KR20180057922A - Apparatus and method of speech-recognition error correction based on linguistic post-processing - Google Patents

Abstract

The present invention relates to a postprocessing-based speech recognition error correction apparatus using linguistic analysis and a method thereof which can correct an error created in speech recognition by postprocessing for the corresponding error. The apparatus comprises: a speech recognition unit to recognize an inputted speech, and output a recognized result in a text form; a linguistic recognition unit to perform morpheme analysis on the outputted text sentence and use a morpheme analysis result to perform syntax analysis and semantic analysis into a unit of a dependency structure for a syntax analysis result; an error candidate search unit to use preset collocation information to search analysis results analyzed by the morpheme analysis, the syntax analysis, and the semantic analysis for an error candidate; a right answer candidate search unit to identify possible right answer candidates for the searched error candidate to finally generate a single right answer result; and a text generation unit to use the generated right answer candidate to generate error-corrected final text for the inputted speech.

Description

BACKGROUND OF THE INVENTION 1. Field of the Invention [0001] The present invention relates to an apparatus and a method for correcting a speech recognition error based on post-

The present invention relates to an apparatus and method for correcting a speech recognition error based on post-processing using language analysis and, more particularly, to an apparatus and method for correcting a speech recognition error based on post- The present invention relates to a speech recognition error correction apparatus and method therefor.

Recently, various application technologies and systems using speech recognition technology such as an interactive system, an automatic interpretation system, and a navigation voice interface have been emerging in accordance with the improvement of recognition performance of speech recognition technology.

Although various speech recognition based services, systems, and techniques require accurate speech recognition results, speech recognition results can not be 100% accurate.

Accordingly, attempts have been made to improve speech recognition performance in various ways. These attempts can be classified into two types.

The first is an attempt to improve the performance of the speech recognition technology itself, and the second is an attempt to improve the performance through post-processing of the speech recognition result.

A typical example of performance improvement of speech recognition technology itself is an attempt to apply DNN (Deep Neural Network) to speech recognition field. As a result, the speech recognition technology achieves an improvement over the existing performance.

However, even if the performance of the speech recognition technology itself is improved as described above, it is not possible to obtain a 100% correct result, and the speech recognition result still includes an error. Typically, unregistered words or technical terms are still difficult to solve with current speech recognition technology.

Also, vocabularies that are not included in the language model used in speech recognition are likely to cause errors in speech recognition results. The language model is the data that judges the possibility of vocabulary to be recognized. It uses the local collocation of the vocabulary. If it is out of the regional peculiarity, the possibility of error occurs.

A typical type of performance improvement through post-processing is a method of constructing an error pattern that occurs in the speech recognition result and correcting the error when matching the corresponding pattern. However, constructing such an error pattern is also time-consuming and costly, and problems arise when a domain for speech recognition is changed.

SUMMARY OF THE INVENTION Accordingly, the present invention has been made in view of the above problems, and it is an object of the present invention to provide a method and apparatus for correcting errors through post-processing, that is, a morphological analysis, a syntax analysis and a semantic analysis The present invention provides a device for correcting a speech recognition error based on a post-processing using a language analysis to identify candidates of speech recognition errors and correct errors based thereon.

According to an aspect of the present invention, there is provided an apparatus for correcting a speech recognition error based on a post-processing using a language analyzer, the apparatus comprising: a speech recognition unit for recognizing an input speech and outputting a recognized result in a text form; A language recognition unit for performing a semantic analysis on the output textual sentence and a semantic analysis on the basis of a dependency structure for the result of the syntactic analysis and the analysis result using the morphological analysis result; An error candidate search unit for searching for an error candidate using predetermined air information about analysis results analyzed through morphological analysis, syntax analysis, and semantic analysis; A correct answer candidate searching unit for recognizing possible correct candidates for the searched error candidate and finally generating one correct answer result; And a text generating unit for generating a final text in which the error is corrected with respect to the input speech using the generated correct answer candidates.

According to the present invention, a post-processing method of recognizing a speech recognition error candidate through language analysis and correcting a corresponding speech recognition error based on air information is performed on speech-recognized text, There is an advantage that improvement can be achieved.

Further, according to the present invention, there is an advantage in that it is possible to correct a speech recognition error in a long-distance context which is difficult to be processed by merely constructing intuitive air information data for a speech recognition engine error.

In addition, according to the present invention, a method for correcting a speech recognition error based on an air information has an advantage that a general person other than an expert can sufficiently participate in error improvement.

