JP2001034287A - Class determining method for language model, voice recognition device, and program recording medium - Google Patents

Abstract

PROBLEM TO BE SOLVED: To generate a class which is free from redundancy and has generality by determining respective classes while making them correspond to respective nodes which do not always have the same layer depth suitable for classing as to thesaurus information wherein words are classified hierarchically in tree structure. SOLUTION: As an initial state, respective words are classified into classes which are in the shallowest layer, i.e., have the highest degree of abstraction (1, 2). At this time, the number of the classes is the least. Then one class is selected out of those classes (3), the layer of the thesaurus is made one layer deeper to increase the number of classes, and entropy is calculated (4). The selected class is put back and similar trial is carried out for all the classes (6); and the class where the entropy decreases most through fractionation is determined and the fractionation is carried out. The respective classes which are thus determined correspond to nodes in layers of different depth of the thesaurus information generally by the classes.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a method for determining a class in a language model used in language analysis, language understanding, speech recognition, and the like, and a speech for performing speech recognition using a language model constructed based on the determined class. It relates to a recognition device.

[0002]

2. Description of the Related Art A method of constructing a class N-gram model which is one of statistical language models for speech recognition will be described below.

In a conventional method of constructing a class N-gram model,
As attributes for creating a class, there are those that use information such as part of speech information and frequency of occurrence of words (for example, Masataki,
Matsunaga, Sakasaka: IEICE Technical Report SP95-73).

According to this method, words are first classified into parts of speech classes, and words having the highest adjacent co-occurrence frequency are separated from the classes in that order to form one class with one word, and the words are treated independently. The aim is to improve performance.
For example, in the example of FIG. 6, it is assumed that "apple", "hope", and "blood" appear as nouns following the word "red" at adjacent co-occurrence frequencies as shown in the figure. Among them, the probability that "apple" co-occurs with "red" is 80%,
Assuming the easiest connection, "apple" is a noun class 7
Will be treated as a single word. In other words, "apple" separated from the noun class forms a class with a single word. A class N-gram model can be created by describing the frequency of co-occurrence of classes created in this way together with class pairs.

[0005] Further, there is a case where a more subdivided class is constructed by using thesaurus information as compared with a class using only part of speech information (Ando, et al .: Transactions of the Acoustical Society of Japan, Fall 1997). . The thesaurus is a systematic classification of words in a tree structure, and it can be considered that words belonging to the same class have similar examples and meanings in any hierarchy.

In this method, the depth of the hierarchy of each branch of the thesaurus, ie, the subdivision level, is all treated as uniform. For example, the depth of the hierarchy of the thesaurus information used for class determination is treated as being constant, for example, a class is determined corresponding to the classification of the third hierarchy of the thesaurus information.

[0007] Even for an unlearned word, an attempt is made to determine the class to which the word belongs from the thesaurus information of the word.

[0008] A class N-gram model can be created by describing the frequency of co-occurrence of a class created from the thesaurus information together with a class pair.

[0009]

The method of using information such as part of speech information and the frequency of appearance of words as an attribute for creating a class requires a single class even if the word appears at a high frequency of adjacent co-occurrence. When many classes that contain only one word are generated, the generality of connection with adjacent words in content words (nouns, verbs, adjectives, adjective verbs, adverbs, adverbs, inflections, conjunctions, etc.) is eliminated. Thus, the speech recognition performance does not improve despite the increase in the number of classes.

Here, the exclusion of the generality of connection with an adjacent word means the following. That is, for words with extremely high adjacent co-occurrence frequency, the prediction performance is improved by constructing one class with one word. However, when class separation is performed up to a predetermined number of classes, words that have never sufficient frequency ( A), one word constitutes one class. For a word or a class adjacent to the word A (referred to as D), it may be more efficient to treat the words A and B, which are connected in the same manner as the word A, as the same class. In other words, by forming the word A into one class alone, the specialized connection between D and the word A is separated, and the generality of D, a word group having a similar connection relationship of A, B, and C, and Adjacent relationships that have are eliminated.

Many other words that are not separated remain in a very large part of speech, but since the part of speech framework was not originally created in consideration of word connectivity, When used for speech recognition to predict words, it becomes very redundant. Therefore, it is necessary to further subdivide the class to eliminate such redundancy.

