KR101079653B1 - Apparatus and method to generate keywords for speech recognition in a navigation device - Google Patents

Abstract

The present invention relates to a technology for generating a voice recognition target keyword in a navigation device. In order to realize POI search by voice recognition in a navigation device, a voice recognition target keyword having a high probability of speech is automatically combined from a POI name. It is characterized in that by generating a voice recognition of the various spoken text of the user. According to the present invention, the navigation device can automatically generate these types of POIs that can be spoken by the user, thereby enabling a search service of interest for voice, thereby increasing user convenience.

Voice recognition, navigation device, POI (Point of Interest), heteromorphic

Description

Apparatus and method for generating voice recognition target keyword in navigation device {APPARATUS AND METHOD TO GENERATE KEYWORDS FOR SPEECH RECOGNITION IN A NAVIGATION DEVICE}

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a navigation device, and more particularly, to speech recognition in a navigation device suitable for automatically generating a voice recognition target keyword for searching a point of interest (hereinafter referred to as POI) in a vehicle or a portable navigation device. An apparatus and method for generating a target keyword.

The present invention is derived from a study conducted as part of the IT growth engine technology development project of the Ministry of Knowledge Economy and the Ministry of Information and Communication Research and Development. [Task management number: 2006-S-036-03, Task name: Large-capacity interactive distribution for new growth engine industries Processing voice interface technology development].

In general, a navigation device is installed in a moving object such as a car or the like, and is used to find and guide a location of a destination and a way to reach a destination while being carried by a user, and the global positioning system is used for map information and a navigation program. And a GPS receiver for displaying a location signal received from a satellite and displaying the same to a user, and a GPS receiver connected to the navigation unit or provided in the navigation unit to receive a GPS signal that is a location signal from a GPS satellite. There is a GPS satellite that transmits a location signal.

As such, the navigation device receives GPS information from the GPS satellites rather than tracking any radio waves, calculates the current location and the movement path, distance, and speed of the destination, and shows the exact location to the user in real time. You will be guided through the service.

Accordingly, the user of the navigation device searches for a desired destination, and searches for a destination search word input by the user through the local information, trade name, and building information DB stored in the navigation unit, and outputs the searched result. When selecting the destination from the results, after analyzing the corresponding location of the selected destination and the current GPS receiving location, it will automatically guide the way to the destination.

However, since the user of the navigation device does not know the exact name of the destination, or the long name is searched by a name that is easily called for convenience, it is likely to fail when searching. In most of these cases, the original POI name is a subword or similar variant. Therefore, in order to realize the voice search for the POI name, it is necessary to generate its subwords and variants from the original POI name to cope with various user utterances to increase the efficiency of the voice search.

For example, if the POI name is "Osan City Unam Jugong Apartment 4", most users will not search for the long POI name as it is. Most likely to ignite.

In another example, "GS Caltex Riverside Gas Station" may also be referred to as "Gangsan GS Gas Station" or "Gangsan GS Caltex".

In the voice search using a navigation device according to the prior art operating as described above, the user's spoken text is likely to fail in voice recognition because various speech variations of the original POI name are not considered. This is caused by a mismatch between the target keyword POI name and the user's voice spoken text, but there is no way for this.

Accordingly, in order to realize POI search by voice recognition in a navigation device, a navigation device capable of automatically recognizing various speech texts of a user by automatically generating a combination of voice recognition target keywords having a high probability of speech from POI names. An apparatus and method for generating a voice recognition target keyword are provided.

In addition, the present invention, in order to realize the POI search by speech recognition in the navigation device, after learning a combination of statistically probable nouns from the POI DB from the learning corpus can automatically generate a new morphology based thereon An apparatus and method for generating a voice recognition target keyword in a navigation device is provided.

According to an embodiment of the present invention, a device for analyzing a POI string of a POI name DB collects air between nouns and constructs a statistical model learning unit for constructing probability values as statistical information, and collects collected statistical information. It includes a variant generation unit for generating utterance variants for the POI name input using.

