JPH05324718A - Proper noun processor - Google Patents

Abstract

PURPOSE:To convert a proper noun of a language which uses the alphabet into an adequate proper noun of KANJI (Chinese character) without generating a syllable-KANJI translation dictionary, an exceptional pronunciation rule, etc. CONSTITUTION:When the proper noun in alphabet description is inputted from an input part 21, a proper noun-Chinese word conversion part 22 converts the input proper noun into a corresponding natural proper noun in Chinese through a neural network 22a. The converted proper noun is outputted from an output part 23. Thus, the proper noun of alphabet characters can be converted into the natural Chinese proper noun without generating the syllable-KANJI translation dictionary, exceptional pronunciation rule, etc., by using the neural network 22a. Further, this processor is provided with a means which decides that an inputted proper noun is a female name and its output is supplied to the neural network 22a to enable conversion into a proper noun suitable to the female name.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a proper noun processing apparatus in a text processing system for a language that uses input alphabetic characters, and more particularly to a proper noun in a character string using a neural network. The present invention relates to a proper noun processing device for converting into words.

[0002]

2. Description of the Related Art In a text processing system for a language that uses alphabetic characters, if there is a proper noun in the input sentence that is not in the dictionary, the proper noun is conventionally output as it is, or
It was converted to Chinese kanji by using a syllable-kanji bilingual dictionary, and when there was ambiguity, it was judged using exception pronunciation rules.

[0003]

By the way, in the above-mentioned conventional technique, in the method of (1), it is necessary for a human to convert an alphabet character into Chinese, and the alphabet character is directly converted into a proper noun of Chinese. If you can't get a natural Chinese sentence. Also,
This method requires humans to create bilingual dictionaries and complicated exception pronunciation rules.

The present invention has been made in order to improve the above-mentioned drawbacks of the prior art. In converting a proper noun of a language using alphabetic characters into a proper proper noun of Chinese kanji, a syllable- Provide a proper noun processing device that does not need to create a Kanji bilingual dictionary, exceptional pronunciation rules, etc., and can convert a proper noun in a processing system into a natural Chinese proper noun. The purpose is to

[0005]

FIG. 1 shows the principle of the present invention. The present invention is configured as shown in FIG. 1 in order to solve the above problems, and the invention of claim 1 of the present invention is
In a proper noun processing device of a text processing system for converting a language using alphabetic characters into Chinese, an input unit 21 for inputting proper nouns expressed by alphabetical characters and a neural net 22a are provided, and are given from the input unit 21. A proper noun-Chinese conversion unit 22 for converting a proper noun in alphabetical character notation into a proper Chinese-Kanji character noun, and a proper noun converted to Chinese-Kanji notation in response to outputs from the proper noun-Chinese conversion unit 22 An output unit 23 for outputting a noun is provided to convert a proper noun of a language using alphabetic characters into a corresponding proper noun of Chinese kanji.

According to a second aspect of the present invention, in the first aspect of the invention, the neural network 22 uses the proper noun of Chinese kanji notation corresponding to the alphabet character as a teacher signal.
means 24 for giving a, the teacher signal and the neural network 2
A means 25 for comparing the outputs of 2a is provided, and the proper noun and the teacher signal in the Chinese kanji notation corresponding to the alphabetic character are output to the neural network 2 of the proper noun-Chinese conversion unit 22.
2a, the neural network 22a is trained, and then proper nouns of a language using alphabetic characters are given.
It is configured to be converted into a proper noun in the corresponding Chinese kanji notation.

In the invention of claim 3 of the present invention, in the invention of claim 1 or 2, it is determined whether or not the input proper noun is a female name, and the input proper noun is a female name. Discriminating means 22b for generating an output in some cases, and discriminating means 2
A neural network 22a in which a unit that reacts to the output of 2b is provided in the input layer, and when the input proper noun is a female name, it is configured to convert it into a proper noun in Chinese kanji notation suitable for a woman Is.

[0008]

In the invention of claim 1, when a proper noun in alphabetical character notation is input from the input unit 21, the proper noun-Chinese conversion unit 22 corresponds to the proper noun in alphabetical character notation input by the neural network 22a. Convert to proper Chinese nouns. The proper noun converted by the proper noun-Chinese conversion unit 22 is output from the output unit 23.

