JP2009198686A - Response generator and program - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide a response generator and a program, capable of generating a proper response to a voice recognition result, and reducing an erroneous response. <P>SOLUTION: This response generator 10 includes: an input part 12 for recognizing an utterance of a user, and generating recognition candidates; a response generating part 14 for generating a response candidate corresponding to each recognition candidate generated in the input part 12; a response selecting part 16 for selecting the response candidate having the largest number of the corresponding recognition candidates generated in the response generating part 14; and an output part 18 outputting the response candidate selected in the response selecting part 16. <P>COPYRIGHT: (C)2009,JPO&INPIT

Description

  The present invention relates to a response generation apparatus and a program for generating a response to a user's utterance.

  As a technique in which the system responds to the user's utterance, for example, as described in Patent Document 1 below, the speech recognition result is corrected based on the speech recognition context recognized so far, or the syntax information of the speech recognition result is used. A device for correcting a speech recognition result is known.

Further, the following paper of Non-Patent Document 1 describes a technique for searching for a response using a plurality of recognition candidates and generating a response. For example, if the speech recognition result candidates are “Where is the toilet”, “Is there a toilet” and “Is there a cafeteria”, all the morphemes contained in these recognition candidates are used (in this case, “Toilet / Has / Where / Is / Kana / Are / Masu / Dining room” 8 morphemes) Search for responses.
JP 2003-223185 A "Operation of a voice information guidance system as an actual environment research platform", IEICE Transactions

  However, in the technique described in Patent Document 1, the accuracy of correction of misrecognition is not sufficient even if context or syntax information recognized so far is used, and it is not always possible to correct misrecognition appropriately. There is a possibility that a false response is generated by generating a response with erroneous recognition. There is also a problem that processing time is required to correct erroneous recognition.

  Further, in the technique described in Non-Patent Document 1, since a plurality of recognition results are integrated to search for a response, an appropriate response cannot always be generated. Specifically, in the above example, an appropriate response is searched using the 8 morphemes “toilet / ha / where / is / ka / a / mas / canteen”, but “toilet”, “canteen”, etc. If there is a recognition result that is considered to be spoken only in one of them, and you search for a response as a recognition result by collecting them together, you will be asked "Where is the toilet?" May generate a response such as

  The present invention has been made to solve the above-described problems, and it is an object of the present invention to provide a response generation apparatus and program that can generate an appropriate response to a speech recognition result and reduce erroneous responses. And

  In order to achieve the above object, the response generation apparatus of the present invention recognizes a user's utterance, generates a recognition candidate, and generates a response candidate corresponding to each recognition candidate generated by the recognition unit. Generating means for selecting, a selecting means for selecting a response candidate having the largest number of corresponding recognition candidates from the response candidates generated by the generating means, and a response output means for outputting the response candidate selected by the selecting means. It is configured to include.

  According to such a configuration, it is possible to generate an appropriate response to the speech recognition result and reduce erroneous responses. For example, even if misrecognition occurs in the speech recognition result of the user's utterance, the present invention outputs response candidates corresponding to many recognition candidates. A response with little contradiction can be generated for each speech recognition candidate without correction.

  The response generation apparatus according to another aspect of the present invention recognizes a user's utterance and generates a recognition candidate, and generates a response candidate corresponding to each recognition candidate generated by the recognition unit. And a numerical value indicating the appropriateness of the response to the user utterance based on the recognition reliability of the recognition candidate for each response candidate generated by the generating means and the appropriateness indicated by the calculated numerical value. Selection means for selecting the highest response candidate from the generated response candidates, and response output means for outputting the response candidate selected by the selection means are configured.

  According to such a configuration, it is possible to generate an appropriate response to the speech recognition result and reduce erroneous responses. For example, even when an erroneous recognition occurs in the speech recognition result of the user utterance, the present invention fails because the response candidate having the highest appropriateness indicated by the numerical value calculated from the reliability of the recognition candidate is selected and output. The dialogue can proceed without any error, and a response with less contradiction can be generated for each speech recognition candidate without correcting the speech recognition error.

  The generating means includes a first generating means for generating a response candidate corresponding to each recognition candidate generated by the recognizing means, and a term included in the response candidate generated by the first generating means. And a second generation unit that regenerates a response candidate by replacing it with a term corresponding to.

  Thus, by replacing with terms corresponding to the superordinate concept, as a result, a plurality of response candidates are aggregated, and response candidates that can correspond to many recognition candidates are generated. Therefore, it is possible to reduce erroneous responses by selecting response candidates to be output from the response candidates.

  The second generation unit may regenerate the response candidate by replacing a term included in the response candidate generated by the first generation unit using a thesaurus with a term corresponding to the superordinate concept.

  According to such a configuration, a term corresponding to a superordinate concept can be easily found and replaced.

  In addition, the generation unit includes a first generation unit that generates response candidates corresponding to each recognition candidate generated by the recognition unit, and a plurality of response candidates generated by the first generation unit that can be aggregated. You may comprise including the 2nd production | generation means which produces | generates these response candidates to the response content of the same content by which the content of these response candidates was aggregated when a response candidate exists.

  According to such a configuration, a plurality of response candidates can be aggregated, and response candidates that can handle many recognition candidates can be generated. Therefore, it is possible to reduce erroneous responses by selecting response candidates to be output from the response candidates.

  The plurality of response candidates that can be aggregated are a plurality of response candidates that are partially the same, and the second generation means includes a plurality of response candidates that are partially the same as the response candidates generated by the first generation means. When there are two response candidates, the plurality of response candidates may be regenerated as response candidates having the same content from which parts other than the same part are deleted.

  The plurality of response candidates that can be aggregated are a plurality of response candidates that are partially the same, and the second generation unit includes a plurality of response candidates that are partially the same as the response candidates generated by the first generation unit. When there are two response candidates, the plurality of response candidates may be regenerated as response candidates having the same contents representing the concept inferred from the same part.

  The plurality of response candidates that can be aggregated are a plurality of response candidates that have the same concept inferred from at least one of the terms included in the response candidates or the entire response candidates, and the second generation unit includes: When there are a plurality of response candidates having the same concept inferred from at least one of the terms included in the response candidates or the entire response candidates in the response candidates generated by the first generation unit, the plurality of response candidates are You may make it produce | generate again the response candidate of the same content by which this inferred concept was expressed. Furthermore, the concept inferred from at least one of the terms included in the response candidate or the entire response candidate may be an emotion or a superordinate concept.

  Note that the present invention can also be applied to a program for realizing the functions of the response generation apparatus by a computer.

  As described above, according to the present invention, it is possible to generate an appropriate response with respect to the speech recognition result, and to obtain an excellent effect that an erroneous response can be reduced.

  Hereinafter, embodiments of the present invention will be described in detail with reference to the drawings.

  [First Embodiment]

  FIG. 1 is a block diagram illustrating a schematic configuration of a response generation apparatus 10 according to the first embodiment. The response generation device 10 includes an input unit 12, a response generation unit 14, a response selection unit 16, and an output unit 18.

  The input unit 12 includes a speech recognition dictionary, recognizes a speech waveform indicating the input user's utterance using the speech recognition dictionary, and outputs a plurality of possible speech recognition candidates. The input unit 12 is a general method that uses acoustic information or linguistic information for speech recognition (for example, Lee Shin-nobu, Tatsuya Kawahara, Kiyohiro Shikano, “Reliability based on fast word posterior probabilities in the two-pass search algorithm” Calculate the reliability of speech recognition according to "Calculation method" Information Processing Society of Japan Research Report, 2003-SLP-49-48, 2003-12 etc.).

  The response generation unit 14 generates response candidates for the user's utterance individually for all candidates that have been voice-recognized. In the present embodiment, the question information in which the speech recognition candidate and the question type are associated with each other is stored in a predetermined storage unit, and a response template is also stored, and the question information and the template are stored. Response candidates are generated individually by reading from the storage means. The question information and template will be described later.

  The response selection unit 16 selects response candidates that can correspond to more recognition candidates from the response candidates generated by the response generation unit 14, and further selects response candidates in consideration of the reliability of speech recognition.

  The output unit 18 includes a speaker, a display, and the like. The output unit 18 displays a character image representing the response candidate selected by the response selection unit 16 on the display, or voice-synthesizes the selected response candidate, and then outputs the voice through the speaker. Output.

  Each component (except for the speaker and display included in the output unit 18) constituting the response generation apparatus 10 described above is realized by a computer including a CPU, a RAM, and a ROM. That is, the CPU implements the functions of the above-described components by executing a program stored in a ROM or a predetermined storage device, and the processing described below is performed. In addition, each component may be configured by a separate computer or a single computer.

  The response generation device 10 configured as described above can realize, for example, the following dialogue example 1 with the user. Hereinafter, the utterance of the apparatus in response to the user utterance is referred to as “system utterance”.

(Dialogue example 1)
User utterance 1: Listen, my mom opened it automatically System utterance 2: What?
User utterance 3: A box in a drawer. System utterance 4: Yes, I got it from a friend

  Hereinafter, the flow of response generation in the response generation apparatus 10 will be described.

  First, the input unit 12 of the response generation apparatus 10 recognizes a user utterance such as the user utterance 1 of the above-described dialogue example 1 and inputs a plurality of speech recognition candidates to the response generation unit 14. For example, as illustrated in FIG. 2, the input unit 12 inputs “heading” that is a recognition result of an actual user utterance and “original form” of the heading to the response generation unit 14 as a speech recognition result. Further, as described above, the reliability of each voice recognition candidate is also input to the response generation unit 14.

  FIG. 3 is a flowchart showing a flow of a response candidate generation processing routine performed by the response generation unit 14.

  In step 100, the response generation unit 14 generates a response candidate for one of the speech recognition candidates recognized by the input unit 12. Here, as described above, response candidates are generated based on the question information and the template stored in advance.

  FIG. 4 shows a specific example of the question information. In FIG. 4, three types of question types “who”, “what”, and “why” are associated with a speech recognition candidate (original form) “open”.

  FIG. 5 shows a specific example of a response template. FIG. 5A shows a response template for a predicate recognition candidate, which is composed of question type + speech recognition candidate (predicate) (*) + temporal final particle (*) + question final particle (*). . The part marked with * is an optional part, and in this embodiment, the speech recognition candidate (predicate) and the tense final particle are always omitted from the * mark. FIG. 5B shows a response template for a noun, which is composed of question type + speech recognition candidate (noun) (*) + question final particle (*). In this embodiment, voice recognition candidates (nouns) are always omitted from the * mark.

  There are various types of tense-term particles and question-term particles, but the tense-term particles and question-term particles are stored in the storage means in advance, like the template. For example, it is registered in advance in a table as shown in FIG. And the response production | generation part 14 reads a required information from a memory | storage means suitably, and produces | generates a response candidate.

  For example, in response to a voice recognition candidate “open”, normally, as shown in FIG. 7, a response “Who opened it?”, “What opened it?”, “Why opened it?” Candidates are generated, but in this embodiment, the “opened” part included in these response candidates is omitted.

  In step 102, the response generation unit 14 determines whether or not the response candidate generation processing for all speech recognition candidates has been completed. If it is determined that the response candidate generation processing has not ended, the process returns to step 100 and is unprocessed. For the voice recognition candidates, response candidates are generated in the same manner as described above. If it is determined in step 102 that the generation of response candidates for all speech recognition candidates has been completed, this routine ends.

  Thereby, as shown in FIG. 8, a response candidate is produced | generated with respect to each speech recognition candidate. For example, for the speech recognition candidate “I opened”, “Who?”, “What?”, “Why?” Are generated from the original predicate “Open”. For the voice recognition candidate “Kicked”, “who” and “what?” Are generated from the predicate “kick”.

  The response selection unit 16 selects an appropriate response candidate from the response candidates generated by the response generation unit 14.

  FIG. 9 is a flowchart showing a flow of a response candidate selection processing routine performed by the response selection unit 16.

  In step 120, the response selection unit 16 counts the number of generations for all response candidates. Here, the number of generations refers to the number of speech recognition candidates to which each response candidate corresponds. In the present embodiment, as shown in FIG. 8, “Who?” “What?” “What?” “What?” “Why?” “When?” Kinds of response candidates have been generated. The number of recognition candidates corresponding to each response candidate is as follows.

“Who?” = 3 speech recognition candidates “something?” = 2 speech recognition candidates “What?” = 4 speech recognition candidates “What?” = 1 speech recognition candidate “when?” = Speech recognition 2 candidates

  In step 122, the response selection unit 16 selects a response candidate having the largest number of generations (here, “What?”).

  In step 124, the response selection unit 16 determines whether there are a plurality of response candidates with the largest number of generations. If it is determined that there are a plurality of response candidates, at step 126 one of the plurality of response candidates is selected at random and output as the final selection result. If it is determined that there are not a plurality, the response candidates selected in step 122 are output as they are as the final selection results (see also system utterance 2 in Dialogue Example 1 above).

  Here, the number of recognition candidates is counted, and the response candidates common to many recognition candidates are selected from the results. However, the response candidates may be selected using the reliability of speech recognition. Good. FIG. 10 is a flowchart showing the flow of a response candidate selection processing routine when selecting a response candidate in consideration of the reliability of voice recognition.

  In step 140, the response selection unit 16 calculates appropriateness of the response to the user utterance for all response candidates. The appropriateness of the response is calculated by the following equation 1 using the reliability of the recognition candidate.

[Formula 1]
Appropriateness of response of response candidate X = Σ {Reliability of recognition of recognition candidate corresponding to response candidate X}

  Using the speech recognition candidates and the reliability shown in FIG. 2, the appropriateness of each response candidate is calculated by Equation 1 as follows.

“Who?” = “Reliability of opening 0.1” + “Reliability of kicking 0.2”
+ "Reliability of multiplication 0.15" = 0.35
“What?” = “Confidence to avoid 0.1” + “confidence to cook 0.9” = 1.0
“What?” = “Reliability of opening 0.1” + “Reliability of kicking 0.2”
+ "Reliability of multiplication 0.15" + "Reliability of avoidance 0.1" = 0.45
“What?” = “Reliability of multiplying 0.15” = 0.15
“When?” = “Reliability of hanging 0.15” + “Reliability of cooking 0.9” = 1.05

  In step 142, the response selection unit 16 selects a response candidate (here, “when?”) Having the highest appropriateness score.

  In step 144, the response selection unit 16 determines whether or not there are a plurality of response candidates having the highest appropriateness score. If it is determined that there are a plurality of response candidates, at step 146, one of the plurality of response candidates is selected at random and output as a final selection result. If it is determined that there are not a plurality, the response candidates selected in step 142 are output as they are as final selection results.

  The output unit 18 displays a character image representing the response candidate selected in this way on the display, or synthesizes the selected response candidate by voice, and then outputs the voice through a speaker.

  In addition, when the number of generations or appropriate scores for all the response candidates generated by the response generation unit 14 are the same, it is not selected from the response candidates, but is simply matched (for example, “That's right”). May be processed.

  As described above, a response candidate for each speech recognition candidate is generated from a plurality of speech recognition candidates, and a response candidate common to many recognition candidates is selected from the generated response candidates, or calculated from the reliability of the recognition candidates. Since the response candidate having the highest appropriateness indicated by the displayed numerical value is selected and output, the dialogue can proceed without failure even if the speech recognition candidate includes a false recognition.

  Conventionally, when a speech recognition error occurs, a response is generated after correcting the misrecognition. However, since the accuracy of correcting the misrecognition is not sufficient, it is not always possible to correct the misrecognition appropriately. In addition, there is a problem that it takes a processing time to correct misrecognition. However, by generating a response as described above, an appropriate response can be output without correcting the speech recognition error even if the speech recognition candidate includes a misrecognition.

  Conventionally, since responses are generated using a plurality of speech recognition candidates, inconsistent responses may be selected for the speech recognition candidates. However, as described above, a common response candidate for many speech recognition results or a response candidate with a high appropriateness score is selected and output, so that a response with less contradiction is generated for each speech recognition candidate. can do.

  Note that the response generation method of the response generation unit 14 is an example, and is not limited to the method described in the present embodiment. For example, you may make it perform using the technique as described in Unexamined-Japanese-Patent No. 2007-206888, the technique as described in Unexamined-Japanese-Patent No. 2006-201870, etc., for example.

  Specifically, in the technique described in Japanese Patent Application Laid-Open No. 2007-206888, an analyzed predicate and a case element corresponding thereto are extracted, and a response for confirming the predicate and the case element is generated. In the technique described in Japanese Patent Application Laid-Open No. 2006-201870, response utterance information is stored in accordance with a rule including a combination of an event and an event, an event and an evaluation, and an evaluation and an evaluation, and the voice recognition result and the stored rule are stored. Based on the above, a response utterance is generated. Here, the combination of an event and an event, and the combination of an event and evaluation are comprised by causal relationship and temporal relationship, for example. Moreover, evaluation is comprised by the emotion etc. which a user utterance means, for example.

  Thus, for various types of response candidates, for example, response candidates for confirming words recognized from user utterances, response candidates for evaluating events included in user utterances with expressions including emotion words, and user utterances. Response candidates representing events resulting from the included events are generated.

  [Second Embodiment]

  FIG. 11 is a block diagram illustrating a schematic configuration of the response generation apparatus 20 according to the second embodiment. The response generation device 20 includes an input unit 12, a response generation unit 14, a response candidate aggregation unit 15, a response selection unit 16, and an output unit 18. Here, in FIG. 11, the same or equivalent parts as in FIG.

  The response candidate aggregating unit 15 includes a thesaurus, and for each of the plurality of response candidates generated by the response generation unit 14, the term included in the response candidate is replaced with another term corresponding to the superordinate concept using the thesaurus. cure. As shown in FIG. 12, the thesaurus refers to a dictionary in which terms are hierarchically classified by synonyms, semantic similarity relationships, inclusion relationships, and the like. In addition, since the response candidates are generated again by replacing the terms, as a result, a plurality of response candidates are aggregated into response candidates having the same contents, so in the following, response candidates are generated by replacing the terms after generating the response candidates. Re-doing is expressed as “aggregation”.

  The response selection unit 16 selects a response that can handle more recognition candidates from the response candidates aggregated by the response candidate aggregation unit 15, and further selects a response in consideration of the reliability of voice recognition.

  The response generation device 20 configured as described above can realize, for example, the following dialogue example 2 with the user.

(Dialogue example 2)
User utterance 1: I took a break at an elementary school during an excursion System utterance 2: What kind of school is it?
User utterance 3: It was a big school, the toilet was clean System utterance 4: That's right

  Hereinafter, the flow of response generation in the response generation device 20 will be described.

  First, the input unit 12 of the response generation device 20 recognizes a user utterance such as the user utterance 1 of the above-described dialogue example 2 and inputs a plurality of speech recognition candidates to the response generation unit 14. In the following, as shown in FIG. 13, the case where three speech recognition results of “elementary school”, “junior high school”, and “parking lot” and their reliability are input to the response generation unit 14 is taken as an example, and the subsequent processing Will be described more specifically.

  As described in the first embodiment, the response generation unit 14 generates response candidates for each of the speech recognition candidates in FIG. As a result, as shown in FIGS. 14A and 14B, the recognition candidate “elementary school” is “what kind of elementary school is?”, And the recognition candidate “junior high school” is “what kind of junior high school is?”, “Parking lot” A response candidate “Where is the parking lot?” Is generated from the recognition candidates.

  Using a thesaurus as shown in FIG. 12, the response candidate aggregating unit 15 replaces words included in the response candidates generated by the response generating unit 14 with words corresponding to the superordinate concept of the thesaurus. In FIG. 12, the word corresponding to the superordinate concept of the speech recognition candidates “elementary school” and “junior high school” shown in FIG. 13 is “school”, and the word corresponding to the superordinate concept of the speech recognition candidate “parking lot” is “storage location”. Is. Therefore, as shown in FIG. 14 (C), “what kind of elementary school?” Is rephrased as “what kind of school?”, “What kind of junior high school student?” Is rephrased as “what kind of school?” "Where is the car?" Can be rephrased as "Where is the storage location?" Note that the degree of superordinate conception is not particularly limited, but may be superordinated according to, for example, the hierarchical level of speech recognition candidates, the type of terms that have been speech-recognized, or the like.

  The response selection unit 16 selects an appropriate response candidate from the response candidates after aggregation in the response candidate aggregation unit 15. The method for selecting a response candidate is the same as that in the first embodiment. For example, in the case where the processing routine of FIG. 9 is executed and a response candidate is selected from the response candidates shown in FIG. 14C, “what school is?” Is selected and the response is confirmed (in the dialogue example 2 above). (See also system utterance 2.)

  In the example described above, the response candidate aggregating unit 15 uses the thesaurus for each of the response candidates generated by the response generation unit 14 to replace the response candidates with the words corresponding to the higher concept. Although an example of generating and selecting an appropriate response candidate from the response candidates after aggregation has been described, the present invention is not limited to this. For example, only a common part among a plurality of response candidates is left from terms included in the response candidates. And an appropriate response candidate may be selected from the response candidates after aggregation.

  For example, as shown in FIGS. 15A and 15B, the response generation unit 14 determines “what kind of elementary school is” from the recognition candidate “elementary school” and “what kind of junior high school?” From the recognition candidate “junior high school”. From the recognition candidate “parking lot”, a response candidate “where is the parking lot?” Is generated.

  The response candidate aggregating unit 15 compares all the response candidates generated by the response generating unit 14, leaves only a part common to a plurality of response candidates among terms included in each response candidate, and deletes other than the common part Thus, the response candidates are collected as shown in FIG. Here, since the word “school” is common in the two response candidates, the parts “small” and “medium” are deleted, and a response candidate “what school is it?” Is generated. In addition, the response candidate “where is the parking lot?” Is not rephrased because there is no common part with other response candidates. Here, the part generated in common by the template, such as a final particle, is not regarded as a common part.

  The response selection unit 16 selects an appropriate response from the aggregated response candidates. For example, when the processing routine of FIG. 9 is executed and response candidates are selected from the response candidates shown in FIG. 15C, “what school is it?” Is selected.

  In addition, the response candidates may be generated again as response candidates representing the concept inferred from a common word included in each response candidate.

  For example, when “100 m” and “200 m” shown in FIG. 16A are speech recognition candidates, the response generation unit 14 makes “100 m?” And “200 m?” As shown in FIG. ”Is generated. Here, the part common to each response candidate is “m (meter)”.

  When the inference rules as shown in FIG. 16D are stored in a database in advance, the response candidate aggregating unit 15 refers to the database and “m (meter)” common to the two speech recognition results. From this, the concept of “length” is inferred. Accordingly, as shown in FIG. 16C, the response candidate aggregation unit 15 can generate a response candidate “so long?” Representing the inferred concept.

  In addition, each response candidate may be regenerated as a response candidate representing emotion.

  For example, as shown in FIGS. 17 (A) and 17 (B), the response generation unit 14 selects “going to elementary school” from the recognition candidate “elementary school” and “goes to junior high school” from the recognition candidate “junior high school”. ? "Or" going to school "from the recognition candidates" going to school "is generated.

  The response candidate aggregating unit 15 regenerates each response candidate generated by the response generating unit 14 into a response candidate expressing the emotion represented by each response candidate. For example, according to a predetermined rule, it is determined whether it corresponds to a positive (desirable) emotion or a negative (undesirable) emotion from a noun or a predicate included in each response candidate. Then, the response candidates determined to correspond to the positive emotion are paraphrased as common response candidates expressing the positive emotion. Response candidates determined to be negative emotions are converted into common response candidates expressing negative emotions.

  Note that positive emotion words (fun, beautiful, laughter, etc.) and negative emotion words (unpleasant, cry, hard, etc.) are registered in advance in a predetermined storage means. A response candidate expressing an emotion is generated by selecting an emotion word expressing a positive emotion from the storage means. FIG. 17C shows a conversion example in which each response candidate is paraphrased as a response candidate expressing emotion.

  The response selection unit 16 selects an appropriate response from the aggregated response candidates. For example, when the processing routine of FIG. 9 is executed and a response candidate is selected from the response candidates shown in FIG.

  As described above, since the generated response candidates are aggregated and an appropriate response candidate is selected from the aggregation result, the same effects as those of the first embodiment can be obtained, and each recognition candidate can be further improved. Responses with less contradiction can be output, and erroneous responses are reduced.

It is a block diagram showing a schematic structure of a response generating device concerning a 1st embodiment. It is a figure which shows an example of the voice recognition result of an input part. It is a flowchart which shows the flow of the response candidate production | generation processing routine which a response production | generation part performs. It is a figure which shows the specific example of question information. It is a figure which shows the specific example of the template for a response. It is a figure which shows an example of the table which registered the tense final particle and the question final particle. It is a figure which shows the specific example of the response candidate produced | generated in a response production | generation part. It is a table | surface which shows the production | generation result when the response production | generation part of 1st Embodiment produces | generates a response candidate with respect to each speech recognition candidate shown in FIG. It is a flowchart which shows the flow of the response candidate selection process routine which a response selection part performs. It is a flowchart which shows the flow of the response candidate selection process routine in the case of selecting a response candidate in consideration of the reliability of voice recognition. It is a block diagram which shows the schematic structure of the response generation apparatus which concerns on 2nd Embodiment. It is a figure which shows the specific example of a thesaurus. It is a figure which shows an example of the voice recognition result of an input part. It is explanatory drawing explaining the specific example in the case of collecting a response candidate using a thesaurus. It is explanatory drawing explaining the specific example in the case of deleting a response candidate except the part common to a some response candidate from the term contained in a response candidate. It is explanatory drawing explaining the specific example in the case of paraphrasing a response candidate into the response candidate expressing the concept inferred from the word common to several response candidates. It is explanatory drawing explaining the specific example in the case of paraphrasing a response candidate to the response candidate expressing the emotion which a response candidate represents.

Explanation of symbols

DESCRIPTION OF SYMBOLS 10 Response production | generation apparatus 12 Input part 14 Response generation part 15 Response candidate aggregation part 16 Response selection part 18 Output part 20 Response generation apparatus

Claims (11)

A recognition means for recognizing a user's utterance and generating a recognition candidate;
Generating means for generating response candidates corresponding to each recognition candidate generated by the recognition means;
A selection means for selecting a response candidate having the largest number of corresponding recognition candidates from the response candidates generated by the generation means;
Response output means for outputting the response candidates selected by the selection means;
A response generation device including: A recognition means for recognizing a user's utterance and generating a recognition candidate;
Generating means for generating response candidates corresponding to each recognition candidate generated by the recognition means;
For each response candidate generated by the generating means, a numerical value indicating the appropriateness of the response to the user utterance is calculated based on the recognition reliability of the recognition candidate, and the response indicated by the calculated numerical value is the most appropriate Selecting means for selecting candidates from the generated response candidates;
Response output means for outputting the response candidates selected by the selection means;
A response generation device including: The generating means includes
First generation means for generating a response candidate corresponding to each recognition candidate generated by the recognition means;
A second generation means for regenerating a response candidate by replacing a term included in the response candidate generated by the first generation means with a term corresponding to the superordinate concept;
The response generation device according to claim 1, comprising: The response generation apparatus according to claim 3, wherein the second generation unit regenerates a response candidate by replacing a term included in the response candidate generated by the first generation unit using a thesaurus with a term corresponding to the superordinate concept. . The generating means includes
First generation means for generating a response candidate corresponding to each recognition candidate generated by the recognition means;
When there are a plurality of response candidates that can be aggregated in the response candidates generated by the first generation means, the plurality of response candidates are converted into response candidates having the same content in which the contents of the plurality of response candidates are aggregated. Second generation means for regenerating,
The response generation device according to claim 1, comprising: The plurality of response candidates that can be aggregated are a plurality of response candidates that are partly the same,
The second generation means, when there are a plurality of response candidates that are the same in part among the response candidates generated by the first generation means, the plurality of response candidates are the same by deleting a part other than the same part Re-generate the content response candidates,
The response generation device according to claim 5. The plurality of response candidates that can be aggregated are a plurality of response candidates that are partly the same,
The second generation means has a concept in which when there are a plurality of response candidates that are the same in part among the response candidates generated by the first generation means, the plurality of response candidates are inferred from the same part. To generate a response candidate with the same content
The response generation device according to claim 5. The plurality of response candidates that can be aggregated are a plurality of response candidates that have the same concept inferred from at least one of terms included in the response candidates or the entire response candidates,
When the second generation means includes a plurality of response candidates having the same concept inferred from at least one of terms included in the response candidates or the entire response candidates in the response candidates generated by the first generation means, Regenerating the plurality of response candidates into response candidates having the same content in which the inferred concept is expressed;
The response generation device according to claim 5. The concept inferred from at least one of the terms included in the response candidate or the entire response candidate is an emotion or a superordinate concept.
The response generation device according to claim 8. Computer
A recognition means for recognizing a user's utterance and generating a recognition candidate;
Generating means for generating response candidates corresponding to each recognition candidate generated by the recognition means;
A selection unit that selects a response candidate having the largest number of corresponding recognition candidates from the response candidates generated by the generation unit; and a response output unit that outputs the response candidate selected by the selection unit;
Program to function as. Computer
A recognition means for recognizing a user's utterance and generating a recognition candidate;
Generating means for generating response candidates corresponding to each recognition candidate generated by the recognition means;
For each response candidate generated by the generating means, a numerical value indicating the appropriateness of the response to the user utterance is calculated based on the recognition reliability of the recognition candidate, and the response indicated by the calculated numerical value is the most appropriate Selecting means for selecting candidates from the generated response candidates; and response output means for outputting response candidates selected by the selecting means;
Program to function as.

Images