JPH08171398A - Data recognizing device, dictionary forming device and learning device - Google Patents

Abstract

PURPOSE: To provide a recognizing device capable of more improving a recognition rate compared with the conventional recognizing device. CONSTITUTION: This device is constituted of an input processing part 2, a reference pattern dictionary 5, a reference pattern comparison processing part 3 comparing the input data with a reference pattern, a data conversion processing part 12, a comparison pattern dictionary 14, a data comparison processing part 13 and an output processing part 16. The data conversion processing part 12 converts the data outputted by the reference pattern comparison processing part 3 as a function to the data for comparison processing with the comparison pattern dictionary 14. The pattern of the data outputted from the reference pattern comparison processing part 3 are stored in the comparison pattern dictionary 14. The data comparison processing part 13 inputs the data converted by the data conversion processing part 12, and compares the pattern existing in the dictionary 14 with the pattern of the input data, and thus, it extracts the pattern with a similarity in the comparison pattern dictionary to output a recognition word.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a data recognizing device, a dictionary creating device and a learning device capable of recognizing input data to be recognized.

[0002]

2. Description of the Related Art A conventional recognition apparatus, as shown in FIG.
It is composed of an input processing unit (2), a standard pattern comparison processing unit (3), a standard pattern dictionary (5), and an output processing unit (4). An input processing unit (2) inputs data and converts it into an internal data format. Then, the standard pattern comparison processing unit (3)
The input data is compared with the standard pattern dictionary (5), and which pattern is input is estimated from the similarity to the standard pattern. At this time, the similarity is output together with the recognition result.

The operation of the conventional recognition apparatus will be described below by taking the case of voice recognition as an example.

The standard pattern comparison processing unit for voices, when uttering "ta" as shown in FIG.
“Ha”, “wa”, “ka”, and “na” are output in the order of candidates. This is the case when the user uttered "ta" but mistakenly recognized it as "fa".

Next, FIG. 9 shows an output result in the case of uttering "Fa".

[0006] In response to the utterance of "Fa", the standard pattern comparison processing unit starts from the first place and selects "Fa"
Is output and the correct recognition result is output.

Further, the conventional learning method has been realized by creating and editing the standard pattern dictionary of the above recognition method.

[0008]

However, in such a conventional technique, the recognition rate is determined by the processing method for checking the similarity between the pattern of the input data and the standard pattern. Therefore, such a method has a problem that there is a limit in improving the recognition rate.

In view of the above conventional problems, it is an object of the present invention to provide a recognition device capable of further improving the recognition rate as compared with the conventional recognition device.

[0010]

According to the present invention of claim 1, a plurality of recognition candidates obtained by a pattern recognition method (A) on the basis of feature data of an object to be recognized and recognition accuracy corresponding thereto. Based on the input means for inputting the information indicating the degree of intensity as new feature data and the new feature data input, the pattern recognition method (B) obtains the recognition result of the object to be recognized. A data recognition device comprising pattern recognition processing means and output means for outputting the recognition result.

According to a second aspect of the present invention, the pattern recognition processing means is a data recognition device constituted by a neural network.

A third aspect of the present invention is a data recognition device comprising a feature data storage dictionary in which feature data used in the pattern recognition method (B) and a recognition result corresponding thereto are stored.

According to the present invention of claim 4, when the pattern recognition processing means determines that the newly selected recognition candidate is erroneous, the object to be recognized which is determined to be erroneous is also The data recognition device is provided with an adjusting means for adjusting the pattern recognition method (B) so that the pattern can be correctly recognized.

According to the present invention of claim 5, when the pattern recognition processing means comprises a neural network, the pattern recognition method (B) is adjusted so that the pattern can be correctly recognized. A data recognition device, which is to apply a predetermined teacher signal to the neural network and adjust the configuration of the neural network so that the neural network can be correctly recognized.

