JP3043625B2 - Word classification processing method, word classification processing device, and speech recognition device - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

The present invention relates to a word classification processing method and a word classification processing device for a speech recognition device, a morphological analysis device and a syntax analysis device, and a speech recognition device provided with the word classification processing device. .

[0002]

2. Description of the Related Art A word classification system is one of very important knowledges for smooth processing in a speech recognition device, a morphological analysis device and a syntax analysis device. As one method for constructing this word classification system, a method using a mutual information amount between words based on a large amount of text data (hereinafter, referred to as a first conventional example) is described in, for example, the related art document. “Peter Brown, et al,“ Class-Based n-gram Models
of Natural Language ", Computational Linguistics, Vo
l. 18, No. 4, pp. 467-479, 1992 ". In the conventional method, a predetermined mutual information amount is calculated using an n-gram model to classify English words.

[0003] However, when English words are classified using the mutual information classification method of the first conventional example, there is a problem that words with low appearance frequency are often inappropriately classified. In order to solve this problem, a word classification processing device and a speech recognition device (hereinafter, referred to as a second conventional example) have been proposed by the present applicant in the patent application of Japanese Patent Application No. 7-056918.

[0004] The second conventional example of the word classification processing device,
A first classifying unit that classifies a plurality of words into a plurality of classes on the basis of assigning words having a high ratio adjacent to the same word to the same class using an n-gram of the word; For a plurality of words classified by the classification means, the frequencies of appearance of all the words are checked, and the frequencies of the first class words and the second class words, which are different from each other, appearing adjacent to each other are determined by the first class. The plurality of words are divided into a binary tree so that a predetermined mutual information representing the ratio of the relative frequency to the product of the frequency of appearance of the words of the class of the second class and the frequency of appearance of the words of the second class is maximized. And a second classifying means for classifying into a plurality of classes. Here, it is preferable that the second classifying unit examines the appearance frequency of all the words with respect to the plurality of words classified by the first classifying unit, and determines the predetermined frequency in order from the word having the highest appearance frequency. By assigning to a plurality of N classes and combining the two classes having the largest mutual information amount among the N classes into one class, (N−
1) Classify into classes, and among words not assigned to a class, assign a new word having the highest appearance frequency as the Nth class, and repeat the above until all words are assigned to N classes. Is repeated, and the two classes having the largest mutual information amount from the current class are combined into one class, and this process is repeated until the classes are combined. Thereby, the word classification can be performed more stably, and when automatically acquiring the word classification system from the text data, a more precise and accurate classification system can be obtained.

[0005]

When English words are classified using the mutual information amount classification method of the first conventional example, words with low appearance frequency are often inappropriately classified. It is considered that this is because the separation result does not have a balanced hierarchical structure.

Further, in the second conventional example, the frequency at which the first class word and the second class word which are different from each other appear adjacently is determined by the appearance frequency of the first class word. The plurality of words are classified into a plurality of classes in the form of a binary tree so that the predetermined mutual information representing the ratio of the relative frequency to the product of the frequency of appearance of the words of the second class is maximized. Therefore, for the words of the first class and the words of the second class, a locally optimized word classification result can be obtained, but a totally optimized word classification result can be obtained. There was a problem that it was not possible.

An object of the present invention is to solve the above problems, to provide a word classification processing method capable of obtaining a word classification result which has a balanced hierarchical structure by the word classification processing and is totally optimized. An object of the present invention is to provide a word classification processing device and a speech recognition device provided with the word classification processing device.

[0008]

According to a first aspect of the present invention, there is provided a word classification processing method comprising: examining text data including a plurality of words; Sort by words that appear frequently,
Assigning to a plurality v of classes;
Less than v words with high appearance frequency (c
+1) storing the words of the class as words of the class in one window in the first storage device, and based on the words of the class in one window stored in the first storage device, Relative to the product of the probability of occurrence of the first class word and the probability of occurrence of the second class word, the probability that the first class word and the second class word that are different from each other appear adjacent to each other The above-mentioned plurality of words are classified into a plurality of c classes in the form of a binary tree so that the predetermined average mutual information amount representing the maximum ratio is maximized. Storing in the second storage device a plurality of c classes of the intermediate layer of the entire tree diagram to be represented; and based on the words of the plurality of c classes of the intermediate layer stored in the second storage device, Above average mutual information is maximum Classifying the words of the plurality of c classes into one class in the form of a binary tree so as to store the classification result in a third storage device as an upper layer of the tree diagram; For each of the plurality c classes of the intermediate layer stored in the second storage device,
On the basis of a plurality of words in each of the plurality of c classes of the intermediate layer, the plurality of words are divided into one class in the form of a binary tree so that the average mutual information is maximized. Classifying each of the classes and storing a plurality of classification results for each class in a fourth storage device as a lower layer of the tree diagram; and a lower layer of the tree diagram stored in the fourth storage device. Is connected to the plurality of c classes of the intermediate layer stored in the second storage device, while the upper layer of the tree diagram stored in the third storage device is connected to the second storage device. The above-mentioned tree diagram including the upper layer, the intermediate layer, and the lower layer is obtained by linking with the plurality of c classes of the intermediate layer stored in the fifth storage device, and is stored in the fifth storage device as a word classification result. And a step.

According to a second aspect of the present invention, in the word classification processing method of the first aspect, the step of storing the plurality of classified c classes in the second storage device includes the step of: There is a class outside one window stored in the first storage device, or
When the number of classes in one window is not c, the words of the class that is outside the current window and has the highest frequency of occurrence are inserted into the window, and then the word classification processing in the form of the binary tree is performed. It is characterized by executing.

[0010] According to a third aspect of the present invention, in the word classification processing apparatus, for text data including a plurality of words, an appearance frequency of all a plurality of v words different from each other is checked.
First control means for sequentially arranging words having a high frequency of occurrence and assigning them to a plurality of v classes; and, among words of the plurality of v classes, words of less than v (c + 1) classes having a high frequency of occurrence, A second control unit for storing the words of the class in one window in the first storage device, and a first class based on the words of the class in one window stored in the first storage device. Represents the relative ratio of the probability that the first class word and the second class word that are different from each other appear adjacent to the product of the occurrence probability of the second word and the occurrence probability of the second class word The plurality of words are classified into a plurality c classes in the form of a binary tree such that the predetermined average mutual information amount is maximized, and the classified plurality c
Control means for storing the plurality of classes in the second storage device as a plurality of c classes in the middle layer of the entire tree diagram representing the word classification result, and the intermediate control means for storing the intermediate classes stored in the second storage device. Based on the words of the plurality c classes of the layer, the words of the plurality c classes are classified into one class in the form of a binary tree so that the average mutual information amount is maximized. Fourth control means for storing the classification result in the third storage device as an upper layer of the tree diagram, and for each of a plurality of c classes of the intermediate layer stored in the second storage device, Based on the plurality of words in each of the plurality of c classes in the intermediate layer, the plurality of words are each converted into a single tree in the form of a binary tree such that the average mutual information is maximized. And classify the results of each class into the above tree. Fifth control means for storing in the fourth storage device as a lower layer of the diagram, and a lower layer of the tree diagram stored in the fourth storage device being stored in the second storage device. While linking with the plurality of c classes of the intermediate layer, the upper layer of the tree diagram stored in the third storage device is combined with the plurality of c classes of the intermediate layer stored in the second storage device. A sixth control unit that obtains the tree diagram including the upper layer, the intermediate layer, and the lower layer by linking with the class, and stores the tree diagram as a word classification result in the fifth storage device. And

According to a fourth aspect of the present invention, there is provided a word classification processing apparatus according to the third aspect, wherein the third control means is configured to execute the processing by the third control means more than one window stored in the first storage device. An outer class exists, or 1
When the number of classes in one window is not c, the words of the class that is outside the current window and has the highest frequency of occurrence are inserted into the window, and then the word classification processing in the form of the binary tree is performed. It is characterized by executing.