According to an embodiment of the present invention, a method of collecting PO between strings of nouns by analyzing POI strings of a POI name DB, constructing probability values as statistical information, and using the constructed statistical information Generating utterance variants for the input POI name.

In the present invention, the effects obtained by the representative ones of the disclosed inventions will be briefly described as follows.

The present invention can enable the search service of the interest for voice by automatically generating this type of POI that can be spoken by the user in the navigation device, thereby increasing user convenience.

Hereinafter, the operating principle of the present invention will be described in detail with reference to the accompanying drawings. In the following description of the present invention, when it is determined that a detailed description of a known function or configuration may unnecessarily obscure the subject matter of the present invention, the detailed description thereof will be omitted. The following terms are defined in consideration of the functions of the present invention, and may be changed according to the intentions or customs of the user, the operator, and the like. Therefore, the definition should be based on the contents throughout this specification.

In order to realize POI search by voice recognition in a navigation device, a voice recognition target keyword having a high utterance possibility is automatically generated from a POI name so that various speech texts of a user can be recognized. After learning the combination of statistically probable nouns from the DB from the learning corpus, a new variant is automatically generated based on this, thereby enabling a search service of interest for speech.

POI names are characterized by a combination of a number of nouns. In most cases, substrings of POI names or strings generated by inversions of these substrings are spoken variants, otherwise nouns with synonyms or synonyms are replaced to create a new variant. It is the case.

Here, this form is to generate substrings and variants having high possibility of user speech in order to generate a speech recognition target word from the POI name list, which is referred to as an original speech 'this form'.

1 is a block diagram showing the structure of a navigation device according to an embodiment of the present invention.

Referring to FIG. 1, the navigation device includes a control unit 100, a GPS receiver 102, a memory 104, an operation panel input unit 106, a display unit 108, a speaker 110, a voice input unit 112, and the like. Include.

In detail, the GPS receiver 102 measures accurate time and distance based on location information provided from three or more GPS satellites, for example, calculates three different distances accurately according to a triangulation method, and calculates the calculated information. Transfer to the control unit 100.

The memory 104 stores an electronic map database including map data for each region, and transmits it to the controller 100 so that the controller 100 performs the road guidance based on the location information and the electronic map database.

The operation panel input unit 106 is a control command input from a user through a keypad or a touch screen included in the navigation, and transmits a key input signal or a touch signal to the controller 100, and in response thereto, the controller 100 transmits the key input signal or touch signal. Each function block is controlled based on the received control command.

The voice input unit 112 delivers the voice input by the user to the controller 100 when performing a voice search.

The control unit 100 controls the overall function in association with each functional element of the navigation device, performs a position measurement, a road guidance service, a speed detection service. In addition, the destination search is performed based on the text information and the voice information input by the user through the operation panel input unit 106 and the voice input unit 112.

The controller 100 may include a statistical model learner and a heterogeneous generator to perform voice recognition of various speech texts of the user.

2 is a block diagram showing the processing relationship between the statistical model learning unit and the shape generation unit according to an embodiment of the present invention.

Referring to FIG. 2, the statistical model learning unit 202 constructs resources for use by the heterogeneous generation unit 208 when generating the POI heteromorphic form. At this time, the generated resource collects co-occurrence between nouns from a large amount of corpus, that is, the input POI name DB 200 and builds their probability values in the statistical information DB 204.

Afterwards, the morphology generator 208 outputs the POI utterance variants 210 that can be generated using the statistical information DB 204 constructed by the statistic model learner 202 with respect to the input POI name 206. The POI name 206 input here is arbitrarily input to output a speech morphology, not input from the outside, and may be all POI names in the POI name DB 200.

3 is a block diagram showing the element processing structure of the statistical model learning unit according to an embodiment of the present invention.

