JP4226942B2 - Accent position estimation method, apparatus and program - Google Patents

Description

[0001]
[Industrial application fields]
The present invention relates to an accent position estimation method, apparatus, and program, and in particular, an accent position estimation rule for estimating the accent position of an unregistered word is identified to maximize the distance between the identification boundary surface and data close to the identification boundary surface. The present invention relates to an accent position estimation method, apparatus, and program for estimating an accent position of a word that is acquired by learning with a learning algorithm and is not registered using a rule obtained by the learning.
[0002]
[Prior art]
A speech synthesizer that synthesizes speech based on text reads out each word in the input text and extracts accent position information, which is information indicating which sound should be distinguished, from the dictionary and synthesizes speech It is used to generate However, in any text, a word that is not registered in the dictionary, that is, an unregistered word appears. Therefore, in speech synthesis of an arbitrary text, it is necessary to estimate probable reading and accent information for an unregistered word.
In this way, both speech and accent position information is required for speech synthesis, but the accent position is compared to the case where reading is given in parenthesis or other words after the word that appears in the input text. Is rarely or rarely given. This situation is not limited to Japanese, and is generally the same in Chinese, English, French, German, and other languages that do not write accent marks in the text that indicate the points that make the sound stand out. Therefore, in the speech synthesis, the estimation of the accent position has an important meaning.
[0003]
In the case of Japanese standard words, for example, a word consisting of N mora is (N + 1) depending on whether accents are placed somewhere in the N mora (N ways) or not (one way). ) Street accent type is possible. Words have one of the accent types. In English, there are a total of (N + 1) cases where accents are placed or not placed anywhere in the N syllables, one of which is determined for each word. ing.
From the above, attempts have been made to solve the problem of accent position estimation as an identification problem of (N + 1) categories of the word.
[0004]
As a conventional technique, a technique for defining a score based on a statistical distribution and estimating an accent type from a score value for each word has been announced (see Non-Patent Document 1).
However, the technique of Non-Patent Document 1 merely performs two simple classifications as accent type estimation. In other words, the three-mora word only deals with two types of identification, either flat or one of the four possibilities, and the four-mora word has five of the possibilities. It only deals with two types of identification: whether it is the two or the two. Actually, there are other shapes, but there is no mechanism to give them an accent type.
[0005]
As a prior art, a result of manually creating and evaluating an accent type estimation rule based on differences in ending reading has been reported (see Non-Patent Document 2).
However, similarly to the technique of Non-Patent Document 1, the technique of Non-Patent Document 2 has an accent type that is not handled, and does not have a mechanism for providing these types.
As a conventional technique, a neural network or a decision tree is used as a classifier for performing classification and identification. These discriminators are obtained as a result of inductive learning using a large number of examples (see Patent Document 1).
However, since a classifier called neural network and decision tree used in the technique of Patent Document 1 is constructed based only on the frequency of examples of learning examples and other statistics, the number of examples is correct but not enough. There is no guarantee that the case will not be fully learned. In other words, there is no guarantee that the position of the identification boundary as a learning result is appropriate even in the distribution of learning data, and there is always a risk that it is not appropriate for unknown data that is not used for learning. ing.
[0006]
In the techniques of Non-Patent Document 2 and Patent Document 1, the input information for estimating the accent position is limited to a kana character string or phoneme symbol string representing a word reading. However, Japanese words have different accents depending on how readings are combined, that is, how they are organized. For example, in the readings of the names “Asaka” and “Asaka” with the same “Asa” reading, there are groups before and after “/”, such as “Asa / ka” and “Asa Saka”. Depending on the difference, the sound is generated with different pitches, such as “low and high” and “high and low”. That is, the accents are different. However, the techniques of Non-Patent Document 2 and Patent Document 1 cannot give different accents to the input information because the input information is limited to phoneme symbol strings. Patent Document 1 describes that additional information can be set, but does not describe what information is set.
