JP3059504B2 - Part of speech selection system - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a natural language processing technique, and more particularly, to a part of speech selection system that automatically selects the type of part of speech of each word in a sentence (text).

[0002] In recent years, research on speech synthesis has been advanced, and it has begun to be used in various fields such as a reading system and a response system. In order to improve the naturalness of synthesized speech, it is necessary to extract linguistic information of a sentence to be output and generate a pattern for generating an accent or intonation according to the linguistic information. The part of speech forms the basis of such linguistic information. The part of speech refers to a type of word such as a noun, a verb, an adverb and the like. Since a voice synthesis system is required to have portability, a system that can easily extract the part of speech of a word in a sentence is strongly desired.

[0003]

2. Description of the Related Art Recently, a method of determining the part of speech of a word in a sentence using statistical information has been used. This is, for example, “A Stochastic Parts
Program and Noun Phrase
Passer for Unrestricted Te
xt "(ACL processings 2nd A
pplied Natural Language P
As shown in (processing), the part of speech of each word in a sentence is determined using a vocabulary probability that is a probability that a given word takes a part of speech and a context probability that indicates the frequency of appearance of a part of speech. .

FIG. 19 is a block diagram showing a configuration example of a conventional apparatus. In the figure, 1 is a vocabulary probability search unit that receives an input text and searches for vocabulary probabilities for each word, 2 is a vocabulary probability dictionary in which vocabularies and the probabilities of the vocabularies are stored, and are searched by the vocabulary search unit 1 Reference numeral 3 denotes a part-of-speech list creation unit that creates a part-of-speech list of each word based on the output of the vocabulary probability search unit 1.

[0005] Reference numeral 4 denotes a context probability detection unit for detecting context probabilities of a plurality of parts of speech based on the vocabulary probabilities retrieved from the vocabulary probability retrieval unit 1; The probability table 6 receives the output of the context probability detection unit 4 and is an evaluation value calculation unit that calculates an evaluation value from the context probabilities of a plurality of part-of-speech lists and the vocabulary probabilities of the first and last words of the part-of-speech list. is there. The operation of the device configured as described above will be described below.

First, a vocabulary probability search unit 1 inputs an input text (sentence) and reads a vocabulary probability for each word from a vocabulary probability dictionary 2. For example, if the input text is I love her very much much. It is assumed that the sentence was. Here, a three-word part-of-speech arrangement is considered as a category of the part-of-speech arrangement of words. The vocabulary probability search unit 1 extracts the vocabulary probabilities of the words constituting the sentence from the vocabulary probability dictionary 2.

The vocabulary probability dictionary 2 has, for example, a configuration as shown in FIG. The contents shown in the figure are obtained by extracting only words of the input text, and in fact, all words are stored in alphabetical order as appearance probabilities for each type of part of speech.

[0008] Here, consider a case where processing is performed after a sentence.

The part-of-speech list creation unit 3 first creates a part-of-speech list (Z ##) for the period “. ##” of the sentence, and gives it to the context probability detection unit 4. Here, # is a symbol indicating the end of the sentence or the beginning of the sentence. The context probability detection unit 4 extracts a context probability of a part of speech sequence (Z ##) from the context probability table 5. The context probability table 5 stores, as context probabilities, the appearance probabilities of all patterns that are combinations of three parts of speech.

FIG. 21 is a diagram showing a configuration example of a context probability table. The figure shows only a part thereof. Also, here, in the case of English as a part of speech,
Word part of speech is used. The probabilities that each combination of parts of speech can take are obtained by statistically calculating the probabilities that can be taken with reference to a large amount of sentences.

The evaluation value calculator 6 multiplies the context probability of the part of speech (Z ##) by the vocabulary probability of (.Z) to obtain an evaluation value. For example, if the context probability of (Z ##) is 0.983,
If the vocabulary probability of (.Z) is 1.00, the evaluation value is 0.983.

