JPH03154985A - Maximum likelihood word recognizing system - Google Patents

Abstract

PURPOSE:To increase the processing speed by summing points of characters included in a character candidate string correspondingly to each word and narrowing down words to a smaller number of words based on the frequency in this summing and discriminating coincident/nonconcident parts of characters to perform recollation of words. CONSTITUTION:Evaluation values which are given to characters coinciding with each candidate word are summed to a pertinent evaluation value table 12 by an evaluation value summing processing part 20, and the frequency in summing of the evaluation value for each word is counted and its value of the frequency in summing is stored in a summing frequency count totalizing table 15. A sort processing part 3 sorts candidate words obtained as the output of the evaluation value summing processing part 2 in the descending order of evaluation values, and the candidate word having the largest evaluation value is outputted as the collation output by a discrimination processing part 4. Thus, the number of candidate words as the object of recollation processing is reduced to increase the processing speed.

Description

[Detailed Description of the Invention] [Objective of the Invention] (Industrial Application Field) The present invention searches for candidate words from an input character string, and recognizes the most appropriate and most likely word from among the candidate words. Regarding the maximum likelihood word recognition method, in more detail, from the N candidate words obtained as an intermediate result of word matching, character extraction of unmatched character portions and character matching are reprocessed to select n candidates to be subjected to rematching. This invention relates to a maximum likelihood word recognition method using a candidate word narrowing down method for extracting words.

(Prior Art) The conventional maximum likelihood word recognition method, as shown in FIG. It has a search processing section 1 that searches for candidate words in which the character exists and extracts all the corresponding candidate words.Then, an evaluation value addition processing section 2 searches for candidate words in which the character exists in the search processing section 1, and extracts all the corresponding candidate words. The evaluation values given to the matching characters are shown in evaluation value table 1.
Add to 2. Next, the candidate words output from the evaluation value addition processing section 2 are sorted by the sorting processing section 3 in descending order of evaluation value to obtain the top N candidate words. Note that the search processing section 1, the evaluation point addition processing section 2, and the sorting processing section 3 constitute a word matching processing section 16.

For each of the N candidate words obtained by the sorting processing section 2, a judgment processing section 4 judges whether all the characters of each word are included in the human character string.

As a result of this determination, if all the characters are not included, the re-matching processing unit 10 performs character extraction and character matching again for the unmatched part of the word, and starts the word recognition process again. There is. This re-verification processing unit 1
0 is an unmatched character position search processing unit 61, a narrowing down unit 13 based on character matching rate, a character re-cutting unit 7, a character re-matching unit 8,
It is composed of a word matching processing section 5 and a determination processing section 9.

Further, as a result of the determination in the determination processing unit 4, if it is determined that all the constituent characters of the candidate word are present in the candidate character string (input character string), the corresponding candidate word is output as a matching recognition result.

In arbitrary pitch character strings, the area for one character is not necessarily divided into areas for one character, such as the area for a one-character sentence being divided into areas for two characters, or conversely, the area for two characters is combined into an area for one character. cannot be identified accurately.

Therefore, if the region is incorrectly identified, the matching state between the character candidate string narrowed down from the input character string and the candidate word cannot be determined by simply comparing the character codes character by character from the first position of each character string. It is not possible to separate the discrepancies. When a simple comparison is made, incorrect matching or non-matching portions are often determined. Therefore, the determination of matching portions and non-matching portions is performed using a method similar to the DP matching method in which the optimal combination is determined by comparing the characters before and after each character candidate. The processing amount for this mismatch determination is the number of candidate words N x length of candidate words.
lx candidate character string length is on the order of m.

The ratio of the number of characters in the matching part obtained as a result of this judgment to the number of characters in the word is called the character matching rate, and by selecting n words (<N) with a large character matching rate, the words to be rematched are narrowed down. .

(Problems to be Solved by the Invention) The conventional method described above has a problem in that the processing load is heavy and speeding up cannot be achieved because processing is performed to determine matching/mismatching portions for N candidates.

The present invention has been made in view of the above, and its purpose is to reduce the number of candidate words to be subjected to re-matching processing, reduce the amount of re-matching processing, and speed up maximum likelihood word processing. The objective is to provide a recognition method.

