JPS60251484A - Word recognition system - Google Patents

Abstract

PURPOSE:To find out simply and efficiently a recognized result by scanning an input pattern over the length of a word to be recognized and recognizing the word while extracting proposed characters in a recognition unit sequence out of the input pattern. CONSTITUTION:A word recognition part 6 extracts the proposed input character strings the same in the number of characters as that of a word to be recognized which are registered in a recognition dictionary 5 out of the input character string. The group of similarity calculated between the mutually corresponding character positions of respective character patterns of the proposed input character string and the dictionary patterns of respective characters of the word to be recognized is found out as word recognized results for the proposed input character string of the word to be recognized. The word recognized results consisting of the group of similarity found out in each proposed input character string out of plural ones extracted from the input character string are mutually compared. Consequently, the proposed input character string which may be the most precise as the word to be recognized is recognized as the input character string corresponding to the word to be recognized.

Description

[Detailed Description of the Invention] [Technical Field of the Invention] The present invention relates to words in a string of characters written continuously. man machine
Various pattern recognition methods have been developed that play an important role as an interface.

Among these, word recognition processing can recognize input patterns as meaningful information (words), so it can be said to be extremely superior as the above-mentioned man-machine interface.

However, conventional word recognition methods, for example,
As exemplified in No. 64936, this is done on the premise that the input cover pattern is divided into words and given. However, there is no guarantee that the input pattern used for word recognition is necessarily divided into word units. For example, when recognizing a character string indicating the address written on a postal item and sorting and processing the postal item according to the recognition result,
The above character strings used for character recognition are not always clearly categorized into "prefecture core", "city/gun name", "town/village name", "street address", etc. For this reason, when trying to recognize input cover patterns such as character strings and sounds using various word recognition methods that have been proposed in the past, it is necessary to identify the existence range (section) of the word to be recognized in the input cover patterns. It was necessary to determine this by some means. However, it has been extremely difficult to determine the range of existence of a target word from a continuously written character string or continuously uttered voice.

[Purpose of the invention]

The present invention has been made in consideration of these circumstances, and its purpose is to recognize the words to be recognized in the entered cover turn even if the words are separated and do not exist in the entered cover turn. An object of the present invention is to provide a word recognition method that can easily and effectively recognize words.

[Summary of the invention]

The present invention recognizes an input pattern consisting of a continuously written character string or continuously uttered voice, etc. in a predetermined recognition unit (
The above-mentioned recognition unit series for each character (or phoneme or syllable) is determined, and recognition unit series candidates with the same number of recognition units as the recognition unit series constituting the recognition target word are selected as the recognition unit series for the above-mentioned kavataan. Each of the recognition unit entry cover patterns of the recognition unit sequence candidates is compared with each recognition unit dictionary pattern of the II target word, and the recognition results for the recognition unit sequence candidates are calculated. are compared with each other to determine a recognition unit sequence candidate in the input bartan that corresponds to the i1m target word.

For example, the similarity between each recognition unit input pattern of the recognition unit sequence candidates extracted from the input pattern and each recognition unit dictionary pattern of the recognition target word is calculated for each Wg recognition unit, and the sum of these similarities ( The recognition results for the recognition unit sequence candidates are calculated as the sum (product) of the product) or the sum (product) of the values obtained by weighting the similarity degrees with a predetermined value, and these recognition results are compared with each other to determine the input bar number corresponding to the recognition target word. The recognition unit sequence candidates in the middle are determined.

〔Effect of the invention〕

Thus, according to the present invention, even for an input pattern in which words are not separated, such as a continuously written character string or continuously uttered speech, a recognition unit sequence with the same number of g* units as the Wlil target word is used. The candidates are extracted, and based on the degree of similarity between the recognition unit input cover pattern of each recognition unit series candidate and the recognition unit dictionary pattern of the recognition target word, the recognition result for the WIN target word of each recognition unit series candidate is input. Since the words in the Kabataan are recognized, it is possible to easily and effectively recognize the words in the Kabataan.

That is, depending on the length of the word you are trying to recognize (recognition.

Since word recognition is performed while scanning the input pattern (unit sequence) and extracting recognition unit sequence candidates from the input pattern, it is possible to easily and efficiently obtain the recognition result, etc. This produces a certain effect.

[Embodiments of the invention]

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

FIG. 1 is a schematic diagram of an embodiment of an apparatus capable of recognizing an address (destination) indicated by a string of characters consecutively written on a piece of mail or the like.

A character string written on a piece of mail or the like is photoelectrically converted and read and input via an input unit 1, and is written into an image memory 2 as an input character string image. The cutting section 3 processes the input character string image stored in the image memory 2 one by one character by character unit, which is the basic unit of recognition processing, and determines the character sequence of the above-mentioned cover pattern. The cutoff process for this character is
For example, we can use already known techniques such as detecting the breaks between characters using frame and pitch information, or detecting each character using timing marks that indicate the position of the characters. This is done by using it appropriately. Note that the character pitch can be detected by, for example, detecting a change in the projected density of a character string pattern.