Here, "redundancy" means the following. That is, for example, the nouns connected to the adjective "brilliant" include "arm" and "male", but it is considered that the same noun "cup" is not connected. When considering classifying by part of speech, the connectivity of "cup" is "arm",
Despite being different from "male", these three nouns are connected to the same class, so that classifying by part of speech is redundant. Of course, nouns can be subdivided into common nouns, number nouns, sa-variant nouns, adjective nouns, and the like. However, the difference between "arms" and "cups" cannot be clarified simply by subdividing the parts of speech. That is, the word “redundant” means that even words that are not connected to adjacent words are assigned to the same class.

In the method using only thesaurus information, the depth of each branch is treated as being uniform. However, when words are actually used in spoken words, written words, etc. There is no guarantee that the depth of the hierarchy is uniform in all branches, and the thesaurus system itself varies depending on the classification method, so there is a problem that an appropriate hierarchy depth must be determined. . for that reason,
One class is subdivided more than necessary and does not have sufficient generality, and the other class is not sufficiently subdivided and includes redundancy.

As described above, according to the conventional method, it is not possible to eliminate redundancy from a class, and since the class is so specialized as one word, only a class lacking generality and versatility is generated. I couldn't do that.

Therefore, there is a need for a method of adding new classifying information for removing redundancy and generating an optimal class having generality.

SUMMARY OF THE INVENTION An object of the present invention is to provide a method of determining a class in a language model for generating a class having generality while eliminating redundancy and in consideration of the above problem, and a speech recognition apparatus. .

[0017]

In order to solve the above-mentioned problem, a first aspect of the present invention (corresponding to claim 1) is that a node of a thesaurus information in which words are hierarchically classified into a tree structure is provided. This is a class determination method in a language model, characterized in that each class is determined in association with each node whose hierarchical depth suitable for classifying is not always the same.

Further, the second invention (corresponding to claim 2)
Is a class determination method for a language model according to the first aspect of the present invention, wherein each class is determined using an entropy value as an evaluation scale.

Further, the third invention (corresponding to claim 3)
Is determined by tracing the thesaurus information to the lower layer side, stopping the tracing based on a predetermined criterion, and associating each unbranched node with each class at that time. The trace expansion is performed by trace-expanding only a node having the smallest entropy of the next branch among all unbranched nodes among the trace-expanded nodes. This is a method for class determination in the language model.

A fourth aspect of the present invention (corresponding to claim 4)
In the third aspect of the present invention, the predetermined criterion is that the number of unbranched nodes is a predetermined number.

Further, a fifth aspect of the present invention (corresponding to claim 5)
Is a language model storing means for storing a class N-gram model created from a class determined by the class determining method in the language model according to any of the first to fourth aspects of the present invention; Means, a voice recognition means for performing voice recognition of the input voice using the language model, and an output means for outputting a result of the voice recognition as text. It is.

[0022]

Embodiments of the present invention will be described below with reference to the drawings.

(First Embodiment) First, a first embodiment will be described.

First, the thesaurus information will be described. FIG. 5 is an example of thesaurus information created by the National Language Institute. These are thesaurus information belonging to the category of "1.3 human activities, spirits and actions". In the example of FIG. 5, each word of the thesaurus information has a five-digit code, and each word is classified into a category by this code. For the sake of simplicity, "1.3823 architecture", "1.3830 transportation / transport / communication", "1.38
Consider only the three words "31 medical care". If you look at these words in the four digits from the top of the code,
Is “1.382”, “Transportation / Transportation / Communication” and “Medical” are “1.383”, which are divided into two categories. If we take one more digit and consider it with five digits, it is divided into three categories. Thus, each word of the thesaurus information is classified into a tree structure by the code.

In the present embodiment, a class is determined corresponding to words belonging to each category of such thesaurus information. That is, in the above example, the codes of up to four digits were divided into two categories. Correspondingly,
Classes consisting of "Architecture" and "Transportation / Transportation / Communication"
A class consisting of "medical care" is obtained. In addition, the codes up to the upper five digits are divided into three categories, and accordingly, three classes are obtained. In the present embodiment, a class is created in such a manner as to correspond to a node of the thesaurus information.

FIG. 1 is a block diagram of a method for determining a class in a language model according to claims 1 to 4 of the present invention. In the present embodiment, a class is determined corresponding to each node of the thesaurus information.

First, as an initial state, each word is classified into the class having the lowest hierarchy, that is, the class having the highest abstraction, using thesaurus information (1, 2). At this time, the number of classes is the smallest. Next, one of the classes is selected in 3 and the hierarchy of the thesaurus is set to 1
Increase the number of classes by increasing the depth, and calculate entropy by 4. Here, the entropy is calculated using H in Equation 1 below.

