JP2016090891A - Response generation apparatus, response generation method, and response generation program - Google Patents

Abstract

PROBLEM TO BE SOLVED: To alleviate discomfort in a dialog which is caused by a uniform response pattern.SOLUTION: A method includes: recognizing a user voice; analyzing a structure of the recognized voice; storing information produced by correlating plural key words respectively with associative word and additional word that are related to each key word; extracting noun and verb from the recognized voice; selecting a key word consistent with the noun or verb; adding the additional word corresponding to the key word to the extracted noun or verb; generating a repeating response sentence by replacing with a related word corresponding to the selected key word; generating a voluntary response sentence on the basis of the analyzed voice structure; outputting the voluntary response sentence after outputting the repeating response sentence; extracting a second noun or verb related to a first noun or verb from given sentence group information; and questioning the user as to whether or not the extracted both words are the same, and if the answer is positive, storing the first noun or verb as a key word, and the second noun or verb as an associative word.SELECTED DRAWING: Figure 1

Description

  The present invention relates to a response generation method, a response generation apparatus, and a response generation program for responding to a user.

  A voice recognition means for recognizing the user's voice, a structure analysis means for analyzing the structure of the voice recognized by the voice recognition means, and a response sentence to the user's voice based on the structure of the voice analyzed by the structure analysis means. There is known a response generation device including a response output unit that generates and outputs the generated response sentence (see, for example, Patent Document 1).

JP 2010-157081 A

The response generation apparatus as described above takes time for the structure analysis of voice and the generation of the response sentence, and waiting for a response occurs. For this reason, there is a possibility that a sense of incongruity may occur in the dialogue. Thus, for example, it is conceivable to simply respond by using the user's voice recognized by the voice recognition means while waiting for the response as a repeated response sentence. In this case, the waiting time for the response is shortened, and the uncomfortable feeling of the dialogue is alleviated, but the response pattern becomes uniform and unnaturalness as the dialogue remains.
The present invention has been made to solve such problems, and provides a response generation method, a response generation apparatus, and a response generation program that can alleviate the uncomfortable feeling of dialogue due to a uniform response pattern. The main purpose is to do.

To achieve the above object, one aspect of the present invention includes a step of recognizing a user's voice, a step of analyzing the structure of the recognized voice, a plurality of keywords and associated words and additional words related to the keywords. Each of the stored information, the step of extracting a noun or verb from the recognized speech, the keyword of the stored information that matches the extracted noun or verb, and the selected keyword A repeated response sentence for repeating the user's voice by adding a corresponding additional word to the extracted noun or verb and replacing the noun or verb with a related word corresponding to the selected keyword. And generating an optional response sentence to the user's voice based on the analyzed voice structure, and generating the repeated response. Outputting a desired response sentence after outputting a sentence, comprising: a first noun or verb; and a second noun or verb related to the first noun or verb Extracting from the predetermined sentence set information, and asking the user whether or not the extracted first noun or verb and the second noun or verb are the same, and Storing the first noun or verb as the keyword and storing the second noun or verb as the associative word when a user's answer to the question is affirmative, This is a response generation method.
In this one aspect, the first noun or verb and the second noun or verb are extracted from the predetermined sentence set information existing on the network, and the extracted first noun or verb and the extracted Using the co-occurrence frequency with the second noun or verb, calculating the similarity between the first noun or verb and the second noun or verb, and when the calculated similarity is equal to or greater than a threshold, When the user is asked whether the extracted first noun or verb and the second noun or verb are the same, and the user's answer to the question is affirmative, One noun or verb may be used as the keyword, and the second noun or verb may be stored as the associative word.
In this aspect, the method may further include a step of limiting the number of associative words and additional words associated with the stored keyword under a predetermined condition.
In this aspect, the method further comprises: analyzing a phoneme of the user's voice; and generating a reciprocal response to the user's voice based on the analyzed phoneme analysis result. Before outputting the repeated response sentence, the generated response of the response may be output.
In order to achieve the above object, one aspect of the present invention includes a speech recognition unit that recognizes a user's speech, a structure analysis unit that analyzes a structure of speech recognized by the speech recognition unit, and a plurality of keywords. Extracted by the storage part for storing the information in which the associative word and the additional word related to the keyword are associated with each other, the part of speech extraction part for extracting the noun or the verb from the speech recognized by the voice recognition part, and the part of speech extraction part. Selecting a keyword in the storage means that matches the noun or verb, adding an additional word corresponding to the selected keyword to the extracted noun or verb, and adding the noun or verb to the selected keyword To the related word corresponding to the above, the repetition generating means for generating a repeated response sentence for repeating the user's voice, and the structural analysis means A response generation unit that generates an arbitrary response sentence to the user's voice based on the analyzed voice structure, outputs the repeated response sentence, and then outputs the optional response sentence An apparatus for extracting a first noun or verb and a second noun or verb related to the first noun or verb from predetermined sentence set information, and extracting by the extracting means When the first noun or verb and the second noun or verb are the same, the question means for asking the user whether or not the answer is positive. A response generation device including: a determination unit that registers the first noun or verb as the keyword and the second noun or verb as the association word in the storage unit. .
One aspect of the present invention for achieving the above object includes: a process for recognizing a user's voice; a process for analyzing the structure of the recognized voice; a process for extracting a noun or verb from the recognized voice; Information associated with each keyword is associated with each keyword and an additional word corresponding to the selected keyword is selected by selecting the keyword that matches the extracted noun or verb. Is added to the extracted noun or verb, and the noun or verb is replaced with a related word corresponding to the selected keyword, thereby generating a repeated response sentence for repeating the user's voice. And generating an optional response sentence to the user's voice based on the analyzed voice structure, outputting the repeated response sentence, and then outputting the optional response sentence. A process for outputting and a response generation program to be executed by a computer, wherein a first noun or verb and a second noun or verb related to the first noun or verb are determined from predetermined sentence set information A process of extracting, a process of asking the user whether or not the extracted first noun or verb and the second noun or verb are the same, and a user's answer to the question is positive A response generating program that causes the computer to execute a process of storing the first noun or verb as the keyword and storing the second noun or verb as the associative word. May be.

