JP2003177786A - Language model creation device and speech recognition device using the same - Google Patents

Abstract

(57) [Summary] [PROBLEMS] To easily generate a language model capable of recognizing example sentences and variations thereof. An example sentence corpus storage unit for storing example sentences in advance, a large-scale corpus storage unit for storing a large-scale corpus, and a sentence similar to an example sentence stored in the example-sentence corpus storage unit are stored in the large-scale corpus storage. Selection unit 13 for selecting text from a large-scale corpus and extracting text
And a language model learning unit 15 that learns a language model using the selected corpus and the example sentence corpus, and uses an example sentence and a large-scale corpus prepared in advance. Thus, it is possible to easily generate a language model capable of recognizing example sentences and variations thereof.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a language model device which is responsible for a language restriction part in speech recognition and a speech recognition device using the language model device.

[0002]

2. Description of the Related Art As a language constraint used in continuous speech recognition, a statistical language model N-gram which stochastically models word concatenation is used. However, in order to construct a language model N-gram, it is necessary to collect a large number of utterance examples according to the environment in which they are used and to obtain word concatenation information. However, it is practically difficult to collect a large number of corpus from one for each application from the viewpoint of cost and time.

For this problem, an attempt has been made to use an existing large-scale corpus. However, it has been pointed out in Japanese Patent Laid-Open No. 2001-134285 that sufficient performance cannot be obtained by simply using an existing corpus, and when adding a large-scale corpus, a sentence to be added depending on the utterance purpose is added. It is disclosed that the selection is effective.

FIG. 4 shows a block diagram of a voice recognition device disclosed in Japanese Patent Laid-Open No. 2001-134285, which will be described below. In the learning data selection unit 1, the language model or corpus used in the speech recognition unit 5 is set as the learning data storage unit 2
Selected from the speech recognition section 5 in the learning data conversion section 3
After conversion into the format used in, the voice recognition unit 5 performs voice recognition by using a language model suitable for the utterance content and the partner.

[0005]

However, in order to select sentences to be added according to the purpose of utterance, it is necessary to add classification information to each sentence of the large-scale corpus in advance. Therefore, it is practically difficult to judge whether or not to select each selected sentence.

[0006] This poses a problem for the development of a system for recognizing a relatively small number of specific sentences (about several thousand sentences) such as voice translation. In order to recognize a specific sentence, it is necessary to build a language model that can absorb the sentence to be input and the variations of each sentence that can be uttered by the user for each sentence. However, it is difficult from a cost point of view to manually determine whether or not each sentence of a large-scale corpus corresponds to its variation.

An object of the present invention is to easily generate a language model in which an example sentence and its variations can be recognized by using the example sentence and a large-scale corpus prepared in advance.

[0008]

In order to solve this problem, the present invention provides an example sentence corpus storage unit that stores example sentences in advance, a large-scale corpus storage unit that stores a large-scale corpus, and the example sentence corpus storage unit. A sentence similar to the stored example sentence is selected from the large-scale corpus of the large-scale corpus storage unit, and a text selection unit for extracting text, a language model learning using the selected corpus and the example sentence corpus. And a language model learning unit for performing.

Thus, by providing a criterion for determining whether or not each sentence of the large-scale corpus is a variation of the prepared example sentence, a method for automatically selecting is provided, and the example sentence and the prepared sentence are prepared in advance. By using a large-scale corpus, it is possible to easily generate a language model that can recognize example sentences and their variations.

[0010]

BEST MODE FOR CARRYING OUT THE INVENTION The invention according to claim 1 of the present invention prepares two kinds of corpora, and selects sentences related to the first corpus from the second corpus based on the first corpus. In the text selecting method, a sentence similar to the sentence in the first corpus is selected from the second corpus, and a language model is created by the selected corpus obtained by the selection or the combination of the first corpus and the selected corpus. By selecting a sentence that cannot be assumed in the creation of the first corpus from the second corpus and using it in the creation of the language model, a more robust language model can be created.

