KR102174148B1 - Speech Recognition Method Determining the Subject of Response in Natural Language Sentences - Google Patents

Abstract

The present invention relates to a speech recognition method for determining whether to respond in a natural language sentence using Natural Language Processing and Deep Neural Network (DNN) technology during natural speech without calling keywords. It provides a method for discriminating responded and non-responsive sentences through voice recognition without using. By configuring language models for each topic, it is possible to discriminate without reflecting all of the numerous proper nouns or terminology in the recognition dictionary. By using the language model for each topic, a domain-specific language model can be easily configured, and maintenance and improvement are possible in the future, and the burden of memory is reduced. Since topic information corresponding to the service area and information on questions or instructions can be obtained together, it can be used usefully when configuring a scenario in an application service company, and can be applied to various services because the topic area can also be expanded and changed. have.

Description

Speech Recognition Method Determining the Subject of Response in Natural Language Sentences

The present invention relates to an interactive speech recognition method, and in particular, responding to natural language sentences using natural language processing and deep neural network (DNN) technologies during natural speech without calling keywords. It relates to a voice recognition method to determine whether or not.

Speech Recognition is a process in which a computer interprets the speech language spoken by a person and converts the contents into text data. It is a technology that is distinguished from speaker recognition, which is used for authentication by comparing it with the speech pattern of a specific person previously recorded. It is widely used in intelligent machines such as telematics and robots, where information communication and automobile industries are fused, to control devices and retrieve information by voice. In order to expand the user range, an acoustic model is constructed by statistically modeling the voices uttered by various speakers, and a language model is constructed through corpus collection.

In order to be able to communicate between humans and machines using'words', the input/output interface of intelligent machines must be voice, and these machines are also called voice recognition devices. As the accuracy of the voice recognition rate of voice recognition devices increases, the application service of voice recognition technology has also been expanded.From the secretary-type voice recognition system of a smartphone to a speaker-type artificial intelligence assistant, the input technology of the Internet of Things It is expanding.

In a conventional voice recognition device, a voice recognition mode is activated using a call word such as "Siri" or "Alexa", or a range of call words is expanded to activate a predetermined keyword. When a keyword is set in this way, since the device cannot be controlled by recognizing the spoken word during natural conversation, the versatility of the device is lowered.

Republic of Korea Patent Publication No. 2014-0073889 relates to a'caller buffering and filling interface for interactive voice recognition', and the present invention provides a natural conversation of the user even if the user does not repeatedly input the caller each time the user inputs a voice. Disclosed is a technology that enables speech recognition for command phrases to be processed from type speech input.

However, the above invention is a natural language speech recognition technology because the user goes through the process of directly inputting a call word with his or her own voice, recognizes the voice waveform input with the call word, and re-recognizes the waveform in the answer when the device asks a question. It is difficult to see.

The input unit of the topic classifier and the intention classifier is a character, and each word is embedded with an n-gram character.

Republic of Korea Patent Publication No. 2014-0073889

In order to solve the above problems, an object of the present invention is to provide a voice recognition method in which a user's speech content is determined through sentence identification without a keyword, and the content that the voice recognition device should respond to the user and the content that is not.

The present invention is a voice recognition method for determining whether to respond in a natural language sentence:

The method includes the steps of inputting a voice uttered by a user into a voice input device in sentence units; Recognizing the input voice sentence for each word by a voice recognizer connected to the voice input device; Classifying the words recognized by the speech recognizer into topics and others of a predetermined class by a topic classifier connected to the speech recognizer; Sending a sentence containing words classified as topics of the predetermined class in the classifying step to an intention classifier, and excluding the remaining sentences from a response target; And classifying the sentence inputted in the intention classifier into a command sentence, a plain sentence, and a question sentence, and determining the command sentence and the question sentence as a response target sentence, wherein the topic classifier and the intention classifier are provided by Fasttext, a natural language processing toolkit. Provides a speech recognition method that determines whether or not to respond in natural language sentences using the Linear Bag of Words Classifier, which is a sentence classification algorithm.