Next, the part-of-speech list creation unit 3 sends the word list "m
uch. # ”For possible combinations of parts of speech (ADJ Z #), (NOUN Z #), (ADV
Z #) is created. The context probability detection unit 4 determines these parts of speech (ADJ Z #), (NOUN Z #), (A
The context probabilities for DVZ #) are shown in the context probability table 5
Extract from Here, ADJ is an adjective, NOUN is a noun, and ADV is an adverb. FIG. 24 shows the classification of parts of speech.
In the case of English, as shown in the figure, there are 20 types of parts of speech, and here, it is assumed that each part of speech uses a part of speech symbol as shown in the figure.

The evaluation value calculation unit 6 determines that if “much” takes ADJ in each part of speech sequence,
The evaluation value is calculated by multiplying the context probability and the vocabulary probability in the case of taking UN and the case of taking ADV by the corresponding cumulative evaluation value.

Next, the part-of-speech list creator 3 generates the word list "ve
ry much. ”(ADV ADJ
Z), (ADV NOUNZ Z), (ADV AD
VZ), (ADJ ADJ Z), (ADJ NOUN
Z Z) and (ADJ ADV Z).

The context probability detection unit 4 detects the part of speech (A
DV ADJ Z), (ADV NOUNZ Z),
(ADV ADV Z), (ADJ ADJ Z),
The context probabilities for (ADJNOUNZ Z) and (ADJ ADV Z) are extracted from the context probability table 5. The evaluation value calculation unit 6 determines, for each part of speech,
For each case in which “very” takes ADV and ADJ, the evaluation value of each part-of-speech list is calculated by multiplying the context probability, the vocabulary probability, and the cumulative evaluation value.

Hereinafter, when the same operation is repeated up to “## I”, five parts of speech are arranged, and I love her very much. The evaluation values for all possible part-of-speech arrangements are obtained. Then, the one having the highest evaluation value is selected as the part of speech sequence. In this case, the part-of-speech sequence to be selected is ## PPRON V PRON ADV ADV Z ##.

[0017]

In the conventional apparatus described above, the context probabilities of the parts of speech are stored in a table (context probability table 5). However, there is a problem that a large amount of memory is required.

The present invention has been made in view of such a problem, and has as its object to provide a part-of-speech selection system capable of reducing the memory capacity.

[0019]

FIG. 1 is a block diagram showing the principle of the present invention. The same components as those in FIG. 19 are denoted by the same reference numerals. In the figure, 1 is a vocabulary probability search unit that receives an input text and searches for vocabulary probabilities for each word, 2 is a vocabulary probability dictionary in which vocabularies and the probabilities of the vocabularies are stored, and are searched by the vocabulary search unit 1 It is.

Reference numeral 10 denotes a context probability detecting unit for detecting a context probability of a plurality of parts of speech based on the vocabulary probabilities retrieved from the vocabulary probability retrieving unit 1. Reference numeral 20 denotes an input pattern provided from the context probability detecting unit 10. To calculate the context probability according to the type of part of speech of the next word, using a neural network.
An evaluation value calculation unit that receives an output of 0 and calculates an evaluation value according to a predetermined procedure from the context probabilities of a plurality of part-of-speech lists and the vocabulary probabilities of the words at the beginning or end of the part-of-speech list.

[0021]

[Operation] The context probability calculation unit 20 is made to previously learn the probability of the part of speech arrangement. For example, in the case of a three-word part of speech,
The neural network inputs the input patterns of the two parts of speech from the back, and the probability according to the type of part of speech of the next word (head word) at that time is given as a teacher pattern in advance, and the teacher pattern and the output of the neural network are output. Train them to be equal.

Thereafter, the context probability calculation unit 20 inputs the pattern of the part-of-speech arrangement given from the context probability detection unit 10 and calculates the context probability according to the type of the part of speech of the next word. For example, in the case of a three-word part-of-speech arrangement, two parts-of-speech arrangements are input as input patterns, and the probability of taking the part-of-speech arrangement of the next word (context probability) is calculated for each type of part of speech. Based on the calculated context probabilities, the evaluation value calculation unit 6 calculates an evaluation value for each pattern of the three-word part of speech.

As described above, according to the present invention, the neural network constituting the context probability calculating unit 20 learns the probability of the part-of-speech arrangement. If a pattern is input as an input pattern, the probability (context probability) of the next third part of speech (the first word of speech) is output for each type of part of speech, so that information on the context probability is provided as a context probability table. It is possible to provide a part-of-speech selection system that eliminates the need and can reduce the memory capacity.

