JPS63268082A - Pattern recognizing device - Google Patents

Abstract

PURPOSE:To automatically correct a pattern recognizing result at a high speed by defining the final result when the coincidence of recognizing results is obtained between a KANJI (Chinese characters) string and its Japanese phonetic syllabary. CONSTITUTION:The KANJI strings recognized by a character recognizer 1 are stored in a KANJI candidate table 2 together with the Japanese phonetic syllabary stored in a Japanese syllabary candidate table 3 respectively. A control part 4 reads successively the candidate words out of the table 2 and refers to a read dictionary 5 storing the reading of each KANJI for candidate words to obtain the reading of all candidate words. Then the part 4 selects the reading coincident with a character number signal 8 for the Japanese phonetic syllable received from the table 3 to send it to a word collating part 6 together with the likelihood degrees of KANJI candidate words. The part 6 calculates the collation degree between the entire reading of the candidate word of the transferred KANJI and the Japanese phonetic syllabary and sends the collation degree to a sorting circuit 7. The circuit 7 sorts the reading of all KANJI candidate words in the order of larger collation degrees and outputs the KANJI candidate word having the largest collation degree and its reading as the recognizing results of the KANJI string and the Japanese phonetic syllabary.

Description

[Detailed Description of the Invention] [Objective of the Invention] (Industrial Application Field) The present invention automatically corrects recognition results obtained by character recognition using an optical character reading device, an online character recognition device, etc. in a simple and highly accurate manner. This invention relates to a recognition processing device that can perform

(°Prior Art) In recent years, with the development of information processing technology, various devices have been developed that recognize handwritten characters and input data.

These devices use pattern recognition technology to recognize individual characters, but currently it is not possible to obtain high recognition accuracy when the handwritten characters are significantly deformed or the quality of the characters is poor. be. Therefore, when inputting information that has meaning as a word, such as a first and last name, address, or product name, instead of outputting a single recognition result for each character, several candidate characters are output. address,
It is common to improve accuracy by outputting candidate character combinations that form meaningful words by comparing them with a word knowledge base such as product names. Such a method is disclosed, for example, in Japanese Patent Laid-Open No. 197974/1983. According to this method, when inputting a limited number of product names or a limited range of addresses, it is possible to obtain a reading rate close to 100% as long as the recognition performance of each character is secured to a certain degree. .

However, in cases where there are 100,000 types of names, such as common surnames and given names, the following two problems arise. First, as the number of words registered in the knowledge base increases, the number of cases of false matching increases. False verification means 1 character or 2
For example, in the case of given names, in order to distinguish between Emiko and Etsuko, it is necessary to correctly recognize Mi and Tsu. It is not possible to utilize the redundancy of words, and the performance of character recognition becomes the performance of word recognition. The second problem is that the number of types of surnames and given names is not disclosed, and it is nearly impossible to construct a knowledge base that includes all surnames and given names. Furthermore, according to "Japanese Surnames" (published by Nihon Keizai Shimbun, 1978), although there are approximately 100 surnames in kana, accounting for 37% of the total, there are 10,000 surnames, accounting for 96.7% of the total. To exceed 99.9%, tens of thousands of surnames would be required, and a vast knowledge base would be required to select candidate characters with high precision. In the case of a surname written in kanji, many more types are required than in kana, and in the case of a given name, there are even more types, which increases the difficulty.

(Problems to be Solved by the Invention) The present invention has been made in view of the above technical situation, and aims to automatically correct the recognition results of a pattern recognition device efficiently and accurately with a small storage capacity. The purpose of this invention is to provide a pattern recognition device that can perform

[Structure of the invention]

(Means for solving the problem) In conventional methods for automatically correcting recognition results, processing is performed only on target character strings. Therefore,
Highly accurate automatic correction cannot be expected when there are few differences between character strings. However, for example, in documents that are read by a character-based character reader, especially in various return letters and application forms, a border is usually added to the name and address. The gist of the present invention is to improve processing efficiency and accuracy by simultaneously processing a plurality of pattern sequences that can be correlated in this manner and utilizing the existing redundancy. Specifically, the pronunciation of each kanji is used to match a kanji character string with its furigana, and the pronunciations of each candidate character obtained as a result of recognition of the kanji string are combined to form a combination of furigana recognition candidate characters. The one that matches will be the final result.

The present means will be explained using an example. FIG. 2 is an explanatory diagram thereof. In Figure 2, numeral 21 is an entry form, in which the surname in kanji and its furigana are entered.