According to a fifth aspect of the present invention, there is provided a speech recognition apparatus, wherein a plurality of words are classified into a plurality of classes by the word classification processing apparatus according to the third or fourth aspect based on an input speech signal. And a speech recognition unit for recognizing the uttered speech by referring to a word dictionary including the word categorized result and a predetermined hidden Markov model.

[0013]

Embodiments of the present invention will be described below with reference to the drawings. FIG. 1 is a block diagram of a speech recognition device according to a first embodiment of the present invention. This speech recognition apparatus classifies words having a relatively low frequency of appearance with respect to words in text data in the text data memory 10 on the basis of assigning words having a high percentage of adjacent words to the same class to the same class. The word classification results are classified into three layers: an intermediate layer, an upper layer, and a lower layer,
Using a predetermined average mutual information, which is a global (overall) cost function for all words in the text data, classify words by hierarchy in the order of middle layer, upper layer, and lower layer And a word classification processing unit 20 that stores the word as a word dictionary in the word dictionary memory 11.

<Word Classification Processing Method> First, the word classification (clustering) method of the present embodiment according to the present invention will be described in comparison with the method of the first conventional example. The method of the present invention is a method in which the method of the first conventional example disclosed in the prior art document is modified and greatly developed and improved. The method of the first conventional example and the method of the present embodiment are modified. The difference from the expression will be described, and then the word classification processing method will be described. Here, in order to compare the first conventional example with the present embodiment, the same notation as that used in the first conventional example will be used.

First, a clustering method using mutual information will be described. Here, a text with the number of words T,
It is assumed that a vocabulary having the number of words V and a vocabulary division function π exist, where the vocabulary division function π is a mapping function representing a division mapping from the vocabulary V to the word class set C in the vocabulary. is there. In the first conventional example, the likelihood L (π) of a bigram class model that generates text data composed of a plurality of words is obtained by the following equation.

[0016]

## EQU1 ## L (π) =-H + I

Here, H is the entropy of the word distribution of the monogram, and I is the average mutual information (Averag) regarding two adjacent classes C ₁ and C ₂ in the text data.
e Mutual Information; A
Notated as MI. ) And can be calculated by the following equation.

[0018]

(Equation 2)

Here, Pr (C ₁ ) is the occurrence probability of a word of the first class C ₁ , and Pr (C ₂ ) is the second class C ₂
Pr (C ₁ | C ₂ ) is the conditional probability that a word of the first class C ₁ will appear after a word of the second class C ₂ has appeared, and Pr (C ₁ | C ₂ ) C ₁ , C ₂ ) is the probability that a word of the _first class C _{1 and} a word of the second class C ₂ appear adjacent to each other. Therefore, AMI is different first mutually Class C ₁ word and the second represented by the number 2
The relative proportions and words of class C ₂ is the probability of occurrence adjacent, divided by the product of the above first occurrence probabilities of the words in the classes C ₁ and a second class of C ₂ probability of occurrence of words Represents

Since the entropy H is a value independent of the mapping function π, the mapping function that maximizes the AMI also maximizes the likelihood L (π) of the text. Therefore, AM
I can be used as an objective function in word class construction.

The first prior art clustering method using mutual information uses a bottom-up merge procedure from the lower layer to the upper layer. Initially, each word is assigned to a class. Next, a search is made for a pair of two classes that gives the smallest AMI reduction among all pairs of classes, and the pairs of the two classes are merged. The number of classes after the merge is determined in advance. The above merging process is repeated until the number reaches c. In the basic method of the first conventional example, for example, the complexity (or the cost of the calculation time) of the calculation time executed by the computer becomes V when the process is directly executed as shown below. ⁵ (the number of words in the vocabulary V to the fifth power), which is denoted as O (V ⁵ ).
Here, the complexity of the calculation time is an index indicating how long the calculation time takes.

<Step A1> The total number of merge processes is (Vc) times, and the complexity of the operation time at this time is of the order O (V) proportional to the number V of words in the vocabulary.
It is. <Step A2> After n merging processes, (V−
n) classes remain, and in the next merge processing stage, the number of combinations _Vn C ₂ (that is, 2
Test or trial of merge processing (number of combinations when taking one class) (hereinafter referred to as trial). Here, merge processing is performed multiple times, but actual merging is performed and reflected in the word classification result. This is one of them, so it is called in this embodiment.). Only one of them will be activated in a later merge process. Therefore, the complexity of the operation time at this time is of the order O (V ² ) proportional to the square V ² of the number of words in the vocabulary. <Step A3> The trial of one merge process in the n-th stage includes an addition operation for (V−n) ² terms or classes for calculating the AMI using the above equation ( ² ). Therefore, the complexity of the operation time at this time is of the order O (V ² ) proportional to the square V ² of the number of words in the vocabulary.

Accordingly, for example, the complexity of the total operation time executed by the computer is on the order O (V ⁵ ) in proportion to the fifth power V ^{5 of the} number of words. However, as will be described later, by eliminating redundant calculations, the complexity of the operation time can be reduced to three times the number of words in the vocabulary.
It is also possible to reduce the order to the order O (V ³ ) proportional to the power V ³ . That is, according to the following method of the present invention, the step A3 can be executed in a fixed time.

<Step B1> The above equation 2 is calculated only for the class whose value has been changed in the merge processing in the previous stage. Therefore, the complexity of the operation time is changed from the order O (V ² ) in the first conventional example to the order O (V ² ).
(V). <Step B2> Store the results of all trials in the previous merge process. Therefore, the complexity of the operation time is the order O in the first conventional example.
From (V), a constant order O (1) independent of the number V of words in the vocabulary is obtained.

For example, starting from a vocabulary of V classes, the merging process has already been performed (V−k) times, and k classes C _k (1), C _k (2),..., C _k ( Suppose k) remains. AMIIk at this stage is calculated by the following equation.

[0026]

(Equation 3)

## EQU4 ## q _k (l, m) = p _k (l, m) log [p
_{k (l, m) / {} pl k (l) pr k (m)}]

Where p _k (l, m) is the class C
_This is the probability that a word in class C _k (m) follows a word in _k (l), and is expressed as: In addition,
In this specification and the drawings, for clarity of indication,
l is also used as l (lowercase letter), and l = l.

[0028]

P _k (l, m) = Pr (C _k (l), C _k (m))

(Equation 6)

In Equation 3, q _k is added over the entire (k × k) class bigram plane table, where the (l, m) cell is represented by q _k (l, m). Now, when searching for a trial of a merge process between the class C _k (i) and the class C _k (j), the amount of decrease in the AMI due to the merge process is represented by L _k (i, j) ≡I _k −I
_k (i, j), where I _k (i, j) is the AMI after the merge processing.

FIG. 3 is a diagram of a class bigram plane table showing the addition area and the addition / subtraction processing in the word classification processing according to the present invention. Here, FIG. 3 and FIGS. 4 and 5 shown below show planes of two classes of bigrams.
As shown in FIG. 3, the addition area P0 in Equation 3 is
From the sum of the partial areas P1, P2 and P3 in FIG.
4 can be represented as a reduced portion. Of the four parts P1, P2, P3, and P4, the added value of the partial area P1 does not change due to the merge processing of C _k (i) and C _k (j). Therefore, the AMI reduction amount L _k (i,
When calculating j), the addition area P0 can be reduced from a two-dimensional area (that is, a square area) to a one-dimensional area (that is, a plurality of lines or lines).
Therefore, the complexity of the calculation time in step A3 is changed from the order O (V ² ) to the order O (V ² ).
(V). When C _k (i + j) is used as a notation representing a class generated by merging the class C _k (i) and the class C _k (j), the AMI reduction L _k (i, j) becomes Given by the expression.

