JP2000047681A - Information processing method - Google Patents

Abstract

(57) [Summary] [Problem] When selecting a feature value corresponding to a case, even when a correct feature value is not selected, a feature value closer to a correct answer can be selected. An information processing method for selecting a feature amount related to a case from a plurality of options in accordance with states of a plurality of attributes related to a case, wherein an estimated value y _k of a feature amount corresponding to a k-th option is set to a j-th option. function w _kj attribute values d _j which is determined by the state of the attribute (d _j) and with the constant w _k0, Equation 1] And selecting a feature value related to the case from the plurality of options based on the estimated value of the feature value.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

[0001] 1. Field of the Invention [0002] The present invention relates to a system for selecting a feature amount corresponding to a case (or classifying a case) from attribute information.

[0002]

2. Description of the Related Art Various studies have been made on the problem of determining which class a case belongs to in the field of artificial intelligence and the like. For example, the document "Data analysis by AI, JR Kinlan, Toppan Co., Ltd."
A method for classifying cases using a decision tree is disclosed.
The classifier based on the decision tree is described by the IF-THEN rules regarding attributes.

For example, FIG. 1 shows an example of a classifier that determines whether a given case belongs to a class of “held” or “cancelled”, with attributes as weather and temperature. Such a decision tree is learned so that a large number of cases whose attribute values and correct classes are already known are used as training cases so as to increase the accuracy of classification for the training cases.

[0004]

In the problem of classifying cases, when the number of classes is three or more, not only the accuracy rate of classification but also how to make an error may be a problem. As an example, a pitch pattern generation process in text-to-speech synthesis that converts text to speech will be described. The pitch pattern generation process is a process of generating a pitch pattern, which is a time-varying pattern of voice pitch, from information such as part of speech, accent type, and number of syllables obtained by analyzing a text. As a method of generating a pitch pattern, prepare a plurality of representative patterns, which are typical pitch patterns in units of accent phrases, and select a representative pattern with attributes such as the part of speech, accent type, and the number of syllables for each accent phrase,
There is a method of generating a pitch pattern using this. In the selection of the representative pattern, it is important that not only the selectivity of the optimal pattern is high, but also that a pattern deviating from the optimal pattern is hard to be selected.

[0005] Assuming that the representative patterns are the four patterns shown in FIG. 2, if the pattern shown in FIG. 2A is the optimum pattern, the pattern shown in FIG. It doesn't matter much,
If the pattern shown in FIG. 2C is selected by mistake, the accent sounds completely different, which is a serious problem.

When a conventional classification method using a decision tree is used for representative pattern selection, learning is performed on the basis of increasing the selectivity of the optimal pattern using a set of an attribute and an optimal pattern number as a training example. For this reason, there is a problem that it is inevitable to select a pattern that is far apart when the optimal pattern is not selected.

Therefore, the present invention has been made in consideration of the above problems, and when selecting a feature amount corresponding to a case (or classifying a case), each option is selected. By estimating the evaluation value based on the attribute information of the case and selecting the feature based on the estimated value, even if the correct feature is not selected, the feature closer to the correct answer can be selected. Another object of the present invention is to provide an information processing method capable of performing the above.

[0008]

According to the information processing method of the present invention, a plurality of attributes related to a certain case (for example, the accent type of accent phrase in text-to-speech synthesis, the number of mora, the destination of the accent phrase, and the according to the state of dependency destination), an information processing method for selecting the feature quantity relating to the case (for example, the representative patterns in text-to-speech synthesis) from a plurality of choices, the estimated value y _k of the feature quantity corresponding to the k-th choice using j th attribute of state function w _kj attribute values d _j which is determined by the (d _j) and constant w _k0,

[0009]

(Equation 2) By selecting a feature amount related to the case from the plurality of options based on the estimated value of the feature amount, when selecting a feature amount corresponding to the case, each option is selected. Since the evaluation value at the time of selection is estimated based on the attribute information and the feature amount is selected based on the estimated value, even if the correct feature amount is not selected, a feature amount closer to the correct answer can be selected.