The invention according to claim 2 is characterized in that, in the language model creating method according to claim 1, the degree of similarity given by hand is used, and a similar sentence is obtained from the second corpus with high accuracy. It has the effect of being able to select well and to create a robust language model.

According to a third aspect of the present invention, in the language model creating method of the first aspect, the target sentence for which similarity is to be determined is divided into units corresponding to morphemes, and the notational correspondence is obtained. The scale of how much the sentence belonging to the first corpus is reproduced in the sentence belonging to the second corpus and the scale of how accurately it is reproduced are used as the degree of similarity. This has the effect of making it possible to create a robust language model.

According to a fourth aspect of the present invention, in the language model creating method according to the first aspect, the target sentence for which similarity is to be determined is divided into units corresponding to morphemes, and the correspondence is determined.
A measure of how much the sentence belonging to the first corpus is reproduced in the sentence belonging to the second corpus and a measure of how accurately it is reproduced are used as a combination similarity with appropriate weights. The selection criterion can be weighted, and the similarity parameter can be unified.

The invention according to a fifth aspect is characterized in that, in the language model creating method according to the third or fourth aspect, the similarity is calculated by weighting each unit corresponding to a morpheme. There is an effect that the similarity can be calculated more precisely.

The invention according to claim 6 is characterized in that, in the method for creating a language model according to claim 5, the similarity is calculated based only on independent words such as nouns and verbs, and the meaning of the sentence. It has an effect of enabling calculation of similarity based on content.

According to a seventh aspect of the present invention, in the language model creating method according to any one of the first to sixth aspects,
The second corpus is characterized by using a mail or a sentence written by the user, and has an effect that a language model adapted to an individual can be created.

The invention according to claim 8 is the same as claims 1 to 7.
The method for creating a language model according to any one of
The feature is that sentences are selected and used from a plurality of corpuses based on the corpora, and the selection corpora can be further expanded by using more corpora.

The speech recognition method according to claim 9 is characterized in that speech recognition is performed using a language model created based on the language model creation method according to any one of claims 1 to 8. The effect of expanding the corpus has the effect that the degree of freedom of the user's utterance can be increased with respect to the speech recognition by the language model learned by the first corpus.

According to the tenth aspect of the invention, an example sentence corpus storage unit for storing an example sentence in advance, a large scale corpus storage unit for storing a large scale corpus, and a sentence similar to the example sentence stored in the example sentence corpus storage unit. Is selected from the large-scale corpus of the large-scale corpus storage unit to extract text, and a language model learning unit that learns a language model using the selected corpus and the example sentence corpus. This has the effect of making it possible to create a more robust language model by selecting sentences that could not be assumed in the creation of the first corpus from the second corpus and using them in the creation of the language model.

According to an eleventh aspect of the present invention, in the language model creating apparatus according to the tenth aspect, the text selection unit divides a target sentence for which similarity determination is performed into units corresponding to morphemes, In addition to obtaining the correspondence, a measure of how much the sentence belonging to the first corpus is reproduced in the sentence belonging to the second corpus and a measure of how accurately it is reproduced are used as the similarity. This has the effect that a robust language model can be created without manpower.

According to a twelfth aspect of the present invention, in the language model creating apparatus according to the tenth aspect, the text selection unit divides a target sentence for which similarity is to be determined into units corresponding to morphemes, and establishes correspondence relationships. In addition, a measure of how much the sentence belonging to the first corpus is reproduced in the sentence belonging to the second corpus and a measure of how accurately it is reproduced are combined with appropriate weights as the similarity degree. It is characterized by being used, and has the effect that the selection criteria can be weighted and that the similarity parameter can be unified.

According to a thirteenth aspect of the present invention, in the language model creating apparatus according to the eleventh aspect or the twelfth aspect, the similarity is calculated by weighting each unit corresponding to a morpheme. This has the effect of making it possible to calculate the degree of similarity more precisely.