[0031]

(Equation 7) here,

(Equation 8)

AMI reduction L of all classes of pairs
After calculating the _k, pairs such as reduction L _k of the AMI is minimized, for example, class C _k (i) and class C _k (j)
(Where i <j) and then the name of the new class when these pairs of classes are merged is C _kl
(I), and further execute the next merge process using the new set of (k-1) classes. Index numbers (indexes) are assigned to all classes except the class C _k (i) and the class C _k (j) in the same manner after the merge processing. That is, the class C _k (m) is _defined as a class C _kl (m), where m ≠ i, j. Here, if j ≠ k, the class C _k (k) is changed to the class C _k
kl (j), and if j = k, C after merge processing
_k (k) is deleted.

All AMs from the previous merge process
By storing the reduced amount L _{k of} I in the storage device, another optimization process can be executed. Here, the class pair (C _k (i), C _k (j)) is selected as a target of the merging process, that is, for all pairs (l, m), L _k (i, j) ≦ Suppose L _k (l, m). In the next merge processing stage, it is necessary to calculate the AMI reduction amount L _kl ^{(i, j)} (l, m) for all (l, m) pairs. Where superscript (i, j)
Means that the class pair (C _k (i), C _k (j)) has been merged in the previous merge processing stage. Here, the AMI reduction amounts L _kl ^{(i, j)} (l, m) and L _k (l,
Note the difference in m). That is, L _kl
^{(i, j)} (l, m) is A due to merging of class l and class m after merge processing of class i and class j.
L _k (l, m) is the AMI reduction amount due to merging of class l and class m without merging class i and class j. Therefore, AMI
Reduction ^{_{L kl (i, j) (}} l, m) and, reduction of AMI L _k of
The difference from (l, m) is the class pair (C _k (i), C _k
It occurs only from terms or classes affected by the merge processing of (j)).

The above process will be described with reference to FIG. 4. The class bigram plane for the AMI decrease L _kl ^{(i, j)} (l, m) and the AMI decrease L _k (l, m) The addition area of the table is as shown in (b) and (a) of FIG. Region {(x, y) | x {i, j, l, m, and y}
Since the added value of i, j, l, m} does not change by the merging process of the class i and the class j or by the merging process of the class 1 and the class m, those regions are shown in the drawing. Not. Here, as is clear from FIG. 4, mh is a class whose area is lost when class m is merged with class l. Further, as described later in detail, {L _kl ^{(i, j)} (l, m) -L
When calculating _k (l, m)}, most of the regions in the diagram of FIG. 4 cancel each other out, leaving only regions of several points. Thus, the complexity of the calculation time in step A3 can be reduced to a constant.

[0035]

L _k (l, m) = I _k −I _k (l, m) and L _k−1 ^{(i, j)} (l, m) = I _k−1 ^{(i, j)} −I _{k −1} ^{(i, j)} (l, m), where L _k−1 ^{(i, j)} (l, m) −L _k (l, m) = − (I _k−1 ^{(i, j) )} (l,
m) −I _k (l, m)) + (I _k−1 ^{(i, j)} −I _k )

AMII _kl ^{(i, j)} (l, m) and AMII
Some parts of the summation region of _k are AMII _kl
It is canceled with a part of ^{(i, j)} or a part of AMII _k (l, m). Where Ih _kl ^{(i, j)} (l, m), I
h _k (l, m), Ih _kl ^{(i, j)} and Ih _k are the AMII _kl after all the cancelable common classes have been cancelled.
^{(i, j)} (l, m), I _k (l, m), I _kl ^{(i, j)} , and I _k . At this time, the following relational expression is given.

[0037]

L _k-1 ^{(i, j)} (l, m) −L _k (l, m) = − (Ih
^{_{k-1 (i, j)}} (l, m) -Ih k (l, m)) + (Ih k-1 (i, j) -I
h _k ) where

## EQU11 ## Ih _k-1 ^{(i, j)} (l, m) = q _k-1 (l + m, i) +
q _k-1 (i, l + m)

## EQU12 ## Ih _k (l, m) = q _k (l + m, i) + q _k (i, l
+ M) + q _k (l + m, j) + q _k (j, l + m)

## EQU13 ## Ih _k-1 ^{(i, j)} = q _k-1 (i, l) + q
_k-1 (i, m) + qk _-1 (l, i) + qk _-1 (m, i)

[Number 14] _{Ihk = q k (i, l} ) + q k (i, m) + q k (j,
l) + q _k (j, m) + q _k (l, i) + q _k (l, j) + q _k (m,
i) + q _k (m, j)

FIG. 5 shows the addition area of Ih in the above equation (10).
Shown in In the first conventional example, the AMI reduction amount L _{k− 1} ^{(i, j)} (l, m) −L _k is used by ignoring the second term on the right-hand side of Expression 10 and using only the first term. It seems that (l, m) is being calculated. In addition, in the third expression of the equation (17) in the related art document corresponding to the first term of the expression 10, the signs are opposite in sign. However, since the second term in Equation 10 has the same weighting factor as the first term, the inventor uses Equation 10 to complete the model of the present invention.

Even when a method having the order of complexity O (V ³ ) of the operation time is used, when the number of vocabulary words is as large as 10 ⁴ or more, it cannot be calculated actually. In any case, since the calculation time of the order O (V) is required in the step A1, it can be considered that only the step A2 can be corrected. In step A2, all possible merging of class pairs can be considered, but in practice, the range of class pairs to be searched can be limited. Regarding this, the first conventional example proposes the following method, and the method of the present invention also adopts this method. First, V single classes are created based on the words in the text data including the number V of words that do not overlap with each other, and are arranged in descending order of frequency to form a "merged area" (in this embodiment, a window called a window). Therefore, in the present specification, the merge area and the window are synonyms.) Is the word of the first c + 1 classes in the class order. Therefore, first,
(C + 1) frequently used words become a merge area. Next, the following processing is performed.

<Step D1> Among all pairs in the merge area, a pair of classes that minimizes the amount of AMI reduction is merged. <Step D2> The class at the (c + 2) th position is inserted into the merge area or window, and each class behind the class at the (c + 2) th position is moved by one to the left. <Step D3> The processes of steps D1 and D2 are repeated until the number of remaining classes reaches a predetermined value c.

In the processing algorithm of the first conventional example, the complexity of the operation time in step A2 is an order O (c ² ) proportional to c ² which is the square of the final class number c. the total calculation time is reduced to the order (c ² V) which is proportional to c ² V.

Next, a method for obtaining a word clustering structure will be described. The simplest method for obtaining a representation of a word in a tree representing the clustering structure of words is to construct a dendrogram (also referred to as a tree diagram or tree diagram) dendrogram as a by-product of the merging process. In general,
To take a record (or history) of the order of merge and create a binary tree based on that record. FIG. 6 shows a simple example using a vocabulary of five words. The merge history (or the merge history) in FIG.
As shown in FIG. Note that the first line of Table 1 means that "the class A and the class B are merged, and the new class after the merge is named A."

[0043]

[Table 1] Merge history Ｍ Merge (A, B → A) Merge (C, D → C) Merge (C, E → C) Merge (A, C → A) ) ─────────────

However, when this method is directly applied to the O (C ² V) algorithm of the first conventional example, the balance of each class becomes extremely poor, and FIG.
As shown in FIG. 7, the tree structure has only a leftward branch. The reason for this is that, in terms of the amount of AMI, rather than merging a class having a relatively large size and a higher frequency after growing a plurality of classes in the merge area to have a certain size, This is because merging classes of a single set having a lower frequency is significantly lower in cost.

