KR101388569B1 - Apparatus and method for adding new proper nouns to language model in a continuous speech recognition system - Google Patents

Abstract

The present invention relates to a technique for adding a proper noun of a language model in a continuous speech recognition system, collecting new proper nouns and classification words, selecting candidate sentences from a text corpus using collected words, and selecting candidate sentences from candidate sentences. Extract the candidate frame, generate an engram of the candidate door frame, score each candidate door frame using statistical calculations, apply the upper door frame to the proper noun, and expand the engram to reflect the language model. Characterized in that. According to the present invention, since a new proper noun that is not in a language model for speech recognition can be added in the form of an engram reflecting various expressions of sentences, a higher speech recognition performance can be obtained than a method of simply adding a proper noun only. have.

Description

Apparatus and method for adding proper nouns of language models in continuous speech recognition system {APPARATUS AND METHOD FOR ADDING NEW PROPER NOUNS TO LANGUAGE MODEL IN A CONTINUOUS SPEECH RECOGNITION SYSTEM}

The present invention relates to a technique for adding a proper noun to a language model of continuous speech recognition, in particular, a continuous word suitable for reflecting independently collected new proper nouns not included in a speech recognition vocabulary dictionary and a training model for language models. The present invention relates to a device and a method for adding a proper noun of a language model in a speech recognition system.

In general, speech recognition methods are divided into several types according to the vocalization behavior. Isolated word recognition, connected word recognition, continuous speech recognition, and keyword spotting are used. have. Unlike the case of isolated word recognition in which individual words are recognized, continuous word recognition is a method for finding sentences or consecutive word strings corresponding to a voice signal. As the number of words in the vocabulary dictionary increases, The more pronounced the number of words due to the variation of pronunciation between the word and the word, the more likely it is to be mistaken for words of similar pronunciation.

The language model in speech recognition is a model constructed by collecting the connectivity between words from the text corpus in a statistical way so that the user's spoken sentence is recognized as the correct sentence. The language model uses a lot of 1-gram, 2-gram, and 3-gram. A unigram uses the probability of a word and does not use a word in the past immediately preceding it. Biagrams and trigrams use probabilities that depend on the first one and two words, respectively. The use of such a language model allows grammatically valid word strings to be recognized, minimizing the search space of words or sentences, thereby improving recognition performance and reducing search time.

The text corpus for training a language model does not contain enough proper nouns, so to speech-recognize natural language sentences that include new proper nouns that are not registered, you can intentionally enter a sentence that contains a new proper noun in the text corpus, or type a proper noun. I had to enter it again and again. However, in the former case, when the number of proper nouns to be added is large, it is very inefficient because each proper noun has to write a sentence covering various expressions. In the latter case, only the proper noun is input in the form of unigram. Since the information about proper nouns and words appearing adjacent to each other is not reflected in the language model, the recognition rate has to fall.

Korean Laid-Open Patent Publication No. 10-2006-0092683, 2006.08.23

In the conventional methods of adding proper nouns as described above, the engram of the proper noun can be extracted from the text corpus for the language model. However, not only most proper nouns appear in the text corpus, but also there are many randomly collected sentences, and various types of proper nouns do not appear evenly, and the context in which each proper noun appears is different.

Accordingly, an embodiment of the present invention can provide an apparatus and method for adding a proper noun for a language model in a continuous speech recognition system capable of adding a new proper noun to an language model in the form of an engram.

In addition, an embodiment of the present invention can provide a device and method for adding a proper proper noun of a language model in a continuous speech recognition system that can be used to generate an engram of proper nouns by extracting a general context that can be used for various proper nouns from a text corpus. have.

In addition, the embodiment of the present invention collects new proper nouns and classification words for the addition of proper nouns to the language model in the continuous speech recognition system, select candidate sentences from the text corpus using the collected words, candidate sentences After extracting candidate door frames from them, generating engrams of candidate door frames, scoring and ranking each candidate door frame using statistical calculations, applying the upper door frames to proper nouns, and then expanding the engrams to language. It is possible to provide an apparatus and method for adding a proper noun for a language model in a continuous speech recognition system that can be reflected in a model.

