JPS5955576A - Word recognizing system - Google Patents

Abstract

PURPOSE:To recognize words with high accuracy, by obtaining the 1st resemblance for each element pattern of input data and then the 2nd resemblance between the order of array of element patterns and the standard array of a word dictionary from the standard element pattern of an element disctionary. CONSTITUTION:The 1st resemblance is obtained by calculating 31 the composite resemblance to each standard element for each element pattern forming the words to be recognized which are given from an input part 1 on the bais of the data stored in an element dictionary 2. This 1st resemblance is obtained. edited 32 and stored 33 in the form of a matrix expression. A word dictionary 4 stores the numbers (category names) of standard element patterns related to each other with each array order of the standard element patterns in terms of each standard word consisting of plural categories. Using the dictionary 4, the 2nd resemblance is calculted from the 1st resemblance expressed in a matrix and then delivered. A deciding part 6 decides the category of the word to be recognized from the degree of the 2nd resemblance. Thus the recognition factor is improved for each words although the 1st resemblance is low, which is suitably applied to recognition of characters or aural words.

Description

DETAILED DESCRIPTION OF THE INVENTION [Technical Field of the Invention] The present invention relates to a word recognition method suitable for character reading devices and voice typewriters.

[Technical background of the invention and its problems]

Kanji and speech recognition technology has been rapidly progressing recently. However, when looking at the recognition rate, it is still not sufficient compared to conventional digit recognition devices. For this reason, word recognition technology that utilizes word redundancy is important. As an example of word recognition technology, for example, Japanese patent application No. 56-13816 filed by the same applicant
No. 3 etc. have been proposed.

On the other hand, as a pattern recognition technique, Japanese Patent Application No. 56-63844 and the like have been proposed by the same inventor. This method used a similarity matrix for pattern recognition.

However, the former invention lacks flexibility in word matching calculations;
Characters or phoneme patterns (hereinafter referred to as element patterns) must be selected as candidate categories within a specified ranking, and similarities between element patterns cannot be used. Furthermore, the latter invention lacked any ingenuity to deal with the special nature of word recognition in that words have a large number of categories for each element pattern.

[Purpose of the invention]

The present invention has been made in consideration of the above circumstances, and its purpose is to provide a highly practical word recognition method that is flexible and can be implemented with a relatively small amount of hardware.

[Summary of the invention]

One of the features of the present invention is that it fully incorporates the matrix concept of the invention disclosed in Japanese Patent Application No. 56-63844.

In other words, the degree to which an element pattern at a certain position (arrangement order) of a word is similar to each element category is expressed as a first degree of similarity, and one of the positions and element categories is expressed as a row, and the other is expressed as a column. Words are expressed using a first similarity matrix. This matrix becomes large when there are many elementary categories, but not all components have luxurious values.

Therefore, as shown in Figure 1 (a), it is not necessary to memorize the entire matrix; instead, as shown in Figure 1 (b), there are cases where it is more efficient to represent the non-zero component and its row or column number as a pair. 0 Then, for word recognition, a second similarity is calculated using the first class (U degree) 0 This second similarity is calculated based on the position of the word element in the word dictionary ( For each position (array order), the sum of the squares of the coefficients of the first similarity for the power category of at least one element is calculated as the sum of the squares of the first similarity for each position and the square of the coefficient. Use the sum of the sum divided by the sum.

〔Effect of the invention〕

This makes it possible to provide flexibility in calculating the degree of matching between words, which was a drawback of the Japanese Patent Application No. 56-138163, and to utilize the similarity of each element pattern.

It is also possible to prevent the matrix from becoming large when there are many element categories.

[Embodiments of the invention]

Specific embodiments of the present invention will be described below with reference to the drawings. Figs. 2(a) and 2(b) show an embodiment of the present invention in flow diagrams and block diagrams. First, the degree of similarity (first degree of similarity) between element patterns at each position of a word is calculated using a conventionally known method. It is currently considered that the use of composite similarity is the most effective here.

The result is a similarity matrix, which may require a lot of memory to store all of it. For example, this is the case when recognizing a word consisting of multiple kanji characters, which requires several thousand words of memory per word. Generally speaking, the first degree of similarity is not calculated for all kanji categories, so most of the elements in the matrix are zero. Therefore, if some effort is made when calculating the second degree of similarity as will be described later, the matrix may be stored in an expression format as shown in FIG. 1(b).

Next, calculate the word-by-word similarity (second similarity). This calculation is, in principle, a similarity calculation between the matrices. Next, calculate the similarity between the two matrices. Calculate a second similarity.

The second degree of similarity is calculated by multiplying the sum of coefficients of the first degree of similarity for at least one element category for each position of the word element in the word dictionary to the second power of the degree of similarity for each position. It is the sum of the sum of the multiplications and the sum of the squares of the coefficients. This can be expressed as a mathematical formula.
, □, ] is the partial similarity at the γ-th position of the h-th word.

Note that R (k> is the length and arrangement rank number of words in the category), fJr is the element pattern at position γ, and α(k) is a set of element types related to (standard) words in the category.