The new method according to the present invention adopted by the inventor is as follows. <Step E1> MI clustering: c classes are created using a mutual information clustering algorithm having a constraint condition of a merge area. As shown in FIG. 19, the c classes constitute the intermediate layer 100 of the dendrogram which is the last tree diagram. <Step E2> External clustering: classifying all words in the text data into class tokens (class t)
(actually, only the bigram table is processed instead of the entire text in the text data), and all classes are merged into a single class by the merge process without the constraints of the merge area. Perform the merge process (binary merge process) in the form of a binary tree until it is. By this process, a dendrogram D _root is created. This dendrogram D _root constitutes the upper layer 101 of the final tree structure, for example, as shown in FIG. <Step E3> Internal clustering: {C ¹ , C ² ,
, C ⁱ ,..., C ^c } are a set of classes (class set) of the intermediate layer 100 obtained in the above step E1. Each i (1 ≦ i ≦ c; i is a natural number)
Perform the following processing.

[0046] except those of <step E3-1> class C ^i, replace all the words in the text and the class token. Determine the new vocabulary V ′ = V ₁ ∪V ₂ . Here, V ₁ = {all words in C ⁱ }, V ₂
= {C ¹ , C ² ,..., C ⁱ⁻¹ , C ^{i + 1} , C ^c }, and C ^j is a j-th class token (1 ≦ j ≦ c). Assign each element of the vocabulary V 'to each class, executes the merge processing in the form of a binary tree by the constrained merging process of merging processing can be only class that contains only the elements of the vocabulary V _1. The process orders the elements of vocabulary V ′ that include the elements (ie, words) of vocabulary V ₁ in the _first | V ₁ | classes in order of frequency, and then
This can be done by first performing the merge process in a merge region having a width | V ₁ | and a width that decreases by one with each merge process. Where |
V ₁ | is the number of words in the vocabulary V ₁ . Repeated merging process until all the elements in <Step E3-2> vocabulary V ₁ is entered into a single class. As shown in FIG. 19, a dendrogram D _sub of the lower layer 102 is created by a merge process for each class. This dendrogram constitutes a subtree of each class in which leaf nodes represent each word in the class.

<Step E4> The node of each leaf of the dendrogram D _root of the upper layer 101 is stored in the corresponding lower layer 102.
By substituting the dendrogram D _sub with the dendrogram, the entire dendrogram can be obtained.

The word clustering method according to the present invention may generate a tree structure having a balanced binary tree form in that words having similar meanings or syntactic features are located in close proximity. it can. FIG. 8 shows a 500 assembled using the method of the present invention for the most frequently used top 70,000 words in the Wall Street Journal (WSJ) corpus.
FIG. 9 shows an example of a lower layer dendrogram D _sub for one of the classes. FIG. Finally, a path from a root node (root node) to a leaf node (leaf node) is traced, and one bit of 0 or 1 representing a leftward branch or a rightward branch, respectively. Can be assigned a bit string (word bits) for each word in the vocabulary.

FIG. 10 is a block diagram showing the configuration of the word classification processing section 20 of FIG. The configuration and operation of the word classification processing unit 20 will be described with reference to FIG. FIG.
, The word classification processing unit 20 is a controller provided with the CPU 50, the CPU 50, a ROM 51 for storing a word classification processing program executed by the CPU 50 and data necessary for executing the program, A work RAM 52 that is a work area required when executing the word classification processing, a RAM 53 having a plurality of memory areas required when executing the word classification processing, and two memory interfaces 5
4 and 55, and these circuits 50 to 55 are connected to each other via a bus 56. Here, the memory interface 54 includes the text data memory 10 and the CPU.
50, the text data memory 10 and C
While the memory interface 55 is an interface circuit for executing interface processing such as signal conversion with the PU 50, the memory interface 55
Between the word dictionary memory 11 and the CP
It is an interface circuit for executing interface processing such as signal conversion with U50. RAM5
3 has a plurality of memory sections divided as follows.

(A) Initialization class word memory 61: stores v words and their classes obtained by initialization processing described later; (b) AMI memory 62: intermediate layer clustering processing, upper layer clustering described later In the processing and the lower layer clustering processing, temporary pairs of all combinations are created among the words of the class in one window, and the average mutual information when each temporary pair is merged is calculated using Equation 2. (C) Intermediate layer memory 63: Stores the words of the c intermediate layer classes obtained by the intermediate layer clustering process described later; (d) Upper layer history memory 64: Upper layer clustering process described later (E) Upper-layer tree memory 65: upper-layer clustering process Storing dendrogram D _root is a tree diagram obtained by physical; (f) the lower layer history memory 66: obtained by the lower layer clustering process, a single tree for each class of the intermediate layer 100 storing a plurality c pieces of dendrogram D _sub diagrams; (g) lower layer tree memory 67: storing dendrogram D _sub is a tree diagram obtained by the lower layer clustering process; (h) Tree memory 67: Stores a dendrogram as an entire tree diagram obtained by connecting one dendrogram of the upper layer 101 and a plurality of c dendrograms of the lower layer 102 via the intermediate layer 100. I do.

FIG. 11 is a flowchart showing the word classification processing of the main routine executed by the word classification processing section 20 of FIG. As shown in FIG. 11, first, in step S1, an initialization process for executing a process of arranging words in order of appearance frequency based on text data is performed, and then, in step S2, words of a class of the intermediate layer 100 are obtained. An intermediate layer clustering process is executed, and further, an upper layer clustering process for obtaining a tree diagram of the upper layer 101 is executed in step S3, and a lower layer clustering process for obtaining a tree diagram of the lower layer 102 is executed in step S4. Finally, the entire tree diagram obtained by connecting one tree diagram of the upper layer 101 and a plurality of c tree diagrams of the lower layer 102 via the intermediate layer 100 in step S5. A data output processing for obtaining a dendrogram and storing the result in the word dictionary memory 11 as a word dictionary To run. Thus, the word classification processing ends. In these tree diagrams, each word is classified into one class, and the connection relationship between the classes is shown.

FIG. 12 is a flowchart showing the initialization processing (S1) of the subroutine of FIG. As shown in FIG. 12, in step S11, based on the text data in the text data memory 10, the appearance frequencies of all the plurality of v words excluding the duplication of the words are checked, and the words having the higher appearance frequencies are checked in order. The words are assigned to a plurality of v classes, the words of the plurality of v classes are stored in the initialization class word memory 61, and the process returns to the original main routine. Here, v is a natural number of 2 or more.

FIG. 13 is a flowchart showing the intermediate layer clustering processing (S2) of the subroutine of FIG. As shown in FIG. 13, first, in step S21, after reading words of a plurality of v classes from the initialization class word memory 61, the words of the class having a high appearance frequency among the words of the plurality of v classes are read first. Less than v (c +
1) Assuming that words of a class are words of a class in one window (or merge area), as shown in FIG.
It is stored in the work RAM 52. Here, 1 <c + 1 <v
It is. Next, in step S22, the work RA
Among the words of the class in one window stored in M52, a tentative pair of all two combinations is created, and the average mutual information when each tentative pair is merged is calculated using Equation 2. Then, each temporary pair and the calculated average mutual information corresponding to the temporary pair are expressed in the form of AMI in the following Table 2.
It is stored in the memory 62.

[0054]

[Table 2] 仮 Provisional pair average mutual information ────────────────── (C ¹ , (C ² ) 0.867678 (C ² , C ³ ) 0.234689 (C ³ , C ⁴ ) 0.125686 ............ (C ^c , C ^{c + 1} ) 0.675642 ── ────────────────

Next, in step S23, FIG.
As shown in (1), one class is reduced by finding the largest temporary pair out of the average mutual information amount of each temporary pair stored in the AMI memory 62 and merging the temporary pair, thereby reducing one class. The word of the class in one window is updated, and the updated word of the class is stored in the work RAM 52. Then, in step S24, it is determined whether there are no classes outside the window and the number of classes in the window is c. If the determination is NO, in step S25, as shown in FIG. The words of the class having the highest frequency of appearance outside the current window are inserted into the window, the words of the class in one window after the insertion are updated, and the words of the updated class are stored in the work RAM 52. After memorizing,
Returning to step S22, the processing after step S22 is repeated.