In the present invention, the topic of the predetermined class is email, house control, weather, and schedule, and the topic classifier is a new topic to the topic of the predetermined class. It provides a speech recognition method for determining whether to respond in a natural language sentence further comprising a topic addition unit for adding a.

In the present invention, the topic classifier includes a word database, the word database includes word and similar word data for each topic, and the word and similar word data for each topic are updated in units of a predetermined period. Provides a voice recognition method for determining whether to respond in a stored, natural language sentence.

In the present invention, the intention classifier further includes a sentence database, and the sentence database includes sentence data for each sentence for classifying the input sentence into an instruction sentence, a plain sentence sentence, and a question sentence, and the sentence sentence, Provides a voice recognition method for determining whether to respond in natural language sentences by updating and storing the plain text and sentence data for each questionnaire in a predetermined period.

In the present invention, the sentence database further includes a topic-specific answer according to the determining step as plain preface sentence data responding to a command sentence and a question, and the determining step includes uttering the responding plain preface sentence by a speaker. It provides a voice recognition method for determining whether to respond in a natural language sentence, further comprising the step.

The present invention also includes a code portion programmed to input a voice uttered by a user into a voice input device in sentence units; A code portion programmed to recognize the input speech sentence for each word by a speech recognizer connected to the speech input device; A code portion programmed to classify words recognized by the speech recognizer into topics of a predetermined class and others by a topic classifier connected to the speech recognizer; A code portion programmed to send a sentence containing a word classified as a topic of the predetermined class in the classifying step to an intention classifier, and exclude the remaining sentences from the subject of response; And a code portion programmed to classify the sentence input from the intention classifier into a command sentence, a plain sentence, and a question sentence, and among them, a command sentence and a question sentence, and a code portion programmed to determine the sentence to be answered, the topic classifier and the intention classifier as a toolkit. Provides a computer-readable storage medium that stores a speech recognition computer program programmed to determine whether or not to respond in natural language sentences using Linear Bag of Words Classifier.

The present invention provides a method capable of discriminating responded and non-responsive sentences through voice recognition without using a pager. By configuring language models for each topic, it is possible to discriminate without reflecting all of the numerous proper nouns or terminology in the recognition dictionary. By using the language model for each topic, a domain-specific language model can be easily configured, and maintenance and improvement are possible in the future, and the burden of memory is reduced. Since topic information corresponding to the service area and information on questions or instructions can be obtained together, it can be used usefully when configuring a scenario in an application service company, and can be applied to various services because the topic area can also be expanded and changed. have.

1 shows an exemplary structure of a method for discriminating response and non-response sentences through a topic classifier and an intention classifier according to the present invention.
2 shows a Linear Bag of Words Classifier, an algorithm used for sentence classification in Fasttext according to the present invention.
3 is a flowchart illustrating a method for determining responded and non-responsive sentences using a topic classifier and an intention classifier according to the present invention.

Various aspects are disclosed with reference to the drawings. In the following description, for illustrative purposes, a number of specific details are disclosed to aid in an overall understanding of one or more aspects. However, it will be appreciated that this aspect can be implemented without the specific details of each. The following description and the accompanying drawings set forth in detail certain illustrative aspects of one or more aspects. However, these aspects are exemplary and some of the various methods may be used in the principles of the various aspects, and the description described is intended to include all such aspects and their equivalents.

Various aspects and features will be presented by a system that may include a number of devices, modules, and the like. It should also be understood and recognized that various systems may include additional devices, parts, components, etc. and/or may not include all of the devices, parts, components, etc. discussed in connection with the drawings.

As used herein, “examples”, “examples”, “aspects”, “examples” and the like should not be construed as having any aspect or design described being better or advantageous than other aspects or designs. The terms used below, such as'system','server' terminal, etc. generally mean a computer-related entity, and may mean, for example, hardware, a combination of hardware and software, and software.

In addition, the term “or” is intended to mean an inclusive “or” rather than an exclusive “or”. That is, unless specified otherwise or is not clear from the context, "X employs A or B" is intended to mean one of the natural inclusive substitutions. That is, X uses A; X uses B; Or, when X uses both A and B, "X uses A or B" may be applied to either of the above cases. In addition, the term "and/or" as used herein should be understood to refer to and include all possible combinations of one or more of the listed related items.

