JP7363307B2 - Automatic learning device and method for recognition results in voice chatbot, computer program and recording medium - Google Patents

Description

The present invention is applicable to interactive voice systems such as AI (Artificial Intelligence) speakers, smart speakers, voice recognition IVR (Interactive Voice Response), and chat IF (Inter-Face). An AI chatbot (chatbot or chatterbot) or voice response system that includes a recognition device and automatically carries out a conversation with the speaker through voice or through voice and text, automatically learns the recognition results of the speaker's utterances. The present invention relates to the technical field of devices and methods, as well as computer programs and recording media that cause a computer to function as such devices.

In recent years, AI chatbots or voice response systems have been commercialized by companies in various countries and Japan, and various proposals have been made to further improve their functionality. For example, Patent Document 1 proposes a device that determines whether the intention of a user's utterance is an inquiry. Patent Document 2 proposes an information processing device that expands knowledge to answer questions that cannot be answered by a dialog system. Patent Document 3 proposes a voice inquiry system that uses knowledge already accumulated in an AI chatbot server.

Japanese Patent Application Publication No. 2019-114141 JP2019-061482A Patent No. 6555838

In an AI chatbot or voice response system, it is desirable to efficiently improve the accuracy of voice recognition, or to improve the accuracy with less work and data processing.

However, for example, in Patent Document 1, a dialogue act indicating the intention of the user's utterance and the utterance subject of the user's utterance are estimated, and the estimated dialogue act indicates an inquiry, and the estimated utterance subject is an answerable inquiry target. In certain cases, the utterance is configured to be determined to be a query. Patent Document 2 is configured to analyze the cause of failure, generate and output a question from dialogue log data according to the cause, and add the answer to the question as new knowledge to the dialogue data. In Patent Document 3, a large number of chatbot server devices perform learning processing in a distributed manner, so that a learning model that has already been accumulated in a large number of chatbot server devices is used to obtain a model that is acquired by an individual AI speaker. The device is configured to answer questions by voice.

Therefore, according to the systems related to these background technologies or existing products, learning data such as acoustic models and language models are required in order to improve the accuracy of speech recognition. Furthermore, learning data such as pronunciations for sounds and notations for pronunciations currently requires humans to provide correct answers, which is extremely time-consuming. Moreover, simply increasing the training data will not necessarily improve accuracy. Of course, there is also the problem that an increase in registered data leads to an increase in the amount of data to be handled or an increase in data processing load.

In this way, due to the nature of speech recognition, it is important to efficiently improve its accuracy, or to improve the efficiency of the user (in other words, the system user or user who is the speaker), or the worker who provides training data such as acoustic models and language models. It is technically difficult to reduce the burden on personnel such as (in other words, a worker who provides correct data) and to improve the accuracy with less manual labor and less data processing.

The present invention has been made, for example, in view of the above-mentioned technical problems, and provides an automatic learning device and method for recognition results in an AI chatbot, and a computer that can efficiently improve the accuracy of speech recognition. An object of the present invention is to provide a computer program and a recording medium that allow the computer to function as a functional device.

One aspect of the automatic learning device for recognition results in an AI chatbot according to the present invention, in order to solve the above-mentioned problems, includes a repeating unit that repeats the recognition results by the AI chatbot regarding utterances from a speaker; a determination unit that determines whether the recognition result is correct or incorrect based on the speaker's reaction to the recognition result; and a determination unit that determines whether the recognition result is correct or incorrect based on the speaker's reaction to the recognition result; and a learning section that extracts learning data related to the utterance based on a difference between the erroneous recognition result and the correct recognition result when the recognition result becomes erroneous and then correct.

One aspect of the automatic learning method for recognition results in an AI chatbot according to the present invention, in order to solve the above-mentioned problems, includes a step of reciting a recognition result by the AI chatbot regarding an utterance from a speaker; a determination step of determining whether the recognition result is correct or incorrect based on the speaker's reaction to the recognition result, and a determination by the determination step during a dialogue between the speaker and the AI chatbot regarding the utterance. a learning step of extracting learning data related to the utterance based on a difference between the incorrect recognition result and the correct recognition result when the recognition result becomes correct after being incorrect.

One aspect of the computer program according to the present invention causes a computer to execute one aspect of the automatic learning method for recognition results in an AI chatbot described above.

One aspect of the recording medium of the present invention is a recording medium on which one aspect of the computer program described above is recorded.

According to one aspect of each of the above-described automatic learning device and method for recognition results in an AI chatbot, as well as the computer program and recording medium, the accuracy of speech recognition can be automatically and efficiently improved without imposing a burden on the user. It becomes possible to improve the performance. More specifically, for example, it is possible to automatically learn by prompting the user to register correct answer data without placing a burden on the user or the speaker, thereby improving recognition accuracy for similar questions. The more it learns, the more it will be possible to reduce interactions with the AI chatbot. Furthermore, since conversion using synonyms can have negative effects, for example, if the automatic learning is performed within a limited range of scenarios, or in other words, if learning data is constructed for each scenario, or if it is divided, each The above-mentioned effect appears more clearly in the scenario.

These effects of the present invention will be made more clear by the embodiments of the invention described below.

BRIEF DESCRIPTION OF THE DRAWINGS FIG. 1 is a block diagram showing the overall configuration of a voice response system as an example of the "AI chatbot" according to the present invention, which includes an automatic learning device for recognition results in the voice response system according to the first embodiment. . 7 is a flowchart showing a process flow of automatic learning by recitation according to the first embodiment. FIG. 2 is a schematic conceptual diagram illustrating a usage image of automatic learning by recitation according to the first embodiment. FIG. 2 is a schematic conceptual diagram showing a voice recognition mechanism according to the first embodiment. FIG. 2 is a schematic conceptual diagram showing a "learning data creation method" in a comparative example of the first embodiment. FIG. 3 is a schematic conceptual diagram showing a "learning data creation method" in another comparative example of the first embodiment. FIG. 2 is a schematic conceptual diagram illustrating a process of correcting a speech recognition result using synonyms according to the first embodiment. FIG. 2 is a schematic conceptual diagram (part 1) illustrating a usage image related to the determination of synonyms using characteristics of utterances according to the first embodiment; FIG. FIG. 7 is a schematic conceptual diagram (Part 2) showing a usage image related to the determination of synonyms using the characteristics of utterances according to the first embodiment; FIG. FIG. 2 is a schematic conceptual diagram showing a usage image after reflecting learning data according to the first embodiment. 12 is a flowchart illustrating a process flow of automatic learning based on evaluation decisions for answers according to the second embodiment. FIG. 7 is a schematic conceptual diagram showing a usage image of automatic learning based on evaluation judgments on answers according to the second embodiment.