On the other hand, if YES is determined in the step S24, the work RAM 52 is determined in a step S26.
Are stored in the intermediate layer memory 63 as the intermediate layer 100. Thus, the intermediate layer clustering process ends, and the process returns to the main routine.

FIG. 14 is a flowchart showing the upper layer clustering process (S3) of the subroutine of FIG. 11, and as shown in FIG.
Is a process of obtaining a tree diagram in the direction of. As shown in FIG. 14, first, in step S31, the intermediate layer memory 6
After reading the words of the c classes in 3, the words of the c classes are stored in the work RAM 52 as the class words in one window. Then, step S
In step 32, as in step S22, the work RAM
Among the words of the class in one window stored in 52, a tentative pair of all two combinations is created,
The average mutual information amount when each of the temporary pairs is merged is calculated using Equation 2, and each temporary pair and the calculated average mutual information amount corresponding thereto are stored in the AMI memory 62 in the format of Table 2 described above. Remember.

Next, in step S33, as in step S23, the largest provisional pair is found out of the average mutual information amount of each provisional pair stored in the AMI memory 62, and the provisional pair is merged. One class is reduced, the words of the class in one window after the merge are updated, and the words of the updated class are stored in the work RAM 5.
Stored in 2. In addition, the history of the merge process having the format shown in Table 1 and indicating which class and which class are merged into a newly named class is stored in the upper layer history memory 64. And step S
At 34, it is determined whether or not the number of classes in the window has reached c. If the determination is NO, the process returns to step S32, and the processes after step S32 are repeated.

On the other hand, if YES in step S34, in step S35, based on the history or history of the upper layer in the upper layer history memory 64, for example, as shown in FIG. A gram D _root is created and stored in the upper layer tree memory 65. Thus, the upper layer clustering process ends, and the process returns to the main routine.

FIG. 15 is a flowchart showing the lower layer clustering processing (S4) of the subroutine of FIG. 11, and as shown in FIG. Arrow 2 for each class C ⁱ
This is a process for obtaining a tree diagram in the direction of 02. is there. FIG.
As shown in FIG. 5, first, in step S41, after reading out words of c classes in the intermediate layer memory 63,
One class is selected from the c classes. Then, in step S42, v in the selected class
_i words as class words in one window,
It is stored in the work RAM 52. Next, step S43
In the same manner as in steps S22 and S32, in the words of the class in one window stored in the work RAM 52, a temporary pair of every two combinations is created, and each temporary pair is merged. The average mutual information is calculated using Equation 2, and each tentative pair and the calculated average mutual information corresponding thereto are represented by A in the format of Table 2 described above.
It is stored in the MI memory 62.

Next, in step S44, as in steps S23 and S33, the largest provisional pair is found out of the average mutual information amount of each provisional pair stored in the AMI memory 62, and the provisional pair is merged. By 1
One class is reduced, the word of the class in one window after the merge is updated, and the word of the updated class is stored in the work RAM 52. Further, the history of the merging process having the format shown in Table 1 and indicating which class is merged with which class into a newly named class is stored in the lower layer history memory 66. And
In step S45, it is determined whether or not the number of classes in the window has reached c. If the determination is NO, the process returns to step S43, and the processing from step S43 is repeated. Here, the processing of steps S43 and S44 is executed for each class of the intermediate layer 100.

On the other hand, when YES is determined in the step S45, all the intermediate layers 10 are determined in a step S46.
It is determined whether or not the processing has been performed for the class of 0. If the determination is NO, an unprocessed class remains, and another unprocessed class of the remaining intermediate layer 100 is selected in step S47. Thereafter, the process proceeds to step S42. On the other hand, when YES is determined in the step S46, in a step S48, the lower layer history memory 6
Based on the history or the history of the lower layer 6, for example, as shown in FIG. 6, and stores the lower layer tree memory 67 to create a tree diagram or dendrogram D _sub of the lower layer.
Thus, the lower layer clustering process ends, and the process returns to the main routine.

FIG. 16 is a flowchart showing the data output processing (S5) of the subroutine of FIG. FIG.
As shown in FIG. 19, first, in step S51, FIG.
As shown in FIG. 7, based on the upper-layer tree diagram in the upper-layer tree memory 65 and the lower-layer tree diagram in the lower-layer tree memory 67, these two tree diagrams are stored in the intermediate layer 100.
Linked via respective class C ^i, that is, while connecting the tree view of the upper layer of the upper layer tree memory 65 in each class C ⁱ of the intermediate layer 100, the lower layer of the lower layer in the tree memory 67 The class at the top of the tree diagram is represented by the middle layer 10.
0 to each class C ⁱ . As a result, an entire tree diagram based on the text data is created, and information of the tree diagram is stored in the tree memory 68. As shown in FIGS. 6 and 8, the tree memory 68 stores the connection relation between words of each class as a word dictionary. Then, in step S52, the information of the tree diagram in the tree memory 68 is output to and stored in the word dictionary memory 11 as a word classification result (or word clustering result).

<First Embodiment> FIG. 1 is a block diagram of a speech recognition apparatus according to a first embodiment of the present invention. In FIG. 1, text data including a plurality of words, for example, English or Japanese, stored in a text data memory 10 is subjected to a plurality of classes by executing the above-described word classification processing by the word classification processing unit 20. And stored in the word dictionary memory 11 as a word dictionary in which the connection relations of the classes are described.

On the other hand, the uttered voice composed of a plurality of words input to the microphone 1 is converted into a voice signal by the microphone 1 and then A / D converted by the A / D converter 2 into a digital voice signal. The digital audio signal is input to the feature extracting unit 3, which performs, for example, LPC analysis on the input digital audio signal to extract feature parameters such as cepstrum coefficients and logarithmic power, and stores the buffer memory 4 in the buffer memory 4. And outputs the result to the voice recognition unit 5. The speech recognition unit 5 refers to the word dictionary stored in the word dictionary memory 11 and refers to the language model stored in the language model memory 12 which is, for example, a phoneme hidden Markov model (hereinafter, referred to as a phoneme HMM). , Perform speech recognition for each word, and output a speech recognition result. Here, the word dictionaries in the word dictionary memory 11 include, for example, (a) 010010010, position; (b) 010010011, location; (c) 110010100, for; It contains information such as the bit string to be represented.

<Second Embodiment> FIG. 2 is a block diagram of a morpheme and syntax analyzer according to a second embodiment of the present invention. In FIG. 2, the text data memory 3
2, consisting of a plurality of words stored in
Each of the two pieces of text data is subjected to the above-described word classification processing by the word classification processing unit 20.
As word dictionaries classified into a plurality of classes and describing the connection relations of the classes, word dictionary memories 41 and 4 are provided, respectively.
2 is stored.

A natural language sentence consisting of a character string of a predetermined language such as Japanese or English and consisting of a plurality of words is input to the morphological analysis unit 21, and the morphological analysis unit 21 converts each word of the input natural language sentence. For the appearance form, the word dictionary memory 41
In addition to dividing the natural language sentence into a plurality of words by referring to the word dictionary stored in the above, information such as part of speech, inflected form, standard expression form, and synonym code is given to each of the appearance forms. Is output to the syntax analyzer 22. Next, the syntax analysis section 22 refers to the word dictionary stored in the word dictionary memory 42, executes a predetermined syntax analysis, adds syntax tree information to the word string, and outputs the result as an analysis result.