[0024]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS The operation of the present invention will be described below in detail with reference to the drawings.

FIG. 2 is a diagram showing a configuration example of the context probability calculation unit 20. 21 is a neural network, the input layer L
1, an intermediate layer L2 and an output layer L3. The input layer L1 is provided with a combination of the number of parts of speech categories (the number of parts of speech arranged minus one). For example, the part of speech is 3
When the number of parts of speech is one, the input layer L1 includes two combinations of the number of parts of speech categories. The output device L3 has one category number combination. The unit of the combination of the number of parts of speech category is, for example, 20 which is the number of parts of speech in English.
(20 units) are used.

Reference numeral 22 denotes a comparing unit for comparing the output value of the output layer L3 with the teacher pattern. The comparison unit 22 includes an output layer L
The weight of the synapse between the input layer L1 and the intermediate layer L2 and the weight between the intermediate layer L2 and the output layer L3 are changed so that the output (probability) of No. 3 becomes equal to the teacher pattern. Eventually, learning ends when the output value of the output layer L3 matches the teacher pattern. At this time, the weight of the synapse between the input layer L1 and the intermediate layer L2 and the weight of the synapse between the intermediate layer L2 and the output layer L2
The weight between the three is fixed to a certain value.

FIG. 3 is a block diagram showing the configuration of an embodiment of the present invention. 1 that are the same as those shown in FIG. In the figure, a context probability detecting unit 10 receives a word from the vocabulary probability searching unit 1 and generates a predetermined number (for example, three) of a part-of-speech list, and outputs a part-of-speech list creating unit 11 and an output of the part-of-speech list creating unit 11 An input pattern creating unit 12 for creating an input pattern in response to the input, and inputting the output of the input pattern creating unit 12 to a context probability calculating unit 20 to receive the context probability for each part of speech obtained from the neural network and output an output value Is selected from the output value selection unit 13 for selecting the value.

The evaluation value calculating unit 6 calculates an evaluation value from the context probability and the vocabulary probability given from the output value selecting unit 13, and outputs a plurality of outputs from the evaluation value calculating unit 31. It comprises a maximum value selection unit 32 for selecting the one with the largest evaluation value from among them. Then, the output of the maximum value selection unit 32 becomes the part of speech string to be obtained. The operation of the device configured as described above will be described below.

The neural network 21 in this embodiment selects a part-of-speech from the last word toward the first word in the input text in the reverse direction and learns from the part-of-speech frequency extracted from a large amount of data in advance. Alternatively, any of the methods of selecting and learning the part of speech in a forward direction from the first word to the last word can be used. FIG. 4 is a diagram showing an example of creating a learning pattern when selecting a part of speech from the last word. Here, it is assumed that a three-word part-of-speech arrangement is used. For example, consider a case where three adverbs are arranged like ADV ADV ADV. When a large amount of data is searched, the frequency of such part-of-speech arrangement is 307, and the context probability is 0.07. This 0.07 is
The frequency 307 is obtained as a value obtained by dividing the last two parts of speech of a three-word part of speech by the sum of the frequencies of the same parts. In this way, the frequency of the sequence of NOT and two adverbs such as the next part of speech sequence NOT ADV ADV is 78, and the context probability is 0.02. Similarly,
The last two are ADV, and the first one is 20 categories
Is shown in FIG. 4 as the context probability. This context probability is calculated using the neural network 21
A method for causing the learning will be described.

Here, the neural network 21 shown in FIG. 2 has two sets of inputs, and the part of speech type is AD
V. This ADV is shown in a state where "1" is set in the first unit of the 20 units. Here, the use of 20 units as the number of part-of-speech categories corresponds to the case where English has 20 kinds of part-of-speech as shown in FIG.

FIG. 5A is a diagram showing an example of the input pattern at this time. Now, the input pattern creation unit 12 causes the neural network 21 to input a pattern as shown in FIG. If, for example, the last two words in the three-word part-of-speech sequence are NOUN ADVs, the input pattern is as shown in FIG. 5B. In the case of NOUN, the number is "4" from FIG. In FIG. 24, since the number starts from 0, “4” indicates the fifth. Therefore, in the case of NOUN, "1" stands fifth from the front.