[0007]
[Non-Patent Document 1]
Hirokawa, et al. "Accent Form Estimation Method for Personal Names (Last Names)", Acoustical Society of Japan, May 1981, pp. 425-426
[Non-Patent Document 2]
Tsuchida, Gao, “Rules for Assigning Japanese Name Accents”, IEICE Transactions, Vol. J67-D No. 5, pp. 625 to 626
[Patent Document 1]
JP 2002-73071 A
[0008]
[Problems to be solved by the invention]
In non-patent literature 1 and non-patent literature 2, it is not a discriminating mechanism that considers all accent types, and the decision tree used in patent literature 1, the discriminator of the neural network itself, and the lack of learning It is a point that cannot handle languages other than Japanese of Non-Patent Document 1, Non-Patent Document 2, and Patent Document 1, and the difference in accents due to the difference in reading of Non-Patent Document 2 and Patent Document 1 is regulated. It is a point that cannot be expressed in.
This invention maximizes the distance between the identification boundary surface represented by the support vector machine (SVM) and the data for identification learning close to the identification boundary surface, instead of using the neural network and decision tree classifier disclosed in Patent Document 1. A rule for estimating the accent position from learning data consisting of a set of linguistic information about the word and the accent position of the word, and using that rule, a new word not included in the learning data is learned. An accent position estimation method, apparatus, and program for estimating an accent position with high accuracy are provided.
[0009]
Patent Document 1 is intended only for Japanese, but by setting a partial character string formed by dividing English words into syllable units as linguistic information, an accent position where English accent position can be estimated. An estimation method, apparatus, and program are provided.
It is another object of the present invention to provide an accent position estimation method, apparatus, and program that can perform more accurate accent position estimation by using linguistic information that expresses a group of readings as identification information.
[0010]
[Means for Solving the Problems]
  An accent position estimation rule storage unit 104 that stores an accent position estimation rule for estimating an accent position using language information about a word as input, a language information setting unit 102 that sets language information about an unregistered word, and the language information setting unit An estimation unit 105 that selects an accent position estimation rule stored in the accent position estimation rule storage unit according to set language information and estimates an accent position of the unregistered word;DoAn accent position estimation device was constructed.
[0011]
  The language information includes the part of speech of the word, all the partial character strings constituting the pronunciation of the word, the total number of the partial character strings, and the presence / absence of coupling between the partial character strings (the number of partial character strings-1). Includes everything with information.
[0012]
  In the above accent position estimation apparatus, partial character strings created by dividing the pronunciation of a Japanese word into mora units are used as all the partial character strings constituting the pronunciation of the word.Accent position estimationapparatusConfigured.
[0013]
  The previous accent position estimation apparatus further includes a learning data storage unit 101 that stores a set of language information related to a word and an accent position of the word, and the language information setting unit 102 also includes language information related to a word for learning. An accent that is set and an accent position is estimated by inputting language information about a word based on learning data selected from the learning data storage unit 101 according to the learning language information set by the language information setting unit 102 An accent position estimation apparatus further comprising a learning unit 103 for learning a position estimation rule was configured.
[0014]
  A language information setting process for setting language information relating to an unregistered word, an estimation process for selecting an accent position estimation rule according to the language information set in the language information setting process, and estimating an accent position of the unregistered word; The linguistic information includes the part of speech of the word, all the partial character strings constituting the pronunciation of the word, the total number of partial character strings,(Partial character string number-1) pieces of information indicating the presence / absence of coupling between partial character strings;An accent position estimation method including all of the above is constructed.
[0015]
  AheadIn the accent position estimation method, the accent position estimation method using partial character strings created by dividing the pronunciation of a Japanese word into mora units as all the partial character strings constituting the pronunciation of the word is configured. . Moreover, the accent position estimation program for functioning a computer as each part of the accent position estimation apparatus mentioned above was comprised.