According to a sixth aspect of the present invention, when the feature data storage dictionary is provided, adjusting the pattern recognition method (B) so that the recognition can be performed correctly means that for the feature data storage dictionary, The data recognition device is to newly add the feature data of the object determined to be erroneous and the correct answer data corresponding thereto.

The present invention of claim 7 is a data recognition device, wherein the object to be recognized is voice information, character information, image information or pattern information.

The present invention of claim 8 is a dictionary creating device for creating the feature data storage dictionary in the data recognizing device according to claim 3, wherein pattern recognition is performed based on the feature data of the learning sample. A plurality of recognition candidates obtained by the method (A) and information indicating the degree of recognition correctness corresponding to them and correct answer data corresponding thereto are registered in a feature data storage dictionary as a creation target, and feature data storage is performed. It is a dictionary creation device provided with a feature data storage dictionary creation means for creating a dictionary.

According to a ninth aspect of the present invention, there is provided a dictionary creating apparatus in which the learning sample is voice information, character information, image information or pattern information.

According to a tenth aspect of the present invention, there is provided a learning device for creating the neural network in the data recognition device according to the second aspect, wherein the pattern recognition method (A ), A plurality of recognition candidates and information indicating the degree of correctness of recognition corresponding to them, and correct data corresponding to them, which is a teacher signal, are given to a neural network that requires learning. The learning device includes a neural network adjusting means for adjusting at least a predetermined coefficient of the neural network so that the output result from the network matches the correct answer data.

The eleventh aspect of the present invention is the learning device, wherein the learning sample is voice information, character information, image information or pattern information.

[0021]

According to the present invention of claim 1, the input means determines a plurality of recognition candidates obtained by the pattern recognition method (A) based on the feature data of the object to be recognized and the recognition correctness corresponding to them. Information indicating the degree is input as new feature data, and the pattern recognition processing means obtains the recognition result of the object to be recognized by the pattern recognition method (B) based on the input new feature data. Obtaining and outputting means outputs the recognition result.

According to an eighth aspect of the present invention, there is provided a dictionary creating device for creating the characteristic data storage dictionary in the data recognizing device according to the third aspect, wherein the characteristic data storage dictionary creating means is for learning samples. A feature data storage dictionary as a creation target of a plurality of recognition candidates obtained by the pattern recognition method (A) based on the feature data and information indicating the degree of correctness of recognition corresponding to the plurality of recognition candidates and the corresponding correct answer. To create a feature data storage dictionary.

According to a tenth aspect of the present invention, there is provided a learning device for creating the neural network in the data recognition device according to the second aspect, wherein the neural network adjusting means is based on the characteristic data of the learning sample. A plurality of recognition candidates obtained by the pattern recognition method (A) and the information indicating the degree of correctness of recognition corresponding to them, and the correct answer data which is a teacher signal corresponding to the recognition candidates are converted into a neural network that requires learning. At least a predetermined coefficient of the neural network is adjusted so that the output result from the neural network matches the correct answer data.

A more specific case will be described.

(1) "Data Recognition Device" For example, the flow of processing of the recognition method in the data recognition device will be described below with reference to FIG.

(B) The input processing section (2) converts the voice into data of a format that can be internally processed.

(B) The standard pattern comparison processing unit (3) compares the input data with the standard pattern dictionary (5) by the pattern recognition method (A) and finds a standard pattern (corresponding to a recognition candidate) close to the input data. It is extracted and a score indicating the degree of similarity with the corresponding word (information indicating the degree of correctness of recognition) is output.

(C) Data conversion processing unit (12) (input means)
Inputs the data output by the standard pattern comparison processing unit (3).

(D) The data conversion processing unit (12) converts the input data into a data format for data comparison processing.

(E) The data comparison processing unit (13) (pattern recognition processing means) uses the pattern recognition method (B) to convert the converted input data and the pattern in the comparison pattern dictionary (14) (feature data storage dictionary). The dictionary patterns (recognition results) having similarities are compared and extracted.