<First embodiment>
A first embodiment will be described with reference to FIGS. 1 to 10. First, the overall configuration of the first embodiment will be described with reference to FIG. FIG. 1 schematically shows the overall configuration of a voice response system 1 having an automatic learning device 100 for recognition results in voice recognition according to a first embodiment. That is, in this embodiment, the voice response system 1 constitutes an example of the "AI chatbot" according to the present invention.

As shown in FIG. 1, the voice response system 1 includes an AI chatbot (QA search) unit 10, a voice capture device 11, a voice recognition device 12, a word dictionary DB 400, an acoustic model DB 401, a language model DB 402, a conversion rule DB 403, and learning. It is configured to include a storage device including a data DB 404 and an automatic learning device 100. The word dictionary DB 400 stores a word dictionary that is a correspondence between the spelling and reading of words. For example, the word dictionary is the word dictionary 400a in FIG. 4. The acoustic model DB 401 stores an acoustic model that is a correspondence relationship between phonemes and pronunciations. For example, the acoustic model is acoustic model 401a in FIG. 4. The language model DB 402 stores language models such as hidden Markov models. For example, the language model is language model 402a in FIG. 4. The conversion rule DB 403 stores conversion rules that are the correspondence between conversion destination and conversion source words. For example, the conversion model is the conversion rule 403a in FIG. 7. The learning data DB 404 stores learning data. For example, the learning data is the learning data 404a in FIG. 1, or the learning data 404b or 404c in FIG. The automatic learning device 100 includes a transmitting section 101, a voice response control section 102, and an automatic learning section 103.

The voice capture device 11 is a device that captures the voice emitted by the user (or system user or speaker) 20 into the microphone of the terminal 21, and captures the voice via a communication network such as the Internet. . Here, the terminal 21 is, for example, a smart phone, a PC (Personal Computer), a tablet, a mobile phone, a wearable device such as a wristwatch type device or a glasses type device. Note that the voice capture device 11 is not limited to the microphone of the terminal 21 as long as it can capture the speech of the user 20, and may be a device that is communicably connected to the terminal 21 by wire or wirelessly. The voice capture device 11 outputs the voice data of the captured voice to the voice recognition device 12.

The speech recognition device 12 processes the speech captured by the speech capture device 11 based on an acoustic model 401 and a language model 402 that have been automatically learned by the automatic learning device 100 and stored in a storage device, as will be described in detail later. Voice recognition. The voice recognition mechanism will be explained in detail later (see FIG. 4, etc.). The speech recognition device 12 outputs the speech recognition result, which is the result of the speech recognition, to the automatic learning device 100.

The AI chatbot (QA search) unit 10 uses various knowledge built in the QA knowledge 10a to perform QA or Q&A (Question and Answer: that is, answers to questions) in response to the questions indicated by the voice recognition results. It is configured to search from the QA knowledge 10a that stores data. More specifically, when the automatic learning device 100 repeats the voice recognition result and determines that the recognition result of the voice from the user is correct, the AI chatbot unit 10 returns an answer to the question from the user. Further, the AI chatbot unit 10 is configured to perform a QA search when the answer to a question from the user is incorrect or there is no answer.

In the automatic learning device 100, under the control of the voice response control unit 102, first, the recognition result by the voice recognition device 12 is converted into a synthesized voice or a combination of synthesized voice and text (hereinafter referred to simply as “synthesized voice, etc.”). The transmitter 101, which together with the voice response controller 102 constitutes an example of the "repeat unit" according to the present invention, repeats the message to the terminal 21 of the user 20. Specifically, first, the voice response control unit 102 receives input of voice recognition results. The voice response control unit 102 generates a synthesized voice to be output to the user 20 based on the voice recognition result. The transmitter 101 transmits the generated synthetic voice data or text data of the synthesized voice to the terminal 21 . The voice response control unit 102 outputs the input voice recognition result to the AI chatbot (QA search) unit 10.

"Repeat" according to this embodiment means to return the recognition result as it is as a synthesized voice (hereinafter also referred to as "repeat synthesized voice") or text (hereinafter also referred to as "repeat text") (that is, to literally repeat it back) ), or it may be returned as a repeating synthesized voice, etc., or repeating text that confirms the recognition result, or that adds an appropriate preface or postscript to the recognition result, or that summarizes the content (i.e., the same content but with different expressions). You can also repeat the phrase by repeating it. For example, in response to the recognition result of "○○", an utterance such as "Is ○○ OK?" may be returned (that is, repeated). For example, when the user 20 says to the terminal 21, ``I want to call a taxi.'', the terminal 21 outputs a repeating synthesized voice or repeating text saying, ``Are you sure that I want to call a taxi?'' . Furthermore, the recognition result is replaced with other synonyms, synonyms, synonyms, or sentences with similar meanings that have the same meaning as the recognition result and are registered in word dictionaries as more versatile words ( In other words, it may be repeated.

Further, in the automatic learning device 100, under the control of the voice response control unit 102, the synthesized speech or text related to the search result performed by the AI chatbot (QA search) unit 10 is performed on the recognition result by the voice recognition device 12. The transmitting unit 101 is configured to transmit to the terminal 21 of the user 20.

In this embodiment, in particular, during these series of dialogues, the automatic learning device 100 uses the recognition results executed by the voice response control unit 102, which constitutes an example of the "judgment unit" according to the present invention, as will be described in detail later. When the first or previous recognition result by the speech recognition device 12 is "incorrect" and the subsequent recognition result is "correct" in the correctness/incorrect determination of The automatic learning unit 103 is configured to extract the difference between the correct and incorrect recognition results as automatic learning data 404a, and register this in a word dictionary.

In other words, in this embodiment, in the correctness determination of the recognition result performed by the voice response control unit 102, there are cases where the recognition result is “correct” from the beginning (that is, there is no incorrect recognition result), and cases where the recognition result is “correct” from the beginning (that is, there is no incorrect recognition result). If the answer is "wrong" (in other words, the correct answer cannot be determined after all), the above-mentioned difference does not exist, so the process of extracting the difference as the automatic learning data 404a is not performed. However, apart from the extraction process related to the difference, recognition results that are "correct" from the beginning or "incorrect" until the end may be used for other statistical data processing, etc. It is optional to store the information as a data file and use it for data analysis in some way after the fact.