[0016]
DETAILED DESCRIPTION OF THE INVENTION
The accent position estimation apparatus according to the present invention is set by a learning data storage unit that stores a set of language information about a word and an accent position of the word, a language information setting unit that sets language information about the word, and a language information setting unit Based on the learning data selected from the learning data storage unit according to the recorded information, the rule for estimating the accent position with the linguistic information about the word as the input, the distance between the identification boundary surface and the data close to the identification boundary surface is maximized A learning unit that learns by the discriminative learning algorithm, an accent position estimation rule storage unit that stores a learning result, and language information setting unit that sets language information related to an unregistered word, and an accent position estimation rule according to the set language information Select the accent position estimation rule stored in the storage unit and estimate the accent position of the unregistered word. It is provided with an estimation unit for.
[0017]
In addition to using SVM as the learning means and the accent position estimation rule, a learning example that uses a probability model estimated by the maximum entropy method, or for the purpose of enhancing identification accuracy represented by bagging or boosting A mixture of a decision tree, a decision list, a neural network, and a linear classifier learned based on the selection algorithm is used.
The accent position estimation method according to the present invention includes a learning data storage unit that stores a set of language information related to a word and an accent position of the word, and includes a language information setting step for setting language information related to a word. Based on the learning data selected from the learning data storage unit according to the language information set in the setting step, the accent position estimation rule for estimating the accent position by inputting the language information about the word, the identification boundary surface and the identification boundary surface Has a learning process that learns by the discriminative learning algorithm that maximizes the distance to data close to, sets language information about unregistered words by the language information setting process, and selects an accent position estimation rule according to the set language information And an accent position estimating step of estimating the accent position of the unregistered word.
[0018]
In addition to using SVM as the above learning process and accent position estimation rule, use of a probabilistic model estimated by the maximum entropy method, or learning examples aimed at enhancing identification accuracy represented by bagging and boosting A mixture of a decision tree, a decision list, a neural network, and a linear classifier learned based on the selection algorithm is used.
[0019]
【Example】
Embodiments of the present invention will be described with reference to the examples of the drawings. FIG. 1 is a block diagram showing an embodiment of an accent position estimating apparatus according to the present invention.
In this embodiment of the accent position estimation apparatus, in order to estimate the accent position of a word that is not registered, an accent position estimation rule is learned from learning data consisting of a set of language information about the word and information on the accent position of the word, The estimation rule obtained as a result of learning is used to estimate the accent position of an unregistered word. For example, it is used in a speech synthesizer that converts arbitrary text into speech. This will be specifically described below.
[0020]
In FIG. 1, reference numeral 101 denotes a learning data storage unit, which stores language information relating to a word for learning an accent position estimation rule and a set of accent positions of these words. Reference numeral 102 denotes a language information setting unit that sets language information related to words at the time of learning and at the time of estimating an accent position. Reference numeral 103 denotes a learning unit, which uses the learning data read from the learning data storage unit 101 and the language information set by the language information setting unit 102 to determine the distance between the identification boundary surface and data close to the identification boundary surface. One or more discriminators that identify accent positions are learned as accent position estimation rules based on a discriminative learning algorithm that maximizes. An accent position estimation rule storage unit 104 stores a learning result by the learning unit 103. An estimation unit 105 selects a rule stored in the accent position estimation rule storage unit 104 based on the information set by the language information setting unit 102 for an unregistered word when estimating the accent position. Estimate the accent position.
[0021]
FIG. 2 is a flowchart showing a learning procedure of an unregistered word accent position estimation rule in the spoken language processing apparatus having the configuration of FIG. First, in language information setting step S201, word language information is set. Next, in the vector expression generation step S202, data corresponding to the language information set in the language information setting step S201 is extracted from the learning data storage unit 101 to generate a vector expression. That is, a vector in which a component is provided for each content of language information, and a vector in which the component of the vector is 1 or 0 is created according to the presence or absence of each information content. Then, a numerical value indicating the accent position is set as the final component of the vector. Subsequently, in the identification learning step S203, the vector created in the vector expression generation step S202 for each language information is learned by an identification learning algorithm that maximizes the distance between the identification boundary surface and data close to the identification boundary surface. Create an accent location estimation rule.