Since the proper nouns are converted using the neural network 22a, proper nouns expressed in alphabetic characters can be converted into proper proper Chinese kanji notations without making a syllable-kanji bilingual dictionary or exceptional pronunciation rules. Can be converted. According to the invention of claim 2,
Means 25 for comparing the teacher signal and the output of the neural network 22a are provided, and the result of the neural network 22a
By learning, it is possible to convert proper nouns written in alphabetical characters into natural nouns of Chinese characters.

Further, according to the third aspect of the present invention, since the discrimination means 22b for generating an output when the input proper noun is a female name is provided, the proper noun of the female name represented by alphabetic characters is It can be converted into proper Chinese kanji proper nouns.

[0011]

FIG. 2 is a diagram showing the structure of a neural network used in the present invention. In the figure, A is an input layer, B is an intermediate layer, C is an output layer, and the output layer C is 1-1.
1 to 378 of 378 output units, and each output unit corresponds to 377 foreign Kanji characters and 1 skip.

The intermediate layer B is composed of 85 units 2-1 to 2-85. The input layer A is composed of seven groups 3a to 3g and one unit corresponding to the female name 3-1. Each group 3a to 3g is composed of 26 units that respond to 26 alphabetic characters and 2 units that respond to breaks in proper nouns, for a total of 28 units,
The input layer A has a total of 197 units 3-1 to 3
-197.

Further, in FIG. 2, 4 is a teacher signal, 5 is a comparing means for comparing the output of the output layer with the teacher signal 4, 6, 7
Is a weight value determined by the output of the comparison means 5. In order to convert a proper noun of an alphabetic character into a proper noun of kanji using the above neural network, first, the proper noun expressed by the alphabetical character is input from the input layer A, and the proper noun of the alphabetic character is given as the teacher signal 4. Input the proper noun in Chinese and Chinese corresponding to and learn using a neural network.

FIG. 3 is a diagram showing an example of the learning data. The proper noun represented by the alphabetical characters in the left column of the figure is input from the input layer A of the neural network and is shown in the right column of the figure. The proper noun in Chinese-Kanji notation is given as the teacher signal 4 to the comparison means 5 provided on the output side of the output layer C of the neural network. When a proper noun written in alphabetical characters in FIG. 3 is input from each group 3a to 3g in the input layer A of the neural network, the corresponding unit in each group 3a to 3g responds to the input alphabetical character.

This signal is converted in each unit of the input layer A, and is output to the output layer C via the intermediate layer B. The comparing means 5 compares the proper nouns written in Chinese and Chinese characters appearing in the output layer C with the teacher signal 4 and changes the weight values 6 and 7 so as to reduce the difference, and the strength of the unit coupling of each layer. change.

At this time, the neural network uses the following backpropagation algorithm for learning. The input X _i (i is the unit of the previous layer) and the output y _j (j is the unit of the output layer C) of the unit in each layer are defined by the following equations (1) and (2).

[0017]

[Equation 1]

Where X _i : input W _ji : weight value (weight for the input from the i-th unit of the previous layer) θ _j : threshold value f _j : output function, example: f (u) = 1 / (1 + exp (-u)) (si
gmoid function) y _j : output u _j : internal function I: number of input units When a certain input pattern p is given, the unit j of the output layer C
The error between the actual output pattern and the desired output pattern is defined as in the following expression (3), where y _ip is the actual output pattern and the desired output pattern (teaching signal) is t _jp .

[0019]

[Equation 2]

Here, n is the n-th learning (update of weight)
Means that. In order to train the network of the neural network, the strength of all the couplings, that is, the weight value W may be changed so as to reduce this error. In the back-propagation method, the weight value W is sequentially updated so that the sum of squares of the error is minimized, and the learning ends when the error between the output pattern in the output layer and the desired output pattern becomes a certain value or less. To do.

The correction amount ΔW of the weight value for each unit of the intermediate layer B and the output layer C is obtained based on the following method. (A) Correction amount Δ of the weight Wji between the intermediate layer B and the output layer C
_p W _ji (n) is calculated by the following equations (4) and (5). In addition,
In the following equation, δ _jp is the backward propagation amount of the error in the unit j of the output layer C.