1 is a block diagram of a post-processing based speech recognition error correction apparatus using language analysis according to the present invention.
2 is a diagram showing an example of a result of a syntax analysis in the present invention.
3 is a flowchart illustrating an operation flow for correcting a speech recognition error based on post-processing using language analysis according to the present invention.

BRIEF DESCRIPTION OF THE DRAWINGS The advantages and features of the present invention and the manner of achieving them will become apparent with reference to the embodiments described in detail below with reference to the accompanying drawings. The present invention may, however, be embodied in many different forms and should not be construed as being limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the concept of the invention to those skilled in the art. And the present invention is defined by the description of the claims. It is to be understood that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention. In the present specification, the singular form includes plural forms unless otherwise specified in the specification. It is noted that " comprises, " or "comprising," as used herein, means the presence or absence of one or more other components, steps, operations, and / Do not exclude the addition.

Hereinafter, an apparatus and method for correcting a speech recognition error based on post-processing using language analysis according to the present invention will be described in detail with reference to the accompanying drawings.

FIG. 1 is a block diagram of a post-processing based speech recognition error correction apparatus using language analysis according to the present invention. FIG. 2 shows an example of a result of a syntax analysis in the present invention FIG.

As shown in FIG. 1, the post-processing-based speech recognition error correction apparatus using language analysis according to the present invention includes a speech recognition unit 10, a language analysis unit 20, an error candidate search unit 30, A candidate search unit 40 and a text generation unit 50. [ Here, the language analyzer 20 may include a morphological analyzer 21, a syntax analyzer 22, and a semantic analyzer 23.

When a voice is input, the voice recognition unit 10 recognizes the input voice and provides the recognized result to the language analysis unit 20 in the form of a text.

The morpheme analysis unit 21 of the language analysis unit 20 analyzes the text sentence provided from the speech recognition unit 10 by morpheme unit and provides the analysis result to the syntax analysis unit 22.

The syntax analysis unit 22 performs a syntax analysis using the morpheme analysis result provided through the morpheme analysis unit 21 and provides a result of the syntax analysis to the semantic analysis unit 23. [

The semantic analysis unit 23 performs semantic analysis on the basis of the syntax analysis result provided from the syntax analysis unit 22 as a unit of the dependency structure and provides the result to the error candidate search unit 40. [

The error candidates search unit 30 searches for error candidates by using predetermined air information on the analysis results analyzed through the respective steps (morphological analysis, syntax analysis, and semantic analysis) of the language analysis unit 20, I.e., an error candidate, to the correct answer candidate search unit 40. [

The correct answer candidate search unit 40 grasps possible correct candidates for the error candidates searched out by the error candidate search unit 30 to finally generate one correct answer result and outputs the generated correct answer candidate to the text generating unit 50 ).

The text generating unit 50 generates and outputs the final text of the input voice, that is, the corrected text, using the correct answer candidate searched by the correct answer candidate searching unit 40.

Hereinafter, a detailed operation of the configuration of the post-processing based speech recognition error correction apparatus using the language analysis according to the present invention will be described.

First, when a voice is input, the voice recognition unit 10 generates a voice recognition result in a text form.

The text generated by the speech recognition unit 10 is analyzed in morphological units through the morphological analysis unit 21, and the syntax analysis unit 22 receives the morpheme analysis result and generates a syntax analysis result.

The semantic analysis unit 23 performs semantic analysis on the syntactic analysis result in units of the dependency structure, and the error candidate search unit 30 searches the analysis result obtained through the language analysis unit 20, And transmits the error candidate to the correct answer candidate search unit 40. [

The correct answer candidate search unit 40 grasps possible correct candidates for the error candidates and finally generates one correct answer result. Here, since the analysis result is the result of correcting the error, the analysis result must be generated again in a text form, and the text form generation unit 50 performs the text form generation. Through these steps, error correction text is finally generated.

Hereinafter, each of the above steps will be described in detail.

The speech recognition unit 10 receives speech as input and generates text. As an example, suppose that the input voice is input as text such as " various non-powered words as educational material. &Quot;

As such, the speech-recognized result texts undergo a series of language analysis processes in case of errors. That is, as shown in FIG. 1, morphological analysis, syntax analysis, and semantic analysis are sequentially performed through the morphological analysis unit 21, the syntax analysis unit 22, and the semantic analysis unit 23.