  ADVANTAGE OF THE INVENTION According to this invention, the response production | generation method, the response production | generation apparatus, and the response production | generation program which can relieve the discomfort of the dialog by a uniform response pattern can be provided.

It is a block diagram which shows the schematic system configuration | structure of the response generation apparatus which concerns on Embodiment 1 of this invention. It is a block diagram which shows the schematic hardware constitutions of the response generation apparatus which concerns on Embodiment 1 of this invention. It is a flowchart which shows the processing flow of the response generation method which concerns on Embodiment 1 of this invention. It is a block diagram which shows the schematic system configuration | structure of the response generation apparatus which concerns on Embodiment 2 of this invention. It is the schematic which shows an example of the structure by which related word information is memorize | stored and updated in a network device.

Embodiment 1
Embodiments of the present invention will be described below with reference to the drawings. FIG. 1 is a block diagram showing a schematic system configuration of a response generation apparatus according to Embodiment 1 of the present invention. The response generation apparatus 1 according to the first embodiment includes a voice recognition unit 2 that recognizes a user's voice, a structure analysis unit 3 that analyzes a voice structure, and a response output that generates and outputs a response sentence to the user's voice. A unit 4, a repetition generation unit 5 that generates a repeated response sentence, a part of speech extraction unit 9, a related word extraction unit 10, a user question unit 11, and a memory determination unit 12 are provided.

  The response generation device 1 includes, for example, a CPU (Central Processing Unit) 1a that performs arithmetic processing and the like, a ROM (Read Only Memory) and a RAM (Random Access Memory) that store arithmetic programs executed by the CPU 1a, control programs, and the like. 2) and a microcomputer including an interface unit (I / F) 1c for inputting / outputting signals to / from the outside, and the like (FIG. 2). The CPU 1a, the memory 1b, and the interface unit 1c are connected to each other via a data bus 1d.

  The voice recognition unit 2 performs voice recognition processing based on the user's voice information acquired by the microphone 6, converts the user's voice information into text, and recognizes it as character string information. The speech recognition unit 2 is a specific example of speech recognition means. The voice recognition unit 2 detects a speech section from the user's voice information output from the microphone 6, and performs voice matching by performing pattern matching on the detected voice information of the speech section with reference to, for example, a statistical language model. Recognize. Here, the statistical language model is a probability model for calculating the appearance probability of a language expression, such as the distribution of the appearance of a word or the distribution of a word that appears after a certain word, and learned the connection probability in units of morphemes. Is. The statistical language model is stored in advance in the memory 1b or the like. Note that the speech recognition unit 2 generates morpheme information with part-of-speech information by adding the part-of-speech type (noun, adjective, verb, adverb, etc.) to each morpheme of the user's speech information. The voice recognition unit 2 outputs the recognized voice information of the user to the structure analysis unit 3 and the part of speech extraction unit 9.