In this case, since word lengths vary, the second similarity tends to be larger for words with larger word lengths, so it may be better to normalize the summation by the word length. be. At this time, the normalized second similarity is as follows. By using such second similarity,
Even if it happens that the similarity for a certain element category has not been calculated, it may be saved by the similarity of similar elements. This is because the degrees of similarity for similar elements are considered to have close values to begin with, so it is possible to set an appropriate value as the coefficient corresponding to the similar elements. Due to the coefficients for the similar elements, the partial similarity can have a suitable value other than O.

Furthermore, as mentioned above, word recognition involves a large number of element categories, so a large amount of memory is required to store all the matrices. According to the present invention, the second similarity can be calculated even when the entire matrix is not stored. Hereinafter, an embodiment in which the matrix is stored in the format shown in FIG. 1(b) will be explained with reference to the flowchart in FIG. In the case of inputting characters, it consists of a microphone, a sound processing circuit, etc., and in the case of character input, it consists of a Kyuden scanning section, a pre-processing circuit, etc.

Reference numeral 2 denotes an element dictionary containing data indicating the periphery of each of a plurality of types of standard element patterns. If the word to be recognized is a phonetic word, for example, the mother's turn (turn or 17) that makes up this sokugoko is used as an element pattern.0 If the word to be recognized is composed of kanji terms, for example, Individual kanji turns that make up a kanji term are used as elements/turns. Therefore, the word to be recognized is composed of a plurality of element patterns, and these iL have a predetermined arrangement order (temporal order or spatial order).

3 is a first similarity calculation unit, which includes a composite similarity NI calculation circuit 31. Editing circuit 32. and a matrix storage circuit 33. For example, the circuit 31 is disclosed in Patent No. 73989.
The composite similarity Mi arithmetic circuit shown in No. 0 is used. For each element pattern constituting the recognized word obtained from the input part l, use the data in the fake dictionary 12 to create a composite similarity with each standard element. The first similarity, which is 0, is edited by the editing circuit 32 and entered into the circuit 33 in a matrix representation as shown in FIG. 1(b).In this case, the editing circuit 32 For example, for each position other than the top n similarity values, the first
For a specific category, edit it as if the similarity is O, or store only the maximum similarity value for a specific category where the difference between the maximum similarity value (IJ degree I111 and the next largest similarity value is extremely large). You can.

4 is a word dictionary that stores, for each standard word consisting of multiple categories, the number (category name) of at least one related standard element pattern for each arrangement order (position) of the standard element patterns constituting the standard word. The information shown is memorized.

5 is a second similarity dividing unit, which calculates the second similarity for each category according to the flowchart shown in FIG.

The second degree of similarity of the recognized word to the standard word of category 'J k is first calculated by the element number tv of rank N from the word dictionary 4,
If the t key is 0, the matrix storage circuit 33
The first similarity s(A[L] of position I and element number l
x is extracted, and partial similarity U, (k)y at position 1 is determined according to equation (2). The following positions γf2, 3.・・・
, R(k) to obtain the partial similarity UγCk) at each position in the same way, and remove the total sum k according to equation (1) to obtain the second similarity 0. In figure (b), 6 is the determination part, and the second similarity! The category of the word to be recognized is determined based on the magnitude relationship of the second degree of similarity obtained by the I calculation unit 5.

Therefore, firstly, according to the present invention, word-by-word recognition can be performed with extremely high accuracy even when the recognition rate of each element pattern is low.

The present invention is not limited to the first embodiment.

For example, the IJ sock storage format shown in FIG. 4 can be used. Further, as the element pattern, a phoneme pattern, a syllable pattern, etc. can be used.

[Brief explanation of the drawing]

Figure 1 is a diagram for explaining the representation format of a matrix.
FIGS. 2(a) and 2(b) are flowcharts and block diagrams for explaining the present invention in detail, and FIG.
FIG. 4 is a flowchart for explaining the similarity analysis method. FIG. 4 is a diagram showing another example of the matrix notation format. 1... Input section, 2... Element dictionary, 3... First similarity calculation section, 4... Word dictionary, 5... Second similarity! l(1
1fil, 6...Decision section. Figure 2 (b)

Claims (3)

[Claims] (1) Means for inputting a recognized word composed of a plurality of element patterns having an arrangement order. An element dictionary containing all data showing the characteristics of each of multiple types of standard element patterns. and a first dividing means for calculating a first degree of similarity indicating the similarity between each input element pattern constituting the entire input word and each standard element pattern using the data in the element dictionary. means for storing each of the first degrees of similarity obtained by this means in association with the arrangement order of the input element pattern and the type of standard element pattern; A word dictionary that stores all information indicating the type of at least one standard element pattern associated with each of the standard words of a plurality of categories for each arrangement order of the standard element patterns constituting the standard word. a second calculation means for sequentially extracting the corresponding first similarity from the storage means according to the information in the word dictionary and calculating a second similarity between the standard word of each category and the input word; , the second calculation means squares the coefficient distribution of at least one first similarity for each arrangement rank, and calculates the sum of squares of the first similarity for each same arrangement rank and the sum of squares of the coefficients. The sum of the values divided by the product of is determined as the second similarity of each category, and the category of the recognized word is determined from the dog/J relationship of the second similarity obtained for each category. A word recognition method featuring (2) The word recognition method according to claim 1, wherein the second calculation means normalizes the total sum by the number of sequence ranks as the second similarity. (3) The word recognition method according to claim 1 or 2, wherein the first reduction means uses a composite similarity as the first similarity.