Thus, the character sequence information of the input pattern obtained through the inspection process is supplied to the character recognition unit 4, and is compared with the dictionary pattern of each character of the recognition target word registered in advance in the recognition dictionary 5. Character recognition processing is performed for each input character. This character recognition process calculates the degree of similarity between each character pattern of the input character string and the dictionary pattern of each character of the recognition target word registered in the recognition dictionary 5, or a distance equivalent to the degree of similarity. This is done by calculating each.

For this character recognition, it is appropriate to use, for example, the well-known composite similarity method.

The word recognition unit 6 extracts input character string candidates having the same number of characters as the recognition target word registered in the recognition dictionary 5 from the input character string, and extracts each character pattern of the input character string candidate and the recognition target word. A set of similarities calculated between mutually corresponding character positions with the dictionary pattern of each character is taken as a word recognition result for the input character string candidate of the recognition target word. Then, the word recognition results consisting of the sets of similarities obtained every 7 meters among the plurality of input character strings extracted from the input character strings are compared with each other, and the most probable date of the recognition target words is compared with each other. new input string candidates,
It is recognized as an input character string that corresponds to the recognition target word.

That is, the word recognition unit 6 recognizes words to be recognized from the input character string in the following manner. For the sake of simplicity, an example will be described here in which the word "Shinyoku-ku" is recognized from an input character string indicating an address.

In this case, the recognition dictionary 5 is prepared in advance with dictionary patterns for each character of the recognition target word ``article'', [kawa], and ``ku''. The information required for word recognition is the degree of similarity of each character pattern of the input character string to each of the dictionary patterns ``1 item'', ``river'', 1 ward''. Now, as shown in Figure 2, "Tokyo Parts Ward Koyama 1-2-
A character string A indicating an address 3J is input, and the degree of similarity C between each of these input character patterns and the dictionary pattern of each character of the recognition target word B registered in the recognition dictionary 5 is as shown in Figure 2. Suppose that it is obtained. In addition, each of these numbers is
It is given as a degree of similarity with the maximum value being (i, o).

However, in this case, the word to be recognized, "Shinyoku-ku", is 3.
Since the character string is made up of characters, three character recognition character string candidates are extracted from the input character string subjected to the cutoff process. This extraction of character string candidates is performed by, for example, extracting three characters from the first character position of the input character string, and then shifting the extraction start character position one character at a time. Specifically, first, the similarity of the three-character string "Tokyo" with the word "Shinyo-ku" is determined. The similarity of this character string "Tokyo" as a word is the similarity calculated by character recognition between the character pattern of the first character "Higashi" and the dictionary pattern of "Shin", the similarity of the second character Similarity between the characters “Kyo” and “Kawa”, the third character “Miyako”
It is considered as a pair of similarity between ``ku'' and ``ku''. The recognition result for the word indicated by this similarity set is, for example, the sum (average) of the name similarities. Incidentally, the recognition result for this word may be obtained as the product of each of the above-mentioned degrees of similarity, or the recognition result for the word may be obtained by multiplying each of the above-mentioned degrees of similarity by a predetermined coefficient and then adding or multiplying them. good. In this way, when the product of similarities is used as the evaluation value, if even one character in the input character string has a questionable character recognition result, it is effective to reject the recognized character string candidate from the recognition processing target. There is.

Furthermore, as described above, performing recognition processing by multiplying each character recognition result (similarity) by a predetermined coefficient is effective when there is a difference in recognition performance for each character. In other words, it is possible to obtain the evaluation value for the word described above, taking into account the recognition results for simple characters and the recognition results for complex characters.

Thereafter, the three character string extracted from the input character string is moved one character at a time, and the word similarity is similarly calculated for each of these character strings. In other words, "Tokyo"
After the recognition results are obtained for the recognition character string candidates,
Recognized character string candidates are extracted in order, such as "Kyo Parts", "Parts", "1 Item Ward", and "Yu Ward Elementary School", and the recognition results for each of these recognized character string candidates are similarly displayed. In the example shown in Figure 2, the degree of similarity (sum of degrees of similarity determined by character recognition) of each recognized character string candidate to the recognition target word is (
1,4), (1,[3), (1,4), (2,6), (
1, 3). The degrees of similarity for each of these recognition target words are compared with each other, and the input character string candidate that is most likely to be the recognition target word is recognized as a word corresponding to the recognition target word. Therefore, in this example, the fourth character string candidate extracted from the input string is
The word is recognized as a word corresponding to the recognition target word registered in the recognition dictionary 5.

Note that the following algorithm may be used to determine word recognition in the input character string. For example, when the similarity of a recognized character string candidate to "Product Ward" exceeds a certain threshold, the W
g Literacy character string candidate If it is recognized as a corresponding word, or if it is known that the corresponding word "inplicity ward" always appears only once in the input character string, the similarity is the highest. A recognition character string candidate may be recognized as a corresponding word.