(F) The output processing section (15) (output means) outputs the word of the extracted dictionary pattern and the degree of similarity.

(2) "Dictionary Creation Device" The flow of processing of the dictionary creation method in the dictionary creation device will be described below with reference to FIG. 2 by taking voice recognition as an example.

(B) The input processing section (2) converts the voice into data of a format that can be internally processed.

(B) The standard pattern comparison processing unit (3)
By the pattern recognition method (A), the input data is compared with the standard pattern dictionary (5), a standard pattern close to the input data is extracted, and a score indicating the corresponding word and the similarity is output.

(C) The data conversion processing unit (12) inputs the data output by the standard pattern comparison processing unit.

(D) The data conversion processing unit (12) converts the input data into a data format for data comparison processing.

(E) The pattern extraction processing unit (17) (feature data storage dictionary creating means) extracts the features of the converted input data, and the comparison pattern dictionary (14) together with the correct answer data (19).
Register in (feature data storage dictionary). By repeating this, the comparison pattern dictionary (14) is created.

(3) Another "data recognizing device" When the pattern recognition processing means determines that the newly selected recognition candidate is erroneous, the object to be recognized which is determined to be erroneous Also, the flow of processing of the data recognition apparatus provided with the adjusting means for adjusting the pattern recognition method (B) so that the pattern can be correctly recognized will be described below with reference to FIG.

(A) The input processing unit (2) converts the voice into data in a format that can be internally processed.

(B) The standard pattern comparison processing unit (3) compares the input data with the standard pattern dictionary (5) by the pattern recognition method (A), extracts the standard pattern close to the input data, and extracts the corresponding word. The score indicating the similarity is output.

(C) The data conversion processing unit (12) inputs the data output by the standard pattern comparison processing unit (3).

(D) The data conversion processing unit (12) converts the input data into a data format for data comparison processing.

(E) The data comparison processing unit (13) compares the converted input data with the pattern in the comparison pattern dictionary (14) by the pattern recognition method (B), and extracts dictionary patterns having similarity. .

(F) The output processing section (15) outputs the word of the extracted dictionary pattern and the degree of similarity.

When the data output by (G) and (G) are incorrect, the correct answer data is input to the pattern extraction processing section (17).

(H) The pattern extraction processing section (107) (adjusting means) inputs the input data from the data comparison processing section (13).

(I) The pattern extraction processing unit (107) extracts the characteristics of the input data, and additionally registers the correct answer data (19) in the comparison pattern dictionary (14) (feature data storage dictionary). In the present specification, the pattern recognition method (A)
The pattern recognition method and the pattern recognition method (B) may be the same pattern recognition method or different pattern recognition methods.

[0048]

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

Here, prior to the description of the data recognizing device which is one embodiment of the present invention, the main difference between the device and the conventional data recognizing device will be described.

In this embodiment, the candidate words and their candidate ranks obtained by the pattern recognition method (A) based on the feature data of the voice data output by the conventional data recognition device are used as new feature data. There is to use. Here, the voice data corresponds to the object to be recognized of the present invention, the candidate words correspond to the plurality of recognition candidates of the present invention, and the candidate ranks thereof are information indicating the degree of correctness of the recognition of the present invention. Correspond.

That is, as described above with reference to FIGS. 8 and 9, in the conventional data recognition device, as a result,
Both "ta" and "fa" are in the standard pattern comparison processing unit of voice,
We recognize as "fa".

However, paying attention to the output pattern of the second or lower rank at this time, although the recognition of "Ta" when uttering "Ta" is the second rank, the recognition of "Ta" when "Fa" is pronounced Recognition is in 5th place, and it can be seen that there are large differences in the output patterns. That is, depending on the uttered words, there is a tendency that there is a large difference in the patterns of the output words after the second place, even if the first recognized words are the same.

Therefore, by paying attention to the difference in the output patterns, it is possible to estimate the correct word, and as a result, the recognition rate can be expected to improve.