[0022]
An example of learning data in the case of Japanese stored in the learning data storage unit is shown in FIG. One line means data related to one word. As linguistic information of a word, the phoneme of each mora as a partial character string constructed by dividing the reading of the word into mora units, the part of speech and the number of mora of the word, the number between the partial character strings, that is, the number of unity Is set. These numbers are “1”, “0”, and “1” from the left for the third row of Tatsuro. This is, for example, a group that exists when Tatsuro is “Tatsuro”, and “Tatsu” and “Taro” are grouped between “T” and “T”, so “1”, “T” and “T” Since there is no unity between “O”, “1” because there is a unity between “RO” and “U”. In the right end, the accent position of the word, that is, the location of the partial character string to which the accent is added is stored. In the example of the icon, it is indicated that the first “a” has an accent position.
[0023]
From this, attribute data used for actual learning of each word is extracted for each group of words for which an accent position estimation rule is created. Here, the word group means the number of pronunciation units, that is, the combination of the number of partial character strings and the part of speech. For example, when a proper noun of 3 mora is grouped to create an accent position estimation rule, the data is as shown in FIG. FIG. 4 shows data in the case of Japanese with 3 component elements, and shows an example in which input data is formed and changed only for part of speech, partial character strings, and the total number for learning. Corresponding pronunciation types and accent positions for the number of partial character strings are used.
[0024]
In some cases, information on the presence / absence of unity between partial character strings corresponding to (number of partial character strings−1) is associated. In this case, information on the presence / absence of grouping between (number of partial character strings−1) partial character strings is indicated by a (partial character string−1) -dimensional vector quantity, and there is a grouping in each dimension element. Describes 1 and 0 if not.
In the language information setting step S201, the part of speech and the number of partial character strings are set so that they can be extracted.
[0025]
In the vector expression generation step S202, in the case of this learning data, all of the three mora positions including a, i, c,..., K, kyu, kyo,. A vector of dimensions indicating whether or not there is a mora (the number of partial character strings 3 × the number of types of all mora) is created. In the case of the first case (Ico) in FIG. 4, the component corresponding to “a” of the first mora is 1, the component corresponding to “i” of the second mora is 1, and “ A vector in which the component corresponding to “co” is set to 1 and all other values are set to 0 is created in the vector expression generation step S202 of FIG. That is, the first mora eye “A” = 1, the first mora eye “I” = 0,..., The first mora eye “N” = 0, the second mora eye “A” = 0, the second mora Eye “I” = 1, second mora eye “U” = 0,..., Second mora eye “n” = 0, third mora eye “a” = 0,. A vector is generated such that “K” = 0, the third mora eye “co” = 1, the third mora eye “sa” = 0,..., The third mora eye “n” = 0.
[0026]
Such a vector and a number indicating where the accent is given, that is, the numerical value at the right end of FIG. 4, are combined to obtain learning data in the identification learning step S203.
As described above, an accent position estimation rule can be learned for each combination of the number of mora and the part of speech based on a large number of created vector data.
The above example of FIG. 4 has only three components, but any number can be stored in the learning data as shown in FIG. For those with 4 or more components, provide a field with 3 or more partial character strings, and change the number of partial character strings to the total number of partial character strings. For those with less than 3, reduce the number of partial character strings. This can be realized by changing the number of partial character strings to the total number of partial character strings.
[0027]
Numeric information representing a group of partial character strings in FIG. 3 can be used in addition to the information in FIG. In other words, information on the presence / absence of unity between (partial character string-1) partial character strings may be used, and a vector x having (partial character string-1) many dimensions may be used. This makes it possible to correctly learn accents of homophones that have different accents due to differences in unity.
In the identification learning step S203, learning is performed using a learning data vector created in the vector expression generation step S202, using an identification learning algorithm that maximizes the distance between the identification boundary surface and data close to the identification boundary surface.
[0028]
FIG. 5 is a flowchart showing a procedure for estimating an accent position of an unregistered word in the spoken language processing apparatus having the configuration of FIG. First, in the language information setting step S501, language information of unregistered words that require estimation of accent positions is set. Next, in the vector expression generation step S502, the linguistic information of the word is converted into a vector expression in the same manner as in the previous vector expression generation step S202. In the identification step S503, the accent position estimation rule acquired in the previous learning is selected based on the language information set in the language information setting step S501, and the accent position is estimated.