In addition, the terms "comprising" and/or "comprising" mean that a corresponding feature, step, action, module, and/or component is present, but one or more other features, steps, actions, modules, components It is to be understood that it does not exclude the presence or addition of elements, and/or this group. In addition, in the present specification, terms such as first and second may be used to describe various elements, but these elements are not limited by these terms. That is, these terms are only used to distinguish between two or more components, and should not be interpreted as meaning order or priority. In addition, unless otherwise specified or when the context is not clear as indicating a singular form, the singular in the specification and claims should be interpreted as meaning "one or more" in general. Hereinafter, embodiments of the present invention will be described with reference to the accompanying drawings.

The robot speech recognition technology without keyword calls can provide a more natural dialogue interface between a person and a robot by judging the contents of a user's speech, unlike a conventional speech recognition interface using a pager. In addition, if the existing system was to perform (calling word recognition -> voice recognition -> task), the present invention can immediately reflect the user's request by performing (voice recognition -> task).

1 shows an exemplary structure of a method for discriminating response and non-response sentences through a topic classifier and an intention classifier according to the present invention. The present invention is a voice recognition method for determining whether to respond in natural language sentences. In one embodiment of the present invention, the method includes the steps of inputting a voice uttered by a user into a voice input device in units of sentences; Recognizing the input voice sentence for each word by a voice recognizer connected to the voice input device; Classifying the words recognized by the speech recognizer into topics and others of a predetermined class by a topic classifier connected to the speech recognizer; Sending a sentence containing words classified as topics of the predetermined class in the classifying step to an intention classifier, and excluding the remaining sentences from a response target; And classifying the sentence input by the intention classifier into an instruction sentence, a plain sentence sentence, and a question sentence, and determining the sentence sentence and the question sentence as a response target sentence. In one embodiment of the present invention, the topic classifier and the intention classifier use the Linear Bag of Words Classifier, which is a sentence classification algorithm of Fasttext, which is a natural language processing tool kit.

2 shows a Linear Bag of Words Classifier, an algorithm used for sentence classification in Fasttext according to the present invention. The analysis procedure of Linear Bag of Words Classifier, a sentence classification algorithm of Fasttext, a natural language processing toolkit according to an embodiment of the present invention, is as follows. First, the speech recognition result sentence is tokenized into several words. One sentence is expressed as N words, and each word is expressed as a vector. In the process of being expressed as vec in Word, based on Distributional Hypothesis, if the distribution of adjacent words of two words is similar, each word considers the meaning and context of the word based on the assumption that “meaning” is similar. Through this process, each word is represented as a vector.

One sentence is expressed as N word vectors, averaging them to create a document vector, and putting this document vector in a linear classifier including one layer of hidden layer and classifying sentences through this Classify into each class. In one embodiment of the present invention, a topic classifier and an intention classifier are made using Fasttext's Linear Bag of Words Classifier. In one embodiment of the present invention, the topic of the predetermined class is email, house control, weather, and schedule, and the topic classifier is applied to the topic of the predetermined class. It further includes a topic addition unit for adding a new topic. The topic classification model makes it possible to construct a distributed language model for each service in a large-capacity speech recognition system. Recently, speech recognizer technology has been applied to many fields, but there is a resource (memory) limitation in reflecting all of the numerous proper nouns or terminology in the recognition dictionary. If a language model for each topic is configured and the corresponding language model (search network) is searched according to the results of topic classification, recognition performance can be improved, and service quality can be improved as a result. In one embodiment of the present invention, the topic classifier includes a word database, the word database includes word and similar word data for each topic, and the word and similar word data for each topic are in units of a predetermined period. Update and save.

In one embodiment of the present invention, the intention classifier includes a sentence database, and the sentence database includes sentence data for each sentence, plain sentence, and question sentence for classifying the input sentence into an instruction sentence, a plain sentence sentence, and a question sentence. The sentence data for each statement, plain preface, and question is updated and stored in units of a predetermined period. The present invention is configured to identify responding/non-corresponding to a 2-pass method, and a topic classifier and an intention classifier are sequentially applied. Therefore, when a response is received, topic (service area) information and intention information can be obtained together. This information can be usefully used when configuring a scenario in an application service company, and can be applied to various services because the topic area can be expanded and changed.