Apparatus for adding a proper noun of a language model in the continuous speech recognition system according to an embodiment of the present invention, a collection unit for collecting new proper nouns and classification words in the proper noun adding apparatus of the language model, and the new proper nouns and classification A candidate sentence selecting unit that selects a match sentence as a candidate sentence by searching for a word in a text scope, a candidate door frame extracting unit extracting a candidate door frame from the candidate sentence, and extracting an engram-type candidate door frame from the candidate door frame. Substituting the proper nouns collected in the candidate door frame generation unit of the engram, the ranking unit for ranking by high score through the calculation of the score for each of the engram-type candidate door frame, An engram extension unit that performs engram expansion, and an engram generated by the engram extension is added to the language model with frequency. Reflecting may include portions.

The collection unit may collect proper nouns that do not appear in the training text corpus, and may assign a category that is conceptually classified from the proper nouns.

The candidate door frame extractor may extract a local context having a predetermined length including the proper noun in the candidate sentence in the form of a door frame.

The ranking unit may measure the suitability of the door frame based on the number of door frames including proper nouns and the number of door frames in the entire corpus for each candidate door frame of the engram type, and the number of door frames and proper state diagrams of the proper nouns. The entropy of the door frame can be calculated based on the number.

The engram expansion unit may select at least one upper score door frame and substitute each collected proper noun at a corresponding position of the door frame.

In the continuous speech recognition system according to an embodiment of the present invention, a method for adding a proper noun of a language model includes: collecting a new proper noun and a classification term in a device for adding a proper noun of a language model; Searching for classification words in a text scope, selecting matching sentences as candidate sentences, extracting candidate door frames from the selected candidate sentences, and extracting engram-type candidate door frames from the extracted candidate door frames. And a process of ranking by high scores by calculating a score for each of the extracted engram candidate door frames, and performing engram expansion by substituting proper nouns collected in the ranked engram door frames. And adding the engram generated by the engram extension with the frequency to the language model.

The collecting process may collect proper nouns that do not appear in the training text corpus, and may assign a category that is conceptually classified from the proper nouns.

The extracting of the candidate door frame may include extracting a local context of a certain length including the proper noun in the candidate sentence in the form of a door frame.

The ranking of the scores may include measuring the suitability of the door frame based on the number of door frames including proper nouns and the number of door frames in the entire corpus, and the door frame including the proper nouns. It may include the process of calculating the entropy of the door frame based on the number of and the state frequency of the door frame.

In the process of performing the engram expansion, at least one upper score door frame may be selected, and each collected proper noun may be substituted at a corresponding position of the door frame.

According to the apparatus and method for adding a proper noun of a language model in the continuous speech recognition system according to the embodiment of the present invention as described above, there are one or more of the following effects.

According to the apparatus and method for adding a proper noun of a language model in the continuous speech recognition system according to an embodiment of the present invention, a new proper noun not included in the language model for speech recognition may be added in an engram form reflecting various expressions of sentences. Because of this, it is more effective to obtain high voice recognition performance of proper nouns than simply adding proper nouns.

1 is a block diagram illustrating a structure of an apparatus for adding a proper noun of a language model according to an embodiment of the present invention;
2 is a flowchart illustrating an operation procedure of an apparatus for adding a proper noun of a language model according to an embodiment of the present invention;
3 is a diagram illustrating an example of a list of new proper nouns and their corresponding classification words according to an embodiment of the present invention;
4 is a diagram illustrating an example of a candidate sentence in which a proper noun or a classification word is searched according to an embodiment of the present invention;
5 is a diagram illustrating an example of a candidate door frame extracted from a candidate sentence according to an embodiment of the present invention;
6 is a diagram illustrating an example of generating an engram type door frame from a candidate door frame according to an embodiment of the present invention;
7 is a diagram illustrating an example of performing engram expansion by assigning a proper noun to an engram according to an embodiment of the present invention.