The invention according to claim 14 is the same as claim 1
The speech recognition is performed by using the language model created by the language model creation device according to any one of 0 to 13, and the effect of expanding the corpus is
It has an effect that the degree of freedom of the user's utterance can be increased with respect to the voice recognition by the language model learned by the first corpus.

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

(Embodiment 1) FIG. 1 is a block diagram of a speech recognition apparatus equipped with a language model creating apparatus according to Embodiment 1 of the present invention.

In FIG. 1, 11 is an example sentence corpus storage unit that stores example sentences in advance, 12 is a large scale corpus storage unit that stores a large scale corpus, and 13 is an example sentence corpus storage unit 11.
The text selection unit for selecting and extracting a sentence similar to the example sentence stored in the large-scale corpus storage unit 12 from the large-scale corpus, 14 is a selected corpus storage unit in which the selected text is stored, and 15 is the selected corpus Language model learning unit for learning a language model using the corpus and example sentence corpus, 17
Is a voice input unit that inputs a voice to be recognized, 16 is a voice recognition unit that performs recognition processing on the voice captured by the voice input unit 17, and 18 is an output unit of the result.

The operation of the speech recognition apparatus configured as above will be described.

The example sentence corpus storage unit 11 is a storage device in which example sentences are stored in advance by a system developer or a user. Regarding the storage form of example sentences, it is desirable to divide each sentence into units (hereinafter, morphemes) corresponding to words in advance, but it is also possible to store example sentences as they are, and the example sentence corpus storage unit 11 When outputting from, a filter such as a morpheme analysis may be used and divided into morphemes.

The large-scale corpus storage unit 12 can use a newspaper database, a travel conversation database, or the like as the large-scale corpus. It is desirable that the storage form is divided into units corresponding to words in advance like the example sentence corpus storage unit 11. However, even if the above database is stored as it is, a filter for morphological analysis or the like is stored in the large-scale corpus storage unit. It may be used when data is output from 12.

The text selection unit 13 receives sentences separated from the example sentence corpus storage unit 11 and the large-scale corpus storage unit 12 in units corresponding to morphemes for determining similarity, and the sentences are separated between them according to a criterion described later. The degree of similarity of. When the degree of similarity satisfies a predetermined criterion, it is determined that the sentence of the determined large-scale corpus is similar to the example sentence, and the sentence is output to the selected corpus storage unit 4.

In the selected corpus storage unit 14, sentences judged to be similar to the example sentence are selected from the large-scale corpus and stored. Based on the selected corpus and the example sentence corpus obtained in the selected corpus storage unit 14, the language model learning unit 15 mixes them with appropriate weights to learn a language model. The language model to be learned may be of any type such as 2-gram or 3-gram.

Based on the language model learned as described above, the voice recognition unit 16 recognizes the voice captured by the voice input unit 17, and sends the result to the data output unit 18 to obtain the data. It is output by the output unit 18. The output method may be presented to the user using a display,
By connecting an application to the end of the output section, the result of voice recognition can be used in another application.

The detailed operation of the text selection unit 13 will be described below. In general, sentences stored in a large-scale corpus and sentences stored in an example sentence corpus have different lengths and different words used. In order to find the similarity between sentences, it is necessary to find the correspondence between them.

FIG. 2 is an example in which the example sentences 28 and 30 and the sentences 27 and 29 on the large-scale corpus side are associated with each other assuming the following two examples.

Such correspondence can be easily realized by using a method such as the DP method (dynamic programming method). The symbols 21 and 22 of C, S, I, and D in FIG.
Is determined to be the same between the two sentences, S24 is determined to have been transposed, and I25 is an extra morpheme on the large-scale corpus side (insert).
The part that was determined to be, D26 escaped to the large corpus side,
In other words, it means the part that was dropped.