Referring to FIG. 3, the statistical model learner 202 includes a compound noun decomposition unit 300, a language model generator 302, a probability value calculator 304, a language model weight calculator 306, and the like. , Receiving the POI name DB 200 and outputs the statistical information 204, wherein the statistical information 204 includes the N-gram probability information DB 308 and the language model weight information DB 310.

Specifically, the compound noun decomposing unit 300 decomposes each unit noun and performs tagging with a specific semantic tag because each input POI name is formed by combining a plurality of nouns. For example, if the POI name is "Sillim 8-dong Post Office," the decomposition results such as 'Sillim / <administrative name>' + '8-dong / <numeric expression>' + 'Mail handling / <agency name>' Tag information is added to the.

On the other hand, each character string (w ₀ w _{1 ...} w _n ) constituting the POI name may have a probability value, which is determined as the product of the probability values of each noun (word, w) as shown in Table 1 below.

P (w ₀ w _{1 …} w _n ) = P (w ₀ ) x P (w ₁ ) x P (w ₂ ) x ... x P (w _n )

In order to obtain a more reliable and accurate probability value, the probability value of each word can be replaced with the N-gram probability value considering the context. A detailed description thereof will be described in the language model weight calculator 306.

The language model generator 302 generates an N-gram for considering the surrounding context of each word constituting the POI string. Here, N-gram defines the probability of the next word from n-1 words in the past.

Six N-gram language models used in the embodiment of the present invention are used. Three of them are N-grams for each noun word resulting from decomposition of compound nouns, and the other three are N-grams using tag information of each noun decomposed.

In the field of statistical language, the former is called the word N-gram and the latter is called the class-based N-gram. The former and the latter use three types of Unigram, Bigram, and Trigram respectively, so six language models are used.

The probability value calculator 304 calculates a probability value for each N-gram string after counting the number of each N-gram generated from the language model generator 302. The calculated probability value is stored in the N-gram probability information DB 308 in the form of a resource file, and the stored N-gram probability value is referenced by the heterogeneous generation unit 208. The N-gram of each noun constituting the POI heteromorphic candidate It is used to refer to the probability value.

The language model weight calculator 306 is a processor for obtaining a more accurate N-gram probability value. The POI name DB 200 used as the learning corpus does not include all nouns that may appear and combination strings between them even if the amount of data is large. Therefore, in the present invention, a probability value obtained by linking six unit N-gram probabilities obtained from the probability value calculator 304 by linear interpolation as shown in Table 2 below is used as the final N-gram probability. Each weight (l) value is determined by learning from the corpus to a value between 0 and 1. The weight of N-gram, which has relatively low data among 6 language models, is calculated to be low. This means that it is relatively less reliable than the model.

P (w _i ) =
λ ₁ (h) * P (w _i ) +
λ ₂ (h) * P (w _i | w _{i -1} ) +
λ ₃ (h) * P (w _i | w _{i -2} , w _{i -1} ) +
λ ₄ (h) * P _c (w _i ) +
λ ₅ (h) * P _c (w _i | w _{i -1} ) +
λ ₆ (h) * P _c (w _i | w _{i -2} , w _{i -1} ) (0 <= λ ₁ +… + λ ₆ <= 1)

Above (Table 2) in the word w (n) and, h is an abbreviation for history w _i w _{i -1} ... _-2 Means. P (x) is the word N-gram probability and P _c (x) is the class N-gram probability. P (w _i ) is the unigram probability, P (w _i | w _{i -1} ) is the bigram probability, and P (w _i | w _{i -2} , w _{i -1} ) is the trigram probability.

The weight l of each N-gram model is calculated for each generated N-gram and stored in the language model weight information DB 310, and is learned by the EM (Expectation-Maximization) algorithm shown in Table 3.