Advantages and features of the present invention and methods for achieving them will be apparent with reference to the embodiments described below in detail with the accompanying drawings. The present invention may, however, be embodied in many different forms and should not be construed as 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 scope of the invention to those skilled in the art. To fully disclose the scope of the invention to those skilled in the art, and the invention is only defined by the scope of the claims. Like reference numerals refer to like elements throughout the specification.

In the following description of the present invention, a detailed description of known functions and configurations incorporated herein will be omitted when it may make the subject matter of the present invention rather unclear. The following terms are defined in consideration of the functions in the embodiments of the present invention, which may vary depending on the intention of the user, the intention or the custom of the operator. Therefore, the definition should be based on the contents throughout this specification.

Each block of the accompanying block diagrams and combinations of steps of the flowchart may be performed by computer program instructions. These computer program instructions may be loaded into a processor of a general purpose computer, special purpose computer, or other programmable data processing apparatus so that the instructions, which may be executed by a processor of a computer or other programmable data processing apparatus, And means for performing the functions described in each step are created. These computer program instructions may also be stored in a computer usable or computer readable memory capable of directing a computer or other programmable data processing apparatus to implement the functionality in a particular manner so that the computer usable or computer readable memory It is also possible for the instructions stored in the block diagram to produce a manufacturing item containing instruction means for performing the functions described in each block or flowchart of the block diagram. Computer program instructions may also be stored on a computer or other programmable data processing equipment so that a series of operating steps may be performed on a computer or other programmable data processing equipment to create a computer- It is also possible that the instructions that perform the processing equipment provide the steps for executing the functions described in each block of the block diagram and at each step of the flowchart.

Also, each block or each step may represent a module, segment, or portion of code that includes one or more executable instructions for executing the specified logical function (s). It should also be noted that in some alternative embodiments, the functions noted in the blocks or steps may occur out of order. For example, two blocks or steps shown in succession may in fact be performed substantially concurrently, or the blocks or steps may sometimes be performed in reverse order according to the corresponding function.

In an embodiment of the present invention, in order to add proper nouns to a language model in a continuous speech recognition system, new proper nouns and classification words are collected, candidate sentences are selected from a text corpus using the collected words, and candidate sentences After extracting candidate door frames from them, generating engrams of candidate door frames, scoring and ranking each candidate door frame using statistical calculations, applying the upper door frames to proper nouns, and then expanding the engrams to language. This is to reflect in the model.

Hereinafter, embodiments of the present invention will be described in detail with reference to the accompanying drawings.

1 is a block diagram illustrating a structure of an apparatus for adding a proper noun of a language model according to an exemplary embodiment of the present invention.

Referring to FIG. 1, the apparatus 100 for adding a proper noun of a language model includes a collector 102, a candidate sentence selector 106, a candidate door frame extractor 110, an engram candidate door frame generator 112, The ranking unit 114 may include an engram extension unit 116, a reflecting unit 118, and the like.

Specifically, the collection unit 102 collects new proper nouns and classification words, and does not appear in the training text corpus 104 as shown in FIG. 3, which shows a list of new proper nouns and corresponding classification words. Proper nouns 301 may be collected independently, and a category name conceptually distinct from them may be constructed as the classifier 300. At this time, some of the proper nouns may not be assigned a classification word, in this case it may be indicated by an X as shown by reference numeral (302).

The candidate sentence selector 106 may search the text corpus 108 for proper nouns and classification words collected through the collection unit 102 and select matching sentences as candidate sentences. 4 illustrates an example of a candidate sentence in which a proper noun or a classification word is searched, and may correspond to a matching proper noun or classification word with a string between '{' and '{'.

The candidate door frame extractor 110 extracts the candidate door frame from the candidate sentence, and the term "door frame" refers to a local context of a certain length including proper nouns.

FIG. 5 is a diagram illustrating an example of a candidate door frame extracted from a candidate sentence according to an exemplary embodiment of the present invention. In the candidate sentence 500, a maximum of two left and right words may be extracted based on a matched string. An example of the extracted candidate door frame is shown at 502. In the reference numeral 502, the part indicated by 'X' means a location where a proper noun or a classification word is matched. And you can use '<s>' or '</ s>' to mark the beginning or end of a sentence in a door frame. The part marked with '<s>' may mark the beginning of the sentence and the end portion of the sentence may be marked with '</ s>'.