The text selection unit 13 reproduces the recall rate (hereinafter, "Corre") of both sentences associated with each other.
And two similarity parameters corresponding to the matching rate (hereinafter referred to as Acc) are calculated, and based on these similarity parameters, those exceeding a preset threshold value Cth, Ath are selected as similar sentences. The calculation of the degree of similarity is performed as follows.

[0037]

[Equation 1]

N is the total number of morphemes on the example side, N _C is the number of morphemes determined to be the same between both sentences, and N _I is the number of morphemes determined to be included in the sentence on the large-scale corpus side not in the example. The number, that is, the number of I.

In the example of FIG. 2, each similarity parameter can be calculated as follows.

Example 1 Corre = 3/5 = 0.6 Acc = 3 / (5 + 1) =
0.5 Example 2 Corre = 4/6 = 0.67 Acc = 4 / (6 + 1) = 0.57 The performance of the finally obtained language model changes by changing the thresholds Cth and Ath for selecting similar sentences. For example, Cth>
Sentences that change the endings of example sentences by setting thresholds such as 0.7 and Ath> 0.7. For example, a sentence such as "Do you have a seat?" Can be selected from a large corpus in response to "Do you have a seat?". Also,
It is also possible to select a sentence in which the words are exchanged, such as "Do you have juice?"

However, when a sentence in which "juice" is replaced with "seat" is selected as in the above example, the types of morphemes that appear will increase. When such a case becomes a problem, information about the semantic similarity between morphemes is held by a thesaurus, etc., and the morphemes that can be replaced according to the similarity are constrained to achieve the desired shape. The sentence can be selected with. It is also possible to calculate the degree of similarity by focusing only on independent words such as verbs and nouns. By weighting the morphemes in this way, it becomes possible to ignore the difference in style and select a sentence only based on the content of the sentence.

FIG. 3 is a flow chart showing the above operation.
become that way. In step 301, variables are initialized. In step 302, it is determined whether or not the following process has been performed for all example sentences (assuming that there are Ny examples), and if so (i> = Ny), the end process of step 311 is performed. If not, the process moves to the process of calculating the degree of similarity with the sentences in all large-scale corpora for the i-th example sentence. In step 304, variables are initialized, and in step 305, it is checked whether all the processing is completed. If not completed, the similarity is calculated in step 307. It is determined whether the similarity calculated in step 308 exceeds a threshold for selecting or not, and if it exceeds, the sentence selected in step 309 is copied to the selected corpus storage unit. After that, 1 is added to the variable j, and similar processing is performed on the sentence j + 1 of the next large-scale corpus. Here, it is possible to avoid the problem of selecting the same sentence many times by adding a flag to the sentence of the large-scale corpus that has been selected once.

In the present embodiment, the example sentence corpus storage unit 11 and the large-scale corpus storage unit 12 which store a relatively small number of sentence examples are assumed. However, it is possible to replace both of each corpus with a large corpus. Similarly, both corpus can be used as a storage unit for storing a small number of example sentences. It is also possible to select a sentence similar to the first corpus from the plurality of corpora by sequentially replacing the plurality of corpora with the second corpus.

It is also possible to use a user's mail, a user-created document file, or a home page viewed in the past as a large-scale corpus. As a result, it is possible to adapt the language model to the expressions and contents uttered by the user.

In the similarity calculation according to the present invention, the value of Corre and the value of Acc can be integrated and a new similarity scale S can be used as follows.

[0046]

[Equation 2]

Here, α is an integrated weight, and it is possible to control which of Corre and Acc is used as the similarity. That is, it is possible to control whether the recall rate or the matching rate is prioritized. Once you have determined the parameters,
It is also possible to judge the similarity with one parameter.

[0048]

By using the method of the present invention, it is possible to improve the performance of speech recognition by selecting and using from the large-scale corpus the linguistic information that is efficient for speech recognition in which example sentences and the like are recognized. It is possible to create various language models.

