JP2009037458A - Response system and response content control method - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide a response system and a response content control method, capable of outputting a proper response sentence. <P>SOLUTION: This response system concerned in one embodiment is a response system for outputting the response sentence with respect to an input input-sentence, and has a microphone 31 for inputting the input sentence, a language understanding part 12 for understanding language information of the input sentence input into the microphone, a dialogue management part 13 for determining a response proposal sentence in response to the language information, a response sentence generating part 14 for generating the response sentence based on the response proposal sentence, a speaker 32 for outputting the response sentence generated by the response sentence generating part 14. The dialogue management part 13 calculates similarity between the response proposal sentence and a prohibited sentence, referring to a prohibited sentence database 16 storing the prohibited sentence. The dialogue management part 13 limits an output of the response proposal sentence in response to a comparison result of the similarity with a determination value. <P>COPYRIGHT: (C)2009,JPO&INPIT

Description

  The present invention relates to a response system and a response content control method, and more particularly to a response system that responds to input language information and a response content control method that controls the response content.

  In recent years, a speech dialogue system has been developed that performs speech recognition processing on speech input to a microphone and interacts with the speech recognition result (Non-Patent Document 1). In addition, in an exhibition facility or the like, a voice dialogue robot for guiding the facility has been developed.

In such a voice interaction process, for example, a voice recognition process is performed on the input voice, and the voice data is converted into text data. Then, the text data is processed in the language understanding module, and the dialogue management unit searches for information to generate a response sentence.
"Spoken Dialogue System" Tatsuya Kawahara and Masahiro Araki Ohmsha (issued on October 15, 2006)

  However, when using a voice interaction system in an exhibition facility, there is a problem that an inappropriate response may be made to the user. That is, an unpleasant response sentence may be uttered to a facility user at a vehicle display facility or the like. This problem will be described with reference to FIG. FIG. 10 is a diagram schematically illustrating how the voice interactive robot responds to a question spoken by the user.

  It is assumed that a user 105 who has had time for vehicle maintenance in the vehicle exhibition facility inquires the voice dialogue robot 104 about the latest news. Here, the facility user 105 asks the voice interactive robot 104 “Tell me the latest news”. At this time, the voice interaction robot 104 downloads the news via the Internet and reads the latest news on the Internet.

  Here, the voice dialogue robot 104 utters a response to the user 105, saying, “Now, the incident of killing the dog was due to the loudness of the neighbor's dog.” When such cruel news is spoken, the user 105 feels uncomfortable.

  As described above, the conventional voice interactive robot 104 may give an inappropriate response that makes the user uncomfortable. Such a problem occurs not only in a voice interactive robot that utters voice, but also in a response system that outputs a response sentence on a monitor or the like.

  The present invention has been made to solve such a problem, and an object thereof is to provide a response system and a response content control method capable of outputting an appropriate response sentence.

  A response system according to a first aspect of the present invention is a response system that outputs a response sentence to an input sentence that has been input, the input unit that inputs the input sentence, and the language of the input sentence that is input to the input part Generated by a language understanding unit that understands information, a management unit that determines a response candidate sentence according to the language information, a response sentence generation unit that generates a response sentence based on the response candidate sentence, and a response sentence generation unit An output unit that outputs the response sentence, and the management unit calculates a similarity between the response candidate sentence and the prohibited sentence by referring to a prohibited sentence database in which the prohibited sentence is stored. The output of the response candidate sentence is limited based on the similarity. As a result, output of a response sentence with inappropriate content is restricted, and therefore an appropriate response sentence can be output.

  A response system according to a second aspect of the present invention is the response system described above, wherein the management unit calculates a determination value of the response candidate sentence based on the similarity, and the determination value and the threshold value Whether or not to limit the output of the response candidate sentence is determined based on the comparison result. Thereby, the output of the response sentence with inappropriate contents can be surely limited.

  A response system according to a third aspect of the present invention is the response system described above, wherein the response candidate sentence is determined based on information updated on a network. Even when a response candidate sentence whose contents are undetermined is determined, an appropriate response sentence can be generated.

  A response system according to a fourth aspect of the present invention is the response system described above, wherein the prohibited text stored in the prohibited text database includes a plurality of morphemes. Thereby, the output of an inappropriate response candidate sentence can be restricted easily and reliably.

  A response system according to a fifth aspect of the present invention is the response system described above, wherein the response candidate sentence and the utterance prohibition sentence are morphologically analyzed, the morpheme included in the response candidate sentence, and the utterance prohibition sentence Is compared with the morpheme included in the morpheme, and the similarity is calculated according to the number of matches of the morpheme. Thereby, the output of an inappropriate response candidate sentence can be restricted easily and reliably.

  A response system according to a sixth aspect of the present invention is the response system described above, wherein the number of coincidence of the morphemes is a number of coincidence of morphemes in a specific part of speech. Thereby, the output of an inappropriate response candidate sentence can be restricted easily and reliably.

  A response system according to a seventh aspect of the present invention is a response system that utters a response sentence to input voice data, the voice recognition unit performing voice recognition processing on the input voice data, and the voice A language understanding unit that understands the language information of the speech data that has undergone recognition processing, determines a response candidate sentence according to the language information, and refers to the prohibited sentence database in which the prohibited sentence is stored, and thus the response candidate sentence A dialogue management unit that restricts the utterance of the response candidate sentence according to the content of the response candidate sentence. As a result, output of a response sentence with inappropriate content is restricted, and therefore an appropriate response sentence can be output.

  A response content control method according to an eighth aspect of the present invention is the response content control method described above, and is a response content control method for controlling the content of a response sentence that is output in response to an input sentence that has been input. A step of understanding linguistic information of the inputted input sentence, a step of determining a response candidate sentence according to the linguistic information, and a step of comparing a determination value determined from the similarity with a threshold value. Preparing a similarity between the response candidate sentence and the prohibited sentence by referring to a prohibited sentence database in which the prohibited sentence is stored, and restricting the output of the response candidate sentence based on the similarity It is. As a result, output of a response sentence with inappropriate content is restricted, and therefore an appropriate response sentence can be output.

  A response content control method according to a ninth aspect of the present invention is the response content control method described above, wherein a determination value of the response candidate sentence is calculated based on the similarity, and the determination value and the threshold value are calculated. Based on the comparison result, it is determined whether to limit the output of the response candidate sentence. Thereby, the output of the response sentence with inappropriate contents can be surely limited.

  A response content control method according to a tenth aspect of the present invention is the response content control method described above, wherein the response candidate sentence is determined based on information updated on a network. . Even when a response candidate sentence whose contents are undetermined is determined, an appropriate response sentence can be generated.

  A response content control method according to an eleventh aspect of the present invention is the response content control method described above, wherein the prohibited text stored in the prohibited text database includes a plurality of morphemes. . Thereby, the output of an inappropriate response candidate sentence can be restricted easily and reliably.

  A response content control method according to a twelfth aspect of the present invention is the response content control method described above, wherein the response candidate sentence and the prohibited sentence are morphologically analyzed, the morpheme included in the response candidate sentence, and the prohibition The morphemes included in the sentence are compared, and the similarity is calculated according to the number of morpheme matches. Thereby, the output of an inappropriate response candidate sentence can be restricted easily and reliably.

  A response content control method according to a thirteenth aspect of the present invention is the response content control method described above, wherein the number of morpheme matches is the number of morpheme matches in a specific part of speech. is there. Thereby, the output of an inappropriate response candidate sentence can be restricted easily and reliably.

  A response content control method according to a fourteenth aspect of the present invention is the response content control method described above, wherein the response content control method controls the content of a response sentence uttered in response to input voice data. Performing speech recognition processing on the input speech data; understanding language information of the speech recognition processed data; determining a response candidate sentence according to the language information; With reference to the prohibited sentence database in which is stored, limiting the utterance of the response candidate sentence according to the content of the response candidate sentence. As a result, output of a response sentence with inappropriate content is restricted, and therefore an appropriate response sentence can be output.

  According to the present invention, it is an object to provide a response system and a response content control method capable of outputting an appropriate response sentence.

  The response system according to the present embodiment is a response system that outputs a response sentence to an input sentence that has been input. The response system includes: an input unit that inputs an input sentence; a language understanding unit that understands language information of the input sentence input to the input unit; and a management unit that determines a response candidate sentence according to the language information; A response sentence generation unit that generates a response sentence based on a response candidate sentence whose determination value is lower than a threshold; and an output unit that outputs the response sentence generated by the response sentence generation unit. Yes. The management unit calculates the similarity between the response candidate sentence and the prohibited sentence by referring to the prohibited sentence database in which the prohibited sentence is stored. Furthermore, the management unit restricts the output of the response candidate sentence based on the similarity.

  A response robot 100 using the response system according to the present embodiment will be described with reference to FIG. FIG. 1 is a block diagram showing the configuration of the response robot 100. The response robot has a processing computer 10, a microphone 31, a speaker 32, and a motor 33. Here, it is assumed that the response robot 100 is a humanoid type voice interaction robot. Further, it is assumed that the response robot 100 is used in, for example, a vehicle exhibition facility. And if the speaker 102 speaks, the response robot 100 will talk autonomously.

  The microphone 31 acquires the voice uttered by the speaker 102. That is, the spoken sentence is received by the microphone 31. This spoken sentence includes language information. For example, the utterance sentence includes language information indicating a question or request to the response robot 100. Therefore, the microphone 31 serves as an input unit for inputting an input sentence including language information. As the microphone 31, for example, a microphone array having a plurality of microphones may be used.

  The voice data of the input sentence input to the microphone 31 is processed by the processing computer 10. The processing computer 10 includes a central processing unit (CPU), a read only memory (ROM), a random access memory (RAM), a communication interface, and the like, and controls various operations of the response robot 100. And the processing computer 10 performs various control according to the control program stored, for example in ROM. That is, the processing computer 10 serves as a control unit that controls the response robot 100 and executes a voice interaction program. Thus, a response sentence is generated in response to the input language information question or request. That is, output data corresponding to the response sentence is generated. The configuration of the processing computer 10 will be described later. Further, the processing computer 10 is not limited to a physically single configuration.

  Then, audio data corresponding to the response sentence generated by the processing computer 10 is output from the speaker 32. Thereby, the response robot 100 reads out the response sentence. That is, the response sentence is output to the outside in a predetermined direction and size. Thus, the speaker 32 serves as an output unit that outputs output data corresponding to the response sentence generated by the processing computer 10. The processing computer 10 is connected to the Internet 103. Then, the processing computer 10 downloads information data on the Internet 103.

  The motor 33 operates each part when outputting a response sentence. For example, the motor 33 operates the lip, neck, arm, foot, joint, wheel, and the like of a humanoid robot. Thereby, when responding, predetermined operations such as hand gestures and gestures are performed.

  The processing computer 10 includes a speech recognition unit 11, a language understanding unit 12, a dialogue management unit 13, a response sentence generation unit 14, a speech synthesis unit 15, an utterance prohibition sentence database 16, a question response database 17, and a motor control unit 20. ing.

  The voice recognition unit 11 performs voice recognition processing. For example, the voice recognition unit 11 converts voice data of an input sentence into text data. Thereby, input data in text format is generated. That is, the voice data is converted into a word string and a character string. For example, the speech recognition unit 11 performs decoding using a language model and an acoustic model. Furthermore, the speech recognition unit 11 performs morphological analysis. That is, the speech recognition result text is divided into morphemes, and part-of-speech information is added to each morpheme. Note that a noise removal process or the like may be performed before the voice recognition process. Such voice recognition processing can be performed using a predetermined voice recognition engine.

  The language understanding unit 12 performs language understanding processing on the speech-recognized data. Thereby, the language information of the input sentence is understood. That is, the language understanding unit 12 understands the contents of questions, requests, etc. in the input sentence. For example, a dependency analysis is performed on text data. It is analyzed which morpheme each morpheme covers. For example, it is determined to which verb a noun (for example, subject, subject, etc.) or adjective verb relates. Alternatively, it is determined which noun the adjective is associated with. In the language understanding process, a predetermined conceptual model may be used. A known processing method can be used for language understanding processing.

  The dialogue management unit 13 performs dialogue management. For example, an utterance candidate sentence is determined with reference to the question answering database 17 for a question whose language is understood. The question answering database 17 stores answer sentences that answer questions as a database. For example, (1) As for the answer to question sentence A “Where is the toilet?”, “As for right front”, (2) As for the answer to question sentence B “What is your name?” "It is AAA." That is, the question answer database 17 stores a question sentence and an answer sentence in association with each other. Then, the dialogue management unit 13 selects a question sentence that is closest to the input text sentence. An answer sentence corresponding to the closest question sentence in the question answering database 17 is selected as an utterance candidate sentence. Thereby, an utterance candidate sentence is determined according to the contents of the input sentence.

  Furthermore, the dialogue management unit 13 performs an intelligent process on a specific utterance. For example, text information is acquired from the Internet 103 for questions that require Internet information. If the question made by the speaker is "Tell me some recent news", get the news from the Internet. Specifically, information is acquired from the Internet 103, and an article portion is extracted using html tag information. Thus, news text data is downloaded. The text sentence of this news is set as an utterance candidate sentence.

  As described above, when the speaker 101 utters “tell me some latest news”, the dialogue management unit 13 downloads the latest news from the Internet 103. The latest news is downloaded from a portal site, for example. Specifically, Yahoo News can be downloaded and used as the latest news. And let the latest news be an utterance candidate sentence.

  Furthermore, the dialogue management unit 13 performs similarity calculation and utterance determination on the utterance candidate sentence. For example, the dialogue management unit 13 refers to the utterance prohibition sentence database 16 and calculates the similarity between the utterance prohibition sentence and the utterance candidate sentence. The utterance prohibition sentence database 16 stores one or more utterance prohibition sentences. Then, the similarity between each utterance prohibited sentence and utterance candidate sentence is calculated. A determination value is calculated based on the similarity, and speech determination is performed. Specifically, the dialogue management unit 13 compares the determination value with a threshold value, and determines whether to limit the output of the response candidate sentence based on the comparison result. That is, the dialogue management unit 13 determines whether the utterance candidate sentence is suitable for utterance. If the utterance determination determines that the utterance candidate sentence is not suitable for utterance, the utterance candidate sentence is changed. For example, a sentence different from an utterance candidate sentence determined to be unsuitable for utterance is set as an utterance candidate sentence. Specifically, a new news next to the latest news is set as an utterance candidate sentence. Then, the similarity and the determination value are obtained, and the speech determination is performed. Then, the process is repeated until an utterance candidate sentence suitable for utterance is obtained. That is, the same processing is repeated until news suitable for utterance appears.

  Thus, the dialogue management unit 13 determines the utterance candidate sentence according to the language information of the input sentence. Then, the dialogue management unit 13 refers to the utterance prohibition sentence database 16 and restricts the output of the utterance candidate sentences according to the contents of the utterance candidate sentences. That is, when it is determined that the content of the utterance candidate sentence is inappropriate based on the content of the utterance prohibited sentence database 16, the utterance of the utterance candidate sentence is prohibited. Note that the processing in the dialogue management unit 13 is not limited to the above processing. For example, you may perform the process which returns the content which the speaker 101 uttered as it is. That is, a sentence uttered by the speaker 101 may be used as an utterance candidate sentence as it is. The selection of the utterance candidate sentence is not particularly limited.

  The response sentence generation unit 14 generates a response sentence based on the utterance candidate sentence. For example, a response sentence is generated by adding a fixed sentence to the latest news. For example, a fixed sentence such as “Information on the Internet” is added before the latest news, or a fixed sentence such as “There was a news called” is added after the latest news. Of course, it may be a fixed sentence different from the above example, and there may be no fixed sentence. Furthermore, the utterance candidate sentence may be appropriately modified. The response sentence generation unit 14 generates a response sentence in a state where output of the response candidate sentence is restricted depending on the utterance determination result.

  The speech synthesis unit 15 performs speech synthesis processing on the response sentence. As a result, the response sentence data in the text format is converted into voice format data (for example, WAV data). Then, the voice data synthesized by the voice synthesizer is output by the speaker 32. A known method can be used for the speech synthesis process. In this way, the latest news combined with the fixed phrase is read out.

  Next, the control method of the response robot 100 according to the present embodiment will be described with reference to FIG. FIG. 2 is a flowchart showing a method for controlling the utterance content according to the present embodiment. Here, the process in the dialogue management unit 13 will be mainly described. First, the response robot 100 determines an utterance candidate sentence (step S101). An utterance candidate sentence is determined based on the linguistic information understood by the language understanding unit 12. As described above, the latest news downloaded from the Internet 103 is set as an utterance candidate sentence.

Next, the similarity A _i (i = 1, 2... M) between the utterance prohibition sentence stored in the utterance prohibition sentence database 16 and the utterance candidate sentence is calculated (step S102). Here, m is the number of prohibited speech statements stored in the prohibited speech database. That is, it is assumed that m utterance prohibition sentences are registered in the utterance prohibition sentence database 16. In this case, m similarities are calculated. That is, the similarity is calculated for each utterance prohibition sentence stored in the utterance prohibition sentence database 16. The utterance prohibition sentence is not limited to a complete sentence but may be a word.

The similarity is calculated by, for example, Equation 1 in FIG. In Formula 1, function F (x, y) is a function that returns 1 when morpheme x and morpheme y match, and returns 0 otherwise. Bk (k = 1,... L) is a morpheme of the utterance candidate sentence, and L is a morpheme number of the utterance candidate sentence. That is, the utterance candidate sentence includes L morphemes. Further, Cij (i = 1, 2,..., M, j = 1, 2,..., N _i ) is a morpheme of the speech prohibition sentence database. However, _ni is the morpheme number of the i-th speech prohibition sentence. That is, the i th utterance prohibition statement contains a n _i number of morphemes.

Thus, the function F (B _k , C _ij ) returns 1 when the morpheme of the utterance candidate sentence is included in the utterance prohibition sentence, and returns 0 when the morpheme of the utterance candidate sentence is not included in the utterance prohibition sentence. return. Therefore, the sum of F (B _k , C _ij ) in one utterance prohibited sentence is the number of coincidence of morphemes in the prohibited sentence. That is, Expression 1 shows a value obtained by dividing the number of morpheme matches between the utterance prohibited sentence and the utterance candidate sentence by the morpheme number of the utterance prohibited sentence. Here, the coincidence number of morphemes is the number of morphemes common to the utterance candidate sentence and the utterance prohibition sentence. Then, based on the number of matches between the morpheme of the utterance candidate sentence and the utterance prohibition sentence, the similarity of the utterance prohibition sentence to the utterance candidate sentence is calculated. That is, the number of morphemes that match the morphemes included in the utterance prohibition sentence among the morphemes included in the utterance candidate sentence is obtained. Then, the similarity A _i is calculated by dividing the number of matching morphemes by the number of morphemes of the utterance prohibition sentence. The similarity is calculated for all utterance prohibited sentences.

  Further, the similarity may be determined by paying attention to a specific part of speech. That is, the similarity may be calculated based on the number of morpheme matches for a specific part of speech. For example, when focusing only on nouns, verbs, and adjectives, 1 is returned when morphemes of nouns, verbs, and adjectives match. Other morphemes such as particles, conjunctions, auxiliary verbs are not counted. In this case, Pi is the number of morphemes of a specific part of speech (noun, verb, adjective) in the i-th speech prohibition sentence. Here, morphemes other than a specific part of speech are not counted. Note that the part of speech of interest is not limited to nouns, verbs, and adjectives. For example, attention may be paid only to nouns and verbs. As described above, the number of morphemes of specific parts of speech that are common to the utterance candidate sentence and the utterance prohibition sentence may be used as the coincidence number. Further, each part of speech may be weighted.

In this way, we obtain the similarity A _i relative to each utterance prohibition statement. The similarity A _i is a scale indicating how similar the utterance prohibited sentence and the utterance candidate sentence are. Therefore, the utterance candidate sentence having a high similarity to the utterance prohibition sentence may include content inappropriate for the utterance.

Then, it calculates a determination value based on the similarity A _i, compares the determined value with a threshold (step S103). Here, since the determination value is the maximum value of A _i , determination value = MAX (A _i ). That is, the largest value among all the similarities is set as the determination value. Then, the determination value is compared with a predetermined threshold value.

  If the determination value does not exceed the threshold value Z, the robot is uttered with the content of the utterance candidate sentence (step S104). That is, a response sentence is generated based on the utterance candidate sentence. For example, as described above, a sentence in which a fixed sentence is combined with the latest news is used as a response sentence. Then, the response sentence is output from the speaker 32. If the determination value does not exceed the threshold value Z, the utterance candidate sentence is not similar to all utterance prohibition sentences, and therefore, it is determined that the content is not inappropriate for utterance. For this reason, a response sentence is generated based on the utterance candidate sentence and output.

  If the determination value exceeds the threshold value Z, the robot is not uttered with the content of the utterance candidate sentence (step S105). That is, when the determination value is equal to or greater than the threshold value Z, it is estimated that inappropriate content is included in the utterance. Therefore, the utterance candidate sentence is not uttered. An utterance sentence is generated in a state where output of the utterance candidate sentence is restricted. Based on the comparison result between the determination value and the threshold value Z, it is determined whether or not the utterance of the utterance candidate sentence is restricted. Then, a response sentence is generated based on an utterance candidate sentence other than the utterance candidate sentence whose output is restricted. This can prevent inappropriate news and the like from being output. Further, as described above, by making another news as a response candidate sentence, it becomes possible to utter news with appropriate contents. That is, news with a determination value equal to or less than the threshold value Z can be set as a response candidate sentence.

  Here, an example of calculating the similarity will be described in detail with reference to FIGS. 3 to 5 are diagrams for explaining the similarity calculation. Here, in FIGS. 3 to 5, the similarity is calculated by changing the utterance candidate sentence and the utterance prohibition sentence. 3 to 5, the utterance candidate sentence is shown on the left side, and the utterance prohibition sentence is shown on the right side.

First, the similarity calculation example in FIG. 3 will be described. Here, “kill (verb)” is an utterance prohibition sentence. That is, the utterance candidate sentence includes “kill (verb)”. Therefore, the similarity A ₁ = 1/1 = 1. The similarity is calculated by paying attention to nouns, verbs, and adjectives. The utterance candidate sentence includes “kill (verb)”. In this case, the “kill (verb)” morpheme is included in both the utterance prohibited sentence and the utterance candidate sentence. Therefore, one morpheme matches between the utterance candidate sentence and the utterance prohibition sentence. That is, the coincidence number of morphemes is 1, and the numerator of Formula 1 is 1. Furthermore, the number of morphemes of a specific part of speech is one among the morphemes of the speech prohibition sentence. Therefore, the denominator of Equation 1 is 1. Therefore, the similarity A ₁ = 1/1 = 1.

Next, the similarity calculation example in FIG. 4 will be described. In the calculation example of FIG. 4, the utterance candidate sentence is the same as the calculation example of FIG. 3, and the utterance prohibition sentence is changed. Here, the similarity is calculated for the utterance prohibition sentence “go out of the noisy”. The similarity is calculated by paying attention to nouns, verbs, and adjectives. The utterance candidate sentence includes “noisy (adjective)”. In this case, a morpheme of “noisy (adjective)” is included in both the utterance prohibition sentence and the utterance candidate sentence. Therefore, one morpheme matches between the utterance candidate sentence and the utterance prohibition sentence. That is, the coincidence number of morphemes is 1, and the numerator of Formula 1 is 1. Furthermore, the number of morphemes of a specific part of speech among the morphemes of the speech prohibition sentence is three. Therefore, the denominator of Equation 1 is 3. Therefore, the similarity A ₂ = 1/3 = 0.33.

Next, a similarity calculation example in FIG. 5 will be described. This calculation example is an example in the case where the dialogue management unit 13 has a function of storing the utterance of the speaker and rephrasing it at a certain timing. For example, an utterance candidate sentence is “out of noisy” which is the utterance of the speaker in the past. In addition, a sentence “get out of noisy” is registered in the utterance prohibition sentence database 16 as an utterance prohibition sentence. In this case, since the three morphemes “Noisy (adjective)”, “Out (verb)”, and “I (verb)” match, the number of matches is 3. Further, the total morpheme number is 3 when focusing on a specific part of speech. Therefore, the similarity A ₂ = 3/3 = 1. Therefore, a response sentence is generated based on another utterance candidate sentence.

  It is preferable to add to the speech prohibition sentence database 16 sentences that have been uttered by the response robot 100 in the past and are rude to the customer. That is, the log of the sentence uttered by the response robot 100 is confirmed, and an inappropriate sentence is added to the utterance prohibition sentence database 16. By doing so, the response robot 100 does not speak a sentence similar to the rude sentence. Therefore, it is possible to appropriately control the utterance content.

  By controlling in this way, a sentence having an appropriate content can be uttered. For example, when the news shown in FIG. 10 is the latest news, the utterance of the news is prohibited. And appropriate news as shown in FIG. 6 can be read aloud. As a result, output of inappropriate contents is restricted, and it is possible to prevent the conversation person from feeling uncomfortable.

  In the above example, an example of reading out Internet news has been described. However, the present embodiment is not limited to this. For example, current traffic information may be uttered. Specifically, when a customer inquires about traffic information at a vehicle exhibition facility or the like, a response sentence having the contents as shown in FIG. 7 is uttered. Furthermore, this control method can also be used in an interpreting robot. That is, the contents that you do not want to translate may be registered as utterance prohibition sentences. In this case, translation software or the like is incorporated into the processing computer 10. Of course, this control method can be used not only for Japanese but also for other languages such as English.

In the above example, the maximum value of the similarity A _i is used as the determination value, but other values may be used as the determination value. For example, an average value of the similarity A _i may be used as the determination value. That is, when there are ten utterance prohibition sentences, the average value of the similarities A _i calculated for the ten utterance prohibition sentences is set as the determination value. Further, the s-th similarity (s is an integer of 2 or more) from the top may be used as the determination value, and the average value of the first to s-th similarity from the top may be used as the determination value. For example, the fifth similarity from the top may be used as the determination value, and the average value of the top five similarities may be used as the determination value.

  Thus, a value other than the maximum value may be used as the determination value. Thereby, accurate speech determination can be performed. For example, when a sentence having an appropriate content is registered in the utterance prohibition sentence database 16, the utterance of an utterance candidate sentence having an appropriate content is limited. By using a value other than the maximum value, even if the similarity with an utterance prohibition sentence having an appropriate content is accidentally increased, the utterance is not prohibited. Therefore, it is possible to speak appropriate content. Further, even when the number of registered utterance prohibition sentences is increased, it can be determined accurately. Therefore, it is possible to easily select and register an utterance prohibition sentence. As described above, the speech determination can be performed accurately by using the determination value based on the similarity.

  The speech prohibition sentence only needs to include one or more morphemes, but it is preferable to provide two or more morphemes. That is, it is preferable to register an utterance prohibition sentence including two or more morphemes in the utterance prohibition sentence database 16. By comparing an utterance prohibition sentence including two or more morphemes with an utterance candidate sentence, the similarity can be calculated more accurately. This will be described with reference to FIG. FIG. 8 is a diagram showing an example in which a morpheme has only one utterance prohibition sentence. In FIG. 8, since only one morpheme is included in the utterance prohibition sentence, the utterance prohibition sentence database is shown as an NG word database. In the NG word database, for example, “Baka”, “Aho”, “Kill”, and “Out” are registered. In FIG. 8, the NG word database is shown on the left side, and the utterance candidate sentences 1 to 3 are shown on the right side. The utterance candidate sentences 1 and 2 are sentence examples that are not desired to be uttered, and the utterance candidate sentence 3 is an example sentence that is desired to be uttered.

  For example, in the case of the utterance candidate sentence 1, since the word “idiot” matches, the utterance is prohibited. Similarly, in the case of the utterance candidate sentence 2, since “out” matches, the utterance is prohibited. Actually, even in the case of the utterance candidate sentence 3 that is desired to be uttered, since “out” matches, the utterance is prohibited. Therefore, it becomes difficult to discriminate between a sentence to be uttered and a sentence not to be uttered. That is, when only one morpheme is included in the utterance prohibition sentence, it can be determined only whether or not a morpheme matching the NG word exists in the utterance candidate sentence. For this reason, the above-mentioned measure of similarity cannot be calculated.

  For this reason, when the utterance prohibition sentence has only one morpheme, it becomes difficult to design the utterance prohibition sentence database 16. That is, it becomes difficult to select words to be registered in the utterance prohibition sentence database 16. Moreover, depending on the registered word, there is a possibility that an adverse effect may occur unexpectedly. That is, there are cases where the utterance of a sentence that is desired to be uttered is prohibited. Note that utterances with inappropriate content can be surely restricted by focusing on specific parts of speech such as nouns, adjectives and verbs.

  Furthermore, the response robot 100 can easily design and maintain the speech prohibition sentence database 16. For example, the response robot 100 can register the utterance prohibition sentence in the utterance prohibition sentence database 16 from the past log of the utterance contents. That is, referring to the past log of utterance contents, an utterance sentence including inappropriate contents from the actually uttered contents is registered as an utterance prohibition sentence. Thereby, the speech prohibition sentence database 16 can be created simply.

  Further, the threshold value can be adjusted with a test set of sentences close to the task performed by the response robot 100. Thereby, the threshold value can be easily adjusted. For example, a test set in which an appropriate response candidate sentence and an inappropriate response candidate sentence are prepared in advance is prepared. Specifically, 50 inappropriate response candidate sentences and 50 appropriate response candidate sentences are prepared. Then, a similarity is calculated for each response candidate sentence, and a determination value is obtained. A threshold value is set so that an appropriate response sentence is output and output of an inappropriate response sentence is restricted. In this way, the threshold value can be set easily. By adjusting the threshold value, it is possible to surely limit the utterances of inappropriate contents and reliably utter appropriate contents.

  For example, in a car showroom, a response robot is assumed to be a robot that responds to questions about a car on display. In an automobile showroom, a response sentence to an assumed question is prepared in advance as a test set. Then, each response sentence and the utterance prohibition sentence registered in the utterance prohibition sentence database 16 are compared. The threshold value is set so that the determination value with an inappropriate response sentence is equal to or greater than the threshold value, and the determination value with appropriate content is smaller than the threshold value. In this way, it is possible to reliably and easily determine whether or not to limit the output of the utterance candidate sentence.

  Next, another control method of the response robot 100 according to the present embodiment will be described with reference to FIG. FIG. 9 is a flowchart showing another aspect of the utterance content control method according to the present embodiment. First, it is determined whether or not there is a voice input (step S201). For example, the presence or absence of voice input can be determined based on the input of the microphone 31. Then, voice recognition processing is performed (step S202). Here, as described above, the speech recognition unit 11 performs speech recognition processing and converts speech data into text data.

Then, the speech recognition unit 12 performs language understanding processing based on the speech recognition processing result (step S203). A dialog management process is performed based on the language understanding process result (step S204). Thereafter, the utterance candidate sentence is determined based on the result of the dialogue management process (step S205). Then, the utterance candidate sentence and the utterance prohibition sentence are compared, and the similarity is calculated (step S206). Then, the maximum value of the similarity is compared with the threshold value Z as a determination value (step S207). If the determination value is smaller than the threshold value Z, the content of the utterance candidate sentence is confirmed as a response sentence (step S208). Since these processing steps are the same as the processing by the processing computer 10 described above, detailed description thereof is omitted.

  If the determination value is greater than or equal to the threshold value Z, the response sentence is changed (step S209). Here, the response robot 100 generates a response sentence “Sorry, please ask another question” to the speaker. Then, as described above, speech synthesis processing is performed on the text data of the response sentence (step S210). Then, the voice is output. The response robot 100 repeats such processing. Thereby, an appropriate content can be output simply. When the threshold value is smaller than the determination, the response robot 100 returns a question to the conversation of the speaker. Thereby, the utterance content can be determined easily.

  Note that the response sentence does not have to be output as audio data from the speaker. For example, it may be output to the speaker by displaying on a display screen or the like. Also, input is not limited to voice input. For example, a response system that responds by inputting text using a keyboard, a pointing device, a touch panel, or the like is also applicable. Thus, the input / output of data is not limited to that using audio data. Of course, application of the response system according to the present embodiment is not limited to application to a robot. Further, the response candidate sentence and the speech prohibition sentence may not be a complete sentence having a subject and a predicate, but may be a lyrics or a word.

  As described above, the response robot 100 downloads information (for example, the latest news) updated on a network such as the Internet 103 and uses it as a response candidate sentence. In such a case, the contents of the response candidate sentence are unconfirmed. That is, unlike the car navigation system or the like, the response robot 100 does not utter predetermined content. The response robot 100 performs intelligent processing for speaking uncertain content. For this reason, in the conventional control method, an inappropriate content may be output depending on the situation. In the control method according to the present embodiment, when the content of the response sentence to be output is inappropriate, the output is limited. When the content of the response sentence is inappropriate, for example, the response robot 100 generates a response sentence from another response candidate sentence. Therefore, it is possible to easily output a response sentence with appropriate contents.

It is a block diagram which shows the structure of the response robot which has a response system concerning embodiment of this invention. It is a flowchart which shows the response content control method concerning embodiment of this invention. In this Embodiment, it is a figure which shows the process example for controlling the content of a response. In this Embodiment, it is a figure which shows the process example for controlling the content of a response. In this Embodiment, it is a figure which shows the process example for controlling the content of a response. It is a figure which shows typically a mode that the robot which has a response system concerning embodiment of this invention is responding. It is a figure which shows typically a mode that the robot which has a response system concerning embodiment of this invention is responding. It is a figure for demonstrating the example of a suitable speech prohibition sentence. It is a flowchart which shows another response content control method concerning embodiment of this invention. It is a figure which shows typically a mode that the robot which has the conventional response system is responding.

Explanation of symbols

10 processing computers, 11 speech recognition units, 12 language understanding units, 13 dialogue management units,
14 response sentence generation part, 15 speech synthesis part, 16 utterance prohibition sentence database,
17 question answering database, 20 motor control unit,
31 microphones, 32 speakers, 33 motors,
100 response robot, 101 speaker, 103 Internet,
104 Spoken Dialogue Robot, 105 Facility User

Claims (14)

A response system that outputs a response sentence to an input sentence,
An input part for inputting an input sentence;
A language understanding unit for understanding language information of an input sentence input to the input unit;
A management unit for determining a response candidate sentence according to the language information;
A response sentence generator for generating a response sentence based on the response candidate sentence;
An output unit that outputs the response sentence generated by the response sentence generation unit,
The management unit calculates a similarity between the response candidate sentence and the prohibited sentence by referring to a prohibited sentence database in which the prohibited sentence is stored, and outputs the response candidate sentence based on the similarity. Response system to limit.   The management unit calculates a determination value of the response candidate sentence based on the similarity, and determines whether to limit output of the response candidate sentence based on a comparison result between the determination value and a threshold value The response system according to claim 1.   3. The response system according to claim 2, wherein the response candidate sentence is determined based on information updated on a network.   The response system according to any one of claims 1 to 3, wherein the prohibited sentence stored in the prohibited sentence database includes a plurality of morphemes.   Morphological analysis of the response candidate sentence and the utterance prohibition sentence, comparing the morpheme included in the response candidate sentence with the morpheme included in the utterance prohibition sentence, and the similarity according to the number of matches of the morpheme The response system according to claim 1, wherein the response system is calculated.   6. The response system according to claim 5, wherein the number of matching morphemes is a number of matching morphemes in a specific part of speech. A response system that utters a response sentence to input voice data,
A voice recognition unit that performs voice recognition processing on the input voice data;
A language understanding unit for understanding language information of the voice data subjected to the voice recognition processing;
A dialogue management unit that determines a response candidate sentence according to the language information and restricts the utterance of the response candidate sentence according to the content of the response candidate sentence by referring to a prohibited sentence database in which the prohibited sentence is stored A response system comprising: A response content control method for controlling the content of a response sentence output for an input sentence,
Understanding the language information of the input sentence,
Determining a response candidate sentence according to the language information;
Comparing a determination value determined from the degree of similarity with a threshold value,
Response content control that calculates the similarity between the response candidate sentence and the prohibited sentence by referring to the prohibited sentence database in which the prohibited sentence is stored, and restricts the output of the response candidate sentence based on the similarity Method.   9. The determination value of the response candidate sentence is calculated based on the similarity, and it is determined whether to limit the output of the response candidate sentence based on a comparison result between the determination value and a threshold value. The response content control method described.   10. The response content control method according to claim 8, wherein the response candidate sentence is determined based on information updated on a network.   The response content control method according to any one of claims 8 to 10, wherein the prohibited text stored in the prohibited text database includes a plurality of morphemes.   A morpheme analysis is performed on the response candidate sentence and the prohibited sentence, the morpheme included in the response candidate sentence is compared with the morpheme included in the prohibited sentence, and the similarity is calculated according to the number of morpheme matches. Item 12. The response content control method according to any one of Items 8 to 11.   13. The response content control method according to claim 12, wherein the number of morpheme matches is the number of morpheme matches in a specific part of speech. A response content control method for controlling the content of a response sentence uttered with respect to input voice data,
Performing voice recognition processing on the input voice data;
Understanding language information of the speech-recognized data;
Determining a response candidate sentence according to the language information;
A response content control method comprising: a step of referring to a prohibited text database in which a prohibited text is stored and restricting the utterance of the response candidate text according to the content of the response candidate text.

Images