The "difference" according to the present embodiment may be a difference in notation or sentence structure between the correct recognition result and the incorrect recognition result. For example, the "difference" in this case is a difference in words. However, the "difference" according to the present embodiment is not limited to this, and in addition to or instead of this, for example, differences in the strength of the voice, differences in the tempo of the voice, and the speaker's emotion estimated from the voice or the content of the utterance. It may also be a difference in characteristics of the utterance, such as a difference in utterances.

As described above, according to the present invention, in any case, as will be described in detail later, the burden placed on personnel such as workers who provide training data (see FIGS. 5 and 6) can be significantly reduced. , automatic learning data 404a can be registered in the word dictionary by machine learning or AI learning. Moreover, such automatic learning according to the result of the correct/incorrect judgment in the voice response system 1 can be performed while interacting in real time, or by recording logs related to the interaction between the user 20 and the voice response system 1. It is also possible to refer to and execute the program after the fact.

In this embodiment, for example, as illustrated in FIG. 1, the automatic learning unit 103 sets the type to one "synonym" through the dialogue between the user 20 and the voice response system 1, and sets it as the "difference" as described above. Automatic learning data 404a is registered in which the extracted "firefly" and "hotel" are made into mutually corresponding "word 1" and "word 2". This is the recognition result that was determined to be incorrect (in other words, incorrect data), "I want to paraphrase firefly", and the recognition result that was later determined to be correct (in other words, correct data), "I want to reserve a hotel". This is automatic learning data based on the difference between Further, automatic learning data 404a is registered in which the type is set to another "synonym" and "Otaru" and "hotel" are set to mutually corresponding "word 1" and "word 2". This is automatic learning data based on the difference between the recognition result that was determined to be incorrect, ``I want to make a reservation for Otaru,'' and ``I want to make a reservation for a hotel,'' which was later determined to be correct. Similarly, "word 1" and "word 2" which correspond to each other with "random translation" and "reservation" are registered as the automatic learning data 404a with the type as other "synonyms". Furthermore, the type is set to "co-occurrence" (that is, one of the combinations or pairs that are likely to occur together in the same conversation), and "hotel" and "reservation" are defined as "word 1" that corresponds to each other. and "word 2" as automatic learning data 404a.

More specifically, as illustrated in FIG. 1, when the user 20 speaks (I) "I would like to reserve a hotel" into the terminal 21, the voice capture device 11 captures this voice. Further, under the control of the voice response control unit 102, the transmission unit 101 repeats (I) “I want to paraphrase firefly”, which is the recognition result by the voice recognition device 12 using the acoustic model 401 and the language model 402. That is, in this example, the synthesized speech data or the text data related to the synthesized speech that is repeated (I) "Is it okay to say 'I want to paraphrase firefly'?" is uttered from the terminal 21, which is the source of the recognition result. The message is configured to be repeated to the user 20 who is the person in question.

In response to this, the user 20 makes the following utterance (II) ``No, I want to reserve a hotel.'' which means that the most recent recognition result is incorrect (NG). After determining that the recognition result is incorrect (NG), (II) an utterance such as a synthesized voice such as "'There was no answer.' Say it again...'" is made to the user 20. Furthermore, the most recent recognition result on the voice recognition device 12 side turns out to be incorrect (NG). A wide variety of interactions occur depending on the thoughts of the user 20, such as the utterance by the user 20, and the result of the voice response control unit 102 determining whether the recognition result of the voice recognition device 12 is correct or incorrect. , (III) ``It's a hotel reservation. The voice response system 1 will utter a recognition result indicating that the reservation date and number of people are correct (OK).

When a series of dialogues regarding hotel reservations occur between the user 20 and the voice response system 1 in this way, the voice response system 1 eventually reaches a positive (OK) recognition result of "I want to reserve a hotel." The AI chatbot (QA search) unit 10 is configured to utilize the QA knowledge 10a to perform a QA search for the question indicated by the positive recognition result. At this time, in cooperation with the QA search, the voice response control unit 102 determines, for example, that the recognition result "I want to reserve Otaru" is incorrect (NG) and the recognition result "I want to reserve a hotel" is correct (OK). It becomes possible to determine that there is. That is, according to the present embodiment, it is possible to at least partially determine whether the recognition result is correct or incorrect based on the user 20's reaction to the QA search result (for example, pressing the "Like" button). This makes it possible to efficiently improve recognition functions.

The automatic learning unit 103 of the automatic learning device 100 extracts the difference from the above-mentioned recognition result and further registers it as automatic learning data 404a, concurrently or in parallel with execution of such QA search, or records it. It is configured to be executable after the fact using logs.

Note that the voice response system 1 that includes such an automatic learning device 100 and the like is shown as a separate body for each device and each part in FIG. 1 for convenience of explanation, but it has a voice capture function, a voice recognition function, etc. function, voice response control function, transmission function, automatic learning function, AI chatbot QA search function, etc., and if it can be accommodated in the same communication network as the terminal 21, one or more computers or terminal devices, one Alternatively, it may be flexibly realized in various forms in hardware or software from a plurality of server devices, one or more databases or storage devices, or the like. Furthermore, at least one or more of the functions described above can be executed in the cloud. It may be realized by a unique computer program that causes a general-purpose computer to perform the automatic learning method according to the present embodiment. Furthermore, it may be realized by reading such a program directly or after downloading it into a computer from a storage medium in which such a program is stored.

Next, with reference to FIGS. 2 to 10, the operational processing of the first embodiment having the configuration shown in FIG. 1 will be described in detail.

In FIG. 2, the voice recognition device 12 performs voice recognition on the voice input from the user 20 at the terminal 21 via the communication network and voice capture device 11, and the recognition result is, for example, “I want to paraphrase the word hotel.” is obtained (step S10). Subsequently, the automatic learning device 100 generates a response under the control of the voice response control unit 102 such as ``Are you sure that I want to paraphrase the word ``hotel''?''. Further, the transmitting unit 101 transmits the generated response to the terminal 21 and repeats it to the user 20 in the form of synthesized speech (or in the form of text, or in the form of both synthesized speech and text). Step S11).

Next, the voice recognition device 12 again performs voice recognition on the voice input from the user 20 at the terminal 21 via the communication network and the voice capture device 11, and generates a recognition result of, for example, “yes” or “no”. is obtained (step S12).

Subsequently, the voice response control unit 102 determines whether the recognition result is correct or incorrect (step S13). In the determination of step S13, if the recognition result is incorrect (step S13: "No"), the voice response control unit 102 generates a response of "please try again" to urge the user 20 to rephrase, The transmitter 101 transmits the generated response to the terminal 21 (step S14). Furthermore, the process returns to step S10, and the subsequent processes are repeatedly executed (steps S10 to S13). That is, the voice response system 1 prompts the user 20 to rephrase the previous utterance in the series of dialogues through the response.