[Structure of the Invention] (Means for Solving the Problems) In order to achieve the above object, the maximum likelihood word recognition method of the present invention extracts characters from an input character string and extracts characters from character candidate strings obtained as a result of character recognition. For each character, search for a word that has the character, add the score of the character as the score of the word, and determine the likelihood that the obtained word has a high degree of similarity based on the higher score. A search method is used to check whether all constituent characters of a predetermined number of words with high likelihood are included in the character candidate string, and a word that includes all the characters is determined to be a word that matches the input character candidate string. However, if it is determined that they do not match, the predetermined number of words are searched for matching/unmatching parts between the character candidate string and the word, and character area identification and character recognition are performed again for the unmatched parts. is a maximum likelihood word recognition method that searches a word dictionary for a word that matches a Japanese character string written at an arbitrary pitch, and has a rematching means for determining the maximum likelihood word. an addition number storage means for storing the number of additions for each word when adding up as a score; and a pseudo character matching rate obtained by dividing the number of additions by the number of characters constituting a word, which is a number smaller than the predetermined number and which is greater than or equal to a predetermined value. The gist of the present invention is to have a re-verification means that performs re-verification only on words.

(Operation) In the maximum likelihood word recognition method of the present invention, the scores of characters included in a character candidate string are added to each word, the number of additions is stored, and the number of words is narrowed down to a small number based on the number of additions. , then identifies matching/mismatching parts of the characters and rematching the words.

(Example) Hereinafter, the present invention will be explained in detail using the drawings.

FIG. 1 is a block diagram showing the configuration of a maximum likelihood word recognition method according to an embodiment of the present invention. The maximum likelihood word recognition method shown in the figure includes a word matching processing section 31, a determination processing section 4,
It is composed of a reverification processing section 33 and a memory section 35. The word matching processing section 31 includes a search processing section 1, an evaluation value addition processing section 20, and a sorting processing section 3. The re-matching processing section 33 includes a narrowing down section 14 based on pseudo character matching rate
, a mismatch character position search processing section 6, a character re-extraction section 7, a re-character matching section 8, a word matching processing section 5, and a determination processing section 9.
, evaluation value table 12 , and addition count aggregation table 15
It is composed of.

In the maximum likelihood word recognition method shown in this example, the same components as those used in the maximum likelihood word recognition method shown in FIG. The difference between the example maximum likelihood word recognition method and the maximum likelihood word recognition method shown in FIG. A narrowing down unit 14 based on pseudo character matching rate is provided to narrow down candidate words with a high probability of being correct from N candidate words, and then search for unmatched character positions to reduce the processing load. This is what we are doing.

Further, as shown in FIG. 2, the pseudo character matching rate narrowing down section 14 is composed of a pseudo character matching rate calculating section 26, a grouping order positioning section 27, and a candidate word acceptance/rejection determining section 28. ing.

Next, the effect will be explained.

The human character string is first processed by the word dictionary 11 in the search processing unit 1.
is compared with each word in the input character string, a candidate word in which that character exists in each character position of the input character string is searched, and all corresponding candidate words are extracted. For each extracted candidate word, the evaluation value added to the matching character is added to the corresponding evaluation value table 12 in the evaluation value addition processing unit 20, and at the same time, each word to which the evaluation value is added is evaluated. Count the number of times a value is added, and use the number of additions as the number of additions table 1
Store in 5. This storage process of the number of evaluation value additions only requires the addition process of the evaluation value addition processing unit per word and one storage process, so the amount of processing is negligible.

Next, the sort processing unit 3 sorts the candidate words obtained as the output of the evaluation value addition processing unit 2 in descending order of evaluation value, and obtains the top N candidate words. Then, the determination processing unit 4 determines the matching status of the constituent characters of each of the N candidate words obtained by the sorting processing unit 3 with the input character string that is the candidate character, and determines whether all of the constituent characters are in the candidate character string. If there are candidate words in the list, the candidate word with the highest evaluation value is output as the matching result.

If there is no candidate word in which all of the constituent characters are present in the candidate character string, N candidate words are sent to the next re-verification processing unit 33.
hand over to. The rematching processing unit 33 executes the processing in the narrowing down unit 14 based on the pseudo character matching rate in order to determine target candidate words for the rematching process from among the N candidate words. Cutout part 7,
Re-character 0 The matching unit 8, word matching processing unit 5, and determination processing unit 9 are executed in sequence to output the matching results.

The narrowing down unit 14 based on the pseudo character matching rate executes the processes in the pseudo character matching rate calculating unit 26, the grouping order positioning unit 27, and the candidate word acceptance/rejection determining unit 28, as shown in FIG.