  The structure analysis unit 3 analyzes the structure of the voice information recognized by the voice recognition unit 2. The structure analysis unit 3 is a specific example of structure analysis means. For example, the structure analysis unit 3 performs morphological analysis on character string information indicating the voice information of the user that has been voice-recognized using a general morphological analyzer, and interprets the meaning of the character string information. The structure analysis unit 3 outputs the analysis result of the character string information to the response output unit 4.

  The response output unit 4 generates a response sentence (hereinafter referred to as an optional response sentence) to the user's voice information based on the structure of the voice information analyzed by the structure analysis unit 3, and outputs the generated optional response sentence To do. The response output unit 4 is a specific example of response output means. For example, the response output unit 4 generates an arbitrary response sentence for the user's voice information based on the analysis result of the character string information output from the structure analysis unit 3. Then, the response output unit 4 outputs the generated response sentence using the speaker 7.

  More specifically, the structure analysis unit 3 extracts the predicate term structure in the character string information “eating tonkatsu” and specifies the predicate “eat” and the case particle “wo”. Then, the response output unit 4 extracts the type of case particles that can be related to the predicate “eat” specified by the structure analysis unit 3 from the deficiency dictionary database 8 in which the correspondence between the predicate and the case particle is stored. To do. The deficiency dictionary database 8 is constructed in the memory 1b, for example.

  The response output unit 4 generates, for example, predicate term structures such as “what to eat”, “where to eat”, “when to eat”, and “who to eat” as an optional response sentence. Further, the response output unit 4 randomly selects from other predicate term structures except the surface case "" that does not match the user's voice in the generated predicate term structure, and selects the selected predicate term. The structure is an arbitrary response sentence. The response output unit 4 selects, for example, a predicate term structure “who did you eat?” And outputs it as an optional response sentence. In addition, the generation method of the arbitrary response sentence mentioned above is an example, It is not limited to this, Arbitrary generation methods can be used.

  The part-of-speech extraction unit 9 extracts nouns and / or verbs from the recognized user's speech information based on the morpheme information with part-of-speech information of the speech information output from the speech recognition unit 2. The part of speech extraction unit 9 is a specific example of the part of speech extraction means. The part-of-speech extraction unit 9 uses, for example, the morpheme information with part-of-speech information “tonkatsu (noun) (verb) ate (verb) yo (particle)” of the speech information output from the speech recognition unit 2 to “tonkatsu (noun)”. Or “eat (verb)”. The part-of-speech extraction unit 9 extracts, for example, Tonkatsu (general noun), Yaba Ton (proprietary noun), voting => voting (sa variable noun) (however, excluding some nouns such as numerals), etc. To do. The part-of-speech extraction unit 9 extracts, for example, voting (sa variation), swimming, and the like as the verb. The part of speech extraction unit 9 outputs the extracted noun or verb to the repetition generation unit 5 and the related word extraction unit 10.

  By the way, the structure analysis of voice information and the generation of a response sentence as described above require time (for example, about 3 seconds), and the processing cost is high. For this reason, there is a possibility that waiting for a response may occur, and the conversation may feel uncomfortable.

  On the other hand, in the response generation device 1 according to the first embodiment, the repeat generation unit 5 repeats a response sentence (hereinafter referred to as a repeat response sentence) from the user's voice recognized by the voice recognition unit 2. ) Is generated easily. And the response output part 4 outputs the voluntary response sentence based on the structure of an audio | voice after making the repeated response sentence produced | generated by the repetition production | generation part 5. FIG.

  Thereby, since the repeated response sentence only repeats the recognized user's voice by returning a parrot, it does not require generation time (for example, about 1 second), and the processing cost is low. Therefore, it is possible to output a repetitive response sentence with a low processing cost while waiting for a response until the optional response sentence with a high processing cost is output. Therefore, it is possible to alleviate the uncomfortable feeling of the dialogue due to the large duration of the dialogue caused by waiting for a response.

