KR20040055292A - System and method for improving in-domain training data using out-of-domain data - Google Patents

Abstract

PURPOSE: A system and a method for reinforcing a domain subordinated training corpus using the corpus out of a domain are provided to improve a performance of a language model for the continuous speech recognition by using the training corpus out of the related domain. CONSTITUTION: The domain corpus(205) has the language model data of a small size for a predetermined domain. A similar contents corpus(202) has the contents similar to the domain corpus but a form different from the domain corpus while having the language model data different from the domain corpus. A form standardization part(204) standardizes the form of the similar contents corpus to the same form as the domain corpus. An adder(209) generates the corpus reinforcing the language module by adding the corpus standardized in the form standardization part to the domain corpus.

Description

Zone-dependent training corpus reinforcement system and its method using out-of-domain corpus {System and method for improving in-domain training data using out-of-domain data}

The present invention relates to Large Vocabulary Continuous Speech Recognition (LVCSR), and more particularly, to a training corpus reinforcement system and method using an out-of-area training corpus.

The most successful language modeling technique for speech recognition systems is statistical language modeling, which requires a training corpus (text database) of sufficient size to have reliable and robust statistics. Also, a statistical language model built on a certain domain may not show proper performance if the domains are different. In other words, if the new domain-dependent training corpus is lacking, the statistical language model does not contribute to speech recognition performance and causes performance deterioration. In addition, there is a problem that area-dependent corpus securing is not easy.

1 is a block diagram schematically illustrating a general continuous speech recognition system.

Referring to FIG. 1, the feature extractor 101 extracts only information useful for recognition from an input voice and converts the information into a feature vector.

The searcher 102 uses the acoustic model database 104, the pronunciation dictionary database 105, and the language model database 106, which are obtained in advance in the learning process from the feature vectors output from the feature extractor 101, and have the highest probability. Find a string of words using the Viterbi algorithm. Here, the recognition target vocabularies form a tree for the recognition of the large vocabulary, and the search unit 102 searches the tree.

Finally, the recognition result output unit 103 outputs the recognized text using the output of the search unit 102.

As in the system shown in Fig. 1, the statistical language model in the dialogue continuous speech recognition system for the purpose of speech recognition in various areas can be constructed very reliable and robust when the training corpus is sufficiently secured. Here, the language model may be referred to as the grammar of the speech recognition system. The continuous speech recognition system does not recognize any sentence, but only sentences that meet certain grammars. By using the language model in the search process of the speech recognition system, it is possible to reduce the search space of the speech recognition system and contribute to the improvement of the recognition rate because it plays a role of increasing the probability of sentences that match the grammar.

As described above, statistical language modeling among the various language modeling techniques has the best performance in recognition of large vocabulary continuous speech. Statistical language model is a grammar in which word-to-word relationships are expressed as probabilities. The n-gram, which is used frequently in statistical language models, is a grammar that defines the probability of the next word from n-1 words in the past. The commonly used n-gram is a bigram (n = 2) and a trigram (n = 3). )to be. The advantage of the statistical language model is that it defines everything as a probability, so it does not require much human knowledge and can be easily implemented with a large amount of corpus. However, if the corpus is small, a reliable probability value may not be obtained, but may cause performance deterioration.

After all, the probability values of statistical language models require a huge amount of corpus in order to have reliable values, and collecting them is a tremendous amount of effort and time. In particular, it is never easy to secure enough dialogue-type sentences that occur in travel counseling or everyday life, and therefore new techniques are needed to build a robust language model from insufficient text data.

The technical problem to be achieved by the present invention is to overcome the problem of performance degradation of the statistical language model due to the lack of the region-dependent training corpus which is commonly generated when constructing the statistical language model for continuous speech recognition, by using the continuous training corpus It is to provide a training corpus reinforcement system that can improve the performance of language model for speech recognition.

Another technical problem to be achieved by the present invention is to provide a corpus reinforcement method performed in the training corpus reinforcement system.

Another technical problem to be achieved by the present invention is to provide a recording medium in which the training corpus reinforcement method is recorded as program code executable on a computer.

1 is a block diagram schematically illustrating a general continuous speech recognition system.

2 is a block diagram schematically illustrating an embodiment of an area dependent training corpus reinforcement system using an out of zone training corpus according to the present invention.

3 is a block diagram illustrating in detail the formal standardization unit 204 in the system shown in FIG. 2.