Here, the attribute is the type of information that can be known when classifying cases with respect to the case. For example, if the case is “human”, various information such as gender, occupation, height, weight, and age can be considered as attributes.

The attribute value is a numerical value indicating the state of this attribute. If the attribute has an order relationship (so-called order scale) such as grades (excellent, good, acceptable, unacceptable) or clothes sizes (LL, L, M, S), the order is followed. An appropriate value can be used as the attribute value. For example, in the case of a grade, excellent: 4, good: 3, acceptable: 2, unacceptable: 1, etc. When the attribute is measured as a numerical value (so-called interval scale or ratio scale) such as height, weight, age, and Celsius temperature, the measured value can be used as the attribute value. Alternatively, a representative value obtained by quantizing the measured value may be treated in the same manner as the ordinal scale. If the attribute has no order relation such as gender, occupation, color, etc. (so-called classification scale), an attribute value can be set by associating an arbitrary value with the category of these classifications. . For example, if it is a color, red: 1, blue: 2, yellow: 3, etc. may be used.

Attribute value d _j for at least one attribute
_{Takes a} finite number (N) of values (d _j1 , d _j2 ,..., D _jN ), preferably, the coefficients (a _kj1 , a _kj2 ,..., A _kjN ) corresponding to each attribute value , The function w _kj (d _j ) of the attribute value corresponding to the k-th option may be represented by w _kj (d _jm ) = a _kjm .

[0013] or an attribute of the state is unknown, the attribute is in such a state that does not assume, if the attribute value is unknown, the function of unknown attribute value d _j w _kj (d _j) is constant c it may be represented by _{w kj (d j) = c} kj with _kj.

[0014] Preferably, the option may be an option of a feature parameter relating to text-to-speech synthesis for converting text to speech. Preferably, the option of the characteristic parameter may be an option of a characteristic parameter of a pitch pattern control model.

Preferably, the pitch pattern control model is such that one pattern is selected from a plurality of representative patterns, and a pattern obtained by subjecting the selected pattern to deformation is used as a pitch pattern. You may.

Preferably, the deformation includes at least a parallel movement on a logarithmic frequency axis. Preferably, the attribute may include the number of mora of the accent phrase, the destination of the accent phrase, and the destination of the preceding accent phrase.

[0017]

Embodiments of the present invention will be described below with reference to the drawings. Whether a person likes ramen, sushi, or curry and rice the most,
The problem estimated from gender and hometown is explained as an example. Here, it is assumed that a questionnaire result as shown in FIG. 3 is used as a training example. FIG. 4 shows an example of a training case (questionnaire result).

In this embodiment, the attribute (j) is the age, gender, and home region, and the feature quantity (k) to estimate how much each of ramen, sushi, and curry and rice is like. From the attributes shown in Fig. 4, such as age, gender, and hometown, an estimated value indicating how much each of the noodles, sushi, and curry and rice is desired is determined, and the food with the highest estimated value of preference is selected to determine which food is best. Estimate what you like best.

As the attribute value of each attribute, the number of the option of the questionnaire is used as it is, and the attribute values of the age, gender and hometown are represented by d ₁ , d ₂ and d ₃ , respectively. Using a feature w (k = 1, 2, 3), an attribute value function w _kj defined for each attribute (j = 1, 2, 3), and a constant w _k0 defined for each feature, ramen , Sushi, curry and rice are modeled as follows in terms of the estimated values y ₁ , y ₂ , and y ₃ .

[0020]

(Equation 3) Here, the function w _kj of the attribute value is defined as follows.

[0021]

(Equation 4)

Next, a method of determining the parameters of the estimation model using training examples will be described. Evaluation function E representing the sum of squares of the estimation error of the preference of ramen, sushi and curry rice for all training cases
₁ , E ₂ and E ₃ are defined as follows.

[0023]

(Equation 5)

Where M is the number of training cases, y
₁ ^i, y ₂ ^i, y ₃ ⁱ is the i-th training cases respectively noodles represent sushi, desirability value of curry and rice (the number of choices questionnaire), d _j ⁱ is the i-th training case Represents the j-th attribute value of. A parameter that minimizes this evaluation function may be searched for. The search for the parameter may use some known optimization technique.