  As described above, the repeat generation unit 5 generates the speech information recognized by the speech recognition unit 2 as a repeat response sentence for performing a parrot return. Here, rather than returning the parrot as it is without changing the user's voice at all, it is better to add a specific additional word to the noun or verb of the user's voice information and return the parrot to improve the naturalness of the conversation. . For example, in response to the user's utterance “I went to the ocean”, the response generation device 1 responds “Umi ka” or “Oh, ka ka” rather than simply responding “I went to the ocean”. The naturalness of dialogue is further improved.

  Therefore, the repeat generation unit 5 according to Embodiment 1 generates a repeat response sentence by adding a specific additional word to the noun or verb extracted by the part of speech extraction unit 9. This makes it possible to give diversity to the repetitive response sentence of the parrot return, so that it is not a uniform response pattern, and the uncomfortable feeling of dialogue can be further alleviated.

  In the memory 1b, for example, additional information (table information or the like) in which a plurality of keywords and a plurality of additional words (beginning, ending, etc.) are associated with each other is stored. The repetition generation unit 5 selects a keyword of additional information that matches the noun or verb based on the noun or verb output from the part of speech extraction unit 9 and the additional information stored in the memory 1b. Then, the repeat generation unit 5 selects an additional word corresponding to the selected keyword. The repeat generation unit 5 generates a repeat response sentence by adding the selected additional word to the noun or verb output from the part of speech extraction unit 9.

  More specifically, the repetition generation unit 5 adds the noun “ramen” corresponding to the noun “ramen” based on the noun “ramen” output from the part-of-speech extraction unit 9 and the additional information stored in the memory 1b. Select the word "ka". The repeat generation unit 5 generates the repeated response sentence “ramen kaaa” by adding the selected additional word “ka” to the noun “ramen” output from the part of speech extraction unit 9.

  Furthermore, in order to improve the uniform response pattern and further improve the naturalness of the dialogue, in the response generation device 1 according to the first embodiment, the repetition generation unit 5 adds a specific addition to the noun or verb. In addition to adding words, the nouns or verbs are replaced with related words to generate a repeated response sentence. Thereby, it is possible to give more diversity to the sensation of the repeated response sentence of the parrot return, and it is possible to further alleviate the uncomfortable feeling of the dialogue.

  In the memory 1b, for example, as shown below, related word information (table information or the like) associated with a plurality of keywords and associated words related to the keywords is stored.

Keyword “tonkatsu” Related word “pork”
Keyword “Steak” Related word “Beef”,
Keyword "A type" Related word "Careful type",
Keyword "O type" Related word "Oraka type"
Keyword "beef" Related word "Meat"
Keyword “Yaba Ton” Related Word “Miso Katsu”
Keyword "vote" Related word "National obligation"
Keyword "swim" Related Words "Swimming"
...

  The related word is a word that the user associates based on the keyword, such as a word similar to the keyword or a word corresponding to a higher concept of the keyword. Further, the related word information is table information in which keywords and related words are associated one to one, but is not limited thereto. For example, the related word information may be table information in which a single keyword is associated with a plurality of related words, or may be tree-like ontology information. Further, the related word information and the additional information may be integrated (table information in which both are associated).

  Based on the noun or verb output from the part-of-speech extraction unit 9 and the related word information stored in the memory 1b, the repetition generation unit 5 selects a keyword of the related word information that matches the noun or verb. . The repetition generation unit 5 selects a related word corresponding to the selected keyword. When a plurality of related words are associated with a single keyword, the repeat generation unit 5 may select one of the plurality of related words randomly or in the order of registration, for example. The repetition generation unit 5 replaces the noun or verb extracted by the part-of-speech extraction unit 9 with the selected related word to generate a repetition response sentence.

  Here, the word used by the user is constantly changing due to the current trend. For this reason, maintenance such as adding a new keyword and a related word is also required for the related word information in the memory 1b described above. On the other hand, the extraction of new keywords and related words involves intellectual judgment, and it is difficult to make simple rules. Therefore, when a person performs maintenance such as adding such new keywords and related words, a heavy load is placed on the person.