In order to achieve the above object, the corpus reinforcement system according to the present invention has a regional corpus having small language model data for a specific region, but has language model data for a region different from the region corpus, but the content is similar to that of the region corpus. Includes format-like corpus with different format, corpus standardized to corpus-like corpus, and adder to add corpus standardized by corpus to corpus which is standardized by corpus, It is desirable to.

In order to achieve the above object, the corpus reinforcement system according to the present invention has a regional corpus having small language model data for a specific region, but has language model data for a region different from the region corpus, but the format is similar to that of the region corpus. The language of the domain corpus by adding weights to the language model of the domain corpus in the adder that adds the form-like corpus of different contents, the form-like corpus to the region corpus, and the corpus reinforced by the language model. It is preferable to include a weight adjusting unit for increasing the probability that the model is selected.

In order to achieve the above object, the corpus reinforcement system according to the present invention has a regional corpus having small language model data for a specific region, but has language model data for a region different from the region corpus, but the content is similar to that of the region corpus. It has the same language model data as the domain corpus, which has the same language format data as the domain corpus. A first adder that adds the corpus normalized by the formal standardizer, the corpus standardized by the formal standardizer to the region corpus, and a second corpus that adds the form-like corpus to the region corpus to generate the corpus enhanced by the language model Language Model of Area Corpus in Adder and Reinforced Corpus Generated by Adder To give a higher weight including a weight adjusting unit to increase the probability that the language models in the corpus area is selected is preferred.

In order to achieve the above another problem, a corpus reinforcement method according to the present invention for reinforcing a corpus having a small language model data for a specific region by using an outer corpus having a language model data for a region different from the region corpus In step (a) of extracting a content-like corpus that is similar in content but different in form to the region corpus from the non-area corpus, and in steps (b) and (b) normalizing the format of the content-like corpus to the same format as the region corpus. And adding (c) the standardized corpus to the region corpus to produce a corpus enriched in the language model.

In order to achieve the above another problem, a corpus reinforcement method according to the present invention for reinforcing a corpus having a small language model data for a specific region by using an outer corpus having a language model data for a region different from the region corpus (A) extracting a form-like corpus that is similar in format to the region corpus but has different contents in the non-domain corpus; (b) generating a corpus reinforced with a language model by adding a form-like corpus to the region corpus; In the corpus generated in step b), it is preferable to include the step (c) of adjusting the weight so that the weight of the language model of the area corpus is increased by giving a weight to the language model of the area corpus.

In order to achieve the above another problem, a corpus reinforcement method according to the present invention for reinforcing a corpus having a small language model data for a specific region by using an outer corpus having a language model data for a region different from the region corpus (A) extracting a similar-like corpus of similar but different format but different-form corpus from a region-based corpus in the out-of-area corpus. In the steps (c) and (c), the corpus reinforced with the language model is generated by adding the corpus normalized in the steps (b) and (b) and the form-like corpus respectively to the region corpus. In the reinforced corpus, the language model of the region corpus is selected by giving a high weight to the language model of the region corpus. To include the (d) step of adjusting the weight so as to be higher the probability is preferred.

Hereinafter, an area-dependent training corpus reinforcement system using the out-of-area training corpus according to the present invention and a method thereof will be described as follows with reference to the accompanying drawings.

2 is a block diagram schematically illustrating an embodiment of an area dependent training corpus reinforcement system using an out of zone training corpus according to the present invention. The training corpus reinforcement system according to the present invention includes an out-of-area corpus 201, a formal standardizer 204, first and second adders 209 and 210, region corpus 205, and first and second reinforced corpus 206 and 207. ) And a unigram value adjusting unit 208.

Referring to Fig. 2, region corpus 205 is a corpus having small language model data for a specific region. Hereinafter, for convenience of explanation, the area corpus 205 is assumed to be an interactive corpus associated with the trip.

The out-of-area corpus 201 is a corpus having language model data for an area other than the area corpus 205. In this case, the out-of-area corpus 201 may be classified into a form-like corpus 203 that is similar in content but similar in form to a content-like corpus 202 whose content is similar to that of the region corpus 205. .

The format standardizing unit 204 extracts the content-like corpus 202 from the out-of-area corpus 201 and converts the extracted content-like corpus 202 into the form of the region corpus 205, that is, into a travel-related conversation form. Change.