3 is a flowchart illustrating a method for determining responded and non-responsive sentences using a topic classifier and an intention classifier according to the present invention. The sentence identification algorithm proposed in the present invention is expressed in a flow chart as follows. In one embodiment of the present invention, a response and non-response classifier is designed so that voice recognition can be performed only when a user makes a query command to a robot. In order to classify the responding sentences without a caller, a two-pass cascade type classifier that separates the topic classifier and the intention classifier is used to determine whether the sentence responds. In one embodiment of the present invention, when a sentence uttered by a user is included in a topic, only sentences classified as requested or questionable by the intention classifier are considered as responses. The technique proposed in the present invention is a natural language conversational speech recognition technique without keyword calls that can be embedded in a robot HW. The purpose is to automatically determine whether the robot responds or does not respond to the user's uttered sentence by using a meaning-based sentence identification algorithm. To achieve this, a topic classification model and an intention classification model were used. In other words, it is a system that classifies the service type of speech recognition sentences in the topic classifier, and classifies only the sentences corresponding to commands and requests in the intention classifier to determine response/non-response status. In one embodiment of the present invention, the sentence database further includes a topic-specific answer according to the determining step as plain preface sentence data corresponding to a command sentence and a question, and the determining step includes the plain preface sentence corresponding to the speaker It further includes the step of igniting. In one embodiment of the present invention, a topic classification model and an intention classification model are designed to distinguish responding and non-corresponding sentences.

The topic classification model classifies which topic (service) corresponds to when a user utters a sentence. Assuming a typical home AI speaker situation, there are specific tasks a speaker can perform. These tasks can generally be categorized into topics. The sentences uttered by the user require specific tasks to be performed on a specific topic. For example, the sentences uttered by the user may be classified by topics such as email, schedule, house control, weather, etc., which the AI speaker can execute. If a sentence is not included in the topic, it is considered non-responsive. In addition, when a sentence is classified within a topic, it is determined whether a response is non-response or not through an intention classifier. In one embodiment of the present invention, the intention classification model is largely divided into a command sentence, a plain sentence sentence, and a question sentence. In the topic classifier, only sentences classified as topics are input into the intention classifier, and only sentences corresponding to the command sentence and the question sentence are classified accordingly.

In one embodiment of the present invention, the method may be implemented as programmed computer code and stored in a computer-readable storage medium. That is, a computer-readable storage medium that stores a speech recognition computer program programmed to determine whether or not to respond in a natural language sentence, the storage medium comprising: a code portion programmed to input a speech uttered by a user into a speech input device in units of sentences; A code portion programmed to recognize the input speech sentence for each word by a speech recognizer connected to the speech input device; A code portion programmed to classify words recognized by the speech recognizer into topics of a predetermined class and others by a topic classifier connected to the speech recognizer; A code portion programmed to send a sentence containing a word classified as a topic of the predetermined class in the classifying step to an intention classifier, and exclude the remaining sentences from the subject of response; And a code portion programmed to classify the sentence input by the intention classifier into an instruction sentence, a plain sentence sentence, and a question sentence, among which the instruction sentence and the question sentence are judged as a response target sentence. In one embodiment of the present invention, the topic classifier and the intention classifier use a sentence classification algorithm Linear Bag of Words Classifier of toolkit Fasttext.

As described above, the present invention relates to a speech recognition method for determining whether to respond in a natural language sentence. This invention can be applied to immediately reflect the user's needs while driving in the automotive field, for example, and can be applied to increase user interface convenience in the Internet of Things environment in the home automation field, and applied to an artificial intelligence assistant. Can be used to increase the convenience of a smart speaker or a user interface of a robot.

Various embodiments described herein may be implemented in a medium that can be read by a computer or a similar device using, for example, software, hardware, or a combination thereof.