In the pseudo character matching rate calculation unit 26, the evaluation value addition processing unit 20
The pseudo character matching rate is calculated by dividing the number of evaluation value additions stored in , by the length of the word, that is, the number of characters constituting the word. Note that the reason why this is referred to as pseudo character matching rate is as follows. The evaluation value addition processing unit 20 examines the range of characters before and after the character position device and performs addition to the word containing the character, in order to correct the effect of character position shift due to a character segmentation error. Therefore, in the case of ``Tokyo J'' and ``Kyoto,'' both are added once due to the second character ``Kyo'' (1), and an accurate matching state may not be expressed. Note that in conventional match/mismatch determinations, such inaccuracies do not occur because the determination is performed taking into account the order of the characters.

In the grouping order positioning unit 27, the pseudo character matching rate is determined in order to select candidate words with a high pseudo character matching rate for each candidate word obtained in the pseudo character matching rate calculation unit 26 as target candidate words for the re-matching process. Sort candidate words in descending order. here,
When there is a candidate word with a small difference in pseudo-character match rate, it is originally impossible to simply use the rank of the pseudo-character match rate as an evaluation measure and use that threshold to distinguish whether to include it as a candidate for re-matching or to exclude it. , it is difficult to determine whether a word is correct or incorrect, and acceptance or rejection is determined based only on the slight difference in pseudo-character matching rate among similar words, which is unreasonable as a judgment method.

Therefore, in the grouping order positioning section 27, similar candidate words with similar pseudo-character matching rates are treated as a group, and a ranking is assigned to the group.

The final process of the narrowing down unit 14 based on the pseudo character matching rate in the candidate word acceptance/rejection determining unit 28 is based on a preset evaluation margin threshold (grouping order 1) among the grouped candidate words. ] 2) Extract higher ranking candidate words as target candidate words for re-verification processing.

The narrowing down unit 14 based on the above-mentioned pseudo character matching rate can narrow down the N candidate words that are likely to be correct from the N candidate words when the determination processing unit 4 determines that re-verification processing is necessary. Note that the narrowing down unit 14 based on the pseudo character matching rate performs exactly the same processing as the conventional narrowing down unit 13 based on the character matching rate, except that the values used to calculate the character matching rate are different.

[Effects of the Invention] As explained above, according to the present invention, the scores of characters included in a character candidate string are added to each word, the number of additions is stored, and the number of words is reduced based on the number of additions. Since the number of characters is narrowed down, the matching/mismatching portions of characters are identified, and the words are re-verified, the amount of processing that cannot be processed can be reduced by the amount of narrowing down, and speeding up can be achieved.

[Brief explanation of the drawing]

Fig. 1 is a block diagram showing the configuration of a maximum likelihood word recognition method according to an embodiment of the present invention, and Fig. 2 is a narrowing down based on the pseudo character matching rate used in the maximum likelihood word recognition method of Fig. 1. FIG. 3 is a block diagram showing the structure of a conventional maximum likelihood word recognition system. 1... Search processing unit 3... Sort processing unit, 4.
... Judgment processing section 6 ... Unmatched character position search processing section 11 ... Word dictionary 12 ... Evaluation value table 14
...Narrowing down section 15 based on pseudo character matching rate...Addition count aggregation table 20...Evaluation value addition processing section 31...Word matching processing section

Claims (2)

[Claims] (1) Characters are extracted from the input character string, and for each character in the character candidate string obtained as a result of character recognition, a word containing the character is searched for, and the score of the character is added as the score of the word. , a likelihood search means that determines that the obtained words have a high degree of similarity based on the words with high scores, and checks whether all constituent characters are included in the character candidate string for a predetermined number of words with high likelihood. , a word containing all the characters is determined to be a word that matches the input character candidate string, and if it is determined that there is no match based on this determination, the matching part of the character candidate string and the word for the predetermined number of words is determined as a word that matches the input character candidate string. /A word dictionary that searches for unmatched parts, performs character area identification and character recognition again for the unmatched parts, and searches for words that match Japanese character strings written at arbitrary pitches. A maximum likelihood word recognition method that searches from inside,
an addition number storage means for storing the number of additions for each word when character scores are added as word scores; 1. A maximum likelihood word recognition method comprising a re-verification means for performing re-verification only for words whose number is less than a predetermined number. (2) The re-matching means sorts the words by the pseudo-character matching rate, groups words with a pseudo-character matching rate within a predetermined step size, and divides words into one group from those with a large pseudo-character matching rate. 2. The maximum likelihood word recognition method according to claim 1, wherein a ranking is given and re-verification is performed for words whose ranking is a predetermined rank or higher and is smaller than the predetermined number.