The first adder 209 may generate the first reinforced corpus 206 in addition to the corpus and the region corpus 205 changed from the form normalization unit 204 to the region corpus 205. That is, the language model data of the area corpus 205 may be changed into a form of the area corpus 205 by using the lack of language model data to enrich the language model data of the area corpus 205. .

The second adder 210 may extract the form-like corpus 203 from the out-of-area corpus 201 and add it to the region corpus 205 to generate the second reinforced corpus 207. That is, when the format is similar to that of the area corpus 205, the contents thereof are added to the area corpus 205 as it is.

The weight adjusting unit 208 gives a weight higher to the unigram probability value of the language model of the region corpus 205 than the form-like corpus 203 added by the second reinforced corpus 207, thereby increasing the weight of the region corpus 205. Increase the probability of selecting a language model. The reason is that in the case of the added type-like corpus 203, the content may be different from that of the area corpus 205, so that the unigram value of the area corpus 205 is adjusted to adjust the unigram value of the area-like corpus 205 rather than the form-like corpus 203. Increase the probability that the language model of.

As described above, when reinforcing the language model data of the area corpus 205 using the out-of-area corpus 201, the corpus (reinforced by the content-like corpus 202 or the form-like corpus 203) is reinforced ( 206 and 207, respectively.

For example, consider the following cases of [Travel Counseling-Reading] and [Travel Counseling-Dialogue].

[Travel Counseling]

Passport issuance documents are two passport photos and ID card.

You can apply for the place of issue at the Passport section of Gangnam-gu Office on the 1st floor of the airport terminal. For details, please contact the Gangnam-gu Office Passport.

[Travel Consultation-Conversation]

A: Uh, I just want to change the money.

B: There is a currency exchange office on the ground floor here, and a duty free shop on the third floor.

As mentioned above, texts with a theme of travel and also have a conversational feature can be used naturally to build a language model of the travel field of the conversation. However, similar [Travel Counseling-Reading] data cannot be used. This is because the [Travel Counseling-Reading] data exhibits different characteristics in form and sentence style than in the conversational field of conversational conversation. As such, the text data, which is the subject of travel but does not have an interactive style, should be converted through the format standardization unit 204 to have an interactive element in terms of style.

When converting the [travel consultation-reading reader] to have an interactive style through the formal standardization unit 204, it is as follows.

[Travel Consultation-Reading: After Conversion]

Passport issuance documents are two passport photos and an ID card.

You can apply at the Gangnam-gu Office Passport Department on the 1st floor of the airport terminal. For more details, contact the Passport Department at Gangnam-gu Office.

As such, by performing the format conversion through the formal standardization unit 204, it is possible to use the [travel consultation-reading body] to construct the language model of the travel field of the conversational body. Specifically, formal standardization is based on morphological tagging (indicating parts of speech for each morpheme vocabulary) on the original text data, and which parts of speech that exhibit interactive features and which parts of speech that exhibit textual features exist. If the part-of-speech is found out of the data beforehand, the part-of-speech part converts the part-of-speech into an interactive part of speech. The formal standardization unit 204 will be described in more detail with reference to FIG. 3.

Continuing with this, let us consider the following cases of [conversation at the airport while traveling] and [conversation at the hotel while traveling], that is, the form is similar as the dialogue but the contents are different.

[Dialogue at the airport while traveling]

A: Uh, I just want to change the money.

B: There, the exchange office is on the first floor, and the duty free shop is on the third floor.

[Dialogue at hotel while traveling]

A: Well, I want to use the health club, but where are the facilities?

B: There is a fitness club, the swimming pool is outdoors, the fitness room is on the ground floor, the sauna is on the ground floor, the hot spring is on the ground floor, the bowling alley is on the ground floor, and the recreation room is on the ground floor.

A: Yes. And where is the restaurant?

B: There is a tavern on the first floor, and there is a Japanese and Western restaurant on this floor.

If you go to the third floor, you will find a folk restaurant and coffee shop, so go to the right place.

If you want to build a language model for [conversation at the airport while traveling] and you don't have enough area corpus, you can use the above [conversation at hotel while traveling] as a corpus outside the area. In other words, in the case of [conversation in a hotel during a trip], the keyword is different from [conversation in an airport during a trip] in terms of the subject, but the pattern or style of the speech is similar to a corpus. As described above, when constructing a language model using a format-like corpus having different contents but similar styles, the weight adjusting unit 208 increases the weight to the probability value of the airport vocabulary shown in [Dialogue at the airport while traveling]. Increase the chance of being selected.

