteristics of a character. A feature vector storage unit; a character clustering unit that classifies the characters into a plurality of character classes by combining character feature vectors that are close to each other into one; and a character based on the character confusion frequency and the character class. A means for estimating the probability of character class confusion indicating how often a character of a class is recognized by a character of a character class, and estimating the total sum of probabilities that a character is recognized by a character that has not been observed so far. The unobserved event probability estimating means and the sum of the unobserved event probabilities are distributed in proportion to the character class confusion probability, and the sum becomes 1. Means for determining the character confusion probability of the unobserved event by normalizing the data, distributing the sum of the probability of the observed event in proportion to the character confusion probability, and normalizing the sum of the total probabilities to be 1. A character recognition device model creation device characterized in that a character confusion probability between any two characters is given based on the graphic similarity between the characters. 16. A process of storing the frequency of appearance of a word, a process of calculating a graphic similarity of a character, and a process of acquiring the frequency of appearance of a word constituting a Japanese sentence given as an input. As a word candidate for acquiring the graphic similarity of the characters constituting the Japanese sentence and correcting the character recognition error included in the Japanese sentence based on the appearance frequency of the acquired word and the graphic similarity of the character. And a process of presenting an arbitrary number of words in the order of the highest probability. 17. A process for generating a word matching a character string included in a Japanese sentence given as an input as a word candidate, a process for generating an unknown partial character string included in the sentence as an unknown word candidate, A process of generating a word similar to the partial character string included in the sentence as a corrected word candidate irrespective of the context of the sentence; and from the combination of the word candidate, the unknown word candidate, and the corrected word candidate, A process of outputting an arbitrary number of word candidate strings in the order of probability as a corrected word string of a character recognition error included in a Japanese sentence. . 18. A process for giving a joint probability of a word string constituting a Japanese sentence given as an input, a process for giving a character confusion probability between any two characters, and a character string of the word to be corrected And the process of searching for words in the dictionary that are similar to the word to be corrected, regardless of the length of the word to be corrected, and whether or not the context information before and after the word to be corrected can be used. A process of presenting an arbitrary number of words as corrected word candidates in the order of higher probability, wherein the Japanese character recognition error correction program is recorded. 19. A process of inputting a character matrix which is a list of character candidates arranged in an order of a character recognition score for each character position of a Japanese sentence, an average word length, and a first word in two consecutive characters. Storing a character double probability indicating a conditional appearance probability of a second character when the character is given; and storing the character based on the stored average word length and the character double probability. A process of obtaining a set of an arbitrary number of unknown words as unknown word candidates and appearance probabilities from the partial character strings included in the matrix in descending order of probability, and a word in which the Japanese sentence is not registered in the dictionary Irrespective of whether or not the Japanese sentence is included, from the combinations of the word candidates including the unknown word candidate, the coincidence probability of the word string constituting the Japanese sentence and the character confusion probability of each character constituting the Japanese sentence are determined. A process of presenting an arbitrary number of word candidate strings in the order of the product of the rate and the rate, in which the Japanese character recognition error correction program is recorded. 20. A process of inputting a character matrix which is a list of character candidates arranged in the order of the character recognition score for each character position of a Japanese sentence, and a dictionary for performing similarity matching with a character string included in the character matrix. Character confusion matrix between two characters estimated from a character confusion matrix indicating which characters are recognized by which characters and how often The process of storing the probabilities and the order in which the character confusion probabilities of the characters constituting the Japanese sentence obtained from the character recognition device model storage means are large, regardless of whether the correct character is included in the character candidates. And a process of presenting an arbitrary number of word candidate strings to a computer. 21. A process for inputting a character matrix which is a list of character candidates arranged in the order of the character recognition score for each character position of a Japanese sentence, and a dictionary for performing similarity matching with a character string included in the character matrix. Character confusion matrix between two characters estimated from a character confusion matrix indicating which characters are recognized by which characters and how often A process of storing probabilities, a process of storing word appearance probabilities, a process of ranking the words based on the character confusion probability and the word appearance probabilities, and whether a correct character is included in the character candidates. Regardless of whether or not the combination of the word candidates constituting the Japanese sentence is selected from the combination of the word candidates including the ranked words, Recording medium storing a Japanese character recognition error correction program, characterized in that it consists of a process for presenting the word candidate sequence of any number of sequentially product is large character confusion probabilities for each character constituting the sentence. 22. A process of inputting a character matrix which is a list of character candidates arranged in the order of the character recognition score for each character position of a Japanese sentence, and a dictionary for collating with a partial character string included in the character matrix Storing a process of searching for a word inside; a mean word length; and a character pair probability indicating a conditional appearance probability of a second character given a first character in two consecutive characters. Process, when the Japanese sentence includes a word that is not registered in the dictionary, an unknown word is selected from among the partial character strings included in the character matrix based on the stored average word length and the probability of a set of characters. A process of obtaining a set of notations and appearance probabilities of an arbitrary number of unknown words in the order of probability as word candidates, and a process for generating a character map when the correct character is not included in the above character candidates. A character confusion matrix that indicates which words are recognized in which dictionaries and how often a certain character is recognized, and a character feature vector that quantifies the graphic features of the characters A process of storing a character confusion probability between two characters estimated from the following, a process of storing a word appearance probability, and ranking the words based on the stored character confusion probabilities and word appearance probabilities. From the combination of the word in the dictionary included in the character matrix, the unknown word candidate, and the word candidate including the ranked word, and the joint probability of the word string constituting the Japanese sentence and the process A process of presenting an arbitrary number of word candidate strings in the order of increasing the product of the probability of confusion of each character constituting the Japanese sentence. A recording medium on which a character recognition error correction program is recorded. 23. A process for storing a character confusion frequency giving which character is recognized as to which character and how often; a process for storing a character feature vector obtained by digitizing the graphic features of the character; A process of classifying the characters into a plurality of character classes by combining character feature vectors that are close to one another, and based on the character confusion frequency and the character class, a character of a character class A process of estimating the probability of character class confusion representing how often it is recognized, a process of estimating the sum of the probabilities that a character is recognized as a character that has not been observed so far, and a process of estimating the probability of the unobserved event. The sum is distributed so as to be proportional to the character class confusion probability, and the characters of the unobserved event are normalized by normalizing the sum to be 1. The same probability is obtained, the sum of the probabilities of the observed events is distributed in proportion to the character confusion probability, and the process is normalized so that the sum of the total probabilities becomes 1. A recording medium storing a character recognition device model creation program, wherein a character confusion probability between any two characters is given based on the program.