FIG. 6 is an explanatory diagram of the learning at this time. The input layer L1 receives an input of a total of 40 units in two sets of 20 units. The presence of “1” in the first unit of each set indicates an adverb. The output is 20 units. On the other hand, the teacher pattern given to the comparison unit 22 of the neural network 21 is as shown in FIG. This teacher pattern uses the context probability shown in FIG. 4 as it is.

That is, since the probability of the part-of-speech arrangement that becomes ADV ADV ADV is 0.07, this value is given as a teacher signal to the first unit of the 20 units in the output layer L3. The next unit is given 0.02, which is the probability of a part-of-speech arrangement that becomes NOT ADV ADV. Less than,
The values in FIG. 7 are provided to the comparing unit 22 as teacher patterns. The last unit is given with 0.08, which is the probability of the part-of-speech arrangement as #ADV ADV. In the configuration of FIG. 6, the output of each unit of the output layer L3 is 0.0
The learning of the neural network 21 is performed by giving the same input pattern many times until the values become 7, 0.02,..., 0.08.

As described above, the case where the context probability according to the type of part of speech taken by the first word when the latter two are ADV and ADV is taken as an example, but the part of speech arrangement is not limited to this. 5 (b), an extremely large number of patterns (for example, in the case of 20 categories, 3 parts-of-speech
There is a learning pattern 400). The learning is performed by the neural network 21 shown in FIG. 6 for each combination of the patterns.

When the input text is input to the device shown in FIG. 3 in a state where the learning of all the three-word part-of-speech sequences has been completed in this way, the context probability calculating unit 20 determines that three contexts corresponding to the input pattern have been input. Output context probability of part of speech. The output value selection unit 13 selects an output value of the probability of the target part-of-speech arrangement from the output of the context probability calculation unit, and gives it to the evaluation value calculation unit 31.

The evaluation value calculating means 31 calculates an evaluation value by multiplying the context probability, the vocabulary probability of the head word, and the cumulative evaluation value of the corresponding part of speech sequence up to that point. Evaluation value calculation means 31
Calculates the evaluation values for all possible cases of the three part-of-speech arrangements, so the maximum value selection unit 32 selects the part-of-speech arrangement having the largest value among these evaluation values, and outputs it as the part-of-speech sequence to be obtained. (Case of Part-of-Speech Selection from First Word) The case of selecting a part-of-speech from the first word can be considered in the same manner as the case of selecting the part-of-speech from the last word described above.

FIG. 8 is a diagram showing an example of creating a learning pattern when selecting a part of speech from the last word. Here, it is assumed that a three-word part-of-speech arrangement is used. For example, consider a case where three adverbs are arranged like ADV ADV ADV. When a large amount of data is searched, the frequency of such part-of-speech arrangement is 307, and the context probability is 0.07. This 0.07 is
Frequency 307 is the first two parts of speech of the three-word part-of-speech arrangement is ADV, AD
It is obtained as a value obtained by dividing by the total sum of the frequencies of the three part-of-speech arrangements as V. Thus, the frequency of the next part-of-speech sequence, such as ADV ADV NOT and the sequence of two adverbs, is 6,
Its context probability is 0.00. Similarly, FIG. 8 shows the probability that the previous two all take ADV and the last one takes any part of speech with 20 categories as the context probability. A method for causing the neural network 21 to learn the context probability will be described.

Here, it is assumed that there are two sets of inputs, and that each of the parts of speech is ADV. This ADV is indicated by “1” in the first unit of the 20 units.
Is shown in an upright position.

FIG. 9 is a diagram showing an example of the input pattern at this time. Now, the input pattern creation unit 12 causes the neural network 21 to input a pattern as shown in FIG.
The configuration of the neural network at this time is the same as in FIG.

The input layer L1 receives an input of a total of 40 units in two sets of 20 units. The presence of “1” in the first unit of each set indicates an adverb.
The output is 20 units. On the other hand, the teacher pattern given to the comparison unit 22 of the neural network 21 is as shown in FIG. This teacher pattern is shown in FIG.
The context probability shown in FIG.