[0028]

(Equation 1)

The selected class is returned to the original state, and the same trial is performed for all the classes (5). In step 6, the class whose entropy is reduced most by subdivision is determined and subdivided.

For example, in FIG. 2, when it is assumed that the current class is subdivided to the state shown by the hatched circle, entropy is reduced most when class 113 is expanded among the classes. Then, as shown in Figure 2, the class
113 is one level deeper and subdivided. Here, in FIG. 2, the number given after the class (for example, 113 of class 113)
) Represents a part of a code given to each word of the thesaurus information.

With this state as the initial state of the next step, the same trial is repeated until the target number of classes is reached.

Each class determined in this way is different from the prior art, and generally corresponds to a node of a hierarchy having a different depth of thesaurus information for each class.

In order to evaluate the performance of the method of the present invention, target words were narrowed down to only nouns, and a comparative experiment was performed with the two conventional methods. In the conventional method, (I) the state where all nouns are put into one class is set as the initial state, and the words with the highest adjacent co-occurrence probability are separated from the noun class one by one from the noun, forming an independent class. Go. (II) The shallowest state of the thesaurus hierarchy of nouns is set as the initial state, and the semantic hierarchy of the thesaurus is made evenly deeper by one hierarchy. The two.

FIG. 3 shows the relationship between the number of classes and perplexity for each method. Perplexity is a measure of the ease with which the next word can be predicted, and is calculated based on entropy. In general, a model with a low number of classes and a low perplexity value indicates a good performance in the evaluation of a language model.

That is, in general, with a class having a large number of classes (an extreme example is a word N-gram model), the frequency of appearance cannot be sufficiently high for all the words, and reliable statistical data is obtained. Requires a large corpus. However, when words are classified by classifying several words at a time and the number of classes is reduced, the number of parameters is reduced, and the frequency of appearance as a class is the sum of the frequencies of appearance of the words included in the class. It is easy to obtain reliable values. Therefore, it can be said that a model showing a low perplexity value with a small number of classes has good performance.

FIG. 3 shows that the perplexity value of the method of the present invention is reduced most efficiently especially in the case of a small number of classes, and that the method of the present invention has better performance than the conventional method.

It can be said that by adding thesaurus information to the classifying, the redundancy of the class is eliminated, and the generality of the class can be maintained by making the depth of the thesaurus not uniform.

That is, the meaning of FIG. 3 will be described using a specific example. For example, by changing the example of FIG. 6, as a member of the noun class following the adjective "red",
It is assumed that "hopping" is connected 100 times, "apple" is connected 20 times, "tomato" is connected 15 times, and "blood" is connected 10 times.

At this time, if words having a high frequency of adjacentness are treated as a single class, the performance of "hopp", which has an extremely large number of occurrences, is improved by forming one class by itself. However, when the class separation is repeated up to a predetermined number of classes, the words "apple" and "tomato", which are by far the most infrequent, will be treated as classes by themselves.
Since the number of occurrences of “apple” and “tomato” is small, it cannot be treated as a statistically reliable probability. Also, the words connected to each word other than “red” are likely to be different depending on the word.

Therefore, it is considered that the performance of such a word having a very small number of appearances is better if a few words are put together to form a class. That is, when classifying is performed in such a large framework, it can be understood that "class A and class B are generally easily connected". like this,
Classes that maintain generality can be generated by making the depth of the thesaurus hierarchy not uniform.

When a class consisting of "apple", "tomato", and "blood" is considered, even if the degree of connection with "red" is similar, another word (for example, "delicious") is used. ) Does not always mean the same connection. It is better to have separate classes for apples, tomatoes, and blood. That is, the class consisting of these three words is considered redundant. However, in the framework of part-of-speech information, these cannot be differentiated.

Such differentiation can be realized by using thesaurus information when determining a class, and the redundancy of the class can be eliminated.

(Second Embodiment) Next, a second embodiment will be described.

In this embodiment, a speech recognition apparatus that performs speech recognition using a language model constructed using the classes created in the first embodiment will be described.

FIG. 4 is a block diagram of a voice recognition device according to claim 5 of the present invention.

The speech recognition apparatus comprises a microphone 41, speech recognition means 42, an acoustic model 43, a language model storage means 44, and an output means 45.