1) E-step
λ _j ^{(t + 1)} (h) = {λ _j ^(t) (h) x P (w _i | h _i )} / ∑ _{k = 1 to 6} {λ _k ^(t) (h) x P (w _i | h _i )}

2) M-step
K ^{(t + 1)} = log ∑ _{k = 1 ~ 6} {λ _k ^{(t + 1)} (h) x P (w _i | h _i )}

3) Convergence check
1) and 2) are repeated until the difference between K (t + 1) and K (t) is smaller than the specified small value.

4 is a flowchart illustrating an operation procedure of a statistical model learning unit according to an exemplary embodiment of the present invention.

Referring to FIG. 4, when the POI name DB is input to the statistical model learner 202 in step 400, first, in step 402, POI names formed by combining a plurality of nouns through the compound noun decomposing unit 300 are included in each unit. Decompose into nouns and tag them with specific semantic tags.

In step 404, the language model generator 302 generates an N-gram to consider the surrounding context of each noun. That is, a class N-gram, which is an N-gram using tag information of each noun word resulting from compound noun decomposition and tag information of each noun decomposed in step 406, is generated.

In step 408, the probability value calculator 304 counts the number of each N-gram generated from the language model generator 302, and calculates a probability value for each N-gram string in step 410. The calculated probability value is stored in the N-gram probability information DB 308 in step 412. In step 414, the language model weight calculator 306 calculates the weight for every generated N-gram, and the calculated result is stored in the language model weight information DB 310 in step 416.

5 is a block diagram showing the element processing structure of the shape generation unit according to an embodiment of the present invention.

Referring to FIG. 5, the heterogeneous generation unit 208 includes a compound noun decomposition unit 500, a heterogeneous candidate generation unit 502, a language model generator 504, a heterogeneous candidate probability calculation unit 506, and a heterogeneous candidate filtering unit. 508, filtering DB 510, heterogeneous synonym generator 516, synonym dictionary 518, and the like, and the POI name is input to the heterogeneous generator 208, and the output is a list of possible heterogeneous forms. .

Specifically, the compound noun decomposing unit 500 decomposes a POI name composed of a combination of a plurality of nouns into a unit noun and tags the tag with a specific semantic tag. The heteromorphic candidate generator 502 generates a heterogeneous candidate in all cases that may occur according to a combination of a plurality of nouns in which a complex noun is decomposed from the input POI name. If you have n nouns, you can get n!

Thereafter, the language model generator 504 generates an N-gram for considering the surrounding context of each word constituting the POI string, the word N-gram for each noun word that is the result of the decomposition of the compound noun, and the decomposed angle. Create class N-gram using tag information of nouns.

The heteromorphic candidate probability calculating unit 506 calculates probability values of all the heterogeneous candidates according to the formulas of Tables 1 and 2, and refers to the statistical information DB 204 generated by the statistical model learning unit 202. Done.

The form candidate filtering unit 508 performs a filtering process when the form candidate corresponds to at least one of the following conditions, and includes a filtering DB including a filtering pattern DB 512 and a representative company name dictionary 514 ( The filtering is performed by referring to 510.

A) when the probability value of the candidate is lower than a certain threshold;

B) If it matches the filtering pattern ('-' pattern)

C) when a representative name in the input POI string does not appear in this type candidate

Here, the filtering pattern in b) is described in words and tags of words. The BNF (Backus-Naur Form) of the pattern is shown in Table 4 below.

<pattern>: = '+'<termlist> | '-'<termlist>
<term list>: = <term> | <term><termlist>
<term>: = '*' | Nouns | Tag Name | Noun '/' tag name

'+' Means to match the pattern candidates must be created if the pattern is matched, '-' means to remove the candidate candidates. Of course, only one of them can be used. <term> can describe as many as you want, and can only describe 'noun' or 'noun / tag name'.

For example, when the POI name is "Gaiyang 2 Security Center", this form candidate "Gayang Security Center 2" can be removed so that the pattern can be described as "-* <number expression>". This means to remove all heteromorphic candidates ending with <numeric expression>.