That is, since the probability of the part-of-speech arrangement that becomes ADV ADV ADV is 0.07, this value is given as a teacher signal to the first unit of the 20 units in the output layer L3. The next unit is given 0.00, which is the probability of the part-of-speech arrangement that becomes ADV ADV NOT. Less than,
The values in FIG. 10 are provided to the comparing unit 22 as teacher patterns. The last unit is given 0.00, which is the probability of the part-of-speech arrangement that becomes ADV ADV #. In the configuration of FIG. 6, the output of each unit of the output layer L3 is 0.0
The neural network 21 is trained by giving the same input pattern many times until the value becomes 7, 0.00,..., 0.00.

As described above, the case where the context probability according to the type of part of speech taken by the last word in the case where the first two words are ADV and ADV is taken as an example, but the part of speech arrangement is not limited to this. There are many patterns. The learning is performed by the neural network 21 shown in FIG. 6 for each combination of the patterns.

When the input text is input to the apparatus shown in FIG. 3 in a state where the learning on all the three-word part-of-speech arrangements has been completed, the context probability calculation unit 20 determines that the three parts corresponding to the input pattern Output context probability of part of speech. The output value selection unit 13 selects an output value of the probability of the target part-of-speech arrangement from the output of the context probability calculation unit, and gives it to the evaluation value calculation unit 31.

The evaluation value calculating means 31 calculates an evaluation value by multiplying the context probability, the vocabulary probability of the last word and the cumulative evaluation value of the corresponding part of speech sequence up to that point. Evaluation value calculation means 31
Calculates the evaluation values for all possible cases of the three part-of-speech arrangements, so the maximum value selection unit 32 selects the part-of-speech arrangement having the largest value among these evaluation values, and outputs it as the part-of-speech sequence to be obtained.

Next, the operation of the apparatus of the present invention will be described using a specific example. Here, it is assumed that the parts of speech classified into the 20 categories shown in FIG. 24 are used, and the parts of speech of each word of the input text are selected from the end to the beginning by using the context probabilities of the three parts of speech. . Here, the part-of-speech selection process is performed for the sentence I appealed to the children to make less noise. The vocabulary probability search unit 1 searches the vocabulary probability dictionary 2 for each word in the input text, and extracts a part of speech category and its probability for each word. FIG. 11 is a diagram showing the part-of-speech categories extracted in this way and their probabilities. Corresponding to the sentence, the part of speech category of each word and its vocabulary probability are shown.

The part-of-speech arrangement creating unit 11 creates a part-of-speech arrangement of three words based on the search result. FIG. 12 is a diagram showing a three-word part-of-speech arrangement created in this way. All possible parts-of-speech sequences that can be realized from the later part of the sentence are created.

The input pattern creating unit 12 creates an input pattern as shown in FIG. 5 from the two parts of speech after these part of speech arrangements, and gives it to the context probability calculating unit 20 using a neural network. The context probability calculation unit 20 outputs the probability of taking the first word from the last two parts of speech as the context probability for each part of speech category.

The output value selection unit 13 selects a value for the first part of speech in the part of speech list for the output of the context probability calculation unit 20 for the given input pattern, and sets it as the context probability of the part of speech list.

The evaluation value calculating means 31 outputs the output value
Multiplied by the context probability Pb obtained from the above and the vocabulary probability Pv for the first part of speech of the part-of-speech sequence searched by the vocabulary probability search unit 1 to the cumulative evaluation value e ′ for the part-of-speech sequence created from the last word of the input text e Is the new evaluation value. The new evaluation value e is represented by the following equation.

E = Pb × Pv × e ′ (1) FIGS. 13 and 14 are diagrams showing the evaluation value calculation results obtained in this way. 13 and FIG. 14 are continuous, and FIG. 14 comes after FIG. In the figure, part A is the context probability obtained by the above-described sequence in the context probability calculation unit 20, B is the vocabulary probability according to the type of part of speech of the first word, and C is the cumulative evaluation value obtained in the process up to that point. “e” and “D” indicate the part-of-speech arrangement corresponding to the cumulative evaluation value, respectively.