For example, a ₁₁₁ = −0.2, a ₁₁₂ = −
0.1, a ₁₁₃ = 0.3, a ₁₁₄ = 0.5, a ₁₁₅ =
Assuming that a coefficient of 0.4 is obtained, the function W _k1 (d ₁ ) is represented as shown in FIG. 5 using the variable d ₁ as the age. Further, instead of the function of Figure 5, the variable d ₁ as age, it may be used functions like Figure 6 obtained by smoothly interpolating function of Figure 5.

In order to estimate the most preferable food (characteristic amount) using the obtained model, the attribute values d ₁ , d
₂ , d ₃ are obtained, and the obtained values are substituted into the equation (1) to obtain estimated values y ₁ , y ₂ , y ₃ of the preference corresponding to each food,
Food the estimated value becomes maximum is estimated to be the most favorite food (if the maximum of the estimated value is y ₁ Ramen, if y ₂ sushi, curry rice if y _3).

Further, using this model, all the favorite foods can be estimated and selected from ramen, sushi, and curry and rice. For example, if the threshold value is set to “3.5” and the estimated value y _k of the preference is 3.5 or more, it may be estimated that the user likes the food. If the estimated value of the preference is y ₁ = 2.5, y ₂ = 4.2 and y ₃ = 3.8, the favorite foods are estimated to be sushi and curry rice.

If the information given as a training example is not information about the preference of each food but only information about which one likes the most, the value of the preference for the most favorite food is set to “5”, and Is set to “1”, the parameters of the model can be obtained and estimated in the same manner as in the above method. For example, if your favorite food is sushi, the estimated value of the noodles in the i-th training case y ₁ ⁱ = _1, the estimated value of sushi y ₂ ⁱ = 5, the estimated value of curry rice y ₃ ⁱ = 1 Becomes

In order to estimate a case where there is an unknown attribute value, the value of the function when _dj is unknown is given by w _kj (d _j ) = c _kj (where _dj is unknown). (6) is defined. c _kj is, for example, a _kjm (m = 1, 2,...)
It may be a perfectly equal value. Alternatively, a weighted average value obtained by weighting with the appearance frequency of the attribute value may be used.

The constant w _k0 may always be set to “0” so that the optimization is not performed. Further, a combination of two or more attributes may be newly defined as one attribute. For example, it is also possible to combine two attributes of “sex” and “hometown” into one attribute, and to assign an attribute value as shown in FIG.

FIG. 8 is a block diagram showing the configuration of an information processing apparatus for realizing the information processing method of the present invention. For example, it realizes a text-to-speech synthesis process for converting an input text 101 into a synthesized speech 110. is there.

The linguistic processing unit 10 performs linguistic processing on the text 101, and the processing result, that is, the accent type 102, the number of moras 103, and the destination 1 of the accent phrase
04, destination 105 of leading accent phrase, phoneme symbol string 1
06 is output as output data for each accent phrase.

For example, in the case of the text "everything has been twisted toward oneself", the accent phrase is "everything", "everything", "all", "oneself", "one", "twisted". The accent type 102, the number of moras 103, the destination 104 of the accent phrase, and the destination 105 of the preceding accent phrase for each of these accent phrases are output.

The prosody generation unit 11 includes an accent type 102,
A pitch pattern, which is a time-change pattern of the fundamental frequency, and a phonological duration are generated from the number of moras 103, the relevant part 104 of the accent phrase, the relevant part 105 of the preceding accent phrase, and the phoneme symbol string 106, and the output data is generated. 1
07 and 108 are output.

The voice signal generation unit 12 generates the phoneme symbol string 10
6. A synthesized speech 110 is generated from the pitch pattern 107 and the phoneme duration 108. Next, the prosody generation unit 11
Will be described with reference to FIG.

The representative pattern selection unit 21 is a language processing unit 1
Accent type 102 output from 0, number of mora 10
3. The representative pattern suitable for the accent phrase is selected from the relevant phrase 104 of the accent phrase and the relevant destination 105 of the preceding accent phrase, and the representative pattern number 201 is output.