It is possible to improve the voice recognition rate by adding the process of estimating the correct word from this output pattern to the output result from the voice standard pattern comparison processing unit (7) of the conventional data recognition device. Becomes This is one of the most important points of this embodiment.

FIG. 1 shows a data recognition device having a function of estimating a correct answer word.

That is, as shown in the figure, the data recognition device includes an input processing section (2) and a standard pattern dictionary (5).
And a standard pattern comparison processing unit (3) for comparing input data with a standard pattern, such as the voice standard pattern comparison processing unit described above, a data conversion processing unit (12), a comparison pattern dictionary (14), and a data comparison processing unit. (13) and the output processing section (16).

The data conversion processing unit (12) has a function of converting the data output from the standard pattern comparison processing unit into data for comparison processing with the comparison pattern dictionary. The comparison pattern dictionary (14) stores patterns of data output from the standard pattern comparison processing unit. The data comparison processing unit (13) inputs the data converted by the data conversion processing unit (12) and compares the pattern in the comparison pattern dictionary with the pattern of the input data. As a result of the comparison, the patterns in the comparison pattern dictionary that have similarities are extracted, and the recognition word is output.

Therefore, the standard pattern comparison processing unit (3)
Even if it is erroneously recognized, since the pattern of the erroneously recognized data is stored in the comparison pattern dictionary (14) as a feature,
The correct recognition word can be output.

The data comparison processing unit (1
As a technique used in 3), a neural network technique for performing pattern recognition or the like is considered, and the data comparison processing unit (13) may be configured by a neural network. However, in this case, the above-mentioned comparison pattern dictionary (14) is not necessary.

On the other hand, the features of the dictionary creating apparatus of another embodiment of this embodiment will be described.

That is, the dictionary creating apparatus of the present embodiment is a comparison pattern dictionary (14) attached to the above-mentioned data recognition apparatus.
Is created, and the learning data pattern is newly registered in the dictionary. Here, the learning data corresponds to the learning sample of the present invention.

The dictionary creating apparatus of this embodiment, as shown in FIG. 2, includes an input processing section (2), a standard pattern dictionary (5),
Standard pattern comparison processing unit (3) and data conversion processing unit (1
2) and a pattern extraction processing unit (17).

The pattern extraction processing section (17) selects candidate data from a plurality of recognition data output from the standard pattern comparison processing section (3) by the pattern recognition method (A) based on the input learning data. Further, the features of the data pattern are extracted according to their ranks and the like, and the features of the extracted data pattern are registered in the comparison pattern dictionary (14) as patterns corresponding to the input correct answer data. By repeating this, the comparison pattern dictionary (14) is created.

By the way, as an embodiment according to the present invention, when the data comparison processing unit (13) is constituted by the neural network described in the above embodiment, the neural network is created. The learning device will be described.

That is, the learning device includes a plurality of recognition candidates obtained by the pattern recognition method (A) based on the feature data of the learning sample and information indicating the degree of recognition accuracy corresponding to the plurality of recognition candidates. Correct answer data that is a corresponding teacher signal is given to a neural network that requires learning, and constants of each element of the neural network are set so that the output result from the neural network matches the correct answer data. Neural network adjusting means (203) for adjusting (FIG. 5)
See). The learning method is essentially the same as the conventional method.

Further, a data recognition apparatus as another embodiment of the present invention is characterized in that the data recognition apparatus described in the above embodiment is provided with an adjusting means and has a learning function.

That is, the data recognition device of this embodiment is
As shown in FIG. 3, an input processing unit (2), a standard pattern comparison processing unit (3), a standard pattern dictionary (5), a data conversion processing unit (12), a data comparison processing unit (13), a pattern extraction processing unit. (107), comparison pattern dictionary (14), output processing unit (1
It consists of 6). Here, the pattern extraction processing unit (107)
Corresponds to the adjusting means of the present invention.