The microphone 41 is a means for inputting voice. The speech recognition unit 42 is a unit that converts the input speech into a continuous word string using the acoustic model 43 and the language model stored in the language model storage unit 44. The output unit 45 is a unit that converts a language symbol into text and outputs the text.

The microphone 41 according to the present embodiment
Is an example of the voice input means of the present invention, and the language model storage means 44 of the present embodiment is an example of the language model storage means of the present invention.

Next, the operation of this embodiment will be described.

First, a class N-gram model is created from the classes determined by the class determination method in the language model described in the first embodiment. That is, the class pair and the frequency at which the class pair co-occurs are described. Further, the class uni representing the total number of occurrences of the class
-Gram value, word information in the class in which words included in the class are represented by IDs, and the number of appearances of the words are also described. The created class N-gram model is stored in the language model storage unit 44 in advance.

Next, the operation for performing voice recognition will be described.
The uttered voice is input from the microphone 41,
It is input to the voice recognition means 42. The voice recognition means 42
Based on the language model stored in the language model storage means 44, recognition word candidates are predicted sequentially in time series.
The sum of the acoustic score based on the distance value between the acoustic model 43 and the input speech that has been learned in advance and the language score based on the class N-gram is used as the recognition score.
A search determines a continuous word string that is a recognition candidate. The continuous word sequence determined in this way is output to the output unit 45.
And output as text.

As described above, by performing speech recognition using the language model created from the class determined by the class determination method in the language model described in the first embodiment, high-performance speech recognition can be performed. I can do it.

It is to be noted that all or a part of the functions of each step or each unit constituting the method for determining a class in the language model or the speech recognition apparatus of the present invention may be realized by hardware, or the software of the computer program may be used. It may be realized in a practical way.

Further, a program for causing a computer to execute all or a part of the functions of each step or each means constituting the method for determining a class in a language model or the speech recognition apparatus of the present invention is stored. Is also included in the present invention.

[0055]

As is apparent from the above description, the present invention according to claims 1 to 4 has two problems, that is, the existence of class redundancy and the lack of generality, which are problems of the conventional method. And a class determination method in a language model capable of constructing a high-performance language model can be provided.

The present invention according to claim 5 can provide a speech recognition device capable of performing high-performance speech recognition.

[Brief description of the drawings]

FIG. 1 is a block diagram of a method for determining a class in a language model according to a first embodiment of the present invention.

FIG. 2 is a diagram showing subdivision of classes according to the first embodiment of the present invention.

FIG. 3 is a diagram comparing the performance of a class determination method in a language model of the present invention with the performance of a class determination method in a conventional language model.

FIG. 4 is a block diagram showing a configuration of a speech recognition device according to a second embodiment of the present invention.

FIG. 5 is a diagram showing an example of thesaurus information;

FIG. 6 is a diagram showing an example in which classes are separated by a class determination method in a conventional language model.

[Explanation of symbols]

 1 Word Classification Step 2 Initial State Setting Step 3 Subdivision Class Candidate Selection Step 4 Entropy Calculation Step 5 Number of Classes Representing the Number of Loops 6 Subdivision Class Determination Step 7 Noun Class and Words in Noun Class 41 Microphone 42 Speech Recognition Means 43 Acoustic model 44 Language model storage means 45 Output means 45

Claims (6)

[Claims] 1. The method according to claim 1, wherein, among nodes of the thesaurus information in which words are hierarchically classified into a tree structure, each class is associated with each node whose hierarchical depth suitable for classifying is not always the same. A method for determining a class in a language model, characterized by determining. 2. The method according to claim 1, wherein each class is determined using an entropy value as an evaluation scale. 3. The determination is made by tracing the thesaurus information to the lower layer side, stopping the tracing based on a predetermined criterion, and associating each unbranched node at that time with each class. 3. The trace expansion method according to claim 2, wherein the trace expansion is performed only on a node having the smallest entropy of the next branch among all the unbranched nodes among the nodes that have been trace expanded so far. Class Determination Method in the Language Model 4. The predetermined criterion is that the number of unbranched nodes is a predetermined number.
Class determination method in the described language model. 5. A language model storage means for storing a class N-gram model created from a class determined by the class determination method in the language model according to claim 1, and a voice for inputting voice. Speech recognition, comprising: input means; speech recognition means for performing speech recognition of the input speech using the language model; and output means for outputting the result of speech recognition as text. apparatus. 6. A program for causing a computer to execute all or a part of the functions of each step or each unit constituting a method for determining a class in a language model or a speech recognition device according to claim 1. A program recording medium characterized by storing a program.