In order from the last word of the input text, for all of the part of speech categories taken by the word, for all of the part of speech parts created so far, a three part of speech part list is created by adding the last two parts of speech and the context Search for the probability Pb. The product of the context probability Pb, the vocabulary probability Pv of the word in the part-of-speech category, and the evaluation value e ′ of the part-of-speech list created up to that time is defined as the evaluation value e of the new part-of-speech list.
As shown in the figure, after the evaluation values of all parts of speech are calculated up to the first word of the input text, the maximum value selection unit 32
Selects the part-of-speech list having the highest evaluation value from these part-of-speech arrangements, and determines the part of speech of each word in the input text.

FIG. 15 is a diagram showing a part of speech selection result obtained in this way. This matches the part-of-speech arrangement at the last stage in FIG.

FIG. 16 is a structural block diagram showing another embodiment of the present invention. The same components as those in FIG. 3 are denoted by the same reference numerals. In the figure, 1 is a vocabulary probability search unit that receives an input text and searches for vocabulary probabilities for each word, 2 is a vocabulary probability dictionary in which vocabularies and the probabilities of the vocabularies are stored, and are searched by the vocabulary search unit 1 Reference numeral 10 denotes an input pattern creation unit as a context probability detection unit that creates an input pattern by receiving the output of the vocabulary probability search unit 1.

Reference numeral 20 denotes a context probability calculation unit using a neural network, which inputs an input pattern given from the input pattern creation unit 10 and calculates a context probability according to the type of part of speech of the next word. 31 denotes a context probability calculation. The evaluation value calculating means 32 receives the output of the section 20 and calculates an evaluation value from the given context probability and vocabulary probability.
A maximum value selection unit 32 that selects the output having the largest evaluation value from among a plurality of outputs from one. The evaluation value calculation unit 6 and the maximum value selection unit 32 constitute the evaluation value calculation unit 6. The operation of the device configured as described above will be described below.

In this embodiment, instead of inputting the input pattern of the latter two parts of speech as "1" or "0" as shown in FIG. 5, each part of speech category of the words constituting the latter two parts of speech is An evaluation value indicating the probability of taking is input.

As in the embodiment of FIG. 3, the vocabulary probability search unit 1
Searches the vocabulary probability dictionary 2 for each word in the input text, and extracts the part of speech category and its probability for each word as shown in FIG. Here, the evaluation value at the time of selecting the part of speech is expressed by the vocabulary probability of the part of speech category for each word, and the adjacent two words or the preceding two words of the part of speech category of the two words before and after to the context probability calculation unit 20 using the neural network. Enter the probability.

FIGS. 17 and 18 are diagrams showing examples of evaluation value calculation. FIGS. 17 and 18 show a series of operations, and FIG. 18 follows FIG. When selecting a part of speech from the last word of the input text, the input pattern creating unit 10 sets the last two words of the three-word part of speech
The evaluation value E of the word is input to the neural network 21, and the obtained output is the context probability Pb.

The probability of each part-of-speech category of the context probability Pb multiplied by the probability of each part-of-speech category of the word obtained by the vocabulary probability search unit 1 is set as the evaluation value E of the word.

The maximum value selection section 32 selects, for each word of the input text, the part of speech category having the largest value in the evaluation value E as the part of speech of the word. For example,
In the case of FIG. 17, when Wn = 6, "make" is evaluated as a word. As for the evaluation value E at this time, the tenth unit from the front has the maximum value of 0.34. The tenth part of speech from the front is a verb (V) from FIG. The result of selecting the part of speech in this way is also shown in FIG.
Will be the same as

According to this embodiment, by inputting the evaluation value indicating the probability of taking the part of speech category to the neural network 21, the number of context probability searches can be reduced, and the part of speech can be selected at high speed. Can be.

In the above embodiment, the case where the present invention is applied to English has been described, but the present invention is not limited to this. The present invention can be applied to Japanese and other sentences.

[0062]

As described in detail above, according to the present invention, by using a neural network to obtain the context probability, it is not necessary to have a context probability table, and the memory capacity can be reduced. Can be provided.

[Brief description of the drawings]

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

FIG. 2 is a diagram illustrating a configuration example of a context probability calculation unit.

FIG. 3 is a configuration block diagram showing one embodiment of the present invention.

