JP2010152119A - Response generation device and program - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To output a response sentence in proper timing so that an input is not inhibited by the output of the response sentence from a device when a user intends to input an input sentence voluntarily or in response to the response sentence from the device. <P>SOLUTION: A language analysis part 20 performs voice recognition to a voice signal input from a microphone 12 to perform morphological analysis, an intention estimation part 22 estimates the intention of user utterance on the basis of analysis results, a response candidate generation part 124 generates response sentence candidates on the basis of a response candidate type associated with the estimated utterance intention, and an output timing measurement part 26 and an utterance waiting time clocking part 126 clocks whether or not waiting time determined for each combination of the utterance intention and the response candidate type. When the waiting time passes without the input of the next utterance from the user, it is decided that output timing of the response sentence comes, and an output part 28 converts the generated response sentence candidates into the voice signal to be output from a speaker 14. <P>COPYRIGHT: (C)2010,JPO&INPIT

Description

  The present invention relates to a response generation device and a program, and more particularly, to a response generation device and a program that output a response sentence at an appropriate timing.

  Conventionally, as a voice interaction device that responds at an appropriate timing, the speech speed as information of the input voice is calculated, and a word string is extracted from the input voice, and the succeeding from the appearance probability of the extracted word string Predicting the input word string, predicting the subsequent input time until the predicted subsequent word string is input based on the calculated speech speed, and responding based on the subsequent input time Has been proposed (see, for example, Patent Document 1). Further, it is described that in this spoken dialogue apparatus, information such as an utterance length, a fundamental frequency, a morpheme number, and a part of speech string may be used as input voice information.

In addition, when there are no words that appear more than a predetermined number of times in the sentence data corresponding to the input voice, a response sentence is output for each segmented voice input, and there are words that appear more than a predetermined number of times. Has also proposed a dialogue apparatus that outputs a response, searches for related contents using the word as a keyword, and outputs a response sentence based on the searched contents (see, for example, Patent Document 2).
JP 2008-241890 A JP 2007-328283 A

  However, in the above-mentioned spoken dialogue apparatus of Patent Document 1, since the output timing of the response sentence is determined based on the input voice information, even if the response sentence to be output is different, immediately after the voice input, etc. However, there is a problem that a response sentence may be output from the device even though the user is voluntarily inputting the next voice.

  Moreover, in the dialogue device of Patent Document 2, when there is a word that has appeared more than a predetermined number of times, a response sentence based on the retrieved content is output after outputting the conflict, but the response sentence output first In other cases, the response text from the device may be output even though the user is trying to input the next voice to respond to the response text output earlier. There is a problem that there is.

  The present invention has been made in order to solve the above-described problem. When a user is trying to input an input sentence voluntarily or in response to a response sentence from the apparatus, the response sentence from the apparatus is An object of the present invention is to provide a response generation device and a program capable of outputting a response sentence at an appropriate timing so that the input is not hindered by the output.

  In order to achieve the above object, a response generation apparatus according to a first aspect of the present invention includes an input unit for inputting an input sentence from a user, and an analysis result obtained by analyzing a structure of the input sentence input by the input unit. An intention estimation unit that estimates an intention represented by an input sentence, a response generation unit that generates at least one response sentence corresponding to the intention represented by the input sentence estimated by the intention estimation unit, and an estimation performed by the intention estimation unit Output a response sentence generated by the response generation means after a waiting time determined by a combination of the received intention and the response sentence generated by the response generation means, and input the input means before the waiting time elapses Control means for controlling so that the response sentence generated by the response generation means is not output when the next input sentence is input.

  The response generation program according to the second aspect of the invention estimates the intention represented by the input sentence from the analysis result obtained by analyzing the structure of the input sentence input by the input means for inputting the input sentence from the user. An intention estimation unit, a response generation unit that generates at least one response sentence corresponding to the intention represented by the input sentence estimated by the intention estimation unit, an intention estimated by the intention estimation unit, and the response generation unit After the waiting time determined by the combination with the response sentence generated in step elapses, the response sentence generated by the response generation means is output, and the next input sentence is input by the input means before the waiting time elapses. In this case, the program is made to function as a control unit that performs control so as not to output the response sentence generated by the response generation unit.

  According to the first invention and the second invention, when an input sentence from the user is input by the input means, the intention estimation means estimates the intention represented by the input sentence from the analysis result obtained by analyzing the structure of the input sentence. The response generation means generates at least one response sentence corresponding to the intention represented by the input sentence estimated by the intention estimation means. Then, the control means controls to output the response sentence generated by the response generation means after the waiting time determined by the combination of the intention estimated by the intention estimation means and the response sentence generated by the response generation means has elapsed. . In addition, the control unit performs control so that the response sentence generated by the response generation unit is not output when the next input sentence is input by the input unit before the waiting time elapses.