As described above, as shown in FIG. 19, the lower layer, the middle layer, and the upper layer are hierarchized, and a plurality of words are classified into a plurality of classes in the form of a binary tree. A word classification result having a balanced hierarchical structure can be obtained by the processing. In addition, in the calculation of the AMI, the AMI is calculated for all classes of words in the lower layer, the middle layer, and the upper layer, so that the calculated AMI is not a local AMI but a whole AMI. A global containing word information is based on AMI,
The clustering process is running. Therefore, a totally optimized word classification result can be obtained. This makes it possible to obtain a more precise and accurate classification system when automatically acquiring a word classification system from text data. Furthermore, by performing speech recognition based on the word dictionary obtained by the word classification unit 20, speech recognition can be performed at a higher recognition rate than in the conventional example.

In the above embodiment, the voice recognition unit 5
The word classification processing unit 20, the morphological analysis unit 21, and the syntax analysis unit 22 are configured by, for example, a digital computer. As shown in FIG. 11, the word classification process of the above embodiment is executed in the order of the intermediate layer clustering process, the upper layer clustering process, and the lower layer clustering process. However, the present invention is not limited to this. It may be executed in the order of the layer clustering process, the lower layer clustering process, and the upper layer clustering process. In the above embodiment, before the initialization process in FIG. 11, a plurality of words are assigned based on the criterion of assigning a word having a high ratio adjacent to the same word to the same class using the n-gram of the word. May be executed.

[0070]

【Example】

<Experiment (Simulation)> Clusters and word bits were created using plain text data of the WSJ corpus for six years. Text size is 5 million words, 10
Million, 20 million, and 50 million words (5 MW, 10 MW, 20 MW, 50 MW, respectively;
W is a word. ). The vocabulary was the top 70,000 words used most frequently throughout the corpus. The final class number c was set to 500. The acquired cluster and word bits are evaluated using the following two measures SS1 and SS2 respectively. (A) The measure SS1 is a perplexity that calculates the perplexity of a class-based trigram model based on the WSJ corpus and a general English treebank, a corpus owned by the applicant. Is the law. (B) The measure SS2 is a labeling (labeling or label) of a partial sound using a decision tree owned by the present applicant.
ling) labeling device (tagger or lab)
eler).

<Perplexity Method> The probability of a word trigram can be rewritten as follows by using a class function G that maps a word to its belonging class.

[0072]

[Number 15] _{P (w i │w i-2} w i-1) = P c (G (w i)) │G (w
_{i-2) G (w i} -1)) P m (w i │G (w i))

Here, P _c is a second-order Markov chain probability, and P _m is a word membership probability. Smoothing of Pc and Pm is performed using Katz backoff and Good Turing formula, respectively. The size of the training text is 1.9MW,
The test text is 150 KW, and both are based on the WSJ corpus. The vocabulary size is 77KW. FIG. 9 shows the relationship between the perplexity of the test text and the text size of the clustering.
The zero in the text size of the clustering represents the perplexity of the word trigram model. As the text size of the clustering increases, the perplexity monotonically decreases. This indicates an improvement in the clustering process. At 50 MW, the perplexity is 18% lower than for the word trigram model. This result is in sharp contrast to the result of the first conventional example in which the perplexity of the class trigram is slightly higher than that of the word trigram model.

<Labeling of Audio Part Using Decision Tree>
A labeling device (tagger) for tagging partial voices using a decision tree owned by the present applicant is known as a SPATTER, for example, D2Ma.
german, “Natural Language Parsing as Stat
istical Recognition ", Doctoral Dissertation,
Stanford University, Stanford, California, 1
994 ". ) Is an integrated module of a decision tree parser owned by the present applicant. The labeling apparatus employs 441 syntactic tags, an order of magnitude larger than that of the University of Pennsylvania Treebank project. The learning text, the test text, and the running text all include the entire sequence of word and label pairs. During the learning phase, events
Is a set of feature values or a set of pairs of questions and their answers. One feature is that the current word w to be processed
any attribute in the context where ord (0) appears,
This is conveniently represented in the form of a question. Labeling is done from left to right. Table 3 shows that the current word "li"
An example of an event using “ke” is shown.

[0075]

[Table 3] ─────────────────────────────────── Event-128: {<word (0), “Like”> <word (-1), “flies”> <word (-2), “time”> <word (1), “an”> <word (2), “arrow”> <tag (- 1), “Verb-3rd-Sg-type3”> <tag (-2), “Noun-Sg-type14”> ............ ...... (Basic Questions) <Inclass? (Word (0), Class295), “yes”> <WordBits (Word (-1), 29), “1”> ........ .......... (WordBits Questions) <Tag, “Prep-type5”>} ──────────────── ───────────────────

The last pair of this event is an answer, a special item indicating the correct label of the current word. The first two lines represent the question about identifying the word around the current word and the label for the preceding word. These questions are called basic questions. The second question type (word bit question) consists of clusters and words that say "is this current word in class 295?" Or "what is the 29th bit in the word bits of the preceding word?"It's about bits.

A decision tree is created from a set of events. The root node of the decision tree represents a set of all events, each containing the correct label for the corresponding word. The probability distribution of the labels for the root nodes can be obtained by calculating the relative frequencies of the labels in the set. By querying the feature values for each event in the set, the set becomes N
(N is a possible value for the feature). Then, conditional on this feature value, it is possible to calculate a conditional probability distribution of the labels for each subset. After calculating the reduction in entropy caused by the set division for each feature, the feature that maximizes the reduction in entropy is selected. By iterating the method and dividing the set into subsets, a decision tree can be constructed in which the leaf nodes include the conditional probability distribution of the tags. Next, the obtained probability distribution is smoothed using the execution data. For details of the smoothing process, refer to the above-mentioned prior art document 2. During the test phase, the system examines the conditional probability distribution for each word in the test text and uses beam search to select the most likely labeling sequence.

The present inventors used WSJ text and a corpus (hereinafter referred to as ATR corpus) owned by the present applicant for the labeling experiment. The WSJ text was manually relabeled using Applicant's syntactic label set. The ATR corpus is a comprehensive sample of literary American English, and its grammar style and settings are very wide and are made up of many different domains. Since the ATR corpus is still in development, the size of the text at hand for this experiment is:
Although the label set is large, it is quite small. Table 4 shows the size of the text used in this experiment.

[0079]

[Table 4] ─────────────────────────────────── Text size (number of words) Learning test Execution ─────────────────────────────────── WSJ text 75,139 5,831 6,534 ATR corpus 76,132 23,163 6,680

FIG. 20 shows the labeling error rates for various clustering text sizes. In this experiment, a basic question and a word bit question are used from two types of question formats. Different experiments were performed to see the effect of introducing the word bit information into the labeling device, in which a randomly generated bit string was assigned to each word (a characteristic bit string was assigned to each word). However, the labeling device also uses the bit string as an identification number for each word being processed.In this control experiment, the bit string assignment is random, but the two words are the same word bit. The random word bits do not give the labeling device any class information other than word recognition.), Basic questions and word bit questions were used.
The results are expressed at a value of zero for the text size of the clustering. For both the WSJ text and the ATR corpus, the labeling error rate is reduced by more than 30% by using word bit information extracted from the 5 MW text, and further increases as the text size of the clustering increases. Decreased. At 50 MW, the error rate dropped by 43%. This also indicates that the quality improvement of the cluster is due to an increase in the text size of the clustering. Overall, the high error rate is due to a very large label set and a small training set. The notable point of this result is that A
Although the text of the TR corpus and the text of the WSJ are very different in area from each other, the introduction of the word bits composed of the text of the WSJ is as much as the text of the ATR corpus as for the text of the WSJ. Was also effective in labeling From this point, it is considered that the acquired hierarchical cluster can move beyond the area.