FIG. 4 is a diagram showing an example of creating a learning pattern when selecting a part of speech from the last word.

FIG. 5 is a diagram showing an example of an input pattern.

FIG. 6 is an explanatory diagram of learning.

FIG. 7 is a diagram showing an example of a teacher pattern.

FIG. 8 is a diagram showing an example of creating a learning pattern when a part of speech is selected from the first word.

FIG. 9 is a diagram illustrating an example of an input pattern.

FIG. 10 is a diagram showing an example of a teacher pattern.

FIG. 11 is a diagram showing extracted part-of-speech categories and their probabilities.

FIG. 12 is a diagram illustrating a created part-of-speech arrangement of three words.

FIG. 13 is a diagram showing evaluation value calculation results obtained by the present invention.

FIG. 14 is a diagram showing evaluation value calculation results obtained by the present invention.

FIG. 15 is a diagram showing a part of speech selection result according to the present invention.

FIG. 16 is a configuration block diagram showing another embodiment of the present invention.

FIG. 17 is a diagram illustrating an example of evaluation value calculation.

FIG. 18 is a diagram showing an example of evaluation value calculation.

FIG. 19 is a block diagram illustrating a configuration example of a conventional device.

FIG. 20 is a diagram showing an example of the internal configuration of a vocabulary probability dictionary.

FIG. 21 is a diagram illustrating a configuration example of a context probability table.

FIG. 22 is a diagram illustrating classification of parts of speech.

[Explanation of symbols]

 Reference Signs List 1 vocabulary probability search unit 2 vocabulary probability dictionary 6 evaluation value calculation unit 10 context probability detection unit 20 context probability calculation unit

Continuation of the front page (56) References Yamaguchi, G., "Phrase Boundary Extraction Function Using Neural Network", Proc. Of the IEICE National Convention, Autumn 1990, p. Language Processing and Phoneme Symbol Generation Using -Application to English Speech Synthesis System- ", Transactions of the Institute of Electronics, Information and Communication Engineers, 1992, Vol. J75-D II, No. 5, p. 852-860 Matsumoto T, Yamaguchi Y, "A Multi-Language Text-to-Speech System Usage Neural Networks, Inc. Field (Int.Cl. ⁷ , DB name) G06F 17/27 G06F 17/28 G06F 15/18 JICST file (JOIS)

Claims (4)

(57) [Claims] 1. A vocabulary probability search unit (1) that receives an input text and searches for a vocabulary probability for each word; a vocabulary and a probability of the vocabulary are stored;
A vocabulary probability dictionary (2) to be searched for; and a context probability detection unit (10) for detecting a context probability of a plurality of parts of speech based on the vocabulary probabilities retrieved from the vocabulary probability retrieval unit (1). A context probability calculation unit (20) using a neural network, which inputs an input pattern given from the context probability detection unit (10) and calculates a context probability according to the type of part of speech of the next word; In response to the output of the context probability detection unit (10), an evaluation value calculation unit (6) calculates an evaluation value according to a predetermined procedure from the context probabilities of a plurality of part-of-speech lists and the vocabulary probabilities of words at the beginning or end of the part-of-speech list. ) Is a part-of-speech selection system. 2. When the context probability calculating section (20) calculates a context probability of an n-word part-of-speech list, a part-of-speech pattern of (n-1) th word from the beginning of the part-of-speech list is input as (n 1) The part-of-speech selection system according to claim 1, wherein a probability of taking each part-of-speech category of the n-th word with respect to the part-of-speech arrangement of the word is output. 3. When the context probability calculating section (20) calculates a context probability of an n-word part-of-speech list, a part-of-speech pattern of a word from the second to the end n of the part-of-speech list is input as (n−1) 2. The part-of-speech selection system according to claim 1, wherein a probability of taking each part-of-speech category of the first word with respect to the word part-of-speech arrangement is output. 4. When the context probability calculating section (20) calculates a context probability of an n-word part-of-speech list, an evaluation value indicating a probability of taking each part-of-speech is input to (n-1) part-of-speech arrangement patterns. 4. The part-of-speech selection system according to claim 2, wherein a probability that each part-of-speech category of the nth or first word is taken is output.