The learning function of this embodiment is to input correct answer data and perform learning by a learning method when the result of the data comparison processing section (13) is erroneous recognition.

Next, the above embodiment will be described more specifically.

With respect to the data recognition apparatus of the above embodiment, an embodiment in which a neural network is used for the data comparison processing unit (13) in the recognition method of voice recognition will be described.

The data recognition device is as shown in FIG. It is composed of an input processing unit (2), a voice standard pattern comparison processing unit (7), a standard pattern dictionary (5), a data conversion process (12), and a neural network (23).

The processing procedure of this embodiment is shown below.

(A) Input processing section (2) The input processing section inputs the vocal sound "a" and converts it into a data format for internal processing.

(B) Standard pattern comparison processing section (7) The standard pattern comparison processing section compares the input data converted in (a) with the standard pattern dictionary by the pattern recognition method (A). As a result of comparison, from the first place in order, "Fa: 908""C:877""C:800""A:799""T:792"
A score indicating the degree of similarity with the recognized word is output, such as “wa: 583”. In this example, when the voice standard pattern comparison processing unit erroneously recognizes, “A” is uttered,
It shows the case of erroneously recognizing.

(C) Data conversion processing section (12) The data conversion processing section (12) uses a neural network as shown in FIG. 9 in order to cause the neural network to recognize the data output in (b). Convert according to the input format of the network. In this case, the maximum score indicating the degree of similarity is set to 255. Converted data (2
2) is a pattern for the neural network to estimate correct data. This pattern corresponds to the new feature data of the present invention.

Here, the case of "A" will be described as an example of the conversion processing method in the data conversion processing section (12).

That is, the maximum value of the score indicating the similarity in the data pattern of the input to the data conversion processing unit (12) is 908 points of "Fa". In addition, the maximum value of the score indicating the similarity of the data pattern at the output stage of the data conversion processing unit (12) is set to 255 as described above.

Therefore, in order to maintain the relative ratio of the similarity of “A” of 799 points to the similarity of other candidate data, “255 × 799 ÷ 908” is calculated,
The degree of similarity of "A" is converted into 224 points. Similarities are converted for other candidate data by the same calculation method.

(D) Neural network recognition processing (23) The neural network inputs the data pattern converted in (c), weights the number of points for each word of the input data, and then performs linear addition. Then
Further, a non-linear function is applied to the addition value plus a constant, and the recognition word is output. In this case, since the input data pattern is similar to the data pattern of "A",
"A" is output.

As described above, accurate data can be output by the data comparison processing even when the standard pattern comparison processing unit makes an erroneous recognition in the procedure (b). Therefore, the recognition rate can be improved.

Next, an embodiment of the learning device according to the present invention will be described, and an example of the learning method for the neural network used in the pattern extraction processing section of the data recognition device will be shown.

As shown in FIG. 5, the learning device of this embodiment has an input processing unit (2), a voice standard pattern comparison processing unit (7), a standard pattern dictionary (5), and a data conversion process (1
2), neural networks that require learning (2
3) and a neural network adjusting means (203) for adjusting constants of each element of the neural network (23).

The learning process of this embodiment will be described below.

(A) Input processing unit (2) The input processing unit (2) inputs the vocal sound "A" as a learning sample and converts it into a data format for internal processing.

(B) Standard pattern comparison processing section (7) The standard pattern comparison processing section (7) compares the input data converted in (a) with the standard pattern dictionary (5). As a result of the comparison, from the first place, "Fa: 889" and "F: 76
7 ”,“ C: 756 ”,“ A: 729 ”,“ T: 657 ”,“ Wa: 583 ”.
Recognized word candidates and the score indicating each similarity are output data (2
6) is output as. In this example, the voice standard pattern comparison processing unit (7) erroneously recognizes, and utters "a", but erroneously recognizes "fa".