The representative pattern storage unit 22 stores a plurality of representative patterns in advance corresponding to the representative pattern numbers. Among them, the representative pattern 20 corresponding to the representative pattern number 201 output from the representative pattern selection unit 21
Read out 3 and output.

The representative pattern is obtained by normalizing the time axis of a typical time-change pattern of the fundamental frequency of a voice in units of mora as shown in FIG. The movement amount generation unit 20 obtains a movement amount 202 when the representative pattern 203 is moved in parallel in the logarithmic frequency axis direction from the input number of moras 103, the destination 104 of the accent phrase, and the destination 105 of the preceding accent phrase. Output.

The phoneme duration generation unit 23 calculates and outputs a phoneme duration 108 of each phoneme according to the phoneme symbol string 106. The pitch pattern generation unit 24 expands and contracts the representative pattern 203 in the time direction according to the phoneme duration 108 so that the length of each mora is equal to the phoneme duration, and moves in parallel on the logarithmic frequency axis according to the movement 202. Then, the pitch pattern 107 is output. FIG. 11 shows an example in which a pitch pattern is generated for an accent phrase of “any”.

FIG. 11A shows the representative pattern 203, FIG. 11B shows the pattern obtained by expanding and contracting in the time direction, and FIG. 11C shows the pattern obtained by moving the pattern in parallel on the logarithmic frequency axis. FIG. However, the vertical axis in FIG. 11 represents the logarithmic frequency, and the moving amount 202 is 4.5.

Next, the detailed operation of the representative pattern selecting section 21 will be described. Here, the accent type 102, the number of moras 103, the destination 104 of the accent phrase, and the destination 105 of the preceding accent phrase are attributes of the accent phrase, and the representative pattern is a feature amount. First, an evaluation value for each representative pattern is estimated from the attribute value, and the number of the representative pattern with the smallest estimated value is output as a representative pattern number 201.

Here, the estimated value of each representative pattern is a pitch pattern generated using each representative pattern,
It represents the distance from an ideal pitch pattern for a combination of attribute states.

A method for obtaining an estimated distance value will be described.
Attribute values for each attribute state are defined as shown in FIG. Estimated distance y _k for the k-th representative pattern
To the attribute value functions w _kj (d _j ), (k = 1, 2 _,.
K) is determined by the following equation using (j = 1, 2, 3, 4) and a constant w _ko .

[0044]

(Equation 6)

The functions w _kj (d _j ), (k = 1, 2,...,
K) (j = 1, 2, 3, 4) is a coefficient a _kjm (k = 1,
, K) (j = 1, 2, 3, 4) (m = 0, 1,
2,...) And c _kj are defined as follows.

[0046]

(Equation 7)

Here, the fact that d _j is unknown means that the state of the attribute cannot be known due to a failure in language processing or the like. Instead of equation (8), a polynomial such as the following equation may be used.

W _kj (d _j ) = b _kj2 d _j ² + b _kj1 d _j + b _kj0 (9) w _{k0 of the} equation (7) and the coefficient a _{kjm of the} equation (8) or the coefficient b _kjm of the equation (9) are , The error of the estimated distance value for the training case is determined to be minimum.

The training examples are data on the distance between the pitch pattern extracted from the actual voice and each representative pattern, and the attributes of the corresponding text (accent type, number of mora, the relevant part of the accent phrase). , A combination of data of the preceding account). The training case is
It is obtained by analyzing a large amount of text and the data of the actual voice that reads it.

The optimization of the coefficient is performed by, for example,
This can be performed by using a known optimization technique using the root mean square error as an evaluation function. Further, the coefficient c _kj in the equation (8) is a
The average value of _kj0 , _akj1, ... or a weighted average value weighted in consideration of the appearance frequency of the attribute value may be used.

In this embodiment, the attributes of the text are the accent type, the number of moras, the destination of the accent phrase, and the destination of the preceding accent phrase. Various information, such as the position of the character and the type of phoneme, can be used as attributes.

[0052]

As described above, according to the present invention,
When selecting a feature amount corresponding to a case, an evaluation value when each option is selected is estimated based on attribute information, and a feature amount is selected based on the estimated value. Even in the case where this is not done, a feature value closer to the correct answer can be selected.