According to hardware implementation, the embodiments described herein include application specific integrated circuits (ASICs), digital signal processors (DSPs), digital signal processing devices (DSPDs), programmable logic devices (PLDs), field programmable gate arrays (FPGAs), It may be implemented using at least one of processors, controllers, micro-controllers, microprocessors, and electrical units for performing other functions, in some cases herein. The described embodiments may be implemented as a management server and/or the system itself.

According to the software implementation, embodiments such as procedures and functions described in the present specification may be implemented as separate software modules. Each of the software modules may perform one or more functions and operations described herein. The software code can be implemented as a software application written in an appropriate programming language. The software code may be stored in a management server and/or a database, and executed by an app.

Meanwhile, the various embodiments presented herein may be implemented as a method, an apparatus, or an article of manufacture using standard programming and/or engineering techniques. The term “article of manufacture” includes a computer program, carrier, or media accessible from any computer readable device. For example, computer-readable media include magnetic storage devices (e.g., hard disks, floppy disks, magnetic strips, etc.), optical disks (e.g., CD, DVD, etc.), smart cards, and flash memory devices. (E.g. EEPROM, card, stick, key drive, etc.), but is not limited to these. In addition, the various storage media presented herein include one or more devices and/or other machine-readable media for storing information. The term “machine-readable medium” includes, but is not limited to, wireless channels and various other media capable of storing, holding, and/or transmitting instruction(s) and/or data.

The description of the presented embodiments is provided to enable any person skilled in the art to use or implement the present invention. Various modifications to these embodiments will be apparent to those of ordinary skill in the art, and the general principles defined herein can be applied to other embodiments without departing from the scope of the present invention. Thus, the present invention is not to be limited to the embodiments presented herein, but is to be construed in the widest scope consistent with the principles and novel features presented herein.

Claims (6)

As a voice recognition method that determines whether or not to respond in natural language sentences:
The method includes the steps of inputting a voice uttered by a user into a voice input device in sentence units;
Recognizing the input voice sentence for each word by a voice recognizer connected to the voice input device;
Classifying the words recognized by the speech recognizer into topics and others of a predetermined class by a topic classifier connected to the speech recognizer;
Sending a sentence containing words classified as topics of the predetermined class in the classifying step to an intention classifier, and excluding the remaining sentences from a response target; And
Including the step of classifying the command from the sentence input in the intention classifier, and determining the command as a response target sentence,
The topic classifier and the intention classifier use Linear Bag of Words Classifier, a sentence classification algorithm of Fasttext, a natural language processing toolkit,
The intention classifier includes a sentence database,
The sentence database includes sentence sentence data for classifying a sentence from the input sentence, and the sentence sentence data is updated and stored in units of a predetermined period,
A voice recognition method that determines whether or not to respond in natural language sentences.
The method of claim 1,
The topics of the predetermined class are email, house control, weather, and schedule,
The topic classifier further comprises a topic adding unit for adding a new topic to the topic of the predetermined class,
A voice recognition method that determines whether or not to respond in natural language sentences.
The method of claim 2,
The topic classifier includes a word database,
The word database includes word and similar word data for each topic, and the word and similar word data for each topic is updated and stored in units of a predetermined period,
A voice recognition method that determines whether or not to respond in natural language sentences.
delete The method of claim 1,
The sentence database further includes an answer for each topic according to the determining step as plain preface sentence data responding to the command sentence and the question,
The determining step further comprises the step of uttering the responding plain preface sentence with a speaker,
A voice recognition method that determines whether or not to respond in natural language sentences.
A code portion programmed to input the voice uttered by the user into the voice input device in sentence units;
A code portion programmed to recognize the input speech sentence for each word by a speech recognizer connected to the speech input device;
A code portion programmed to classify words recognized by the speech recognizer into topics of a predetermined class and others by a topic classifier connected to the speech recognizer;
A code portion programmed to send a sentence containing a word classified as a topic of the predetermined class in the classifying step to an intention classifier, and exclude the remaining sentences from the subject of response; And
Including a code portion programmed to classify the statement from the sentence input in the intention classifier, and determine the statement as a response target sentence,
The topic classifier and the intention classifier use Fasttext's Linear Bag of Words Classifier as a toolkit,
A computer-readable storage medium storing a speech recognition computer program programmed to determine whether to respond in natural language sentences.

Images