[Brief description of drawings]

FIG. 1 is a block diagram of a voice recognition device including a language model creation device according to a first embodiment of the present invention.

FIG. 2 is a diagram showing correspondence between sentences.

FIG. 3 Flowchart for text selection

FIG. 4 is a block diagram of a conventional speech recognition apparatus.

[Explanation of symbols]

11 Example sentence corpus storage 12 Large-scale corpus storage 13 Text selection section 14 Selected corpus storage unit 15 Language model learning department 16 Speech recognition unit 17 Voice input section 18 Data output section

   ─────────────────────────────────────────────────── ─── Continued front page    (72) Inventor Mitsuru Endo             1006 Kadoma, Kadoma-shi, Osaka Matsushita Electric             Sangyo Co., Ltd. F term (reference) 5D015 GG01 HH00

Claims (14)

[Claims] 1. A text selection method for preparing two types of corpora and selecting a sentence related to the first corpus from a second corpus based on the first corpus, wherein a sentence in the first corpus A language model creating method characterized by selecting similar sentences from a second corpus and creating a language model from the selected corpus obtained by the selection or the first corpus and the selected corpus combined with appropriate weights. . 2. The language model creating method according to claim 1, wherein the degree of similarity given by hand is used. 3. The language model creating method according to claim 1, wherein the target sentence for which similarity is to be determined is divided into units corresponding to morphemes, and a notational correspondence is obtained, and the sentence belongs to the first corpus. A method for creating a language model, characterized in that a measure of how much a sentence is reproduced in a sentence belonging to the second corpus and a measure of how accurately it is reproduced are used as the degree of similarity. 4. The language model creating method according to claim 1, wherein the target sentence for which similarity is to be determined is divided into units corresponding to morphemes, the correspondence is obtained, and the sentence belonging to the first corpus is first divided. A method for creating a language model, characterized in that a measure of how much is reproduced in a sentence belonging to the corpus 2 and a measure of how accurately it is reproduced are combined with appropriate weights and used as similarity. 5. The language model creating method according to claim 3 or 4, wherein the similarity is calculated by weighting each unit corresponding to a morpheme. 6. The language model creating method according to claim 5, wherein the similarity is calculated only based on independent words such as nouns and verbs. 7. The language model creating method according to claim 1, wherein a mail or a sentence written by a user is used as the second corpus. 8. The language model creating method according to claim 1, wherein a sentence is selected from a plurality of corpora based on the first corpus. 9. A speech recognition method, wherein speech recognition is performed using a language model created based on the language model creation method according to any one of claims 1 to 8. 10. An example sentence corpus storage unit for storing example sentences in advance, a large scale corpus storage unit for storing a large scale corpus, and a sentence similar to the example sentence stored in the example sentence corpus storage unit for the large scale corpus storage unit. Language selection unit for selecting a text from the large-scale corpus and extracting a text, and a language model learning unit for learning a language model using the selected corpus and the example sentence corpus. Model creation device. 11. The text selection unit divides a target sentence for which similarity is to be determined into units corresponding to morphemes, obtains a notational correspondence, and determines that a sentence belonging to the first corpus is a second sentence. 11. The language model creating device according to claim 10, wherein a scale of how much the sentence belonging to the corpus is reproduced and a scale of how accurately the sentence is reproduced are used as the similarity. 12. The text selecting unit divides a target sentence whose similarity is to be determined into units corresponding to morphemes, obtains correspondences, and a sentence belonging to the first corpus becomes a second corpus. 11. The language model creating apparatus according to claim 10, wherein a scale of how much the sentence is reproduced and a scale of how accurately the sentence are reproduced are combined and used with appropriate weights as the similarity. 13. The language model creating apparatus according to claim 11, wherein the similarity is calculated by weighting each unit corresponding to a morpheme. 14. A voice recognition device, which performs voice recognition using a language model created by the language model creation device according to any one of claims 10 to 13.