[Brief description of the drawings]

FIG. 1 is a diagram showing an example of a classifier using a conventional decision tree.

FIG. 2 is a diagram showing a specific example of a representative pattern in pitch pattern generation processing.

FIG. 3 is a diagram showing a specific example of a questionnaire for collecting training cases.

FIG. 4 is a diagram showing a specific example of a training case.

FIG. 5 is a diagram showing a specific example of an attribute value function w _k1 (d ₁ ).

FIG. 6 is a diagram showing a specific example of a function w _k1 (d ₁ ) of an attribute value, which is obtained by smoothly interpolating the function of FIG. 5;

FIG. 7 is a diagram showing an example of a table in which a combination of two attributes is set as one attribute and attribute values are determined.

FIG. 8 is a diagram showing an example of the configuration of an information processing apparatus for executing a text-to-speech synthesis process according to an embodiment of the present invention.

FIG. 9 is a diagram showing a configuration example of a prosody generation unit.

FIG. 10 is a diagram schematically illustrating a representative pattern stored in a representative pattern storage unit.

FIG. 11 is a view for explaining a process of generating a pitch pattern from a representative pattern selected from a plurality of options.

FIG. 12 is a diagram showing an example of a table in which attribute values for attribute states are determined.

[Explanation of symbols]

 DESCRIPTION OF SYMBOLS 10 ... Language processing part 11 ... Prosody generation part 12 ... Speech signal generation part 20 ... Moving amount generation part 21 ... Representative pattern selection part 22 ... Representative pattern storage part 23 ... Phonological duration length generation part 24 ... Pitch pattern generation part

 ──────────────────────────────────────────────────続 き Continued on the front page (72) Inventor Masahiro Morita 8-6-26 Motoyama Minamicho, Higashinada-ku, Kobe-shi, Hyogo F-term in Toshiba Kansai Research Institute Co., Ltd. 5D045 AA09

Claims (8)

[Claims] 1. An information processing method for selecting a feature amount related to a case from a plurality of options according to a state of a plurality of attributes related to a case, wherein an estimated value y _k of a feature amount corresponding to a k-th option is j Function w _{kj of} attribute value d _j determined by the state of the th attribute
Using (d _j ) and a constant w _k0 , And selecting a feature amount related to the case from the plurality of options based on the estimated value of the feature amount. 2. An attribute value d for at least one attribute.
_j takes a finite number of values (d _j1 , d _j2 ..., d _jN ), and coefficients (a _kjl , a _kj2,.
2. The information processing method according to claim 1, wherein a function w _kj (d _j ) of the attribute value corresponding to the k-th option is represented by w _kj (d _jm ) = a _{kjm using} a _kjN ). . 3. If the attribute value d _j is unknown, the function w _kj (d _j ) of the attribute value corresponding to the k-th option is a constant C
_2. The information processing method according to claim 1, wherein w _kj (d _j ) = c _kj using _kj . 4. The information processing method according to claim 1, wherein the option is an option of a feature parameter relating to text-to-speech synthesis for converting text into speech. 5. The information processing method according to claim 1, wherein the option is an option of a special feature parameter of a pitch pattern control model relating to text-to-speech synthesis that converts text to speech. 6. The option is an option of a special pattern parameter of a pitch pattern control model relating to text-to-speech synthesis for converting text into speech, and the pitch pattern control model selects one pattern from a plurality of representative patterns. 2. The information processing method according to claim 1, wherein a pattern obtained by deforming the selected pattern is used as a pitch pattern. 7. The option is an option of a special pattern parameter of a pitch pattern control model relating to text-to-speech synthesis for converting text to speech, and the pitch pattern control model selects one pattern from a plurality of representative patterns. 2. The information processing method according to claim 1, wherein a pattern obtained by subjecting the selected pattern to a deformation including at least a parallel movement on a logarithmic frequency axis is used as a pitch pattern. 8. The method according to claim 1, wherein the attributes include at least the number of mora of an accent phrase, a destination of an accent phrase, and a destination of a preceding accent phrase in text-to-speech synthesis for converting text into speech. Information processing method.