As mentioned above, by using out-of-area training corpus to solve the problem of lack of domain-dependent training corpus which causes the performance degradation of statistical language model for continuous speech recognition, the language model for speech recognition will be more reliable and robust. Can be.

3 is a block diagram illustrating in detail the formal standardization unit 204 in the system shown in FIG. 2. Referring to FIG. 3, the formal standardization unit 204 includes a morpheme analysis unit 302, a part-of-speech information analysis unit 303, and a text normalization unit 304 and an out-of-area corpus 305.

Referring to FIGS. 2 and 3, the morpheme analysis unit 302 analyzes the morphemes of the content-like corpus 202 that is similar in content to the area corpus 205 among the out-of-area corpus 201 to form parts of speech for each morpheme word. Morphological tagging is used.

The part-of-speech information analysis unit 303 analyzes which parts of speech are represented as interactive features and which are parts of speech which are read aloud by using the result of the morpheme tagging by the morpheme analysis unit 302.

The text normalization unit 304 generates the out-of-area corpus that is transformed by using the parts-of-speech analysis part 303 and the part-of-speech transformed to have a format similar to that of the area corpus 205. For example, if the format of the area corpus 205 is a conversational form and the content-like corpus 202 of the out-of-area corpus 201 is a reading form, the text normalization unit 304 is part-of-speech-analyzed by the part-of-speech information analyzer 303. In the result, the reading part-of-speech is transformed into a conversational part-of-speech to change the format of the content-like corpus 202 to the form of the area corpus 205.

Through the morphological analysis, the part-of-speech information analysis, and the text standardization process as described above, the content-like corpus 202 becomes similar to the region corpus 205 as well as its contents.

The invention can also be embodied as computer readable code on a computer readable recording medium. The computer-readable recording medium includes all kinds of recording devices in which data that can be read by a computer system is stored. Examples of computer-readable recording media include ROM, RAM, CD-ROM, magnetic tape, floppy disk, optical data storage, and the like, which are also implemented in the form of a carrier wave (for example, transmission over the Internet). It also includes. The computer readable recording medium can also be distributed over network coupled computer systems so that the computer readable code is stored and executed in a distributed fashion.

The best embodiments have been disclosed in the drawings and specification above. Although specific terms have been used herein, they are used only for the purpose of describing the present invention and are not used to limit the scope of the present invention as defined in the meaning or claims. Therefore, those skilled in the art will understand that various modifications and equivalent other embodiments are possible from this. Therefore, the true technical protection scope of the present invention will be defined by the technical spirit of the appended claims.

As described above, according to the corpus reinforcement system and the method using the out-of-area corpus, the out-of-area training to solve the problem of the region-dependent training corpus shortage that causes the performance degradation of the statistical language model for continuous speech recognition By using corpus, the language model for speech recognition can be more reliable and robust.

Claims (15)