[0029]
Information used as an input for estimating the accent position of an unregistered word corresponds to information other than the accent position at the right end in one line of FIG. That is, as long as it is the same as the language information used for learning, not only FIG. 4 but also information on how to assemble partial character strings and other various language information can be set as input information at the time of estimation.
Learning unit 103, accent position estimation rule storage unit 104, estimation unit 105, identification learning step S203, identification step S503, that is, identification learning algorithms and learning results in learning unit 103, accent position estimation rule storage unit 104, and estimation unit 105 When SVM is used as the accent position estimation rule, the method shown in Reference Document 1 below can be used as the learning method. Select the accent position estimation rule of the learning result using the total number of partial strings of unknown words and the part of speech as a trigger, and a vector that represents the partial strings of unknown words and, in some cases, a group of partial strings, as input to it Is also entered to determine the accent type. SVM is a binary classifier. For example, a 3 mora word can take 4 accent types, so it needs to be identified at least 3 times. That is, when all the accent types are 0 type, 1 type, 2 type, 3 type,
Identification of type 0 or not,
Identification of type 1 or other than type 0,
Of type other than 0 type 1 type, identification of type 2 or other type 3
The accent type is determined by performing binary classification at least three times. Vector X indicating input information for identification learning example identification on page 129 of reference 1_iCorresponds to the portion of the vector created in S202 of the present invention excluding the position of the accent. Y indicating the two classes to be identified in the identification learning example on page 129 of Reference Document 1._iAccent type corresponds to. In each binary classification described above, Y_iIs associated with one of the binary values as y = {− 1, 1} on page 129 of Reference Document 1.
[0030]
A lot of such data is prepared, and the identification boundary surface is learned from them. That is, Expression 5.17 indicating the objective function for obtaining the identification boundary surface on page 130 of Reference Document 1 using these data is expressed by Expression 5.17 indicating the first constraint condition when solving this objective function. 18 and solve the mathematical programming problem of maximizing under the constraint of Equation 5.19 which shows the second constraint. As a result, a coefficient vector α whose elements are coefficients associated with each vector of the learning examples described on page 131 of the reference 1.₀ Is obtained. This component corresponds to each of the cases used for learning. A case corresponding to a non-zero coefficient among these is a support vector, which is a vector of a plurality of learning cases that are close to the boundary surface as indicated by a thick circle and a bold x in the figure on page 128. As a result, an identification boundary surface is formed.
[0031]
Further, in the estimation of the accent position of the unknown word, the vector generated in S502 describing the information of the unknown word is set to X on page 131 of the reference document 1, and the binary classification is performed at least once. This is estimated by using Equation 5.20 which means a function for discrimination determination using the discrimination boundary described on page 131 of 1.
As described above, the support vector represents information necessary for accent position estimation in the present invention.
In addition, in the learning of the rule for estimating the accent position of the previous 3 mora word, the classification other than type 1 is further separated from 2 and 2, and the classification other than 2 is further separated from 3 and 3, in order. By defining the face, the accent position estimation rule information can be obtained.
[0032]
<Reference Document 1> Vladimir N. Vapnik, The Nature of Statiistic Learning Theory, Springer (1995), p.129-p131,5.5 Constructing the Optimal Hyperplane.
Learning unit 103, accent position estimation rule storage unit 104, estimation unit 105, identification learning step S203, identification step S503, that is, identification learning algorithms and learning results in learning unit 103, accent position estimation rule storage unit 104, and estimation unit 105 When a probability model estimated by the maximum entropy method is used as the accent position estimation rule, the method shown in Reference Document 2 below can be used as the learning method. The part excluding the numerical value indicating the position of the accent in the vector created in the vector expression generation step S202 of the present invention corresponds to h on page 36 of Reference Document 2, and w represents the accent position to be estimated. The numerical value shown corresponds.