The Dictionary of Representative Names for A) is a resource describing trade names such as 'LG', 'Samsung' and 'Venigans'. If such a representative name appears in the POI text, the candidate must be included in the POI text.

The heteromorphic synonym generator 516 replaces synonyms by a noun unit using a synonym dictionary 518 for unfiltered heteromorphic candidates. For example, if the candidate for POI variant is "Dongkuk Ambassador Middle School", the "Intermediate Ambassador Subsidiary" is replaced by "Affiliated Middle School" and "Affiliated Middle School" in synonyms. In addition to "Middle School", it is also output as "Dongguk Ambassador" and "Dongguk Ambassador".

6 is a flowchart illustrating an operation procedure of the shape generating unit according to an exemplary embodiment of the present invention.

Referring to FIG. 6, when a POI name is input in step 600, it is transmitted to the compound noun decomposing unit 500 of the heterogeneous generation unit 208, where a plurality of nouns are combined in the compound noun decomposing unit 500 in step 602. Decompose POI names consisting of unit nouns and tag them with specific semantic tags. In operation 604, the POI name input to the heteromorphic candidate generator 502, that is, the number of all cases that may occur according to a combination of a plurality of nouns decomposed from a character string as a heterogeneous candidate is generated.

Thereafter, in step 606, the language model generator 504 generates a word N-gram for each noun word that is a result of decomposition of the compound noun, and a class N-gram using tag information of each noun decomposed in step 608. Probability values of all heteromorphic candidates are calculated through the heterogeneous candidate probability calculation unit 506 based on the formulas of Tables 1 and 2.

In operation 612, the heteromorphic candidate filtering unit 508 filters the exclusion of the heterogeneous candidate, even if one of the preset conditions is applicable, by referring to the filtering pattern DB 512 and the representative business name dictionary information 514 of the filtering DB 510. . In operation 614, the synonymous substitution is performed in the noun unit with respect to the unfiltered heteromorphic candidate through the heteromorphic synonym generation unit 516. In this case, the synonym dictionary 514 is substituted.

In operation 616, the information transmitted from the heteromorphic synonym generator 516 is output as a list of speech forms.

As described above, in order to realize POI retrieval by voice recognition in a navigation device, the present invention may automatically generate a combination of voice recognition target keywords having a high probability of speech from POI names so that various voices of a user may be recognized. As an implementation, a combination of statistically probable nouns from the POI DB is learned from a learning corpus and then automatically generated a new variant.

Meanwhile, in the detailed description of the present invention, a specific embodiment of an apparatus and a method for generating a voice recognition target keyword in a navigation device has been described, but various modifications are possible without departing from the scope of the present invention.

In other words, the present invention is not limited to a navigation device, and can be used in all devices and user devices that automatically generate these types of POIs that can be spoken by a user to enable a search service using voice. Therefore, the scope of the present invention should not be limited to the described embodiments, but should be defined not only by the scope of the following claims, but also by those equivalent to the scope of the claims.

1 is a block diagram showing the structure of a navigation device according to an embodiment of the present invention;

2 is a block diagram showing a processing relationship between the statistical model learning unit and the shape generation unit according to an embodiment of the present invention;

3 is a block diagram showing the structure of the element processing unit of the statistical model learning unit according to an embodiment of the present invention;

4 is a flowchart illustrating an operation procedure of a statistical model learning unit according to an embodiment of the present invention;

5 is a block diagram showing the structure of the element processing unit of the shape generating unit according to an embodiment of the present invention,

6 is a flowchart illustrating an operation procedure of the shape generating unit according to the embodiment of the present invention.