  Thus, when the user is going to input the input sentence spontaneously by waiting for the output of the response sentence until the waiting time determined by the combination of the estimated utterance intention and the response sentence elapses, from the device The response text can be output at an appropriate timing so that the input is not hindered by the output of the response text.

  According to a third aspect of the present invention, there is provided a response generation apparatus that estimates an intention represented by an input sentence from an input unit that inputs an input sentence from a user and an analysis result obtained by analyzing a structure of the input sentence input by the input unit. Intention estimation means, response generation means for generating at least one response sentence corresponding to the intention represented by the input sentence estimated by the intention estimation means, and at least one response generated by the response generation means When one response sentence is output from the sentences and there is an unoutput response sentence, after a waiting time determined by the combination of the intention estimated by the intention estimation means and the output response sentence has elapsed, When one other response sentence is output from the non-output response sentences, and the next input sentence is input by the input means before the waiting time elapses, the other one response Sentence It is configured to include a control means for controlling so as not to force.

  The response generation program according to the fourth aspect of the invention estimates the intention represented by the input sentence from the analysis result obtained by analyzing the structure of the input sentence input by the input means for inputting the input sentence from the user. Intention estimation means, response generation means for generating at least one response sentence corresponding to the intention represented by the input sentence estimated by the intention estimation means, and at least one response generated by the response generation means When one response sentence is output from the sentences and there is an unoutput response sentence, after a waiting time determined by the combination of the intention estimated by the intention estimation means and the output response sentence has elapsed, When another response sentence is output from the unoutput response sentences and the next input sentence is input by the input means before the waiting time elapses, Is a program for functioning as a control means for controlling so as not to output the response sentence.

  According to the third and fourth inventions, when an input sentence from the user is input by the input means, the intention estimation means estimates the intention represented by the input sentence from the analysis result obtained by analyzing the structure of the input sentence. The response generation means generates at least one response sentence corresponding to the intention represented by the input sentence estimated by the intention estimation means. Then, the control means outputs one response sentence from at least one or more response sentences generated by the response generation means, and when there is an unoutput response sentence, it is estimated by the intention estimation means After a waiting time determined by the combination of the intention and the output response text has elapsed, control is performed so that another response text is output from among the non-output response texts. Further, when the next input sentence is input by the input means before the waiting time elapses, control is performed so as not to output another response sentence.

  In this way, after outputting one response sentence, the user waits for the output of the next response sentence until the waiting time determined by the combination of the estimated utterance intention and the response sentence elapses. When an input sentence is to be input in response to a sentence, the response sentence can be output at an appropriate timing so that the input is not hindered by the output of the response sentence from the device.

  Further, the control means of the third and fourth inventions outputs a response sentence that prompts input of an input sentence when the non-output response sentence does not exist and the waiting time has elapsed. Can be controlled. Thereby, the dialogue can be further advanced.

  In addition, the waiting time when the response sentence that asks or states the content not included in the input sentence is combined with the intention estimated by the estimation means is compared with the intention estimated by the estimation means. Thus, it is possible to make the waiting time longer than that when the response sentence, the answer, the understanding, the repetition, or the response sentence to be confirmed are combined.

  When outputting a response to a question or statement about content not included in the input text, the topic will be changed to content not intended by the user, but the waiting time until the output of the response text is increased. By doing so, if there is an input from the user during the waiting time, the conversation can be continued without changing the topic. In addition, by increasing the waiting time after output, it is possible to wait sufficiently for the response from the user to the output response sentence. It is possible to avoid output.

  Further, the waiting time is shortened according to an elapsed time from when the input sentence is first input, or based on a silence time from the past response sentence output by the user to the next input sentence input. Thus, the longer the silence time, the longer it can be. In this way, when the user has become accustomed to interacting with the device by dynamically changing the waiting time, or when there are individual differences such as whether the user responds carefully or responds quickly In response, the response sentence can be output at an appropriate timing.

  As described above, according to the response generation device and the program of the present invention, when the user intends to input an input sentence spontaneously or in response to a response sentence from the apparatus, the response sentence from the apparatus There is an effect that a response sentence can be output at an appropriate timing so that the input is not hindered by the output.

  Hereinafter, embodiments of the present invention will be described in detail with reference to the drawings. In the present embodiment, a case will be described in which the present invention is applied to a response generation apparatus that executes a predetermined process and outputs a voice by using an utterance from a user as an input.

  As illustrated in FIG. 1, the response generation device 10 according to the first exemplary embodiment includes a microphone 12 that collects user utterances to generate an audio signal, a speaker 14 that outputs audio, a microphone 12, and a speaker 14. And a computer 16 that executes a predetermined process for outputting a response sentence at an appropriate timing.

  The computer 16 includes a CPU that controls the entire response generation apparatus 10, a ROM as a storage medium that stores various programs such as a response generation process and an intention estimation process described later, a RAM that temporarily stores data as a work area, and various information Includes an HDD as storage means, an I / O (input / output) port, and a bus connecting them. A microphone 12 and a speaker 14 are connected to the I / O port.

  In addition, when the computer 16 is described by functional blocks divided for each function realizing means determined based on hardware and software, the voice signal input from the microphone 12 is recognized as shown in FIG. The intent of the user utterance is estimated based on the analysis result of the language analysis unit 20 that performs morphological analysis on the character string information indicating the user utterance that has been voice-recognized using a general morphological analyzer. The intention generation unit 22, the response generation unit 24 that generates a response sentence based on the response candidate type associated with the intention estimated by the intention estimation unit 22, and the output timing of the response sentence generated by the response generation unit 24. The output timing measurement unit 26 that determines whether or not the output timing is received, and the response generated when the output timing measurement unit 26 determines that the output timing has been reached. It can be a converted to a voice signal represented by a configuration including the output unit 28 to output from the speaker 14.

  As shown in FIG. 2, the intention estimation unit 22 includes “statement”, “statement type question”, “Y / N question”, “statement type answer”, “Y / N answer”, “ It is estimated whether the utterance intention class is “OK” or “Other”. The classification of the intention of the user utterance is determined based on a move in the discourse analysis (a structural unit that bears a discourse function such as a question and an answer). When classifying a user utterance into any of the above intentions, a classifier that best represents the linguistic characteristics of the class is designed in a classification task of classifying the utterance into a plurality of classes according to the linguistic characteristics of the user utterance.

  When the utterance intention class is “statement type question”, for example, a sentence such as “What time is it now?” Is a sentence that asks something and takes the form of a question sentence. Similarly, in the case of the “Y / N question”, for example, a question sentence format such as “Can you make a gratin?” Is used, and the question is Yes or No. On the other hand, in cases other than the “statement type question” and the “Y / N question”, generally, a plain text is used. When the utterance intention class is “statement”, for example, a sentence having a certain meaning and a clause is included, such as a sentence “a friend will come to play”. In addition, for example, even if a verb is omitted, such as “Friends to play”, a sentence complemented with an omitted part (coming) has a feature that includes a clause. Statement ".

  Also, when the utterance intention class is “statement type answer”, for example, it becomes an answer sentence “Please make a gratin” to the question “What did you make?” Question of 5W1H type On the other hand, it is a sentence that answers a question in a format other than Yes and No, and has a feature of including a clause. In addition, for example, in the case of an utterance in which a verb is omitted, such as “Gratin”, the sentence supplemented with the omitted part (make me) includes a clause. The utterance intention class is “statement type answer”. In addition, when the utterance intention class is “Y / N answer”, for example, it becomes an answer sentence “yes” to the question “Would your friend come to play?” On the other hand, the answer is in the form of Yes, No.

  In addition, when the intention of the utterance is “OK”, for example, it is a sentence that agrees or disagrees, or a sentence that understands, such as “Sokka ~”, “Yeah”, and “Yes”. When the utterance intention class is “other”, for example, a sentence equivalent to laughter or “Ye” corresponding to the utterance of the other party “I am a fine man”.

  The response generation unit 24 generates a response sentence based on the response candidate type selected from the response candidate types associated with the utterance intention class estimated by the intention estimation unit 22.

  As shown in FIG. 3, when the utterance intention class is “statement”, “acknowledge”, “repetition”, “confirmation”, “Fukahori question”, and “statement” are associated as response candidate types. Further, when the utterance intention class is “declaration type question”, when the utterance intention class is “Y / N question”, “Y / N answer” and the utterance intention class are “declaration type”. Type answer ”,“ OK ”, when the speech intention class is“ Y / N answer ”,“ OK ”, and when the speech intention class is“ OK ”,“ OK ”and“ statement ” Are associated as response candidate types and stored in a predetermined storage area. When multiple response candidate types are associated with the utterance intention class, select a response candidate type that is not randomly selected from multiple response candidate types or that has not appeared in the past utterance history. To select one or more response candidate types.

  Further, the response generation unit 24 generates a response sentence as shown in “response sentence example” in FIG. 3 based on the selected response candidate type, for example. Specifically, when the user utterance is “Friend comes to play.”, The utterance intention class is estimated to be “statement”, and when “OK” is selected as the response candidate type for this, , A response sentence such as “Soka ~” is generated. Further, when the response candidate type is “repeat”, a response sentence such as “a friend comes to play” is generated that repeats the contents of the user utterance. Further, when the response candidate type is “confirmation”, a response sentence such as “Would a friend come to play?” For confirming the content of the user utterance is generated. When the response candidate type is “Fukahori question”, a response sentence such as “When are you coming to play?” Is generated to ask a question about items not included in the content of the user utterance. Further, when the response candidate type is “statement”, a response sentence such as “I want a friend” is generated as a new response to the user's utterance content.

  The response sentence is generated by applying a word extracted from the user utterance to a format determined for each response candidate type in advance, or according to a predetermined response sentence generation rule according to the attribute of the word included in the user utterance, etc. This can be done using known techniques.

  The output timing measurement unit 26 measures the duration of the state in which no user utterance is input, and determines that it is the response statement output timing when the duration has exceeded the “response statement output waiting time”. When the next utterance is input from the user before the response sentence output waiting time elapses, the measurement of the duration time is interrupted and the response sentence output timing is not determined.

  The response sentence output waiting time used for this determination is determined for each combination of the estimated utterance intention class and the response candidate type associated therewith. For example, as shown in FIG. 3, in the case of a combination of the utterance intention class “statement” and the response candidate types “accepted”, “repeated”, and “confirmed”, the response sentence output waiting time is set to “0 (waiting time) None) ”, and in the case of a combination of the speech intention class“ statement ”and each of the response candidate types“ Fukahori question ”and“ statement ”, the response sentence output waiting time is“ T1 (for example, 7 seconds) ”. When a response sentence based on the response candidate types “Fukahori question” and “statement” is output for a user utterance of the utterance intention class “statement”, the response sentence output waiting time is set to T1 in the user's utterance content. This is because the dialogue will develop in another direction, such as asking questions that are not included or making a new turnaround. This avoids a situation in which a smooth conversation is hindered by a response sentence that expands the conversation in another direction from the device even though the user continues to voluntarily input the utterance. be able to.

  Next, a response generation processing routine in the response generation device 10 of the first exemplary embodiment will be described with reference to FIG. This routine is performed by the CPU executing a response generation program stored in the ROM.

  In step 100, it is determined whether or not a user utterance has been input from the microphone 12. If a user utterance has been input, the process proceeds to step 102. If not, the determination in this step is repeated until it is input. . Here, it is assumed that the user utterance “Friend comes to play” is input.

  In step 102, the input voice signal indicating the user utterance is recognized as voice string information, and morphological analysis is performed on the string information. Next, in step 104, intention estimation processing is executed based on the analysis result of morphological analysis.

  Here, the intention estimation processing routine will be described with reference to FIG.

  In step 200, it is determined whether the intention of the user utterance is “question” based on the analysis result of the morphological analysis. If the sentence indicating the user's utterance is in the question sentence format, it is determined to be a “question”, and the process proceeds to step 202. If not, the process proceeds to step 208.

  In step 202, it is determined whether or not the question sentence indicating the user utterance is a “Y / N question” that is answered in the form of Yes or No. In the case of “Y / N question”, the process proceeds to step 204 to store “Y / N question” in the variable i. If it is not a “Y / N question”, it is determined that it is a “statement type question”, and the process proceeds to step 206 to store “statement type question” in the variable i.

  If it is determined in step 200 that the utterance intention is not “question” and the process proceeds to step 208, it is determined whether or not the intention of the user utterance is “statement”. If the sentence indicating the user utterance or the sentence supplemented with the omitted part of the user utterance includes a clause, the sentence is determined to be “declaration type”, and the process proceeds to step 210. Proceed to 216.

  In step 210, it is determined whether or not the intention of the user utterance is “statement answer”. Reference is made to the previous response sentence output from the device before the input user utterance, and if the previous response sentence is a question sentence, it is determined that the answer is a “statement answer”, and the process proceeds to step 212. Go ahead and store “statement answer” in variable i. If it is not “statement answer”, the process proceeds to step 214 to store “statement” in the variable i.

  If it is determined in step 208 that the utterance intention is not “statement” and the process proceeds to step 216, it is determined whether or not the intention of the user utterance is “Y / N answer”. Refer to the previous response sentence output from the device before the input user utterance. If the previous response sentence is a question sentence, it is determined that the answer is “Y / N answer”, and the step Proceeding to 218, "Y / N answer" is stored in variable i. If it is not “Y / N answer”, the process proceeds to step 220 to store “OK” in the variable i.

  Here, since the user utterance “Friend comes to play” is a plain text, it is denied in step 200, and because it includes a clause, it is affirmed in step 208 and is not an answer to the previous response sentence. In step 210, i = “statement” is stored. Next, in step 222, the variable i indicating the estimated utterance intention class is output, and the process returns.

  Returning to step 106 of the response generation process (FIG. 4), one or more response candidate types are selected from the response candidate types corresponding to the utterance intention class i output in the intention estimation process. Here, since i = “statement”, referring to FIG. 3, one or more of response candidate types “accepted”, “repeated”, “confirmed”, “Fukahori question”, and “statement” are selected. . When selecting a plurality of response candidate types, the response sentence output waiting time is “0 (waiting time) rather than selecting a plurality of response candidate types whose response sentence output waiting time is“ 0 (no waiting time) ”. None) It is preferable to select a combination of the response candidate type “” and the response candidate type “T1” as the response sentence output waiting time. Here, “OK” and “Fukahori Question” are selected.

  Next, in step 108, a response sentence based on the selected response candidate type is generated. For the response candidate type “OK”, for example, a response sentence such as “Soka ~” is generated, and for the response candidate type “Fukahori question”, a response sentence such as “when do you come to play?” Is generated. The

  Next, in step 110, it is determined whether or not there is a response sentence output waiting time of “0 (no waiting time)” among the selected response candidate types. If it exists, the process proceeds to step 112, and the response sentence “So ~~” based on the response candidate type whose response sentence output waiting time is “0 (no waiting time)” among the response sentences generated in step 108 above. ”Is voice-synthesized, and the voice is output from the speaker 14, and the process proceeds to Step 114. If none of the selected response candidate types has a response sentence output waiting time of “0 (no waiting time)”, the process proceeds to step 114 as it is.

  In step 114, it is determined whether or not there is a response sentence output waiting time of “T1” among the selected response candidate types. When it exists, it progresses to step 116, and when it does not exist, a process is complete | finished.

  In step 116, it is determined whether or not a next utterance has been input from the user. When the next utterance is input, the response text based on the response candidate type of the response text output waiting time “T1” generated in step 108 is not output, and the process returns to step 102. If the next utterance is not input, the process proceeds to step 118, and it is determined whether or not the duration time T1 in which no utterance is input has elapsed since the previous user utterance was input. If T1 has not elapsed, the process returns to step 116 to repeat the processing. If T1 has elapsed, it is determined that the response sentence output timing is based on the response candidate type of the response sentence output waiting time “T1”. Then, the process proceeds to step 120, from which the response sentence “when will you come to play?” Based on the response candidate type whose response sentence output waiting time is “T1” is synthesized from the response sentences generated in step 108 above. Then, sound is output from the speaker 14, and the process is terminated.

  As described above, according to the response generation device of the first exemplary embodiment, the dialogue develops in another direction, such as asking about a matter that is not included in the user's utterance content or making a new return. In the case of outputting a response sentence that shows the response sentence, it is possible to output the response sentence at an appropriate timing by providing a waiting time determined by the combination of the utterance intention and the response candidate type before the response sentence is output. For this reason, it is possible to avoid a situation in which a response is output from the device and smooth dialogue is hindered even though the user continues to voluntarily input an utterance.

  In the first embodiment, a case has been described in which a response sentence based on the selected response candidate type is generated and then output after waiting until an output timing. However, an output waiting time corresponding to the selected response candidate type is described. It is also possible to generate and output a response sentence based on the selected response candidate type after waiting for elapse.

  In the first embodiment, the case where the response sentence output waiting time is set to “0 (no waiting time)” and “T1” has been described. More specifically, three or more waiting times are set. Alternatively, different waiting times may be set for each combination of the utterance intention class and the response candidate type.

  Also, the case where the speech intention class is classified into one of “statement type question”, “Y / N question”, “statement answer”, “statement”, “Y / N answer”, and “OK” is explained as an example However, since a filler or a tweet that cannot be recognized as an intention may be input as a user utterance, a classification of “others” may be added. In this case, it may be determined as “others” when it cannot be classified into any of the above classes.

  Next, a second embodiment will be described. The second embodiment is different from the first embodiment in that an utterance waiting time for waiting for an utterance from the user is provided after a response sentence is output from the apparatus. In addition, about the structure and process similar to 1st Embodiment, the same code | symbol is attached | subjected and description is abbreviate | omitted.

  As illustrated in FIG. 6, the response generation device 110 according to the second embodiment includes a microphone 12 that collects user utterances to generate an audio signal, a speaker 14 that outputs audio, and the microphone 12 and the speaker 14. And a computer 16 that executes a predetermined process for outputting a response sentence at an appropriate timing.

  When the computer 16 is described with function blocks divided for each function realization means determined based on hardware and software, as shown in FIG. 6, the language analysis unit 20, the intention estimation unit 22, and the intention estimation unit 22 perform estimation. A response candidate generation unit 124 that generates a response sentence candidate based on the response candidate type associated with the intended intention, a response candidate storage unit 30 that temporarily stores the response sentence candidate generated by the response candidate generation unit 124, and a response sentence The utterance waiting time from the user after the output is measured, and the utterance waiting time measuring unit 126 that determines whether or not the output timing of the next response sentence is reached, and the utterance waiting time measuring unit 126 determines that the output timing is reached. The output unit 28 selects one of the response sentence candidates stored in the response candidate storage unit 30 and converts it into an audio signal and outputs it from the speaker 14. It can be represented by the inclusive configuration.

  The response candidate generation unit 124 generates a response sentence candidate based on the response candidate type associated with the utterance intention class estimated by the intention estimation unit 22. The response candidate types associated with each of the utterance intention classes and the example response sentences based on each of the response candidate types are the same as in the case of the first embodiment.

  In addition, the response candidate generation unit 124 determines a response sentence selected from the generated response sentence candidates as a response sentence to be output, and transmits the determined response sentence to the output unit 28. On the other hand, the response sentence candidates not selected are stored in the response candidate storage unit 30. When selecting one response sentence candidate from the generated response sentence candidates, it is selected by selecting at random or selecting a response sentence candidate that does not appear in the past utterance history.

  The utterance waiting time measuring unit 126 measures the duration of the state in which the user utterance is not input after outputting the first response sentence, and if the duration exceeds the “utterance waiting time”, the following It is determined that the response sentence output timing has come. When the next utterance is input from the user before the utterance waiting time elapses, the measurement of the duration time is interrupted and it is not determined that the next response sentence is output.

  The speech waiting time used for this determination is determined for each combination of the estimated speech intention class and the response candidate type associated therewith. For example, as shown in FIG. 7, in the case of a combination of the utterance intention class “statement” and the response candidate types “accepted”, “repeat”, and “statement”, the utterance waiting time is set to “T2”, In the case of a combination of the class “statement” and the response candidate types “confirmation” and “Fukahori question”, the speech waiting time is set to “T3”. For example, T3 is set to be longer than T2, such as T2 = 8 seconds and T3 = 15 seconds. This is because the content of the response sentence to be output is such as asking the user for an answer, so that the user has enough time to think about the answer. Further, T3 may be set to ∞ and wait until there is a next utterance from the user. Accordingly, it is possible to avoid that the smooth response is hindered by the next response text being output from the device even though the user considers the response to the response text output from the device.

  Next, a response generation processing routine in the response generation apparatus 110 according to the second embodiment will be described with reference to FIG. This routine is performed by the CPU executing a response generation program stored in the ROM.

  In step 100, it is determined whether or not a user utterance has been input. Next, in step 102, a voice signal indicating the input user utterance is recognized as character string information, and morphemes are applied to the character string information. Next, in step 104, the intention estimation process is executed based on the analysis result of the morphological analysis, and the utterance intention class i is output. Here, as in the first embodiment, it is assumed that the utterance intention class i = “statement” is output for the user utterance “Friend comes to play”.

  Next, in step 300, a response sentence candidate based on the response candidate type corresponding to the utterance intention class i output in the intention estimation process is generated. When a plurality of response candidate types corresponding to the utterance intention class i are associated, response sentence candidates based on all the associated response candidate types are generated. Here, based on the response candidate types “OK”, “Repeat”, “Confirm”, “Fukahori Question”, and “Statement” corresponding to i = “Statement”, “Soka ~” (OK), “Your friends come to play” (repeated), “Would your friends come to play?” (Confirmation), “When do you come to play” (Fukahori question), “I want friends too” (statement) Response sentence candidates such as are generated.

  Next, in step 302, one response sentence candidate is selected from the response sentence candidates generated in step 300, voice-synthesized, and output from the speaker 14. Here, it is assumed that “when do you come to play?” (Fukahori question) is selected and output. Next, in step 304, the remaining response sentence candidates that have not been output are temporarily stored in the response candidate storage unit 30.

  Next, in step 306, it is determined whether or not a next utterance has been input from the user. If the next utterance is input, the response sentence candidate stored in step 304 is not output, and the process returns to step 102. If the next utterance is not input, the process proceeds to step 308.

  In step 308, after the first response sentence is output in step 302, the utterance waiting time associated with the response candidate type of the response sentence in which the duration of the state in which no utterance is input is output is set. Judge whether or not it has passed. Here, the speech waiting time “T3” of the response candidate type “Fukahori question” for the speech intention class i = “statement”. If T3 has not elapsed, the process returns to step 306 to repeat the process. If T3 has elapsed, it is determined that the output timing of the next response sentence is reached, and the process proceeds to step 310.

  In step 310, it is determined whether or not there is a response sentence candidate stored in the response candidate storage unit 30. If there is a response sentence candidate, the process proceeds to step 312, and the response sentence candidate is set to 1 from the response candidate storage unit 30. For example, “I also want a friend” (statement) is selected and output, and the process returns to step 306. When there is no response sentence candidate stored in the response candidate storage unit 30, the process proceeds to step 314 to generate and output a prompt response prompting the user to speak, and the process ends. The prompt response sentence is, for example, a response sentence such as “I want to hear more stories”, “Is there something else?”, “So then”, “Can I talk to you?”.

  As described above, according to the response generation device of the second exemplary embodiment, when the content of the response sentence to be output is asking the user for an answer, the utterance is output after the response sentence is output. By providing a waiting time determined by the combination of the intention and the response candidate type, it is possible to give the user sufficient time to think about an answer and output the next response sentence at an appropriate timing. For this reason, it is possible to avoid the fact that the next response text is output from the device and smooth interaction is hindered even though the user considers an answer to the response text output from the device.

  In the second embodiment, when there are a plurality of response candidate types corresponding to the utterance intention class, a case has been described in which response sentence candidates are generated for all of them. A response sentence candidate may be generated by selecting such a response candidate type. For example, for the utterance intention class i = “statement”, the response candidate types “OK” and “Fukahori question” are selected, and “Wow ~” and “When do you come to play?” Are generated as response sentence candidates. You may do it.

  In the second embodiment, a case has been described in which a response sentence candidate based on a response candidate type is generated and then stored in the response candidate storage unit. However, the response candidate type corresponding to the utterance intention class is temporarily stored. In addition, when the output timing is reached, one of the stored response candidate types may be selected, and a response sentence based on the selected response candidate type may be generated and output. .

  Further, in the second embodiment, the case where “T2” and “T3” are set as the speech waiting time has been described, but more than three stages of waiting time may be set in more detail. A different waiting time may be set for each combination of the response candidate type.

  Next, a third embodiment will be described. In the third embodiment, the response sentence output waiting time and the utterance waiting time are used together, and the response sentence output waiting time and the utterance waiting time are dynamically changed as in the first and second embodiments. Is different. In addition, about the structure and process similar to 1st and 2nd embodiment, the same code | symbol is attached | subjected and description is abbreviate | omitted.

  As illustrated in FIG. 9, the response generation apparatus 210 according to the third embodiment includes a microphone 12 that collects user utterances to generate an audio signal, a speaker 14 that outputs audio, and the microphone 12 and the speaker 14. And a computer 16 that executes a predetermined process for outputting a response sentence at an appropriate timing.

  When the computer 16 is described with function blocks divided for each function realizing means determined based on hardware and software, as shown in FIG. 9, a language analysis unit 20, an intention estimation unit 22, a response candidate generation unit 124, A response candidate storage unit 30, an output timing measurement unit 26, an utterance waiting time measurement unit 126, an output unit 28, and a waiting time control unit 32 that performs control to dynamically change the response sentence output waiting time and the utterance waiting time are included. It can be expressed in configuration.

  In the third embodiment, when outputting one response sentence candidate selected from the generated response sentence candidates or the response sentence candidates stored in the response candidate storage unit 30, the response candidates corresponding to the response sentence candidates Wait for the response message output waiting time for the type to be output.

  For example, a response based on the response candidate types “OK”, “Repeat”, “Confirmation”, “Fukahori Question”, and “Statement” to the user utterance “Friend comes to play” (i = “Statement”) When a sentence candidate is generated and “when do you come to play?” (Fukahori question) is output first, in the second embodiment, the first response sentence is output as soon as it is selected. Then, after waiting for the elapse of the response sentence output waiting time “T1” of “Fukahori question” to i = “statement”, it is output. When T1 elapses without the user's next utterance, for example, “I also want a friend” (statement) is selected and output from the remaining response sentence candidates.

  Here, when outputting the second and subsequent response sentences, the response sentence output waiting time of the response sentence to be output competes with the utterance waiting time of the previously output response sentence. For example, in the above example, the utterance waiting time “T3” is set for the first response sentence “When do you come to play?” (Fukahori question), and the second response sentence “I am also a friend? The response sentence output waiting time “T1” is set for “I want” (statement). In this case, the time from outputting “When do you come to play?” To outputting “I want a friend”, T1 has priority, T3 has priority, or the average time of T1 and T3, etc. However, it is preferable to select the longer waiting time in view of the fact that the user's utterance is not hindered by the output of the response sentence from the device.

  The waiting time control unit 32 dynamically changes the response sentence output waiting time and the utterance waiting time. For example, a long waiting time is set at the start of the dialog when the user is unfamiliar with the dialog, and when a predetermined time elapses from the start of the dialog, the waiting time is changed to a waiting time shorter than the initially set waiting time. Also, for each user, a history of silence time in the dialogue is taken, the user is authenticated by using an image obtained by photographing the user's face, and the waiting time is determined by referring to the history of the user's silence time. May be. Specifically, the average of the silence time until the user utterance input after the response sentence output based on the response candidate types “statement”, “repeat”, “statement answer”, “Y / N answer”, and “OK” is T2 The average of the silence time until the user utterance input after the response sentence output based on the response candidate types “confirmation” and “Fukahori question” can be T3, and the average of the silence time of the entire dialogue can be T1.

  As described above, according to the response generation device of the third exemplary embodiment, by dynamically changing the waiting time, when the user has become accustomed to the dialogue with the device, or when the user thinks carefully It is possible to output a response sentence at an appropriate timing in response to a case where there is an individual difference such as a type to be answered or a type to be answered immediately.

  In the first to third embodiments described above, the case where voice output by a speaker is performed has been described as an example. However, the present invention is not limited to this, and a response sentence may be displayed on a display. Moreover, although the case where audio | voice was input into the microphone from the user was demonstrated to the example, you may make it a user input the text as an input sentence using a keyboard etc.

It is a block diagram which shows schematic structure of the response generation apparatus which concerns on 1st Embodiment. It is a table | surface which shows the speech intention class and the speech example. It is a table | surface which shows a speech intention class, a response candidate type, a response sentence output waiting time, and a response sentence example. It is a flowchart which shows the response production | generation process routine of 1st Embodiment. It is a flowchart which shows the intention estimation process routine. It is a block diagram which shows schematic structure of the response generation apparatus which concerns on 2nd Embodiment. It is a table | surface which shows a speech intention class, a response candidate type, a speech waiting time, and a response sentence example. It is a flowchart which shows the response production | generation process routine of 2nd Embodiment. It is a block diagram which shows schematic structure of the response generation apparatus which concerns on 3rd Embodiment.

Explanation of symbols

10, 110, 210 Response generation device 12 Microphone 14 Speaker 16 Computer 20 Language analysis unit 22 Intention estimation unit 26 Output timing measurement unit 28 Output unit 30 Response candidate storage unit 32 Wait time control unit 124 Response candidate generation unit 126 Speaking waiting time measurement Part

Claims (7)

An input means for inputting an input sentence from the user;
An intention estimation unit that estimates an intention represented by the input sentence from an analysis result obtained by analyzing a structure of the input sentence input by the input unit;
Response generation means for generating at least one response sentence corresponding to the intention represented by the input sentence estimated by the intention estimation means;
After the waiting time determined by the combination of the intention estimated by the intention estimating means and the response sentence generated by the response generating means has elapsed, the response sentence generated by the response generating means is output and the waiting time has elapsed. Control means for controlling not to output the response sentence generated by the response generation means when the next input sentence is input by the input means before
A response generation device including: An input means for inputting an input sentence from the user;
An intention estimation unit that estimates an intention represented by the input sentence from an analysis result obtained by analyzing a structure of the input sentence input by the input unit;
Response generation means for generating at least one response sentence corresponding to the intention represented by the input sentence estimated by the intention estimation means;
One response sentence is output from at least one response sentence generated by the response generation means, and when there is an unoutput response sentence, the intention estimated by the intention estimation means is output. After the waiting time determined by the combination with the response sentence elapses, another response sentence is output from the non-output response sentences, and the next input sentence is input by the input means before the waiting time elapses. Is inputted, control means for controlling not to output the other one response sentence,
A response generation device including:   3. The response generation according to claim 2, wherein the control means controls to output a response sentence that prompts input of an input sentence when the non-output response sentence does not exist and the waiting time has elapsed. apparatus.   For the intention estimated by the estimation means, the waiting time when a response sentence that asks or states a question about the contents not included in the input sentence is combined with the intention estimated by the estimation means, The response generation device according to any one of claims 1 to 3, wherein the response generation device is longer than the waiting time when a response sentence, an answer, an understanding, a repetition, or a confirmation response is combined.   The waiting time is shortened according to the elapsed time from when the input sentence was first input, or based on the silence time from the past response sentence output by the user to the next input sentence input, The response generation device according to claim 1, wherein the response generation device is configured to be longer as the silence time is longer. Computer
An intention estimation unit that estimates an intention represented by the input sentence from an analysis result obtained by analyzing a structure of the input sentence input by an input unit that inputs an input sentence from a user;
Response generation means for generating at least one response sentence corresponding to the intention represented by the input sentence estimated by the intention estimation means;
After the waiting time determined by the combination of the intention estimated by the intention estimating means and the response sentence generated by the response generating means has elapsed, the response sentence generated by the response generating means is output and the waiting time has elapsed. Control means for controlling not to output the response sentence generated by the response generation means when the next input sentence is input by the input means before
A response generator to make it function. Computer
An intention estimation unit that estimates an intention represented by the input sentence from an analysis result obtained by analyzing a structure of the input sentence input by an input unit that inputs an input sentence from a user;
Response generation means for generating at least one response sentence corresponding to the intention represented by the input sentence estimated by the intention estimation means;
One response sentence is output from at least one response sentence generated by the response generation means, and when there is an unoutput response sentence, the intention estimated by the intention estimation means is output. After the waiting time determined by the combination with the response sentence elapses, another response sentence is output from the non-output response sentences, and the next input sentence is input by the input means before the waiting time elapses. Is inputted, control means for controlling not to output the other one response sentence,
A response generator to make it function.

Images