On the other hand, if the recognition result is positive in the determination in step S13 (step S13: "Yes"), after "No" is determined at least once in step S13 in the series of dialogues (i.e., the recognition result is incorrect). After it is determined that there is a sentence), the automatic learning unit 103 determines whether the user 20 has reworded the sentence. That is, in a series of dialogues between the user 20 and the voice response system 1, if the voice response control unit 102 determines that it is correct after determining that it is incorrect, the recognition result that was determined to be incorrect and the recognition result that was determined to be correct are different. It is determined whether there is learning data that can be extracted as a difference from the recognized recognition result (step S15).

In the determination in step S15, if there is no rewording (step S15: NO), there is no learning data to be extracted, so the series of processes is ended. On the other hand, in the judgment in step S15, if the wording has been reworded (step S15: YES), if learning data to be extracted exists (i.e., in a series of dialogues, if the judgment is correct after being incorrect). Therefore, the automatic learning unit 103 extracts the difference as learning data (step S16), and ends the series of processing.

Next, with reference to FIG. 3, an explanation will be given of an image of the use of an automatic learning method in which automatic learning is performed by repetition as described above, along with a specific example. In FIG. 3, first, corresponding to steps S10 to S14 shown in FIG. 2, dialogues C10A to C14A are performed between the user 20 and the voice response system 1 in order from top to bottom in the figure. Note that the processes of determining whether or not there is learning data (step S15 in FIG. 2) and extracting learning data in the case where there is training data (step S16 in FIG. 2) may be executed in real time, or may be performed in real time. The processing may be executed after the fact from the record log showing the processing of Steps S10 to S14 of No.2.

In FIG. 3, first, a voice C10A input from the user 20, using the terminal 21, saying "I would like to reserve a hotel" is accepted. Here, as an example, it is assumed that the word "reservation" cannot be clearly uttered by the user 20 or cannot be clearly captured due to noise or noise.

In response to this, the voice response system 1 goes through the process explained in FIG. 2 (i.e., mainly the process of steps S10 and S11) and then asks the question, ``Is it OK to translate ``hotel''? ” dialogue C11A is transmitted to the user 20. The dialogue C11A here may be not only a synthesized voice but also a text message in addition to or in place of the synthesized voice. In this case, the voice response system 1 transmits the message to the terminal 21 of the user 20 (for example, in the form of an application such as SMS or LINE).

In response to this, the user 20 performs a dialogue C12A via the terminal 21 saying "No." That is, a dialogue C12A is performed to the voice response system 1 indicating that the dialogue C11A is incorrect (that is, incorrect data). In response to this, the voice response system 1 goes through the process explained in FIG. 2 (i.e., mainly the processes of steps S12, S13, and S14) and then sends the message ``Please try again.'' ” dialogue C14A is performed with the user 20. The dialogue C14A here may be transmitted not only as a synthesized voice but also as a text message in addition to or in place of the synthesized voice.

In response to this, the user 20 performs a dialogue C10B via the terminal 21 saying, "I would like to reserve a hotel." Here, as an example, it is assumed that the word "reservation" was clearly uttered by the user 20 or was clearly captured without being affected by noise or noise. In response to this, the voice response system 1 goes through the process explained in FIG. 2 (i.e., mainly the process of steps S10 and S11), and then asks, ``Are you sure you want to reserve a hotel?'' ” dialogue C11B is performed with the user 20. The dialogue C11B here may be transmitted not only as a synthesized voice but also as a text message in addition to or in place of the synthesized voice.

In response to this, the user 20 performs a dialogue C12B via the terminal 21 saying "Yes." That is, a dialogue C12B indicating that the dialogue C11B is correct (that is, correct data) is performed to the voice response system 1. In response to this, the voice response system 1 goes through the process explained in FIG. 2 (i.e., mainly the process of steps S12 and S13), and then sends a confirmation message, ``Accepted.'' ” dialogue C13B is performed to the user 20. The dialogue C13B here may be transmitted not only as a synthesized voice but also as a text message in addition to or in place of the synthesized voice.

In the above series of dialogues, the dialogue C11A is incorrect (i.e., incorrect data) and the dialogue C11B is correct (i.e., correct data). Therefore, in the voice response system 1, the learning data extraction process (i.e., The process of step S16 in FIG. 2) is executed.

More specifically, the automatic learning unit 103 of the voice response system 1 extracts the difference between the incorrect data in the dialogue C11A and the correct data in the dialogue C11B. This extraction is performed by decomposing the automatic learning data 404c corresponding to the entire dialogue into a plurality of words (i.e., words 1 to 4) that make up the automatic learning data 404c. This is done by using the automatic learning data 404b as the "random translation" and "reservation" that make up the dialogue part).

Furthermore, the automatic learning unit 103 of the voice response system 1 corresponds to the automatic learning data 404b the words ``parallel translation'' and ``reservation'' related to the word 3 extracted in this way, with the types of ``synonyms'' mutually. It is registered in the conversion rule DB 403 as a conversion rule 403b (in the form of “reservation” as “correct” and “random translation” as “incorrect”). In parallel or in parallel with the registration of this conversion rule 403b, a conversion rule 403a corresponding to the automatic learning data 404a (see FIG. 1) is registered in the conversion rule DB 403. Note that registration of the automatic learning data 404b based on such differences may be performed in real time without delay, or may be performed after the fact by using a recording log.

Next, with reference to FIG. 4, the voice recognition mechanism in the voice recognition device 102 (see FIG. 1) in the voice response system 1 described above (i.e., the voice recognition process shown in FIG. 2 (processing method in step S10 or S12) ) will be explained using specific examples.

In FIG. 4, first, the speech recognition device 12 converts a sound waveform 301 captured by the speech capture device 11 (see FIG. 1) into a sound waveform 301, which includes data associating phonemes and pronunciations, such as a large number of acoustic models 401a. By referring to the acoustic model DB 401, it is converted to reading 305 (i.e., "I want to serve you as a service" as a specific example here) (step STEP 1 shown in the upper part of FIG. 4). .

In response to this, the speech recognition device 12 breaks down the user's utterances into words, such as a word dictionary 400 including a number of word dictionaries 400a, including correspondence data of notation and pronunciation, and a number of language models 402a, and converts them into kana and kanji. By referring to the language model DB 402 that contains the information in the associated form, the reading 305 is converted to the kana-kanji notation 306 (i.e., "I would like to apply for a service" as a specific example here) (see FIG. 4). Step STEP2) illustrated in the lower row.

As shown in FIG. 4 above, by converting the sound waveform into reading and further converting it into kana-kanji notation, speech recognition in one specific example of this embodiment (for example, steps S10, S12 in FIG. 2, etc.) is performed. ) is executed.

Next, with reference to FIGS. 5 and 6, a description will be given of the remarkable effects or great advantages of the learning data extraction method in step S16 of FIG. 2, in other words, the learning data creation method.

According to the method of creating learning data by transcribing the speech recognition results in the comparative example illustrated in FIG. The operator who provides the data or the operator who provides the correct answer data manually converts it into text. For example, the telephone operator (OP) says, ``Thank you for calling.'' The user or customer (CU) says, ``My computer is broken.'' The operator (OP) says, ``What are your symptoms? ", and after the user or customer (CU) has a dialogue such as "The power is not turned on," the voice recognition result 306 converted into text is sent to the operator from the acoustic data 301 including the dialogue. Created by 30 people. At this time, unknown words in the text are registered in the word dictionary 400, and a language model 402 is constructed by adding rules for word appearance.

According to the comparative example, the manual work for converting into text that the worker 30 must perform takes about 10 times as long as the call time. That is, in this example, a long time of 500 to 1000 hours of manual labor (that is, manual text conversion work) is required.

According to the method of creating learning data by transcribing the recognition results in the comparative example illustrated in FIG. 6, the operator 30 manually searches for and corrects the erroneously recognized portions. For example, the operator (OP) says, ``Thank you for calling.'' The user or customer (CU) says, ``I was asked for a computer.'' The operator (OP) says, ``What kind of award is this? ” and the user or customer (CU) uttered “The power is not turned on.” , parts that are misrecognized manually are recognized as ``begged'' or ``certificates.'' As a result, a speech recognition result 306b that is correctly converted into text is created. Furthermore, appearance rules such as ``begged for'' and ``certificate'' are actually ``broken'' and ``symptom'' and registration of unknown words is performed manually. Then, unknown words (for example, "symptoms") in the text are registered in the word dictionary 400, and a language model 402 is constructed by adding rules for word appearance.

According to the comparative example, the manual work that must be performed by the worker 30 still requires long hours of labor that are time consuming and stressful.

As is clear from the comparison with the comparative example shown in FIGS. 5 and 6, there is no need for the harsh manual labor of transcribing the recognition results by the operator in this embodiment described above (see FIGS. 1 to 4). The effect of automatically extracting the difference between incorrect data and correct data during a series of dialogues and automatically registering this as learning data is that it can be done efficiently and with high precision without any human intervention. This is significantly advantageous in increasing the amount of automatic learning data. In particular, this embodiment is very advantageous because the task of transcribing from the recognition results shown in FIG. 6 can be executed extremely efficiently through automatic learning.

Next, with reference to FIG. 7, a description will be given of correction of speech recognition results using synonyms in this embodiment. In FIG. 7, the upper row shows an example of the speech recognition process that is executed when there is a similar sound by the speech recognition device 12 (see FIG. 1), and the lower row shows a similar sound by the recognition result correction device 12c. An example of a correction process that is executed when there is a problem is shown. The correction is performed by a method that automates the generation of conversion rules, and it is possible to correct the correction correctly by preparing a unique conversion rule under a limited scope of use.

Note that in FIG. 7, the recognition result correction device 12c is shown separately from the speech recognition device 12 for convenience of explanation, but in reality, the recognition result correction function is included in the speech recognition function of the speech recognition device 12. It's enough if you can. That is, in terms of hardware, the recognition result correction device 12c may be included in the speech recognition device 12. Similarly, the conversion rules 403 that are referenced or registered by the recognition result correction device 12c and registered in the storage device are constructed in a form that is included in the word dictionary 400, the acoustic model 401, and the language model 402. good.

In FIG. 7, as shown in the upper row, when voice data is passed, the voice recognition device 12 executes voice recognition by referring to an acoustic model 401, a language model 402, and a word dictionary 400. In particular, the more general-purpose speech recognition equipment is, the more likely it is that erroneous recognition will occur regarding similar words and homophones. It is assumed that a speech recognition result 306a converted into text is output.

Therefore, as shown in the lower part of FIG. 7, when the speech recognition result 306a containing an error and converted into text "What kind of award is it?" is passed, the recognition result correction device 12 refers to the conversion rule 403. By doing so, the recognition results are corrected. In particular, given the limited scope of use, we have created our own conversion rules to ensure correct correction. Therefore, as a correction result, a positive speech recognition correction result 306b that is converted into text "What kind of symptoms do you have?" is output. These unique conversion rules are prepared for a limited scope of use, such as scenarios such as "hospital," "medical care," and "overseas travel." Various data 403a are prepared that specify conversion rules such as giving medicine (Yoyaku) as conversion source 1, free translation (yaku) as conversion source 2, summary (finally) as conversion source 3, etc. for (Yoyaku), etc. has been done. Alternatively, conversion rules 403a are prepared such that the conversion source 1 is a certificate, the conversion source 2 is a little, etc., for the symptoms as conversion destinations.

In this way, if the method shown in Figure 7 that automates the generation of unique conversion rules related to correction is adopted, conversion rules for a limited range of use can be prepared, making it relatively easy and highly accurate. It is possible to make correct corrections.

Next, with reference to FIGS. 8 and 9, a description will be given of the determination of similar words using characteristics of utterances in this embodiment. In the present embodiment, the "difference" that the automatic learning unit 103 extracts as automatic learning data may be a difference in notation or sentence structure between the correct recognition result and the incorrect recognition; If the wording changes due to differences in terms of target or sentence structure, a situation may arise where it is not possible to determine what was wrong, that is, what was wrong and what was correct.

As shown in FIG. 8, word 3, word 4, word 5, word 7, and word 8 are inconsistent with each other as the learning data 404d, but the wording is simply changed or reworded. Therefore, words that were not misrecognized are mixed in with these mismatched words. Therefore, it is desirable that the determination of correctness is based not only on matching and mismatching, but also on some other conversion rule. In this example, "inquiry" and "confirmation", which are correct or incorrect in the recognition results as words, cannot be determined to be correct or incorrect.

As shown in FIG. 9, the language model 402a is configured to include learning data 404e having the types "sound waveform" and "emphasis" in addition to the type "match". Then, since "confirmation" is an "emphasized" word for "inquiry" which is a correct word, it can be determined that it is a correct word. Whether or not the part is "emphasized" can be determined by the user 20 who is the speaker based on the "sound waveform" and the difference in volume, tempo, emotion (e.g. anger) of the restated utterance, etc. This can be determined as the highlighted part.

As can be seen from FIGS. 8 and 9, when the user 20 rephrases the sentence, the incorrect part is naturally emphasized and uttered. If adopted as a conversion rule, it is possible to determine which words are correct or incorrect among a plurality or a large number of words that do not match in the recognition results. Alternatively, when retelling, the user 20 may be trained in advance in a usage manual set in advance for the voice response system 1, so that the user 20 should emphasize the incorrect part when speaking. This may be communicated to the user 20 in advance using a synthesized voice message at an appropriate time. In either case, when the user 20 rephrases the sentence, the parts that were originally incorrect are emphasized, and the parts that are correct or incorrect in the recognition result can be determined, so this embodiment is very advantageous in practice.

Next, the learning data automatically learned through the various motion processes described above (see FIGS. 2 to 9) is reflected in the voice response system 1 (more specifically, the word dictionary 400, the acoustic model 401, the language The usage image of the first embodiment after the model 402, conversion rule 403, etc. have been updated by machine learning will be described with a specific example with reference to FIG. 10. In FIG. 10, dialogues C10A to C13B are exchanged between the user 20 and the voice response system 1 in order from top to bottom in the figure, corresponding to steps S10 to S13 shown in FIG.

In FIG. 10, first, a dialogue C10A is performed from the user 20 via the terminal 21 saying, "I would like to reserve a hotel." Here, as an example, it is assumed that the word "reservation" is not clearly uttered by the user 20 or cannot be clearly captured due to the influence of noise. The process up to this point is the same as the case of dialogue C10A in FIG. 3, which is before learning data is reflected, including the fact that the user is not able to speak clearly.

However, in response to this, the voice response system 1 performs learning that includes the learning data 404f, etc., which has already been learned based on the initial incorrect recognition result by the voice recognition device 12 such as "I want to paraphrase firefly". At this stage, the data is converted and the correct answer, ``I want to book a hotel,'' is derived. As a result, after going through the process explained in FIG. 2 (i.e., mainly the process of steps S10 and S11), the user is asked, ``Are you sure you want to reserve a hotel?'' ” dialogue C11B is performed with the user 20. The dialogue C11B here may be transmitted not only as a synthesized voice but also as a text message in addition to or in place of the synthesized voice.

In response to this, the user 20 performs a dialogue C12B via the terminal 21 saying "Yes." That is, a dialogue C12B indicating that the dialogue C11B is correct (that is, correct data) is performed to the voice response system 1. In response to this, the voice response system 1 goes through the process explained in FIG. 2 (i.e., mainly the process of steps S12 and S13), and then sends a confirmation message, ``Accepted.'' ” dialogue C13B is performed to the user 20. The dialogue C13B here may be transmitted not only as a synthesized voice but also as a text message in addition to or in place of the synthesized voice. In the case of the use shown in FIG. 10, since the dialogue C11B is positive from the beginning, the voice response system 1 does not execute the learning data extraction process (that is, the process of step S16 in FIG. 2).

As described in detail above, according to the first embodiment, automatic learning is performed by repetition, and the process of extracting the above-mentioned differences as learning data is used, so learning is performed relatively easily and efficiently. Can be executed. Furthermore, the more the user learns, the more opportunities for interaction with the voice response system 1 can be reduced. Note that conversion using synonyms has a negative effect such as an unnecessary increase in the number of synonyms, so it can only be used in a limited scenario or scope of use (for example, in the first embodiment). By implementing the first embodiment (by limiting the scope of use of the voice response system 1 to a specific industry for each scenario, etc.), more significant effects will be obtained.

<Second embodiment>
A second embodiment will be described with reference to FIGS. 11 and 12. The overall hardware configuration of the voice response system 1 according to the second embodiment is the same as that of the first embodiment (see FIG. 1), so FIG. 1 will be referred to, and the description thereof will be omitted as appropriate. . The operation processing and usage image of the voice response system 1 according to the second embodiment are different from those of the first embodiment (see FIGS. 2 and 3), so they will be described in detail below. Here, FIG. 11 is a flowchart according to the second embodiment having the same meaning as FIG. 2 of the first embodiment, and FIG. 12 is a diagrammatic flowchart according to the second embodiment having the same meaning as FIG. 3 of the first embodiment. It is a conceptual diagram. In FIGS. 11 and 12, in the second embodiment, an evaluation judgment is performed on an answer to a user 20 (see FIG. 1) in a QA search by an AI chatbot (QA search) unit 10 (see FIG. 1).

Specifically, in FIG. 11, the user first undergoes the same dialogue as the voice recognition (step S10 in FIG. 1), response (step S11 in FIG. 1), and voice recognition (step S12 in FIG. 1) in the first embodiment. 20 is executed by the speech recognition device 12, automatic learning device 100, etc. (step S20). Here, for example, a question such as "I would like to know how to inquire about reservations" is extracted.

Next, within the voice response system 1, a QA search is executed by the AI chatbot (QA search) unit 10 (step S21). Here, for example, a QA search result such as ``Is it okay to say, ``A letter of introduction is required at the outpatient reception desk?'''' can be obtained. Subsequently, in the automatic learning device 100, the voice response control unit 102 generates a response that is the result of the QA search, ``Is it okay to say, ``A letter of introduction is required for the outpatient reception?''''. Further, the transmitter 101 transmits the QA search result to the user 20 in the form of synthesized speech (or in the form of synthesized speech and text) by transmitting it to the terminal 21 (step S22).

In response to this, the voice recognition device 12 again performs voice recognition on the voice input from the user 20 at the terminal 21 via the communication network and the voice capture device 11, for example, “yes” or “no”. A recognition result is obtained (step S23).

Subsequently, the voice response control unit 102 determines whether the recognition result is correct or incorrect (step S24). In the determination of step S24, if the recognition result is incorrect (step S24: "No"), under the control of the voice response control unit 102, a response of "please repeat" to prompt the user 20 to rephrase the phrase is sent. , is transmitted from the transmitter 101 (step S25). Furthermore, the process returns to step S20, and the subsequent processes are repeatedly executed (steps S20 to S23). That is, the voice response system 1 prompts the user 20 in the series of dialogues to rephrase the utterance that includes the previous question sentence, using the response.

On the other hand, if the recognition result is positive in the determination in step S24 (step S24: "Yes"), after "no" is determined at least once in step S24 during the series of dialogues, the automatic learning unit 103 It is determined whether the user 20 has reworded the statement. That is, in a series of dialogues between the user 20 and the voice response system 1, when the voice response control unit 102 makes a determination that is incorrect and then becomes correct, the incorrect recognition result and the correct recognition result are separated. It is determined whether or not there is learning data that can be extracted as a difference from the recognition result obtained (step S26).

In the determination of step S26, if there is learning data to be registered, that is, if there are no errors in the QA search results up to the current stage (step S26: NO), there is no learning data to be extracted, so the process continues as is. The process ends. On the other hand, in the determination in step S26, if there is learning data to be registered, that is, if there has been an error in the QA search result at least once up to the current stage (step S26: YES), the automatic learning unit 103: The difference between the question sentence previously determined to be incorrect and the question sentence currently determined to be correct is extracted as learning data (step S27), and the series of processes ends. Note that the method of understanding or handling the "difference" and the method of registering the extracted learning data in step S27 are the same as in the first embodiment.

Next, with reference to FIG. 12, an explanation will be given of an image of the use of an automatic learning method in which automatic learning is performed by evaluating the QA answers as described above, using a specific example. Note that the processes of determining whether or not there is learning data (step S26 in FIG. 11) and extracting learning data if there is training data (step S27 in FIG. 11) may be performed in real time, or may be performed in real time. It may be executed after the fact from the record log showing the processing of Steps S20 to S25 of No. 11.

In FIG. 12, first, a dialogue C20 is performed from the user 20 via the terminal 21 saying, "I would like to know how to inquire about reservations." In response to this, is it okay to say ``I want to know how to inquire about reservations''? ” dialogue C21 is performed with the user 20. The dialogue C21 here may be transmitted not only as a synthesized voice but also as a text message in addition to or in place of the synthesized voice.

In response to this, the user 20 performs a dialogue C22 via the terminal 21 saying "Yes." That is, a dialogue C22 is performed to the voice response system 1 indicating that the dialogue C21 is correct (that is, the utterance or pronunciation itself is determined to be correct because it is not wrong).

In response, the voice response system 1 executes the QA search process explained in FIG. 11 (i.e., step S21 in FIG. A dialogue C23 such as "I have a letter of introduction..."..."Did you solve the problem?" is performed to the user 20. The dialogue C23 here may be transmitted not only by a synthesized voice but also by a text message in addition to or in place of the synthesized voice.

In response to this, the user 20 performs a dialogue C24 via the terminal 21 saying, "No. I would like to know how to confirm the reservation." In other words, the voice response system 1 receives a dialogue C24 indicating that the dialogue C21 was incorrect (that is, the utterance or pronunciation itself was not wrong, but based on the QA search results, it was actually a mistake or incorrect data). It is done against. At this time, the user 20 is more or less conscious of the existence of homonyms, and uses the word "confirmation" instead of "inquiry" in the dialogue C24. In other words, the user 20 intentionally rephrases the dialogue C20 using a similar expression rather than the same as the previous dialogue C20. Note that "similar expressions" do not necessarily mean synonyms.

In response to this, is it okay to say "I want to know how to confirm my reservation"? ” dialogue C25 is performed with the user 20. The dialogue C25 here may be transmitted not only by synthesized voice but also by a text message in addition to or in place of the synthesized voice.

In response to this, the voice response system 1 executes the QA search process explained in FIG. 11 (i.e., step S21 in FIG. 11) by the AI chatbot (QA search) unit 10, and as a result, "Reference from the reservation site... "..."Did you solve the problem?" dialogue C27 is performed to the user 20. The dialogue C27 here may be transmitted not only as a synthesized voice but also as a text message in addition to or in place of the synthesized voice.

In response to this, the user 20 performs a dialogue C28 via the terminal 21 saying "Yes." That is, a dialogue C28 indicating that the dialogue C25 is correct (that is, correct data) is performed to the voice response system 1.

In this embodiment, instead of or in addition to the input in the dialogue C28, the corresponding recognition result is displayed based on a positive message such as "like" sent when the user 20 is satisfied with the QA search result. It is also possible for the voice response system 1 to determine that the answer is correct (that is, it is correct data).

In the above series of dialogues, the dialogue C21 is incorrect (i.e., incorrect data) and the dialogue C25 is correct (correct data), so the voice response system 1 performs the learning data extraction process (i.e., the 11) is executed.

More specifically, the voice response system 1 extracts the difference between the incorrect answer data in dialogue C21 and the correct answer data in dialogue C25. This extraction constructs different dialogue parts (dialogue parts whose "judgment" is "x" in Figure 12) by breaking down the entire dialogue into multiple words (namely, words 1 to 4). The "inquiry" and "confirmation" are performed using the learning data 404g. Furthermore, the voice response system 1 sets "inquiry" and "confirmation" related to the word 3 extracted in this way as automatic learning data 400e (by changing "confirmation" to "correct answer") as mutually "synonymous" types. and register it in the word dictionary (in the form of "incorrect answer" instead of "inquiry"). Note that registration of the automatic learning data 404h based on such differences may be performed in real time without delay, or may be performed after the fact by using a recording log. Furthermore, "confirmation", "introduction", and "inquiry" may be stored in association with each other as learning data. Further, as learning data, correct answer data "inquiry" and incorrect answer data "introduction" regarding the same sound words may be stored in association with each other.

As described in detail above, according to the second embodiment, the process of extracting the above-mentioned differences as learning data while using the evaluation judgment for QA answers is used, so that the process can be performed relatively easily, efficiently, and automatically. Can perform learning. Furthermore, the more the user learns, the more opportunities for interaction with the voice response system 1 can be reduced. Note that conversion using synonyms has a negative effect such as the number of synonyms becoming unnecessarily large, so it can only be used in limited scenarios or scope of use (for example, in the second embodiment). By implementing the second embodiment (by limiting the scope of use of the voice response system 1 to a specific industry for each scenario, etc.), more significant effects will appear.

Additional Notes Regarding the embodiments described above, the following additional notes are further disclosed.

[Additional note 1]
The automatic learning device for recognition results in an AI chatbot according to appendix 1 includes a repeating unit that repeats a recognition result by the AI chatbot regarding an utterance from a speaker, and a reaction of the speaker to the read recognition result. A determination unit that determines whether the recognition result is correct or incorrect based on the utterance, and during a dialogue between the speaker and the AI chatbot regarding the utterance, the determination by the determination unit becomes incorrect and then becomes correct. and a learning unit that extracts learning data related to the utterance based on a difference between the erroneous recognition result and the correct recognition result in the above case.

[Additional note 2]
In the automatic learning device according to appendix 2, the repeating unit may use the recognition result as it is or replace the recognition result with another word having the same meaning at least partially, and then reproduce the recognition result from the utterance from the AI chatbot. The automatic learning device according to appendix 1 is characterized in that the automatic learning device repeats the content by outputting the content in an output format that can be recognized by the speaker.

[Additional note 3]
In the automatic learning device according to appendix 3, the determination unit determines, as the speaker's reaction to the repeated recognition result, further utterances by the speaker or in a format that can be detected, identified, or recognized by the AI chatbot. The automatic learning device according to appendix 1 or 2, wherein the automatic learning device determines whether the recognition result is correct or incorrect based on the input content.

[Additional note 4]
In the automatic learning device according to Supplementary note 4, the learning unit extracts the learning data, and if the speaker makes a further utterance as the reaction and the determination regarding the further utterance is correct, the learning unit extracts the learning data. Automatic learning according to any one of appendices 1 to 3, characterized in that a word that is a difference between the recognition result that is positive and the recognition result that is positive is registered in the dictionary as a synonym. It is a device.

[Additional note 5]
In the automatic learning device according to appendix 5, when the speaker makes a further utterance as the reaction, the AI chatbot recognizes the speaker's emotion from the audio of the further utterance, and Automatic learning according to any one of appendices 1 to 4, characterized in that the part causing the error in is identified, and the further utterance is recognized on the premise that the identified part is an error. It is a device.

[Additional note 6]
In the automatic learning device according to appendix 6, the determination unit performs the determination by the AI chatbot in accordance with the recognition result by the AI chatbot, in addition to or in place of the speaker's reaction to the repeated recognition result. The automatic learning device according to any one of appendices 1 to 5, wherein the automatic learning device determines whether the recognition result is correct or incorrect based on the speaker's reaction to the search result of the QA search performed.

[Additional note 7]
The automatic learning method of recognition results in an AI chatbot according to appendix 7 includes a repeating step of repeating a recognition result by the AI chatbot regarding an utterance from a speaker, and a reaction of the speaker to the read recognition result. a determination step of determining whether the recognition result is correct based on the utterance; and a determination step that determines whether the recognition result is correct after the determination by the determination step is incorrect during a dialogue between the speaker and the AI chatbot regarding the utterance. and a learning step of extracting learning data related to the utterance based on a difference between the erroneous recognition result and the correct recognition result in the above case.

[Additional note 8]
The computer program described in Appendix 8 is a computer program that causes a computer to execute the model construction method described in Appendix 7.

[Additional note 9]
The recording medium described in Appendix 9 is a recording medium on which the computer program described in Appendix 8 is recorded.

The present invention can be modified as appropriate within a scope that does not go against the gist or idea of the invention as read from the claims and the entire specification, and the automatic learning device and method of recognition results in an AI chatbot that involve such modifications, Furthermore, computer programs and recording media are also included in the technical idea of the present invention.

1...Voice response system (AI chatbot)
10...AI chatbot (QA search) unit 11...voice capture device 12...speech recognition device 20...user 21...terminal 100...automatic learning device 101...transmission unit 102...voice response control unit 103...automatic learning unit 400...word dictionary DB
401...Acoustic model DB
402...Language model DB
403...Conversion rule DB
404...Learning data DB

Claims (8)

An automatic learning device for recognition results in an AI chatbot,
a repeating unit that repeats the recognition result by the AI chatbot regarding the utterance from the speaker;
a determination unit that determines whether the recognition result is correct or incorrect based on the speaker's reaction to the repeated recognition result;
In a dialogue between the speaker and the AI chatbot regarding the utterance, when the judgment by the judgment unit becomes incorrect and then becomes correct, the incorrect recognition result and the correct result are determined. and a learning unit that extracts learning data related to the utterance based on the difference between the recognition result and the recognition result .
In addition to or in place of the speaker's reaction to the recited recognition result, the determination unit is configured to determine the response to the response to the search result of the QA search performed by the AI chatbot in response to the recognition result by the AI chatbot. An automatic learning device for recognition results in an AI chatbot, characterized in that it determines whether the recognition results are correct or incorrect based on a person's reaction . The recitation unit is configured to reproduce the recognition result as it is or after replacing the recognition result with another word having the same meaning at least in part, and outputting the recognition result as an utterance from the AI chatbot or an output recognizable by the speaker. The automatic learning device for recognition results in an AI chatbot according to claim 1, wherein the recognition result is recited by outputting it in a format. The determination unit determines the recognition result based on further utterances by the speaker or input content in a format that can be detected, identified, or recognized by the AI chatbot, as the speaker's reaction to the recited recognition result. The automatic learning device for recognition results in an AI chatbot according to claim 1 or 2, wherein the device determines whether the recognition results are correct or incorrect. When the speaker makes a further utterance as the reaction and the determination regarding the further utterance is correct, the learning unit extracts the learning data by extracting the erroneous recognition result and the correct result. 4. The automatic learning device for recognition results in an AI chatbot according to any one of claims 1 to 3, wherein words that are different from the recognition results obtained are registered in the dictionary as synonyms. When the speaker makes a further utterance as the reaction, the AI chatbot recognizes the speaker's emotion from the audio of the further utterance, and identifies the part that causes the error in the utterance. 5. The automatic learning device for recognition results in an AI chatbot according to any one of claims 1 to 4, wherein the further utterance is recognized on the assumption that the identified portion is incorrect. An automatic learning method for recognition results in an AI chatbot, the method comprising:
a repeating step of repeating the recognition result by the AI chatbot regarding the utterance from the speaker;
a determination step of determining whether the recognition result is correct or incorrect based on the speaker's reaction to the repeated recognition result;
In a dialogue between the speaker and the AI chatbot regarding the utterance, when the judgment in the judgment step becomes incorrect and then becomes correct, the incorrect recognition result and the correct result are determined. and a learning step of extracting learning data related to the utterance based on the difference between the recognition result and the recognition result obtained .
The determining step includes, in addition to or instead of the speaker's reaction to the recited recognition result, the response to the response to the search result of the QA search performed by the AI chatbot in response to the recognition result by the AI chatbot. A method for automatically learning recognition results in an AI chatbot , the method comprising determining whether the recognition results are correct or incorrect based on a person's reaction . A computer program that causes a computer to execute the method for automatically learning recognition results in an AI chatbot according to claim 6 . A recording medium on which the computer program according to claim 7 is recorded.