<Explanation of Signs of Major Parts of Drawings>

100: control unit 102: GPS receiving unit

104: memory 106: operation panel input unit

108: display unit 110: speaker

112: voice input unit 202: statistical model learning unit

204: Statistical information 208: Shape generation unit

Claims (10)

A statistical model learning unit for collecting air between nouns by analyzing POI strings of a POI name DB and constructing probability values and weights thereof as statistical information; A morphology generation unit for generating utterance variants for the POI name input using the constructed statistical information. Apparatus for generating a voice recognition target keyword in the navigation device comprising a. The method of claim 1, The statistical model learning unit, A compound noun decomposing unit for decomposing nouns constituting the POI string; A language model generator for generating N-gram language model strings from the decomposed nouns; A probability value calculator which calculates counts and probability values of the generated N-gram language model strings and records N-gram probability information; Language model weight calculation unit for recording the language model weight information by calculating the weight of the generated N-gram language model string Apparatus for generating a voice recognition target keyword in a navigation device comprising a. 3. The method of claim 2, The N-gram probability information and the language model weight information, The apparatus for generating a voice recognition target keyword in the navigation device, characterized in that the statistical information. The method of claim 1, The heterogeneous generation unit, A compound noun decomposing unit for decomposing nouns constituting the POI string; A heteromorphic candidate generator for generating a heteromorphic candidate by combining all possible nouns from the decomposed nouns; A language model generator for generating an N-gram language model string from nouns constituting the heteromorphic candidate; A heterogeneous candidate probability calculator which calculates counts and probability values of the generated N-gram language model strings and records N-gram probability information with reference to the statistical information DB generated by the statistical model learner; A heterogeneous candidate filtering unit for removing the heterogeneous candidates as a heterogeneous result if predetermined conditions are satisfied; The morphological synonym generation unit generating morphological forms by substituting the corresponding synonyms in the case of nouns having a synonym relationship among the nouns constituting the morphological candidates. Apparatus for generating a voice recognition target keyword in a navigation device comprising a. The method of claim 4, wherein The heterogeneous candidate filtering unit, The probability value of the heterologous candidate is lower than a certain threshold, The heterogeneous candidate is matched by the filtering pattern, If the representative cross name in the POI string does not appear in this type candidate, The apparatus for generating a speech recognition target keyword in a navigation device, characterized in that when any one of the above, removes as a result of the final deformation. Analyzing the POI strings in the POI name DB to collect air between nouns and constructing probability values and weights as statistical information; A process of generating speech forms for the input POI name using the constructed statistical information Method of generating a voice recognition target keyword in the navigation device comprising a. The method of claim 6, The process of building with the statistical information, Decomposing nouns constituting the POI string, Generating an N-gram language model string from the decomposed nouns; Calculating the count and probability of the generated N-gram language model string and recording the N-gram probability information; Process of recording weight of language model weight information by calculating weight of generated N-gram language model string Method of generating a voice recognition target keyword in a navigation device comprising a. The method of claim 7, wherein The N-gram probability information and the language model weight information, The method of generating a voice recognition target keyword in the navigation device, characterized in that the statistical information. The method of claim 6, The process of generating the utterance variants, Decomposing nouns constituting the POI string, Generating a heteromorphic candidate by combining all possible nouns from the decomposed nouns, Generating an N-gram language model string from nouns constituting the heteromorphic candidate; Calculating a count and a probability value of the generated N-gram language model string with reference to the statistical information DB generated from the statistical model learner, and recording N-gram probability information; If the predetermined condition is satisfied for each of the generated heterogeneous candidates, removing them as a heterogeneous result; A process of generating a morphology by substituting the synonyms in the case of nouns having a synonym relationship among the nouns constituting the heteromorphic candidates Method of generating a voice recognition target keyword in a navigation device comprising a. The method of claim 9, The process of removing as a result of this morphology, The probability value of the heterologous candidate is lower than a certain threshold, The heterogeneous candidate is matched by the filtering pattern, If the representative cross name in the POI string does not appear in this type candidate, The method of generating a voice recognition target keyword in a navigation device, characterized in that when any one of the above, removes as a result of the final deformation.

Images