In this way, [[
It becomes possible to easily and effectively recognize the word ``article ward J''. It goes without saying that words such as "Tokyo" and "Oyama" can be similarly recognized.

In addition, after the word ``Shinyoku'' has been recognized, in order to recognize the following word ``Koyama'', the character string starting from the next character after the character for which the word recognition result was obtained is as described above. All you have to do is perform recognition processing.

In addition, if there are multiple words to be recognized at the same time, such as ``Hinnyou-ku'' and ``Ota-ku j'', the degree of similarity for each of these words to be recognized is measured in parallel, and the similarity is calculated for each of these words to be recognized. By comparing the degrees of similarity between the most likely recognition character string candidates, you can select a recognition target word and find the corresponding word in the input character string for the selected recognition target word. good.

As described above, according to the present invention, even if the input cover turn is not divided into words, it is possible to recognize the words to be recognized in the input cover turn, and the effect is extremely large.

Note that the present invention is not limited to the embodiments described above. For example, it can be applied to word recognition in continuously uttered speech. Figure 3 shows an example of this.
Continuous voice is input through the input voice 2a, and this input voice pattern is stored in the input voice memory 2a. This speech pattern is divided into, for example, -syllables (or phonemes) by a cutting unit 3a to obtain a syllable sequence (phoneme sequence). Thereafter, the syllable recognition unit 4a refers to the recognition dictionary 5a and calculates the similarity between each syllable pattern and each syllable dictionary pattern of the word to be recognized using a method such as a composite similarity method. Thereafter, the word recognition unit 6a may perform word recognition processing in the same manner as in the previous embodiment. Note that, compared to the previous embodiment of address recognition, the input section 1a and the extraction section 2a need to be configured specifically for speech processing, but the syllable recognition section 4a and the word recognition section 6a are The structure may be basically the same as that of the character recognition section 4 and the word recognition section 6 in the embodiment. However, it goes without saying that the configuration of the recognition dictionary 5a is different.

Furthermore, it is not necessary to calculate all the similarities in pattern units in advance. For example, in synchronization with the movement of the position of a word to be recognized, only the similarity of patterns necessary to calculate the similarity of the word may be calculated. To do this, for example, as shown in FIG. 4, the ID number of the required pattern is fed back to the character recognition unit 4, and only the similarity of this pattern is input to the word recognition unit 6. good.

In short, the present invention can be implemented with various modifications without changing the gist thereof.

[Brief explanation of the drawing]

FIG. 1 is a schematic diagram of a device according to an embodiment of the present invention, FIG. 2 is a diagram for explaining the operation of word recognition processing in the embodiment, and FIGS. 3 and 4 are diagrams showing other embodiments of the present invention. FIG. 1 is a schematic configuration diagram of an apparatus showing an example. 1.1a...input section, 2...image memory, 2a...
- Voice memory, 3, 3a... Cutting section, 4... Character recognition section, 4a... Syllable recognition section, 5.58... Recognition dictionary, 6, 6a... Word recognition section. Applicant's agent Patent attorney Takehiko Suzue

Claims (6)

[Claims] (1) A means for determining the recognition unit series of the input cover patterns by examining the input cover patterns for each predetermined recognition unit, and a recognition unit series candidate having the same number of recognition units as the recognition unit series constituting the recognition target word. Recognizing the recognition unit series candidate by comparing each recognition unit input coverturn of the recognition unit series candidate with each recognition unit dictionary pattern of the recognition target word by means of extracting each of the recognition unit/place sequences of the entered cover patterns. a means for obtaining results; and recognition of the input pattern corresponding to the recognition target word by comparing each recognition result obtained for each recognition unit series candidate extracted from the recognition unit series of the input pattern. 1. A word recognition method comprising: means for determining unit sequence candidates. (2) The recognition result for a recognition unit sequence candidate is the sum of the degrees of similarity calculated for each recognition unit between each recognition unit input cover pattern of the recognition unit sequence candidate and each recognition unit dictionary pattern of the recognition target word, or the above 2. The word recognition method according to claim 1, wherein the word recognition method is determined as a sum of values obtained by assigning a predetermined weight to the degree of similarity. (3) The recognition result for a recognition unit sequence candidate is the product of the similarity calculated for each recognition unit between each recognition unit entry cover pattern of the recognition unit sequence candidate and each recognition unit dictionary pattern of the recognition target word, or the above 2. The word recognition method according to claim 1, wherein the word recognition method is determined as a product of a value given a predetermined weight to the degree of similarity. (4) The word recognition method according to claim 1, wherein the input cover pattern consists of a string of consecutively written characters, and the recognition unit thereof consists of letters. (5) A heker pattern consists of continuously uttered sounds,
2. The word recognition method according to claim 1, wherein the recognition unit consists of a phoneme or a syllable. (6) The word recognition method according to claim 1, wherein the recognition unit sequence candidates are extracted sequentially from a predetermined recognition unit position of the recognition unit sequence of the input pattern by shifting one recognition unit at a time.