An area corpus with small language model data for a particular area; A content-like corpus having language model data for a region different from the region corpus, the content being similar to the region corpus but having a different format; A format standardizing unit for standardizing a format of the content-like corpus into a format such as an area corpus; And And an adder for adding a corpus normalized by the formal standardization unit to the region corpus to generate a corpus reinforced with a language model. The method of claim 1, wherein the formal standardization unit A morpheme analysis unit configured to analyze morphemes of the content-like corpus and to tag parts of speech for each morpheme vocabulary; A part-of-speech information analysis unit for analyzing which parts of speech representing a formal feature are found by using the result of the morphological tagging by the morphological analysis unit; And And a text normalization unit configured to generate an out-of-area corpus by transforming the parts-of-speech to have a form similar to the area corpus by using the parts-of-speech analysis part of speech analysis part of the part-of-speech information analyzing unit. An area corpus with small language model data for a particular area; A format-like corpus having language model data for a region different from the region corpus, the format being similar to the region corpus but having different contents; An adder for adding the form-like corpus to the region corpus to generate a corpus reinforced with a language model; And And a weight adjustment unit for assigning a high weight to the language model of the area corpus in the reinforced corpus generated by the adder to increase the probability of selecting the language model of the area corpus. The method of claim 3, wherein the weight adjusting unit A corpus reinforcement system, characterized in that to give a high unigram probability value to the language model of the area corpus. An area corpus with small language model data for a particular area; A content-like corpus having language model data for a region different from the region corpus, the content being similar to the region corpus but having a different format; A format-like corpus having language model data for a region different from the region corpus, the format being similar to the region corpus but having different contents; A format standardizing unit for standardizing a format of the content-like corpus into a format such as an area corpus; A first adder for adding a corpus normalized by the formal standardizer to the region corpus to generate a corpus reinforced with a language model; A second adder for adding the form-like corpus to the region corpus to generate a corpus reinforced with a language model; And And a weight adjustment unit for assigning a high weight to the language model of the area corpus in the reinforced corpus generated by the adder to increase the probability that the language model of the area corpus is selected. The method of claim 5, wherein the formal standardization unit A morpheme analysis unit configured to analyze morphemes of the content-like corpus and to tag parts of speech for each morpheme vocabulary; A part-of-speech information analysis unit for analyzing which parts of speech representing a formal feature are found by using the result of the morphological tagging by the morphological analysis unit; And And a text normalization unit configured to generate an out-of-area corpus by transforming the parts-of-speech to have a form similar to the area corpus by using the parts-of-speech analysis part of speech analysis part of the part-of-speech information analyzing unit. The method of claim 5, wherein the weight adjusting unit A corpus reinforcement system, characterized in that to give a high unigram probability value to the language model of the area corpus. A corpus reinforcement method in which a corpus reinforcement having a small language model data for a specific region is corpus reinforced using an out-of-region corpus having language model data for a region different from the region corpus. (a) extracting a content-like corpus in the out-of-area corpus, the content of which is similar to the area corpus but in a different format; (b) normalizing the format of the content-like corpus to the same format as the region corpus; And (c) adding a corpus normalized in the step (b) to the region corpus to generate a corpus reinforced with a language model. The method of claim 8, wherein step (b) (b1) analyzing the morphemes of the content-like corpus and performing morpheme tagging to mark parts of speech for each morpheme vocabulary; (b2) analyzing any parts of speech representing formal features using the morphologically tagged result in step (b1); And and (b3) generating a corpus transformed out-of-area corpus by forming a part-of-speech transformation to have a form similar to that of the area corpus by using the part-of-speech analysis of the part-of-speech analysis in step (b2). A corpus reinforcement method in which a corpus reinforcement having a small language model data for a specific region is corpus reinforced using an out-of-region corpus having language model data for a region different from the region corpus. (a) extracting a type-like corpus in the out-of-area corpus that has a similar format to the area corpus but has different contents; (b) adding the form-like corpus to the region corpus to generate a corpus reinforced with a language model; And and (c) a weighting step of giving a weighting value of the language model of the area corpus in the corpus generated in step (b) to increase the probability of selecting the language model of the area corpus. Reinforcement method. The method of claim 10, wherein step (c) A corpus reinforcement method characterized by giving a high unigram probability value to the language model of the region corpus. A corpus reinforcement method in which a corpus reinforcement having a small language model data for a specific region is corpus reinforced using an out-of-region corpus having language model data for a region different from the region corpus. (a) extracting a content-like corpus in the out-of-area corpus that is similar in content but different in form to the area corpus, and a form-like corpus in which the form is similar to but different in content from the area corpus; (b) normalizing a format of the content-like corpus to a form such as a region corpus; (c) adding a corpus normalized in the step (b) and the form-like corpus to the region corpus, respectively, to generate a corpus reinforced with a language model; And (d) weighting the language model of the area corpus in the reinforced corpus generated in step (c) to increase the probability of selecting a language model of the area corpus; Corpus reinforcement method. The method of claim 12, wherein step (b) (b1) analyzing the morphemes of the content-like corpus and performing morpheme tagging to mark parts of speech for each morpheme vocabulary; (b2) analyzing any parts of speech representing formal features using the morphologically tagged result in step (b1); And and (b3) generating a corpus transformed out-of-area corpus by forming a part-of-speech transformation to have a form similar to that of the area corpus by using the part-of-speech analysis of the part-of-speech analysis in step (b2). The method of claim 12, wherein step (d) A corpus reinforcement method characterized by giving a high unigram probability value to the language model of the region corpus. The recording medium which recorded the corpus reinforcement method of any one of Claims 8-14 with the program code which can be executed by a computer.