(C) Data conversion processing unit (12) The data conversion processing unit (12) uses a neural network adjusting means for causing the neural network (23) to learn the data (26) output in (b). Convert according to the input format to (203). In this case, the maximum score for similarity is set to 255. This converted data (2
7) is the data pattern learned by the neural network (23).

(D) Learning by Neural Network Adjusting Means (203) The neural network adjusting means (203) applies to the neural network (23) that needs learning,
The data pattern converted by (C) and the correct answer data "A" as a teacher signal corresponding to the data pattern are input. Then, each constant of the elements forming the neural network (23) is adjusted so that the correct data “A” is output from the input data pattern.

The neural network (23) can be completed by repeating the above procedure using a plurality of learning samples.

Next, with respect to the data recognition apparatus of another embodiment, an example in which a neural network is used for the data comparison processing unit and the pattern extraction processing unit will be specifically described.

When the pattern recognition processing means determines that the newly selected recognition candidate is erroneous, the apparatus of this embodiment also correctly recognizes the object to be recognized, which is determined to be erroneous. The feature is that an adjusting means for adjusting the pattern recognition method (B) is provided so as to be able to do so.

As shown in FIG. 6, the data recognition apparatus of the present embodiment has an input processing section (2), a voice standard pattern comparison processing section (7), a standard pattern dictionary (5), a data conversion processing (1
2), neural network (23), adjusting means (200
It consists of 3).

The processing procedure centering on the adjusting means of this embodiment will be described below.

(B) Input processing section (2) The input processing section inputs the vocal sound "a" and converts it into a data format for internal processing.

(B) Standard pattern comparison processing unit (7) The standard pattern comparison processing unit compares the input data converted in (a) with the standard pattern dictionary. As a result of the comparison, from the first place, "Fa: 908", "C: 877", "C: 800"
A score indicating the recognition word and the similarity is output, such as “A: 799”, “T: 792”, and “Wa: 583”. In this example, the voice standard pattern comparison processing unit erroneously recognizes "a", but erroneously recognizes "fa".

(C) Data conversion processing section (12) The data conversion processing section, as shown in the figure, causes the neural network to recognize the data output in (b) in order to cause the neural network to perform recognition processing. Convert according to. In this case, the maximum score for similarity is set to 255. The conversion data (22) shown in FIG.
The neural network becomes a pattern for estimating correct data.

(D) Recognition processing of the neural network (23) The neural network (23) inputs the data pattern (27) converted by (c) and weights the count peculiar to each word of the input data. And then
Linear addition, and a constant added to the addition value,
A non-linear function is applied and the recognition word is output. in this case,
Since the input data pattern is similar to the "fa" data pattern, the "fa" score is output the highest. However, this data is a false positive.

(E) Neural network learning processing by adjusting means (2003) (23) The adjusting means (2003) is a neural network (23).
, Input the data pattern converted by (d) and the correct answer data "a", and output the correct answer data "a" from the input data pattern. Adjust. In this way, the adjusting means (2003) trains the neural network (23).

As shown in the above procedure, the recognition rate is further improved by adding the learning function to the data recognition device.

[0099]

As is clear from the above description, the present invention has an advantage that the recognition rate can be further improved as compared with the prior art.

[Brief description of drawings]

FIG. 1 is a configuration diagram of a data recognition device according to an embodiment of the present invention.

FIG. 2 is a configuration diagram of a dictionary creating apparatus according to an embodiment of the present invention.

FIG. 3 is a configuration diagram of a data recognition device according to another embodiment of the present invention.

FIG. 4 is a block diagram of a data recognition device which is an embodiment of the present invention and uses voice as input data, and shows a data flow when "A" is uttered.

FIG. 5 is a configuration diagram of a learning device according to an embodiment of the present invention,
The data flow when "A" is uttered is shown.

FIG. 6 is a configuration diagram of a data recognition device according to another embodiment of the present invention, showing a data flow when "A" is uttered.

FIG. 7 is a block diagram of a conventional data recognition device.