  On the other hand, the response generation device 1 according to the first embodiment has a function of automatically updating the related word information by adding new keywords and related words to the related word information in the memory 1b. . Thereby, a user's load at the time of maintaining related word information in the memory 1b can be significantly reduced. Furthermore, the number of related words is increased as appropriate, and the repeated response sentences can provide diversity, which can further alleviate the uncomfortable feeling of dialogue.

  The related word extraction unit 10 extracts a first noun or verb and a second noun or verb related to the first noun or verb from predetermined sentence set information. The related word extraction unit 10 is a specific example of word extraction means. The related word extraction unit 10 uses, for example, a first noun or verb (hereinafter referred to as the first word) and a second noun or verb (from the sentence set information stored in a server, terminal device, database, etc. on the network. Hereinafter, the second word) is extracted. The sentence set information includes, for example, text information such as newspapers, magazines, novels, conversations, and hierarchical corpus information.

  Note that the related word extraction unit 10 may extract the first and second words from the document set information directly input to the response generation device 1 or stored in the memory 1b in advance. For example, the related word extraction unit 10 may extract the first and second words based on the user's voice information input via the microphone 6.

For example, based on the co-occurrence frequency of the first word and the second word extracted from the sentence set information, the related word extraction unit 10 calculates the similarity between the first word and the second word by the following formula.
sim ("first word", "second word")
= {Co-occurrence frequency of “first word” and “second word”} ² / {frequency of “first word” × frequency of “second word”}

  The related word extraction unit 10 determines that the extracted first word and the second word are related when the calculated similarity is equal to or greater than the threshold. In addition, although the related word extraction part 10 is calculating the similarity using the co-occurrence frequency of a 1st word and a 2nd word, it is not limited to this. The related word extraction unit 10 may calculate the similarity between the first word and the second word using an arbitrary calculation method.

  The related word extraction unit 10 outputs the first word and the second word to the user question unit 11 when it is determined that the similarity is equal to or greater than the threshold and the first word and the second word are related.

  The user question unit 11 asks the user whether or not the first word and the second word output from the related word extraction unit 10 are the same. The user question part 11 is a specific example of a question means. For example, the user question unit 11 asks the user questions by voice using the speaker 7 or by displaying characters using a display device. More specifically, the user question unit 11 asks the user that “Tonkatsu is pork?”.

  When the user's answer to the question from the user question unit 11 is affirmative, the memory determination unit 12 registers the first word as a keyword and the second word as an associative word in the related word information in the memory 1b. The memory determination unit 12 is a specific example of a determination unit.

  The memory determination unit 12 determines that if the user's answer to the question of the user question unit 11 is a predetermined positive pattern such as “Yes”, “Yes”, “Yes”, the answer is positive Judge. On the other hand, the memory determination unit 12 determines that the answer is negative when the user's answer to the question of the user question unit 11 is other than a predetermined positive pattern.

  More specifically, when the memory determination unit 12 determines that the user's answer to the question “Tonkatsu is pork?” In the user question unit 11 is a positive pattern “Yes”, the answer is Judgment is positive. The memory determination unit 12 registers the first word “tonkatsu” as a keyword and the second word “pork” as an associative word, and registers the associated word information in the memory 1b.

  Here, the memory determination unit 12 registers the first word “tonkatsu” as a keyword and the second word “pork” as an associative word in the related word information, but does not perform reverse registration of the first and second words. That is, the memory determination unit 12 does not register the related word information with the first word “pork” as a keyword and the second word “tonkatsu” as an associative word.

  Furthermore, when the memory determination unit 12 registers the new first and second words in the related word information, the same keyword as the new first word may already be registered in the related word information. In this case, the memory determination unit 12 may register the new second word and the already registered related word in the related word information in association with the same keyword. For example, the keyword “tonkatsu” and the related words “pork” and “meat” corresponding to this keyword are registered in the related word information. As described above, a plurality of related words may be associated with a single keyword and registered in the related word information. This leads to saving of the storage capacity of the memory 1b.

  As described above, the storage determination unit 12 may register or register a plurality of related words for a single keyword in the related word information by executing or canceling the overlapping registration according to the user's selection. Good.

  For example, when a plurality of related words are registered with respect to a single keyword, the storage determination unit 12 uses the speaker 7 or the like to make a question for selecting execution or cancellation of the overlapping registration. The memory determination unit 12 executes or cancels the duplication registration for the question according to the user's selection input by the microphone 6 or the like.