Δ _p W _ji (n) = α δ _jp (n) y _ip (n) + β Δ _p W _ji (n-1) (4) δ _jp (n) = t _jp −y _ip (n) f '( u _ip (n)) (5) (b) Correction amount Δ of weight W _jh between input layer A and intermediate layer B
_p W _jh (n) is _calculated by the following equations (6) and (7). In addition,
δ _ip is the amount of backward propagation of the error in the unit i of the hidden layer B.

_Δp W _ih (n) = αδ _ip (n) y _hp (n) + β _Δp W _ih (n-1) (6)

[0024]

[Equation 3]

The second term in the formula for calculating the modification amount of the weight value is added in order to reduce the error oscillation and accelerate the convergence. Further, α is called a learning constant, and β is called a learning rate constant. The algorithm described above corrects the weight value of the neural network in the direction in which the desired output pattern is output. Note that the calculation order of the above mathematical formulas is called the back-propagation method because it proceeds in the reverse direction from the output layer C to the intermediate layer B to the input layer A.

Anyone can learn the neural net as long as he / she has knowledge of the problem to be applied, and it is not necessary to be an expert of the neural net. As described above, the neural network input is given a proper noun in alphabetical characters and the Chinese kanji proper noun corresponding to the input signal is given as a teacher signal to train the neural network and then input the neural network. Give the proper noun of the alphabetic character notation you want to convert, and convert it to the corresponding Chinese kanji.

FIG. 4 is a diagram showing a first embodiment of the present invention. In FIG. 4, the male name "Bridges" is input from the input layer of the learned neural network and the Chinese character "Nuri" is input. An example in which the proper noun "kiseki" is converted is shown. In the figure, the same components as those shown in FIG. 2 are designated by the same reference numerals, 10 is an input unit, 11 is a discrimination unit, A is an input layer, 3-1 is a unit corresponding to a female name. 3a to 3g are groups, B is an intermediate layer, C is an output layer, 12 is an output unit, and steps S1 to S are performed.
Reference numeral 8 indicates a conversion procedure.

In the figure, "B" cut out from the sentence
When a proper noun of a male name in alphabetical notation "ridges" is given from the input unit 10, the determination unit 11 determines whether or not the female name flag is attached.
In the case of the present embodiment, since it is a male name, the female name flag is not attached, and no input is given to the unit 3-1 corresponding to the female name of the input layer A.

Then, as shown in step S1 in the figure, "B" is input to the center group 3d of the input layer A, "r" is input to the group 3e, "i" is input to the group 3f, and "d" is input to the group 3g. To be done. The neural network responds to the above input by inputting "B" to the cloth "B" input to the central group 3d.
And is output from the output unit 12.

Next, as shown in step S2 of FIG.
"Bridgeges" shifts one character to the left, and Group 3
"B" for c, "r" for central group 3d, group 3
“I” is input to e, “d” is input to the group 3f, and “g” is input to the group 3g. In this case, the "r" input to the central group 3d is converted together with the next "i", so that the neural network outputs the skip signal from the output unit 12.

In step S3, again, "Brid
“Ges” shifts to the left by one character, “B” for group 3b, “r” for group 3c, “i” for central group 3d, “d” for group 3e, “g” for group 3f, group 3g. "E" is input. The neural network converts “r” input to the group 3c and “i” input to the central group 3d into “ri” according to the above input, and outputs it from the output unit 12.

As described above, from step S1 to step S8, the input proper nouns in alphabetical notation are converted into Chinese characters by shifting leftward by one character, and in step S8, the end mark "/ When "" is input to the central group 3d, the conversion is completed, and the Chinese kanji proper noun "Furi Kiseki" is obtained. FIG. 5 is a diagram showing a second embodiment of the present invention, in which the female name "Bridged" is input from the input layer of the neural network to input the Chinese kanji female name "Nunrei-Kyote". It shows an example of converting into a proper noun.

In the figure, the same components as those shown in FIG. 4 are designated by the same reference numerals, and step S
Reference numerals 1 to S8 indicate the same conversion procedure as in FIG.
In the figure, “Bridgege” cut out from the sentence
When the proper noun of the female name in alphabetical notation "t" is given from the input unit 10, the determination unit 11 determines whether or not the female name flag is attached. In the case of the present embodiment, since it is a female name, a female name flag is attached, and the input is given to the unit 3-1 corresponding to the female name of the input layer A.