First, when the morphological analysis is performed on the text in the speech recognition unit 10 through the morphological analysis unit 21, the speech-recognized text is analyzed as morpheme-delimited words and parts of words, Data / N + / J various / ADJ + C / E non-power / N times / N + / J cheap / V + Here, N is a noun, J is a survey, ADJ is an adjective, E is a mother, and V is a verb.

Then, the syntax analysis unit 22 grasps the syntactic relationship between the words using the morpheme analysis result in the morpheme analysis unit 21 described above. 2 is a diagram showing an example of a result of a syntax analysis analyzed by the classification analyzer 22.

Then, the semantic analysis unit 23 determines the meaning of each word on the basis of the dependency and price relation obtained through the above-described parsing. In the above example, "boat" has various meanings such as "ride", "fruit", "body part" and the like. Therefore, the semantic analysis unit 23 determines one of the meanings that is appropriate for the context. Here, the method for determining the meaning can be a vocabulary and semantic pneumatic information, a machine learning method based on a tagged learning corpus, and can be determined in various ways. Here, the method of solving with the air information can be confirmed through the example shown in Table 1 below.

Pack + ship Non-powered ship meaning Air information meaning Air information $ Ride + cheaper 30 Powerless $ ride 300 $ Fruit + chew 200 Powerless $ Fruit 5 $ Some body parts 50 Non-powered $ body part 3

The above air information can be divided into lexical air information and semantic air information. The lexical air information indicates the frequency of the lexical data between the vocabularies, and the semantic air information indicates the frequency with respect to the air data between the semantics.

In the above Table 1, only the semantic air information is shown as an example because the information required by the semantic analyzer 23 is semantic air information. However, the vocabulary air data is also used for the " cheaper + cheaper " .

The vocabulary air data is used in the error candidate search unit 30 and the correct answer candidate search unit 40 below. Such air information can be automatically constructed from a large number of corpuses or existing established data may be used.

As can be seen from Table 1 above, when "bait + cheaper" is used as a criterion, "bait" is the most frequently used in the meaning of "fruit" will be.

However, when referring to "non-powered ship", the meaning of "ship" can be decided as "ride", and when considered in its entirety, the meaning of "ship" is decided as "mount".

The error candidate search unit 30 searches for a portion determined to have a speech recognition error using the lexical and semantic air information. Basically, error candidate search is based on syntax structure.

Although the lexical n-gram can be searched by applying the lexical n-gram to the continuous vocabularies based on the morpheme analysis result in the morpheme analysis unit 21, Therefore, in the present invention, an error candidate search is performed on the basis of a syntax structure in order to consider discontinuous vocabularies.

Table 2 below is an example of air information survey based on the phrase structure of the sentence " Various non-motorized ships are used as training materials. &Quot;

Lexical air information Vocabulary air information frequency Meaning air information Meaning air information
Frequency Training materials 300 Education $ Materials 450 Bundled with material + 2 Wrapped in $ data + 10 Various + b ship 7 Various + r $ ride 350 Non-powered ship 30 Powerless $ ride 200 Chew 3 $ Ride + cheaper 30

In Table 2, the speech recognition error candidates judged by the lexical air information criterion may be "cheaper with data +", "various + b +", and "cheaper with cheaper +".

Lexical air information has high reliability for high frequency vocabulary, but low reliability for high frequency vocabulary.

Therefore, when the speech recognition error candidates are judged only by the lexical air information standard, a problem occurs in the reliability, so the semantic air information is also considered in the present invention. That is, semantic air information is more reliable than lexical air information for each entry. The candidates of speech recognition errors that are finally judged by considering the lexical air information and semantic air information are "cheaper with data +" and "cheaper with cheaper +".

In the case of " various ships ", as shown in Table 2, the frequency of vocabulary air information is low, but " various voyages " are excluded from speech recognition error candidates because the meaningful information frequency is very high. Here, the reason for not using only the semantic air information is that the reliability of each type of air information entry without considering the frequency is higher than the lexical air information, and the meaning determined at the semantic analysis step may not be accurate It is because.

In addition, speech recognition error candidates can be determined using only lexical air information. For example, in the case of "non-powered watermelon", if the frequency of "non-power" and the frequency of vocabulary of "watermelon" are high for the entire vocabulary air information, the air frequency of " The "non-powered watermelon" may be judged as a speech recognition error candidate.

On the other hand, the correct answer candidate search section 40 searches the candidate candidates for possible lexical air information determined to be an error candidate in the error candidate search section 30, and determines the final correct answer having the highest reliability among the searched correct candidates do.

The correct answer candidate search unit 40 can search for the correct answer candidates by using a combination of several methods.