FIG. 8 is a block diagram of a conventional data recognition device that uses voice as input data, and shows the flow of data when "ta" is uttered.

FIG. 9 is a block diagram of a conventional data recognition device that uses voice as input data, and shows the flow of data when "Fa" is uttered.

[Explanation of symbols]

1 input data 2 input processing unit 3 standard pattern comparison processing unit 4, 16 output processing unit 5 standard pattern dictionary 6 output data by standard pattern comparison processing 7 standard pattern comparison processing unit for voice 8 vocalized word "ta" 9 "ta" Standard pattern comparison processing result 10 Spoken word “Fa” 11 Standard pattern comparison processing result of “Fa” 12 Data conversion processing unit 13 Data comparison processing unit 14 Comparison pattern dictionary 15, 22, 27 Output data of data conversion processing 17,107 patterns Extraction processing unit 18 Output data 19 Correct answer data 20 Spoken word “A” 1 21, 26 Output data of standard pattern comparison processing unit of speech 23 Neural network 24 Similarity score 25 Spoken word “A” 2 28 Correct answer data “A” 203 Neural Network adjustment means 2003 Adjustment means

Claims (11)

[Claims] 1. A plurality of recognition candidates obtained by a pattern recognition method (A) based on feature data of an object to be recognized and information indicating the degree of correctness corresponding to them are newly added as new feature data. Input means for inputting, as pattern recognition processing means for obtaining a recognition result for the object to be recognized by the pattern recognition method (B) based on the new feature data input, and outputting the recognition result. A data recognition apparatus comprising: an output unit. 2. The data recognition apparatus according to claim 1, wherein the pattern recognition processing means is composed of a neural network. 3. The data recognition device according to claim 1, further comprising a feature data storage dictionary in which feature data used in the pattern recognition method (B) and a recognition result corresponding to the feature data are stored. . 4. When the pattern recognition processing unit determines that the newly selected recognition candidate is erroneous, the pattern can be correctly recognized even for the object to be recognized which has been erroneously determined. 4. The data recognition device according to claim 1, further comprising adjusting means for adjusting the recognition method (B). 5. The pattern recognition method (B), wherein when the pattern recognition processing means is composed of a neural network, the pattern recognition processing means can correctly recognize the pattern.
5. The adjusting means means adjusting the configuration of the neural network so that the teacher can give a predetermined teacher signal to the neural network and correctly recognize the neural network.
The described data recognition device. 6. When the feature data storage dictionary is provided, the pattern recognition method (B) so that the pattern can be correctly recognized.
5. The adjustment of means is to newly add the target feature data determined to be the error and the correct answer data corresponding thereto to the feature data storage dictionary. The described data recognition device. 7. The data recognition device according to claim 1, wherein the object to be recognized is voice information, character information, image information or pattern information. 8. A dictionary creation device for creating the feature data storage dictionary in the data recognition device according to claim 3, wherein the dictionary is obtained by a pattern recognition method (A) based on the feature data of the learning sample. A plurality of recognition candidates that are stored, information indicating the degree of correctness of recognition corresponding to them, and correct answer data corresponding thereto are registered in a feature data storage dictionary as a creation target, and a feature data storage dictionary is created. A dictionary creation device comprising a data storage dictionary creation means. 9. The dictionary creating apparatus according to claim 8, wherein the learning sample is voice information, character information, image information or pattern information. 10. A learning device for creating the neural network in the data recognition device according to claim 2, comprising a plurality of patterns obtained by a pattern recognition method (A) on the basis of feature data of learning samples. Information indicating the degree of correctness of recognition corresponding to the recognition candidates and the correctness data that is a teacher signal corresponding thereto is given to a neural network that requires learning, and the output result from the neural network is A learning device comprising a neural network adjusting means for adjusting at least a predetermined coefficient of the neural network so as to match the correct answer data. 11. The learning device according to claim 10, wherein the learning sample is voice information, character information, image information, or pattern information.