Then, as shown in FIG. 4, in step S1 of the figure, the group 3d at the center of the input layer A is "B", the group 3e is "r", the group 3f is "i", and the group 3g is. "D" is input. The neural network converts “B” input to the central group 3d into “cloth” according to the input, and outputs it from the output unit 12.

Similarly to the case shown in FIG. 4, the inputted proper nouns in the alphabetical notation are converted to Chinese characters by shifting them one character to the left from step S1 to step S8. In this case, since the input is given to the unit 3-1 corresponding to the female name, "r
"i" is converted into "Rei" suitable for the female name, and
"Dge" is converted to "odd" and "t" is converted to "special",
Obtained the proper noun "Chinese character" of Chinese kanji.

[0036]

As is clear from the above description,
In the present invention, in a text processing system for a language that uses alphabetic characters, a proper noun of a language that uses alphabetical characters is converted into a proper noun of Chinese kanji notation using a neural network, so that the following effects are obtained. be able to. Even if there is an unregistered proper noun in the processing system, it can be converted into a natural Chinese proper noun, and the quality of processing can be improved. It is possible to convert proper nouns expressed in alphabetic characters into proper proper Chinese kanji notation without making a syllable-kanji bilingual dictionary or exceptional pronunciation rules. By learning a neural network using learning data, even if a proper noun other than the learned proper noun is input, it can be converted into an appropriate proper noun in Chinese-Kanji notation. By providing a unit that reacts to a female name in the input layer of the neural network, when a female name is input, it can be converted into a proper noun suitable for a female.

[Brief description of drawings]

FIG. 1 is a principle diagram of the present invention.

FIG. 2 is a diagram showing a configuration of a neural network used in the present invention.

FIG. 3 is a diagram showing an example of learning data.

FIG. 4 is a diagram showing a first embodiment of the present invention.

FIG. 5 is a diagram showing a second embodiment of the present invention.

[Explanation of symbols]

 1-1 to 1-378 Output Layer Units 2-1 to 2-85 Intermediate Layer Units 3-1 to 3-197 Input Layer Units 3a to 3g Input Layer Group 4,24 Teacher Signals 5,25 Comparing Means 10, 21 Input unit 11 Discrimination means 12, 23 Output unit 22 Proper noun-Chinese conversion unit 22a Neural network 22b Female name discrimination means A Input layer B Intermediate layer C Output layer

Claims (3)

[Claims] 1. A proper noun processing device of a text processing system for converting a language using alphabetic characters into Chinese, an input unit (21) for inputting proper nouns expressed by alphabetic characters, and a neural network (22a). A proper noun-Chinese conversion unit (22) configured to convert a proper noun in alphabetical character notation given from the input unit (21) into a proper noun of Chinese Kanji
And an output unit (23) that outputs the proper noun converted to Chinese-Kanji notation in response to the output of the proper noun-Chinese conversion unit (22), and the proper noun of the language that uses alphabetic characters is provided. , A proper noun processing device characterized by converting to a proper proper noun of Chinese kanji. 2. A neural network (2) using a proper noun in Chinese-Kanji notation corresponding to an alphabet character as a teacher signal.
2a) means (24), teacher signal and neural network
A means (25) for comparing the output of (22a) is provided, and the proper nouns and teacher signals in Chinese kanji notation corresponding to the alphabetic characters are sent to the neural network (22a) of the proper noun-Chinese converter (22). The proper noun processing apparatus according to claim 1, wherein the proper nouns of a language using alphabetic characters are converted into corresponding proper nouns of Chinese kanji notation after being given and learned by the neural network (22a). .. 3. A discriminating means (22b) for discriminating whether or not the input proper noun is a female name and producing an output when the input proper noun is a female name, and a discriminating means (22b). A neural network (22a) with a unit that responds to output is provided in the input layer, and when the input proper noun is a female name, it is converted into a proper noun in Chinese Kanji notation suitable for a woman. The proper noun processing apparatus according to claim 1 or 2.