The candidate door frame generator 112 of the engram generates an engram candidate door frame necessarily including 'X' in the candidate door frame extracted from the candidate door frame extractor 110.

FIG. 6 illustrates an example of generating an engram-type door frame from a candidate door frame according to an embodiment of the present invention. For example, for example, refer to a candidate door frame such as 600 or 604 of FIG. 6. Unigrams (1-grams), bigrams (2-grams), and trigrams (3-grams) around 'X' may be generated, such as numbers 602 and 606.

The ranking unit 114 calculates a score for each candidate door frame of the engram generated through the candidate door frame generation unit 112 of the engram, and performs ranking in descending order. Can be formulated and calculated.

First, the suitability of the door frame. How much is suitable as a door frame is measured, as shown in <Equation 1>.

Here, t _i may represent the number of door frames i including the Any proper noun, and m _i may represent the number of door frames i in the entire corpus.

Second, the generality of the door frame. It measures how the door frame is used between various types of proper nouns, as shown in Equation 2 below.

Where T _ij is the number of door frames i including the proper noun j, f _ij is the relative frequency of door frame ij, and G _i is the entropy of door frame i (0 ≤ G _i ≤ 1). .

As a whole formula, the score (Score _i ) is as shown in Equation 3 below.

The engram extension unit 116 may select each of the top ranked N frame frames, assign each collected proper noun to the 'X' position of the door frame, and perform engram expansion.

7 is a diagram illustrating an example of performing engram expansion by assigning a proper noun to an engram according to an embodiment of the present invention, and assigning a proper noun 702 to the door frame 700, respectively, with reference numeral 704. Can be extended as:

In the case of a three-gram trigram (three-gram) door frame, such as the door frame 700, when a proper noun is substituted, a trigram having a length of three may be generated as shown by reference numeral 704. In the case of a 2-gram door frame having a length of 2, such as the door frame 706, an engram generated by substituting proper nouns is also a bygram.

The reflector 118 adds the engrams generated by the engram extender 116 to the language model with frequency.

2 is a flowchart illustrating an operation procedure of an apparatus for adding a proper noun of a language model according to an exemplary embodiment of the present invention.

Referring to FIG. 2, the apparatus 100 for adding a proper noun of a language model in the continuous speech recognition system collects new proper nouns and classification words through the collection unit 102 in step 200, and in step 202, a candidate sentence selecting unit. The proper nouns and classification words collected at 106 are searched in the text scope 108 to select matched sentences as candidate sentences.

In step 204, the candidate door frame extractor 110 extracts the candidate door frame from the selected candidate sentence, and in step 206, the candidate door frame generator 112 of the engram must include an 'X' in the extracted candidate door frame. It will create a candidate frame in gram format.

In operation 208, the ranking unit 114 ranks each of the generated engram-type candidate door frames by high score, and performs a score calculation based on the suitability and generality of the door frame.

After that, the engram expansion unit 116 selects at least one of the at least one top-ranked door frame N and assigns the collected proper nouns to the 'X' position of the door frame and expands the engram. In operation 212, the reflector 118 reflects the engrams generated by the engram extension 116 to the language model 120 together with the frequency.

As described above, the apparatus and method for adding a proper noun of a language model in the continuous speech recognition system according to the embodiment of the present invention are classified into a new proper noun and classification for adding a proper noun to the language model in the continuous speech recognition system. Collect words, select candidate sentences from the text corpus using the collected words, extract candidate door frames from candidate sentences, generate engrams of candidate door frames, and then score each candidate door frame using statistical calculations. And ranking, apply the upper door frame to proper nouns, and expand the engram to reflect in the language model.