The first method is to search for air information having similar pronunciation as the corresponding error candidate air information. The criterion to judge similar pronunciation is phoneme, syllable, word / phrase. For example, the similar pronunciation data in the Korean standard is as follows.

First, similar phonetic pronunciation, "ㅔ / ㅐ, ㅖ / ㅒ, ... ", And the similar pronunciation of syllable is" comb / debt / light, crab / dog, ber / ship, day / sick / ", Words / phrases similar pronunciation" beyond / beyond, value / together, weighing / improving, development / development, object / object, ... "In the case of English, the similar pronunciation" for / four, hour / our, whole / hole, whether / weather, air / heir, which / witch, ate / eight, here / hear, knight / night ... &Quot;

Both Korean and English are phonetic characters, and for phonetic characters, even though the pronunciation is the same, there is a case where the surface form is different.

Moreover, the frequency of occurrence of these homonyms varies from language to language, and English is more common than Korean. Instead of constructing a similar pronunciation dictionary as described above for pronunciation of similar pronunciation, pronunciation may be used. For example, you can create a phonetic symbol for each word / string and search for a word / string that has the same pronunciation as the phonetic symbol. That is, it can be judged by phonetic symbols as in the example of Table 3 below.

hear [hir] here [hir] Beyond [n∧m∧] [N ∧ m ∧]

In many cases, words with similar pronunciation are difficult to recognize correctly only by the language model, which is a continuous contiguous context. Therefore, it can be solved by using the syntactic air information which is a long distance context as in the present invention. By using such similar pronunciation data, candidate candidates for speech candidate candidates appearing in the sentence are generated, and speech recognition errors can be improved by applying air information that is a long-distance context to these candidates.

However, it is not possible to grasp and correct all of the speech recognition errors only by constructing the similar pronunciation. Speaking by voice has a lot of individual differences, and there are many cases where the same character is uttered differently. In order to consider such a personalized utterance deviation, the present invention searches for a second correct answer candidate, identifies a pronunciation error type for each user, and applies the error type to a similar pronunciation pronunciation candidate. Let's take the case of pronouncing "rice" as "sal".

Some local people in Korea often pronounce "ㅆ" pronounced at the beginning of a word as "ㅅ". Typically, they often pronounce "rice" as "sal." As in this case, the correct candidate for the speech recognition error of the present invention can be extended by searching for the correct candidate by varying the error pronunciation candidate for each user or region.

The difference of the speech recognition error may be different for each individual, and the error type differs for each speech recognition system. This is due to differences in learning data and learning techniques applied to the development of speech recognition technology. In the present invention, the third correct answer candidate is searched. By considering the difference between the speech recognition systems, the correct answer candidate for the additional speech recognition error can be extended.

For example, "writes" is "wrapped", "written" is "wrapped", ... An example of the occurrence of an error by recognizing "write" as "cheap" is one of the most common speech recognition errors that occur in currently applied speech recognition systems. When considering the characteristics of such a food recognition error, the candidate for the speech recognition error candidate can be identified as shown in Table 4 below.

Errors Candidates Candidates for answers Bundled with material + Write into the material + Chew Spend the boat +

Table 5 below shows an example of obtaining the frequency of air information for the candidate candidates thus identified.

Lexical air information Lexical air information
Frequency Meaning air information Meaning air information
Frequency Bundled with material + 2 Wrapped in $ data + 10 Write into the material + 25 Write to $ data + 150 Chew 3 $ Ride + cheaper 30 Spend the boat + 20 Spend $ ride + 100

As shown in Table 5, when the frequency of air information is changed from "cheap" to "stroked", it can be seen that the frequency is greatly increased, so that it can be understood that "write" is used as a more correct vocabulary.

Therefore, "cheap" generated from the original speech recognition result is corrected to "write". At this time, the air frequency can be simply used as the reliability of the candidate for the answer, and various reliability measurement methods combining the lexical air information and the semantic air information can be applied. Simple reliability through simple summation of frequency can be expressed as Equation 1 below.

Using Equation (1), the total reliability of the air information is obtained by simply adding the lexical-air-information reliability and the semantic-air information frequency to the reliability score.

If the frequency of occurrence of the individual air information of each lexical air information is considered, the following Equation 2 can be used.

In addition, the method of determining the reliability of the air information can be variously configured in consideration of the respective characteristics of the air information data.