[0033]
The w-type accent type estimation rule is P (w | h). w takes a value from 0 to N. These probability values are provided for each total number of partial character strings constituting the part of speech of the word and the pronunciation of the word.
When estimating the accent type of an unknown word, the vector created in S502 with the same information set during learning of the accent type estimation rule is set to h. Using h as a trigger, an accent type having the maximum probability value in w is assumed to be an estimation result.
<Reference 2> Ronald Rosenfeld, "Adaptive Statistical Language Modeling, p.34-p.37", A Maximum Entropy Approach ", Ph.D. thesis, Computer Science Department, Carnegie Mellon University, TRCMU-CS-94-138 April (1994).
[0034]
The learning unit 103, the accent position estimation rule storage unit 104, the estimation unit 105, the identification learning step S203, the identification step S503, that is, the identification learning algorithm and the learning result of the learning unit 103, the accent position estimation rule storage unit 104, and the estimation unit 105 When using a mixture of decision trees, decision lists, neural networks, and linear discriminators learned based on learning case selection algorithms for the purpose of enhancing accuracy represented by bagging and boosting as accent location estimation rules The methods shown in Reference Document 3 and Reference Document 4 below can be used.
[0035]
X1 and Xm of Reference Document 3 and Xn of Reference Document 4 are partial character strings of words, and in some cases, those including a vector representing a unit between partial character strings, and an identifier when Y is an accent type An accent estimation rule is expressed by. When the discriminator is a decision tree, a decision is made with reference to each of X1, Xm of reference document 3 and Xn of reference document 4, and the accent type is determined when the leaf is reached. In the case of a neural network, each of the neural networks that receives X1 and Xm of Reference Document 3 and Xn of Reference Document 4 and obtains an accent type as an output result of the output layer corresponds to a rule. Y is estimated using X1 and Xm of Reference Document 3 and Xn of Reference Document 4 as triggers.
[0036]
X1 and Xm indicating input data in identification of training data in FIG. 1 of Reference Document 3 correspond to portions of the vector created in the vector expression generation step S202 of the present invention excluding the accent position. The binary classification is the same as in the case of the SVM. However, this identification algorithm can handle multi-level classification of two or more values depending on the learning device used.
Xn in Chapter 1 on page 1 and page 2 of Reference Document 4 corresponds to a portion of the vector created in the vector expression generation step S202 of the present invention excluding the accent position. The binary classification is the same as in the case of the SVM. However, this identification algorithm can handle multi-level classification of two or more values depending on the learning device used.
[0037]
<Reference 3> Joab Freund et al., “Introduction to Boosting, pages 11, 12”, Journal of the Japanese Society for Artificial Intelligence, September issue (1999).
<Reference Document 4> Leo Breiman, "Bagging Predictors," Machine Learning, 24, 2 (1996), p. 123-140.
[0038]
Although the above embodiments are directed to Japanese, the present invention can be applied to languages other than Japanese. FIG. 6 is a diagram showing the input and learning data of the embodiment when the present invention is applied to English having three components. This is a substring consisting of parts of speech of linguistic and ignorance, the number of substrings, and phonetic symbols. Originally, it should be described with phonetic symbols, but ASCII characters were substituted here. In FIG. 6, the part of speech of the English word, the partial character string created by dividing the pronunciation symbol string constituting the word into syllables, and the total number of the partial character strings are set.
[0039]
The experimental result about the accent position estimation of a Japanese 3 mora person name (Taro, Hanako, etc.) is shown. Five sets of experimental data were prepared using a 2156-word mora / accent position pair as learning data and a 539-word mora not included in the test set. It was evaluated by the average of the correct answer rate. The structure of the data used for learning is the same as in FIG.
As a result, the average accuracy rate of the prior art based on the decision tree was 88.22%, but the average accuracy rate in the example using the SVM was 89.80%, which is an improvement of 1.58 points. . This confirmed the effectiveness of using an identification learning algorithm that maximizes the distance between the identification boundary surface and data close to the identification boundary surface.