22 is the result of recognition by the character recognition device, in which candidate characters up to second place for kanji characters and candidate characters up to third place for furigana characters are restored. 23
is a table of readings of individual kanji, and only kanji that have been suggested as candidates are displayed. 24 lists possible readings for all combinations of kanji candidates. However, only the readings with the same number of characters as the Fushigana are listed. (Dakuten counts as one digit.) Of the 22 possible combinations of kana, only ``usui'' is included in 24, so the kanji would have been written as ``usui''. In general, it is possible that multiple pronunciations are included, but in this case, it is ranked by matching degree as a word based on the candidate ranking and the similarity and distance scale when recognizing individual characters. The final result can be

Another example is shown in Figure 3: the recognition of addresses written in Una postal codes and Kanji characters. Among the addresses corresponding to the candidates written as a result of recognizing the postal code, those included in the combination of Kanji candidates are taken as the postal code recognition result.

(Operation) As described above, according to the present invention, by simply storing only the reading information of individual kanji as a knowledge base, the recognition results of both kanji strings and furigana that indicate the same content can be automatically ( 6. Can do the correct thing.

(Example) Hereinafter, an example of the present invention will be described with reference to the drawings.

FIG. 1 is a block diagram showing a schematic configuration of the present invention. In FIG. 1, a character recognition device 1 is a pattern recognition device such as an optical character recognition device or an online character recognition device, and the recognition results of kanji strings and furigana are output to a kanji character candidate table 2 and a furigana candidate character table 3, respectively. be remembered. Here, the table is a list of candidate characters arranged in a table format as shown in FIG. 22, and each candidate is encoded and stored in a memory. The control unit 4 sequentially reads candidate words from the kanji candidate table 2. Candidate words can be read out in any order as long as one candidate character is selected from each character position to form a word, and all possible combinations are read out without exception. Also, the likelihood of the candidate word is calculated by adding the scores determined by the candidate ranking over all character positions. Furthermore, the control section 4
generates the reading of the entire candidate word by referring to the reading dictionary 6 which stores the reading of each kanji for the read candidate word. At this time, since each kanji may have multiple readings, the entire candidate word generated here may have multiple readings. Signal 8 is the number of letters in Fushigana υ, and the number of characters among the readings of all candidate characters is signal 8.
Only the words that match are selected and sequentially transferred to the word matching unit 6 along with the likelihood of the Kanji candidate word. The word matching unit 6 calculates the degree of matching between the pronunciation of the entire candidate word of the kanji transferred from the control unit and the Japanese kana. The degree of matching is JP-A-59-
197974, which is the sum of scores determined by candidate rankings for all character positions, can be used.

Furthermore, by calculating the sum or product of the obtained score and the likelihood of the Kanji candidate word transferred from the control unit, this is used as the degree of matching of the pronunciation of the Kanji candidate word. The kanji candidate words, their pronunciations, and matching degrees are transferred to the sorting circuit 7. The sorting circuit sorts the pronunciations of all kanji candidate words into terms with the highest matching degree, and outputs the kanji candidate word with the highest matching degree and its pronunciation as the recognition result of the kanji string and Japanese alphabet.

If there are 4000 types of kanji, the capacity of a reading dictionary is 4O.
The memory capacity is approximately 1,000 K bytes, which is significantly smaller than the several 100 bytes required to store 100,000 surnames. Additionally, the processing time is significantly reduced compared to comparing 100,000 surnames one by one.

〔Effect of the invention〕

As described above, according to the present invention, it is possible to automatically correct pattern recognition results at high speed with a small memory capacity, and moreover, the target word types include all combinations of kanji readings, and the number of kanji types is 1000. If it is more than 1 character, it can be said that it is practically infinite.

Moreover, the present invention can be more effective by combining with other methods. For example, words that appear relatively frequently can be compared using a word matching method such as that shown in Japanese Patent Application Laid-Open No. 197974/1985, and the present invention can be applied only to words that cannot be matched. . In the case of a surname of 1000, it accounts for 80% of the total, and by matching 1000 words, 80 words can be matched, and by applying the present invention only to the remaining 20%, it is possible to further speed up the processing. It is possible to improve recognition accuracy.

[Brief explanation of drawings]

FIG. 1 is a block diagram showing a schematic configuration of one embodiment of the present invention, and FIGS. 2 and 3 are diagrams for explaining the operation of one embodiment of the present invention. 1...Character recognition device 2...Kanji candidate character table 3...Fushigana candidate character table 5...Yomi dictionary 6...Word collation unit 7...Sort circuit agent Patent attorney Chika Nori Ken Yudo Mitsuru Matsuyama Figure 1 Figure 2 Roro Roro Kasuke F Figure 3

Claims (1)

[Claims] pattern identification means for identifying individual patterns in a device that recognizes both first and second pattern sequences indicating the same content; and patterns that may exist in the first pattern sequence;
means for storing a correspondence relationship with patterns that may exist in the second pattern sequence; and means for storing a correspondence relationship between the third pattern sequence and the second pattern sequence obtained by converting the identification result of the first pattern sequence using the correspondence relationship; It has a means for comparing the identification result with the identification result, and one or both of the identification result of the second pattern string with the highest matching degree and the identification result of the first pattern string corresponding to the third pattern string is used as the final result. A pattern recognition device characterized by: