KR100976999B1 - Address Recognition Method and the system by Grammar Moduling - Google Patents

Abstract

The present invention relates to an address recognition method and apparatus therefor using grammar modularization. More specifically, in a method or apparatus for transmitting information through voice as a main system, such as an ARS system, it is used to analyze and recognize information recorded and input by voice from a user, in particular, address information. The present invention relates to a method and apparatus for improving the recognition rate by facilitating the configuration of grammar contents through the modularization and facilitating the recognition of the address provided by the user.

The present invention is configured by combining the voice according to the local unit in the structure of the current address through the stepping and modularization, in preparation for the speech recognition result, has the effect of increasing the recognition rate and reducing the error rate.

Postal code, address, grammar, modularization, voice recognition, automatic recording

Description

Address recognition method and device using grammar modularization {Address Recognition Method and the system by Grammar Moduling}

The present invention relates to an address recognition method and apparatus therefor using grammar modularization.

More specifically, in a method or apparatus for transmitting information through voice as a main system, such as an ARS system, it is used to analyze and recognize information recorded and input by voice from a user, in particular, address information. The present invention relates to a method and apparatus for improving the recognition rate by facilitating the configuration of grammar contents through the modularization and facilitating the recognition of the address provided by the user.

In call centers such as home shopping, telecommunications companies, large corporations, and banks, ARS, which is an automatic recording system, is frequently used to input order information or other information. In such an automatic recording system, there is a lot of work to receive the caller's address. If this is a simple task, the call center staff must be connected to each call directly, which is very cumbersome for both the operator and the caller. There was a problem such as an increase in waiting time for connecting employees.

For example, if a user wants to provide a user's address to a company or change an address registered in an existing DB, the user tells the address while a telephone is connected with a service provider who can use the DB, and the service provider provides the address. After receiving and writing down, access the DB and change the address information recorded in the DB, or change the address information recorded in the DB while the telling address is directly connected to the DB.

Since this method requires a real-time dictator and must write down one by one, there is a problem that a long time is required to register an address or an address may be recorded incorrectly if a person who inputs the address is wrong.

In order to overcome this problem, the user first records the address according to the user's voice, and the service provider receives the recorded files collectively, analyzes them, and then manually inputs them to the DB. It does not solve the above problem at the source in that it requires simple repetitive tasks.

Therefore, there is a need for a voice recognition system that can automatically recognize the recorded address and make it into a DB, and it is necessary to introduce a method of automatically finding and matching the address from the DB by recognizing the address information that the user has called into the system. . Of these, the latter automatic matching method can reduce costs by reducing the amount of transcription required for a person to write down, save manpower, and reduce the time to write down.

In order to use the automatic matching method having such an advantage, it is necessary to properly recognize the voice of the Korean address system, and a method of increasing the recognition rate by using a unique address system is required.

The present invention has been made in order to solve the problems of the prior art through the speech recognition method, address speech recognition method, system for automatically recognizing the voice address, step-by-step vocabulary configuration unique to the address of Korea, step by step grammar group It is an object of the present invention to provide an address recognition method and apparatus using a grammar modularization that can reduce the error of speech recognition and improve the recognition rate through a combination and a combination of the grammar groups for each module.

In order to achieve the above object, an address recognition method comprising the steps of: (1) separating the mailing address contents by five steps of speech location; (2) generating a grammar group for each regional unit generated in the separation step for each vocal position; (3) generating a module-specific grammar group from the generated gradual group of grammars; And (4) storing the generated step-by-step, module-specific grammar group information as address recognition information.

In addition, in the step (2), it is characterized in that when the regional unit is generated in a step-by-step grammar group consisting of a top-down combination of grammar group.

In the step (2), when generating the step-by-step grammar group, it characterized in that it comprises an abbreviation of the middle word.

In the step (2), when generating the step-by-step grammar group, it characterized in that the end vocabulary of the step is included.

In the step (3), when generating the module-specific grammar group, it characterized in that the content defined in the step-by-step grammar group to be included in the front end.

In addition, the address recognition grammar modularization apparatus, comprising: a separation module for separating the postal address contents for each of five levels of vocalization positions; A combination module for generating a regional unit into a grammar group for each step and a grammar group for each module in the mail address contents separated by the utterance location by the separation module; And a storage module for storing the step-by-step, module-specific grammar group information generated by the combination module as information for address recognition.

In addition, the combination module, characterized in that to be made in the grammar group of the top-down combination (top-down) when generating the regional unit in the step by step grammar group.

The combination module may include abbreviations of intermediate words when generating the step-by-step grammar group.

The combination module is characterized in that the end vocabulary of the step is included when generating the step-by-step grammar group.

The combination module, when generating the step by step grammar group, characterized in that the content defined in the step by step grammar group to be included in the front end.

The apparatus may further include a recognition module that recognizes the voice recording information for each step and each module and provides a result.

The present invention is configured by combining the voice according to the local unit in the structure of the current address through the stepping and modularization, in preparation for the speech recognition result, has the effect of increasing the recognition rate and reducing the error rate.

Hereinafter, the present invention will be described in detail so that those skilled in the art can easily understand the present invention, a preferred embodiment of the present invention will be described with reference to the accompanying drawings.

Generally, to provide a voice service through voice input / output based voice recognition technology, recognition technologies such as GRXML (Grammar eXtensible Markup Language), GSL (Grammar Specification Language) and NGO (Nuance Grammar Object) can be used. The tag sets the voice recognition grammar. Speech recognition grammar refers to a candidate vocabulary that a user can speak, and speech recognition refers to finding a candidate vocabulary most similar to the speech uttered by the user. This method is mainly used when the user can limit the vocabulary that can be spoken, and it is difficult to use in situations where the user cannot predict which vocabulary to speak.

In the case of address speech recognition, since the type of vocabulary is limited, such a grammar tag can be used. Looking at the current state of the postal address system in Korea, it is divided into 15 cities / provinces, and it is composed of local level vocal combinations. For example, it consists of "do-gun-eup-ri-ga-individual address". The following grammar components are necessary for accurate and efficient speech recognition of address speech in Korea.

First, we need to distinguish regional grammar. In other words, it is necessary to separate the grammar by the unit of time. Because in the current postal address system, there are local units of the same name in other regions such as "Galhyeon-dong, Gwacheon-si, Gyeonggi-do," and "Galhyeon-dong, Gyeyang-gu, Incheon." It needs to be recognized within.

Secondly, it should be possible to reduce the scope of perception grammar. That is, it is necessary to separate the grammar in five stages of vocalization. For example, in the case of "13, Dongseon-dong 4-ga, Seongbuk-gu, Seoul", the speech is recognized as a grammar that includes only three levels of regional unit vocabulary combinations of "Seoul," "Seongbuk-gu", and "Dongseon-dong 4-ga."

Third, one slot configuration is used. In the 5-stage voice grammar configuration, each stage is configured as a multi-slot, while in the one-slot configuration, a trigram rule is applied to double the recognition rate. For example, due to the one-slot configuration, "Yongsan-gu Galwol-dong" is prevented from being mistaken as "Yongsan-gu Galhyeon-dong". .

Based on these essential elements, the configuration of the present invention will be described. The present invention relates to a method and apparatus for recognizing such address information through a grammar modularization method.

As shown in FIG. 1, the method for applying the grammar module to which the present invention is applied is separated from the postal address content and combined with the depth and module-specific grammar groups.

The postal address consists of six levels, of which the top five levels are separated for each level. The addresses of the upper five levels are combined into the grammar groups for each step (s110), and then combined into the grammar groups for each module (s120). After being combined into the step and module-specific grammar group is stored as address recognition information (s130). That is, in the grammar group combination, the present invention uses a method of combining first step by step and second module.

The address recognition grammar modularization apparatus to which the present invention is applied is divided into mailing locations by the separation module 100 and the mailing location separated by the separation module 100 as shown in FIG. 2. Combination module 200 for combining the regional unit in the postal address content step by step, module-specific Grammar group, and stores the step-by-step Grammar group information and module-specific Grammar group information combined by the combination module 200 as address recognition information. It consists of a storage module 300.

Combination module 200 is to be made for the grammar group of the top-down method combination when generating or combining the regional unit in the step-by-step grammar group. In addition, the combination module 200 may include an abbreviation of an intermediate word, so that the end vocabulary of the corresponding step is necessarily included and combined. In addition, when combining and generating the step-by-step grammar group, the grammar group recognized in the previous step is included in the front end.

Separation module 110 functions to separate the address by the speech step. The address of Korea consists of the top 5 levels + address or detailed address, and the separation module plays a role of separating the top 5 levels, and sets each level for the separated addresses.

The separation module 110 distinguishes the address vocabulary level, and distinguishes the depth of the address data for each speech step. In addition, it separates the number and the same data. The number and the same data are separated by “constant”, “dong”, and “1” to distinguish address depth of address vocabulary. .

The combination module 120 combines the separated addresses step by step, and also combines the modules by grammar group. In this case, referring to FIG. 3 with respect to the depth, the regional unit in the address structure of Korea is used by formalizing the vocalization of a maximum of five (depth) except the last stage. After the contents are separated by five stages of vocalization, the grammar group is combined and created as the depth group by the top-down combination as possible.

In the first level (1 depth), specify the largest classification, such as Gyeonggi-do and Seoul, and in the second level (2 depth), specify the first-level subaddress, such as Gyeonggi-Suwon, Gyeonggi-Seongnam, and the third. This is a way to set a subaddress that is one step smaller.

The grammar configuration thus determined allows the grammar range to be reduced. At this time, the step-by-step grammar configuration according to the top-down method to include the abbreviation of the middle word. For example, the "Gyeongbuk Pohang-si, Buk-gu, Cheonghae" grammar should include the "Gyeongbuk, Buk-gu, Cheonghae" grammar. And each step grammar must include the end vocabulary of that step. For example, if the fourth-level Gramma-gun is designated, "Buk-gu, Pohang-si, Gyeongsangbuk-do" is not included for "Buk-gu, Pohang-si, Pohang-si, Gyeongbuk." Because "Buk-gu, Pohang-si, Gyeongbuk" is a three-degree lexical group.

An example of the composition of each vocabulary by depth is as follows. The address vocabulary is categorized by vocalization step like "Chunghae Seojeong-ri, Buk-gu, Pohang-si, Gyeongbuk."

division Contents example 1 Depth Vocabulary group containing only level (01) content Gyeongbuk 2 Depth Include level (01 + 02) content Pohang, Gyeongbuk 3 Depth Include level (01 + 02 + 03) content Buk-gu, Pohang-si, Gyeongbuk 4 Depth Include level (01 + 02 + 03 + 04) content Cheonghae, Buk-gu, Pohang-si, Gyeongbuk 5 Depth Include level (01 + 02 + 03 + 04 + 05) content Seojeong-ri, Cheongmyeon, Buk-gu, Pohang-si, Gyeongbuk

That is, except for detailed addresses such as ~~ apartments, ~~ building, the address vocalization is divided into the top five levels and used mainly for this. The combination module 120 classifies the depth of the address vocabulary as described above, preprocesses the numeric or special character data included in the divided address vocabulary, and deletes the irregular data. That is, data such as "1 building", "2 buildings", "1.3 buildings", and the like are processed in a recognizable form, and irregular data such as "in ** land" and "* block * block" is deleted.

And the combination module 120 is added to the standard local address, the local address is spoken for regional or historical reasons, statistically, often used in practice. For example, the standard local address is "Gwangju Metropolitan City", but adds similar vocabulary data such as "Gwangju Metropolitan City" to the corresponding DB of the storage module 130.

That is, the combination module 120 adds a local address consisting of a standard local address and a similar vocabulary to the separated postal code and address book information, and then divides the depth of the address data in each utterance step, and separates the address data. Process the number or special characters included in the list, extract the vocabulary included in each depth, and generate the vocabulary group list, and then use the vocabulary group for each level to determine the top-grammar of the subgroup. ) To form a grammar file.

Combination module 120 may separate the number, the same data when separating the depth (depth) of the address vocabulary, and when processing the special character, it may extract and delete random data. In addition, the combination module 120 extracts only the vocabulary included in each depth and generates a list of lexical groups which are duplicated, while forming a lexical group for each depth, and includes a grammar name, a grammar content, and a recognition. Automatically generate the result slot value.

The combination module 120 constructs a vocabulary group for each module composed of sub-group top-grammars using the generated vocabulary groups for each step, and forms a vocabulary group for each module. Form into a file. The module-specific grammar is used to detect the grammar by depth and then search for the hem.

When constructing the vocabulary group for each module in the above step, as in the case of constructing the vocabulary group for each depth, the grammar name, the grammar contents, and the slot value are automatically generated.

The vocabulary group for each module starts with the vocabulary of the corresponding step of the step-by-step grammar group, and is a grammar group that combines the later step vocabulary in a middle-down manner. Referring to FIG. 4, module 3 starts with Gyeongbuk-Pohang-Buk-gu, which is a third step, and combines Heunghae-eup, Cheongmyeon, Namsong-ri, Masan-ri, and the like as one module. Module 4 starts with Gyeongbu-Pohang-Buk-gu-Heunghae-eup, which is in four stages, and combines Namsong-ri and Masan-ri below. In the present invention, the module is configured from step 3, and the module is not formed in steps 1 and 2.

That is, the module-specific grammar configuration generates another grammar group including the contents defined in the step-by-step grammar group at the front end, and this is called the module-specific grammar. For example, after recognizing “Busan-Geumjeong-gu-Bugok-il” using 3 steps, the module 3 graph includes the contents of step 3 at the front end, such as Busan, Geumjeong-gu, Bugokil-dong, Sampalpal, 20, and 4 steps. Form a group of vocabulary to be applied. Grammar is organized by module, resulting in a reduction in the range of the grammar. The module configuration and the selection according to the module are shown in the order of the module as shown in the table below, and the items indicated by L01 and L02 indicate the steps for each address.

division Contents example Module 2 With L01_L02 + (ex) Seojeong-ri, Cheongmyeon, Buk-gu, Pohang, ...
Namsan Donghae-myeon Balsan-ri ... Module 3 With L01_L02_L03 + (ex) Seojeong-ri, Cheongmyeon, Buk-gu, Pohang, ...
Namsong-ri, Heunghae-eup ... Module 4 With L01_L02_L03_L04 + (ex) Seojeong-ri, Cheongmyeon, Buk-gu, Pohang, ...
Wolpori ... Module 5 With L01_L02_L03_L04_L05 + (ex) A Building, Seojeong-ri, Cheong-myeon, Buk-gu, Pohang-si, Gyeongbuk
B apartment

Since the present invention configures and uses the module from step 3, the module 2 is not used, and only modules 3, 4, and 5 are used. Looking at Module 3, you will have a module vocabulary group such as "Buk-gu Cheonghae ~, Buk-gu, Heunghae-eup ~". Eventually, the provided address information is combined in each step and module form according to the above process and form, and stored as a grammar file.

The grammar file (* .grammer) generated in the above step includes contents by depth, all vocabulary inclusions, regions, and top grammars.

The grammar group generated by the combination module 120 is shown in the table below.

Yeah.  Kinds Maximum possible Yeah.  Count O (n)
Explanation
Pile  Contents
By depth
Yeah. 5 Grammar groups for each depth containing all local content Depth01.h
Depth02.h, .... 5 * 15 Grammar groups for each level divided by 15 regions Depth01_Gangwon.h ...
One A group of grammars containing content for every depth A.grammar 1 * 15 Grammar groups, broken down by region, with content for every depth A.grammar
Per module
Yeah. 3 * n Grama-group for each module, containing all local content A.grammar 3 * n * 15 Grama group for each module divided by region Module03.h
Module04.h N = number of combinations of vocabulary that can be generated and deduplicated

The storage module 130 stores the similar vocabulary and the grammar file of the local address in a DB form under the control of the combination module 120.

The recognition module 140 recognizes the voiced recorded information, compares it with the grammar file DB of the storage module 130, and selects any specific data to apply the recognition result.

A method of practically recognizing and applying this after grammar stepping and modularization will be described with reference to FIG. 5. All address voice files will perform a total of four voice recognitions.

That is, by reading the voice recording information, it recognizes the grammar for each step (depth) (S200), and checks the recognition result to determine whether the recognition is possible (S210). If it is possible to recognize the result of the determination, the recognition of the next step is attempted. If the recognition is not possible, the process is terminated.

If it is possible to recognize the previous step is to try to recognize the four-step grammar group (S220). Similarly, it is determined whether the recognition is possible (S230), and proceeds to the next step.

If it is possible to recognize the fourth stage grammar group, this time, it tries to recognize the fifth stage grammar group (S240). Similarly, it is determined whether or not it can be recognized (S250), and the process proceeds to the next step.

If it is possible to recognize the 5th stage grammar group, it tries to recognize it as the grammar group including all contents of the 3rd stage, 4th stage, and 5th stage grammar group (S260), after which the recognition accuracy is the best among the previous recognition results. A high result is selected (S270).

Thereafter, recognition is performed only on the module grammar matching the content of the step grammar obtained in the above four recognition processes to recognize the matching module grammar (S280), and the final recognition result is applied. On the other hand, if the recognition is not possible in each of the above steps, the procedure is terminated.

6 is a diagram showing the content structure of a file automatically generated by the present invention. Referring to FIG. 6, the "#include higher level grammar file" is a part in which a reference file is specified when referencing between files, and the part below the Depth / Module grammar name is a combination content of the corresponding grammar group, a content of a recognition result, Contains the slot value. Combination contents are composed of detailed grammar names, and definition and lexical contents of detailed grammar names are described at the bottom of the corresponding grammar file. The recognition result contents are composed of a combination of the recognition address contents and the lexical search grammar name. Recognition result includes the next lower level grammar name, so that it can be searched by the lower level module without any other search procedure after recognition.

In the part below "Detailed Grammar Name", the detailed Grammar Name and its contents constituting the combination content of the Grammar Group are defined. The detailed grammar name and its contents defined here refer to this content not only in the grammar group of the stage but also in another grammar group of the lower stage.

According to the method and process of automatically generating grammar, it is possible to generate grammar more easily. According to the method of classifying grammar for each region, it shows a higher value than the existing recognition rate. This is because the range is limited to only the vocabulary of addresses that exist below the gradated region, thereby reducing the misperception of the vocabulary.

As mentioned above, although preferred embodiments of the present invention have been described in detail, those skilled in the art to which the present invention pertains may make the present invention without departing from the spirit and scope of the present invention as defined in the appended claims. It will be appreciated that various modifications or changes can be made.

1 is a view for explaining the address recognition grammar modularization method according to the present invention.

2 is a view for explaining the address recognition grammar modularization apparatus according to the present invention.

3 is a view for explaining the configuration of the grammar for each step (depth) applied to the present invention.

4 is a diagram for explaining a module-specific grammar configuration applied to the present invention.

5 is a view for explaining a method for recognizing a grammar generated according to the present invention.

<Description of the symbols for the main parts of the drawings>

100: separation module

200: combination module

300: storage module

400: recognition module

Claims (11)

In the address recognition method, (1) dividing the mailing address contents into five speech positions; (2) generating a grammar group for each regional unit generated in the separation step for each vocal position; (3) generating a module-specific grammar group from the generated gradual group of grammars; And (4) storing the generated grammar group information for each step as address recognition information; and a method for address recognition using grammar modularization, comprising: a. The method of claim 1, In the step (2), when generating the area unit in the step by step grammar group address recognition method using a grammar modularization, characterized in that the top-down combination consisting of the grammar group. The method according to claim 1 or 2, In the step (2), when generating the step-by-step grammar group, address recognition method using a grammar modularization, characterized in that it comprises an abbreviation of the intermediate word. The method according to claim 1 or 2, In the step (2), when generating the step-by-step grammar group, address recognition method using a grammar modularization, characterized in that the end vocabulary of the step is included. The method according to claim 1 or 2, In the step (3), when generating the module-specific grammar group, address recognition method using a grammar modularization, characterized in that the content defined in the step-by-step grammar group to be included in the front end. In the address recognition grammar modular device, Separation module for separating the postal address content by five levels of utterance position; A combination module for generating a regional unit into a grammar group for each step and a grammar group for each module in the mail address contents separated by the utterance location by the separation module; And Address recognition device using a grammar modularization comprising a; a storage module for storing the step-by-step, module-specific grammar group information generated by the combination module as information for address recognition. The method of claim 6, The combination module is an address recognition device using a grammar modularization, characterized in that when generating a regional unit in a step-by-step grammar group consisting of a grammar group of top-down combination content. The method of claim 6, The combination module, when generating the step-by-step grammar group, address recognition device using a grammar modularization, characterized in that it comprises an abbreviation of the intermediate word. The method of claim 6, The combination module, when generating the step-by-step grammar group, address recognition device using a grammar modularization characterized in that the end vocabulary of the step is included. The method of claim 6, The combination module, when generating the step by step grammar group, address recognition device using a grammar modularization, characterized in that the content defined in the step by step grammar group to be included in the front end. The method of claim 6, Address recognition device using a grammar modularization characterized in that it further comprises a recognition module for recognizing the voice recording information for each step (depth), each module, and provides the result.

Images