As described above, the inventor has proposed an algorithm for hierarchical clustering division of a plurality of words, and has conducted an experiment of cluster division using large text data from 5 MW to 50 MW. The high quality of the acquired cluster has been confirmed by two types of evaluation methods. The perplexity of the class-based trigram model is 18% lower than that of the word-based trigram model. By introducing word bits into the labeling device for the speech part of the decision tree owned by the Applicant, the rate of labeling errors is 43%.
% Also decreases. The hierarchical clustering partitioning process derived from WSJ text was also found to be effective for labeling ATR corpus texts in a completely different range from WSJ text.

[0082]

As described in detail above, according to the word classification processing method according to the first aspect of the present invention, all the plural v data different from each other are applied to text data including a plurality of words.
Examining the frequency of occurrence of the words, arranging the words in order from the word having the highest frequency, and assigning the words to a plurality of v classes;
Storing, in the first storage device, (c + 1) classes of words having a high appearance frequency of less than v among words of the plurality of v classes as words of a class in one window; Different products of the first class and the product of the probability of occurrence of the word of the first class and the probability of occurrence of the word of the second class based on the words of the class in one window stored in the storage device of the first class. In order to maximize the predetermined average mutual information representing the relative proportion of the probability that the word and the second class word appear adjacent to each other, the plurality of words are divided into a plurality of c Classifying the plurality of c classes into classes, and storing the classified plurality of c classes in the second storage device as the plurality of c classes in the intermediate layer of the entire tree diagram representing the word classification result; and the second storage Middle layer stored in the device Based on the words of the plurality of c classes, the words of the plurality of c classes are classified into one class in the form of a binary tree so that the average mutual information amount is maximized. In the third storage device as an upper layer of the tree diagram, and for each of the c classes of the intermediate layer stored in the second storage device, Based on a plurality of words in each class of the class, the plurality of words are classified in a binary tree form into one class so that the average mutual information is maximized. Storing a plurality of classification results for each of the plurality of classification results in the fourth storage device as a lower layer of the tree diagram; and storing the lower layer of the tree diagram stored in the fourth storage device in the second storage device A plurality of c classes of the intermediate layer stored in the device. While connecting the upper layer of the tree diagram stored in the third storage device with the plurality of c classes of the intermediate layer stored in the second storage device. And obtaining the tree diagram including the intermediate layer and the lower layer, and storing the tree diagram as a word classification result in the fifth storage device. Therefore, the lower layer, the middle layer, and the upper layer are hierarchized, and a plurality of words are classified into a plurality of classes in the form of a binary tree. Obtainable. In the calculation of the average mutual information, the average mutual information is calculated for all the classes of words in the lower layer, the middle layer, and the upper layer. The global processing including the information of the entire word, not the local average mutual information, performs the clustering process based on the average mutual information. Therefore, a totally optimized word classification result can be obtained. This makes it possible to obtain a more precise and accurate classification system when automatically acquiring a word classification system from text data.

According to a second aspect of the present invention, in the word classification processing method of the first aspect, the step of storing the plurality of classified c classes in the second storage device comprises the step of: There is a class outside one window stored in the first storage device, or
When the number of classes in one window is not c, the words of the class that is outside the current window and has the highest frequency of occurrence are inserted into the window, and then the word classification processing in the form of the binary tree is performed. It is characterized by executing. Therefore, it is possible to obtain an intermediate layer having predetermined plural c classes in an optimized form.

The word classification processing apparatus according to claim 3 of the present invention examines the appearance frequency of all the plurality of v words different from each other with respect to text data containing a plurality of words,
First control means for sequentially arranging words having a high frequency of occurrence and assigning them to a plurality of v classes; and, among words of the plurality of v classes, words of less than v (c + 1) classes having a high frequency of occurrence, A second control unit for storing the words of the class in one window in the first storage device, and a first class based on the words of the class in one window stored in the first storage device. Represents the relative ratio of the probability that the first class word and the second class word that are different from each other appear adjacent to the product of the occurrence probability of the second word and the occurrence probability of the second class word The plurality of words are classified into a plurality c classes in the form of a binary tree such that the predetermined average mutual information amount is maximized, and the classified plurality c
Control means for storing the plurality of classes in the second storage device as a plurality of c classes in the middle layer of the entire tree diagram representing the word classification result, and the intermediate control means for storing the intermediate classes stored in the second storage device. Based on the words of the plurality c classes of the layer, the words of the plurality c classes are classified into one class in the form of a binary tree so that the average mutual information amount is maximized. Fourth control means for storing the classification result in the third storage device as an upper layer of the tree diagram, and for each of a plurality of c classes of the intermediate layer stored in the second storage device, Based on the plurality of words in each of the plurality of c classes in the intermediate layer, the plurality of words are each converted into a single tree in the form of a binary tree such that the average mutual information is maximized. And classify the results of each class into the above tree. Fifth control means for storing in the fourth storage device as a lower layer of the diagram, and a lower layer of the tree diagram stored in the fourth storage device being stored in the second storage device. While linking with the plurality of c classes of the intermediate layer, the upper layer of the tree diagram stored in the third storage device is combined with the plurality of c classes of the intermediate layer stored in the second storage device. Sixth control means for obtaining the tree diagram including the upper layer, the intermediate layer, and the lower layer by connecting to the class, and storing the tree diagram as a word classification result in the fifth storage device. Therefore, the lower layer, the middle layer, and the upper layer are hierarchized, and a plurality of words are classified into a plurality of classes in the form of a binary tree. Obtainable. In the calculation of the average mutual information, the average mutual information is calculated for all the classes of words in the lower layer, the middle layer, and the upper layer. A global containing the information of the whole word, not the local average mutual information, is based on the average mutual information,
The clustering process is running. Therefore, a totally optimized word classification result can be obtained. This makes it possible to obtain a more precise and accurate classification system when automatically acquiring a word classification system from text data.

According to a fourth aspect of the present invention, there is provided the word classification processing device according to the third aspect, wherein the third control means is configured to execute the processing based on one of the windows stored in the first storage device. An outer class exists, or 1
When the number of classes in one window is not c, the words of the class that is outside the current window and has the highest frequency of occurrence are inserted into the window, and then the word classification processing in the form of the binary tree is performed. Execute. Therefore, it is possible to obtain an intermediate layer having predetermined plural c classes in an optimized form.

According to the voice recognition device of the fifth aspect of the present invention, based on the voice signal of the input uttered voice,
A voice for recognizing said uttered voice by referring to a word dictionary including a word classification result in which a plurality of words are classified into a plurality of classes by the word classification processing device according to claim 3 or a predetermined hidden Markov model. A recognition unit is provided. Therefore, by performing speech recognition based on a word dictionary that has a well-balanced hierarchical structure and is totally optimized, obtained by the above-described word classification processing device, the recognition rate is higher than in the conventional example. Can recognize voice.

[Brief description of the drawings]

FIG. 1 is a block diagram of a speech recognition device according to a first embodiment of the present invention.

FIG. 2 is a block diagram of a morpheme and syntax analyzer according to a second embodiment of the present invention.

FIG. 3 is a diagram of a class bigram plane table showing an addition area and an addition / subtraction process in a word classification process performed by the word classification processing unit in FIGS. 1 and 2;

FIG. 4A is a diagram of a class bigram plane table showing an addition area with respect to an AMI reduction amount L _k (l, m) in the word classification process, and FIG. 4B is a diagram illustrating an AMI reduction amount L in the word classification process. It is a figure of the class bigram plane table which shows the addition area to _k-1 ^{(i, j)} (l, m).

FIG. 5A is a diagram of a class bigram plane table showing the AMI amount Ih _k after the merge in the word classification process, and FIG. 5B is a diagram showing the AMI amount Ih _k (l, FIG. 8C is a diagram of a class bigram plane table showing m), FIG. 9C is a diagram of a class bigram plane table showing AMI amount Ih _k−1 ^{(i, j)} after merge in the word classification process, and FIG. Is the AMI amount Ih _k-1 ^{(i, j)} (l,
FIG. 6 is a diagram of a class bigram plane table showing m).