On the other hand, the text generation unit 50 shown in FIG. 1 performs an operation of converting the error correction result of the morpheme unit into the error correction result of the character string unit. This step is necessary in the case of a pseudo-word that combines two or more morphemes, such as Korean, to form a single word. Especially, it is an essential step in the case of Korean, in which the result of morpheme-combining is morphological transformation rather than simple morphological combination result. In the above example, when the text generation is performed on the result of the error correction, the text may be generated as " write = write ".

In other words, it is necessary to combine the words "Tsu" and "tsu" to form a simple morpheme of "writing" rather than "writing". If spaces are used as morpheme units without morpheme combination like English, the text generation module becomes much simpler since it only needs to correct morpheme vocabulary itself considering sex, number, tense, and personality.

In the case of Korean, morphing occurs when combined with irradiation and mother, mostly when combined with mother. That is, the following examples are the same.

a1: I / N + / J => I

a2: I / N + / J => I

b1: Write / V + Go => Write

b2: Write / V + => Write

c1: Go / V + Go => Grind

c2: Going / V + = Going

c3: Go / V + Go => Go

As shown in the above example, when the morpheme-based deformation occurs, only the last syllable (1L) of the first morpheme where the combination is made and the first syllable (2F) of the second morpheme are related. That is to say, "I / N + B / J, W / V + B / E, G / V +

Therefore, if morphological deformation occurs by morpheme-binding, constructing the morphological join table as shown in Table 6 below considering only the last syllable of the first morpheme and the first syllable of the second morpheme, (The first phoneme can be used instead of the first syllable in the second morpheme).

First stem
ultima Second morpheme First syllable Combined string I N I Writing N written Go N liver Go Nee Gani

First, in the case of a combination of nouns and surrogates, if the first syllable of the mother and the first syllable of the search has a prefix, simply combine the morpheme string.

Thus, in Korean, the application of morpheme deformation is as follows: when the first syllable of the mother and the first syllable of the survey are absent, the morpheme combination in which deformation takes place by constructing the last syllable of the preceding syllable and the first syllable of the backward syllable Can be processed.

3 is a flowchart illustrating a method of correcting a speech recognition error based on post-processing using language analysis according to the present invention, which corresponds to the operation of the post-processing based speech recognition error correction apparatus using language analysis according to the present invention. Let's take a step-by-step look at it.

FIG. 3 is a flowchart illustrating a method of correcting a speech recognition error based on post-processing using language analysis according to the present invention. Here, the detailed operation of each step shown in FIG. 3 will be omitted because it is the same as the operation of each component in the above description of the apparatus.

As shown in FIG. 3, it is determined whether a voice is input first (S301). If a voice is inputted, the voice is recognized and the recognized result is outputted in text form (S302).

Next, the outputted text sentence is analyzed in morphological unit (S303), and syntax analysis is performed using the morpheme analysis result (S304).

Then, the semantic analysis of the dependency structure is performed on the result of the syntax analysis (S305).

In step S306, error candidates are searched using predetermined air information on the analysis results analyzed through steps S303 to S305 (morphological analysis, syntax analysis, and semantic analysis).

Then, a candidate correct answer candidate is searched for the searched error candidate, and one correct answer result is finally generated (S307). Then, the final text of the inputted speech, that is, the error is corrected And generates and outputs a text (S308).

The present invention is not limited to the specific embodiments but may be modified and changed without departing from the scope of the present invention. And various modifications, alterations, and changes may be made without departing from the scope of the present invention.

Therefore, the embodiments described in the present invention and the accompanying drawings are intended to illustrate rather than limit the technical spirit of the present invention, and the scope of the technical idea of the present invention is not limited by these embodiments and accompanying drawings . The scope of protection of the present invention should be construed according to the claims, and all technical ideas within the scope of equivalents should be interpreted as being included in the scope of the present invention.

10:
20: Language Analysis Department
21: Morphological Analysis Department
22: parsing section
23: Semantic Analysis Department
30: Error candidate search unit
40: Correct candidate candidate search section
50:

Claims (1)

A voice recognition unit for recognizing the inputted voice and outputting the recognized result in the form of a text;
A language recognition unit for performing a semantic analysis on the output textual sentence and a semantic analysis on the basis of a dependency structure for the result of the syntactic analysis and the analysis result using the morphological analysis result;
An error candidate search unit for searching for an error candidate using predetermined air information about analysis results analyzed through morphological analysis, syntax analysis, and semantic analysis;
A correct answer candidate searching unit for recognizing possible correct candidates for the searched error candidate and finally generating one correct answer result; And
And a text generator for generating an error corrected final text for the input speech using the generated correct answer candidates.

Images