[0040]
The present invention can be applied to either a system composed of a plurality of devices or an apparatus composed of a single device. A storage medium storing software program codes for realizing the functions of the above-described embodiments is supplied to a system or apparatus, and an electronic computer (CPU, MPU), a function expansion board, and a dedicated circuit of the system or apparatus are provided. It can also be implemented by reading and executing the program code stored in the storage medium. In this case, the program code itself read from the storage medium realizes the functions of the above-described embodiment, and the present invention can be configured by a storage medium that records the program code.
[0041]
For example, a floppy (registered trademark) disk, hard disk, optical disk, magneto-optical disk, CD-ROM, CD-R, DVD, magnetic tape, nonvolatile memory card, or ROM is used as a storage medium for supplying the program code. be able to.
The functions of the above-described embodiment are not realized by executing the program code read by the computer, and the OS running on the computer is actually processed based on the contents of the program code. The function of the above-described embodiment can be realized by performing part or all of the above-described processing.
[0042]
【The invention's effect】
As described above, the present invention includes 1) a configuration capable of giving all accent types assumed to words that need accent type estimation, and 2) theoretically safer identification. 3), and 3) using the information representing the grouping between the partial character strings constituting the pronunciation, it is possible to accurately estimate the accent position for the unregistered word.
Further, the present invention has the effect that 4) the accent position can be estimated regardless of the language as long as the language is the same as the data used for learning by not assuming the information depending on the language.
[Brief description of the drawings]
FIG. 1 is a block diagram illustrating an embodiment.
FIG. 2 is a flowchart showing a learning procedure.
FIG. 3 is a diagram showing a configuration of Japanese learning data.
FIG. 4 is a diagram showing the structure of Japanese data used during learning and estimation.
FIG. 5 is a flowchart showing a procedure for estimating an accent position of an unregistered word.
FIG. 6 is a diagram showing the structure of English data used during learning and estimation.
[Explanation of symbols]
101 learning data storage unit 102 language information setting unit
103 Learning unit 104 Accent position estimation rule storage unit
105 Estimator

Claims (6)

An accent position estimation rule storage unit for storing an accent position estimation rule for estimating an accent position by inputting language information about a word;
A language information setting section for setting language information relating to unregistered words;
Selecting an accent position estimation rule stored in the accent position estimation rule storage unit according to the language information set in the language information setting unit, and estimating an accent position of the unregistered word, and
The language information includes the part of speech of the word, all the partial character strings constituting the pronunciation of the word, the total number of the partial character strings, and the presence / absence of coupling between the partial character strings (the number of partial character strings-1). An accent position estimation apparatus characterized by including all of the information . The accent position estimation apparatus according to claim 1 ,
An accent position estimation apparatus using a partial character string created by dividing a pronunciation of a Japanese word into mora units as all the partial character strings constituting the pronunciation of the word. In the accent position estimation apparatus according to claim 1 or 2 ,
A learning data storage unit that stores a set of linguistic information about the word and the accent position of the word;
The language information setting unit also sets language information regarding a learning word,
Based on learning data selected from the learning data storage unit according to the learning language information set by the language information setting unit, learn accent position estimation rules for estimating accent positions using language information about words as input. An accent position estimation apparatus, further comprising: a learning unit that performs the learning. Language information setting process for setting language information about unregistered words,
Selecting an accent position estimation rule according to the language information set in the language information setting process, and estimating an accent position of the unregistered word,
The language information includes the part of speech of the word, all the partial character strings constituting the pronunciation of the word, the total number of the partial character strings, and the presence / absence of coupling between the partial character strings (the number of partial character strings-1). The accent position estimation method characterized by including all with information . In the accent position estimation method according to claim 4 ,
An accent position estimation method using a partial character string created by dividing the pronunciation of a Japanese word into mora units as all the partial character strings constituting the pronunciation of the word. The accent position estimation program for functioning a computer as each part of the accent position estimation apparatus in any one of Claims 1 thru | or 3 .

Images