While the present invention has been described in connection with what is presently considered to be the most practical and preferred embodiment, it is to be understood that the invention is not limited to the disclosed embodiments, but is capable of various modifications within the scope of the invention. Therefore, the scope of the present invention should not be limited to the described embodiments, but should be determined by the scope of the appended claims, and equivalents thereof.

100: device for adding a proper noun of a language model 102: collector
104: training text corpus 106: candidate sentence selection unit
108: text corpus 110: candidate door frame extraction unit
112: engram candidate door frame generation unit 114: ranking unit
116: engram expansion unit 118: reflecting unit
120: language model

Claims (10)

A collecting unit for collecting new proper nouns and classification words from the proper noun adding device of the language model,
A candidate sentence selection unit for searching the new proper nouns and classification words in a text scope and selecting matching sentences as candidate sentences;
A candidate door frame extracting unit extracting a candidate door frame from the candidate sentence;
A candidate door frame generation unit of an engram extracting a candidate door frame having an engram form from the candidate door frame;
A ranking unit that ranks each of the high scores by calculating a score for each of the engram-type candidate door frames;
An engram expansion unit that performs engram expansion by assigning the proper nouns collected in the candidate door frames of the engram format,
Reflecting unit for adding the engram generated by the engram expansion to the language model with a frequency
Apparatus for adding a proper noun of a language model in a continuous speech recognition system comprising a.
The method of claim 1,
Wherein,
Apparatus for adding a proper noun of a language model in a continuous-word speech recognition system, characterized in that it collects proper nouns that do not appear in the training text corpus, and assigns categorical words to categories that are conceptually separated from the proper nouns.
The method of claim 1,
The candidate door frame extraction unit,
Apparatus for adding a proper proper noun of a language model in a continuous speech recognition system, characterized in that for extracting the local context of a certain length including the proper noun in the candidate sentence in the form of a door frame.
The method of claim 1,
The ranking unit,
The suitability of the door frame is measured based on the number of door frames including proper nouns and the number of door frames in the entire corpus for each candidate door frame of the engram format,
Apparatus for adding a proper noun of a language model in a continuous speech recognition system, characterized in that to calculate the entropy of the door frame based on the number of door frames including the proper nouns and the state frequency of the door frame.
The method of claim 1,
The engram expansion unit,
Apparatus for adding a proper noun of a language model in the continuous speech recognition system, characterized in that the at least one top score door frame is selected, and each collected proper noun is substituted at a corresponding position of the door frame.
Collecting new proper nouns and taxonomy from the proper device for adding proper nouns of the language model,
Searching for the collected new proper nouns and classification words in a text scope and selecting matching sentences as candidate sentences;
Extracting candidate door frames from the selected candidate sentences;
Extracting an engram-type candidate door frame from the extracted candidate door frame;
Ranking the high scores by calculating a score for each of the extracted engram candidate door frames;
Performing the engram expansion by assigning the proper nouns collected in the ranked engram-type doorframe,
The process of adding the engram generated by the engram extension with the frequency to the language model.
How to add a proper noun of a language model in a continuous speech recognition system comprising a.
The method according to claim 6,
The collecting process includes:
A method for adding a proper noun for a language model in a continuous speech recognition system, comprising collecting proper nouns that do not appear in a training text corpus, and assigning a category as a category.
The method according to claim 6,
The process of extracting the candidate door frame,
A method of adding a proper noun for a language model in a continuous-word speech recognition system, characterized in that for extracting a local context having a predetermined length including the proper noun from the candidate sentence in the form of a door frame.
The method according to claim 6,
The ranking process by the score,
Measuring the suitability of the door frame based on the number of door frames including the proper nouns and the number of door frames in the entire corpus for each candidate door frame of the engram type;
The process of calculating the entropy of a door frame based on the number of door frames that contain proper nouns and the state frequency of the door frame
How to add a proper noun of a language model in the continuous speech recognition system characterized in that it comprises a.
The method according to claim 6,
The process of performing the engram expansion,
A method of adding a proper noun of a language model in a continuous speech recognition system, characterized in that at least one upper score door frame is selected and each collected proper noun is substituted at a corresponding position of the door frame.

Images