FIG. 6 is a diagram showing an example of a dendrogram (tree system diagram) obtained by the word classification processing.

FIG. 7 is a diagram illustrating an example of a left-side branch tree obtained by the word classification processing.

FIG. 8 is a diagram showing an example of a subtree for one class obtained by the word classification processing.

FIG. 9 is a graph showing perplexity with respect to text size, which is a simulation result in the speech recognition device of the present invention.

FIG. 10 is a block diagram illustrating a configuration of a word classification processing unit 20 of FIG. 1;

FIG. 11 is a flowchart illustrating a word classification process of a main routine executed by the word classification processing unit 20 of FIG. 1;

FIG. 12 is a flowchart showing an initialization process (S
It is a flowchart which shows 1).

FIG. 13 is a flowchart showing an intermediate layer clustering process (S2) of the subroutine of FIG. 11;

FIG. 14 is a flowchart showing an upper layer clustering process (S3) of the subroutine of FIG. 11;

FIG. 15 is a flowchart showing a lower layer clustering process (S4) of the subroutine of FIG. 11;

FIG. 16 is a flowchart showing a data output process (S5) of the subroutine of FIG. 11;

FIG. 17 is a diagram showing a process of step S21 in the intermediate layer clustering process (S2) of the subroutine of FIG. 11, and showing a set of word classes.

FIG. 18 is a diagram showing a process of steps S23 and S24 in the intermediate layer clustering process (S2) of the subroutine of FIG. 11, and showing a set of word classes.

19 is a diagram showing a process in the word classification process of FIG. 11 and a hierarchical structure obtained by the process.

FIG. 20 is a graph showing the error rate of the labeling after the clustering process with respect to the text size, which is a simulation result of the speech recognition device of the present invention.

[Explanation of symbols]

 DESCRIPTION OF SYMBOLS 1 ... Microphone, 2 ... A / D converter, 3 ... Feature extraction part, 4 ... Buffer memory, 5 ... Voice recognition part, 10, 31, 32 ... Text data memory, 11, 41, 42 ... Word dictionary memory, 12 ... language model, 20 ... word classification processing unit, 21 ... morphological analysis unit, 22 ... syntax analysis unit, 50 ... CPU, 51 ... ROM, 52 ... work RAM, 53 ... RAM, 54,55 ... memory interface, 61 ... initial Classified word memory, 62 AMI memory, 63 Middle layer memory, 64 Upper layer history memory, 65 Upper layer tree memory, 66 Lower layer history memory, 67 Lower tree memory, 68 Tree memory , 100: middle layer, 101: upper layer, 102: lower layer, S1: initialization processing, S2: middle layer clustering processing, S3: upper layer Rastering process, S4 ... lower layer clustering process, S5 ... data output process.

Continuation of the front page (56) References JP-A-3-131967 (JP, A) JP-A-3-111972 (JP, A) JP-A-63-172372 (JP, A) Hideki Kashioka, et al., “7G-5” Processing Method of Words with Low Appearance in Word Classification Using Mutual Information ", IPSJ 49th (Late 1994) National Convention, Proceedings (3), September 28-30, 1994 Days, p. 3-185 to 3-186 (58) Fields investigated (Int. Cl. ⁷ , DB name) G06F 17/27 G10L 3/00

Claims (5)

(57) [Claims] 1. A method of examining text data including a plurality of words, the appearance frequencies of all v words different from each other, arranging the words in descending order of frequency, and assigning the words to the v classes. Storing, in the first storage device, (c + 1) classes of words having a high appearance frequency of less than v among words of the plurality v of classes as words of a class in one window; Different products of the first class and the product of the probability of occurrence of the word of the first class and the probability of occurrence of the word of the second class based on the words of the class in one window stored in the storage device of the first class. In order to maximize the predetermined average mutual information representing the relative proportion of the probability that the word and the second class word appear adjacent to each other, the plurality of words are divided into a plurality of c Kula And storing the classified plurality of c classes in the second storage device as the plurality of c classes in the intermediate layer of the entire tree diagram representing the word classification result; and the second storage device The words of the plurality of c classes are converted into one class in the form of a binary tree such that the average mutual information amount is maximized based on the words of the plurality of c classes of the intermediate layer stored in And storing the classification result as an upper layer of the tree diagram in a third storage device. For each of a plurality of c classes in the intermediate layer stored in the second storage device, On the basis of a plurality of words in each of the plurality of c classes of the intermediate layer, the plurality of words are divided into one class in the form of a binary tree so that the average mutual information is maximized. Each class is classified and a plurality of Storing a similar result as a lower layer of the tree diagram in a fourth storage device; and storing the lower layer of the tree diagram stored in the fourth storage device in the second storage device. And linking the upper layer of the tree diagram stored in the third storage device with the plurality of c classes of the intermediate layer stored in the second storage device. And obtaining the tree diagram including the upper layer, the intermediate layer, and the lower layer by storing the tree diagram as a word classification result in a fifth storage device. Classification processing method. 2. The method according to claim 2, wherein the classifying step includes the step of storing the classified plurality of c classes in the second storage device, wherein there is a class outside one window stored in the first storage device, or If there are not c classes in the one window, it is outside the current window,
2. The word classification processing method according to claim 1, wherein after the word of the class having the highest appearance frequency is inserted into the window, the word classification processing in the form of the binary tree is executed. 3. A first method of examining the appearance frequency of all of a plurality of v words different from each other with respect to text data including a plurality of words, arranging the words in descending order of appearance frequency, and assigning the words to the plurality of v classes. A control unit that stores, in the first storage device, (c + 1) classes of words having a high appearance frequency and less than v in the plurality of v classes in the first storage device as words of a class in one window
Based on the word of the class in one window stored in the first storage device, the product of the probability of appearance of the word of the first class and the probability of appearance of the word of the second class Dividing the plurality of words into two so that the predetermined average mutual information representing the relative proportion of the probability that the first class words and the second class words different from each other appear adjacent to each other is maximized. Classifying into a plurality of c classes in the form of a tree, and storing the classified plurality of c classes in the second storage device as a plurality of c classes in the middle layer of the entire tree diagram representing the word classification result. Third control means, based on the words of the plurality of c classes in the intermediate layer stored in the second storage device, so that the average mutual information amount is maximized, Until the words become one class in the form of a binary tree Classified, fourth storing the classification result in the third storage device as the upper layer of the tree diagram
Control means for each class of the plurality of c classes of the intermediate layer stored in the second storage device, based on the plurality of words in each class of the plurality of c classes of the intermediate layer, The plurality of words are classified in the form of a binary tree until they become one class so that the average mutual information amount is maximized. Fifth control means for storing in the fourth storage device as: a lower layer of the tree diagram stored in the fourth storage device, a plurality of intermediate layers stored in the second storage device linking the upper layer of the tree diagram stored in the third storage device with the plurality of c classes of the intermediate layer stored in the second storage device while linking the c classes. Obtains the above tree diagram including the upper layer, the middle layer, and the lower layer. Word classification processing apparatus characterized by comprising a sixth control means for storing in the fifth storage device as a word classification result Te. 4. The method according to claim 1, wherein the third control means is configured to determine whether there is a class outside one window stored in the first storage device or when there are not c classes in the one window. , Outside the current window,
4. The word classification processing device according to claim 3, wherein after the word of the class having the maximum appearance frequency is inserted into the window, the word classification processing in the form of the binary tree is executed. 5. A word dictionary containing a word classification result in which a plurality of words are classified into a plurality of classes by the word classification processing device according to claim 3 or 4, based on an input speech signal of a uttered voice, A voice recognition unit for recognizing the uttered voice by referring to the hidden Markov model.