FIG. 3 is a flowchart showing a processing flow of the response generation method according to the first embodiment.
The related word extraction unit 10 extracts the first word and the second word related to the first word from the corpus information on the network (step S101).

  The related word extraction unit 10 calculates the similarity between the extracted first word and second word (step S102). The related word extraction unit 10 determines whether or not the calculated similarity is equal to or greater than a threshold value and the extracted first word and the second word are related (step S103).

  When the related word extraction unit 10 determines that the similarity is equal to or greater than the threshold and the first word and the second word are related (YES in step S103), the first word and the second word are output to the user question unit 11. To do.

  The user question part 11 asks whether or not the first word and the second word output from the related word extraction part 10 are the same using the speaker 7 or the like (step S104). The memory determination unit 12 determines whether or not the user's answer to the question from the user question unit 11 is affirmative (step S105). When it is determined that the user's answer is affirmative (YES in step S105), the memory determination unit 12 registers the first word as a keyword and the second word as an associative word in the related word information in the memory 1b (step S105). S106).

  The voice recognition unit 2 performs voice recognition of the user's voice information acquired by the microphone 6 (step S107), and outputs the recognized user's voice information to the structure analysis unit 3 and the part of speech extraction unit 9.

  The part-of-speech extraction unit 9 extracts a noun or a verb from the recognized user's voice information based on the morpheme information with part-of-speech information of the voice information output from the voice recognition unit 2 (step S108). The part-of-speech extraction unit 9 outputs the extracted noun or verb to the repetition generation unit 5.

  Based on the noun or verb output from the part-of-speech extraction unit 9 and the additional information in the memory 1b, the repetition generation unit 5 selects the keyword of the additional information that matches the noun or verb output from the part-of-speech extraction unit 9. select. Then, the repetition generation unit 5 adds an additional word corresponding to the selected keyword to the extracted noun or verb (step S109).

  Based on the noun or verb output from the part-of-speech extraction unit 9 and the related word information in the memory 1b, the repetition generation unit 5 generates related word information that matches the noun or verb output from the part-of-speech extraction unit 9. Select a keyword. Then, the repeat generation unit 5 replaces the noun or verb output from the part-of-speech extraction unit 9 with a related word corresponding to the selected keyword (step S110), and generates a repeat response sentence (step S111). ).

  The repeat generation unit 5 outputs the generated repeat response sentence to the response output unit 4. The response output unit 4 outputs the repeated response text output from the repeated generation unit 5 from the speaker 7 (step S112).

  In parallel with the above (Step 108) and (Step 112), the structure analysis unit 3 analyzes the structure of the speech information recognized by the speech recognition unit 2 (Step S113), and returns the analysis result of the character string information as a response. Output to the output unit 4.

  The response output unit 4 generates an optional response sentence based on the analysis result of the character string information output from the structure analysis unit 3 (step S114), and outputs the generated optional response sentence from the speaker 7 (step S115). Here, the response output unit 4 outputs an arbitrary response sentence with a high processing time after outputting a repeated response sentence with a low processing time.

  Note that the above-described processing (step S101) to (step S106) (related word information update processing) is executed before (step S107) to (step S115) (actual user interaction processing). However, the present invention is not limited to this, and arbitrary timing can be executed. For example, the processes of (Step S101) to (Step S106) may be executed after (Step S107) to (Step S115), in the middle, or before and after.

  As described above, in the response generation device 1 according to the first embodiment, the repetition generation unit 5 selects the keyword in the memory 1b that matches the noun or verb extracted by the part-of-speech extraction unit 9, and corresponds to the selected keyword. An additional word is added to the extracted noun or verb. Further, the repetition generation unit 5 replaces the extracted noun or verb with a related word corresponding to the selected keyword of the memory 1b, and generates a repeated response sentence. As a result, it is possible to give diversity to the repetitive response sentence of the parrot return, and to further alleviate the uncomfortable feeling of dialogue. Furthermore, the related word information required when replacing with the related word can be automatically updated. For this reason, while reducing the user's load, it is possible to provide a variety of repeated response sentences to further alleviate the uncomfortable feeling of dialogue.

Embodiment 2. FIG.
FIG. 4 is a block diagram illustrating a schematic system configuration of the response generation apparatus according to the second embodiment of the present invention. In addition to the configuration of the response generation apparatus 1 according to the first embodiment, the response generation apparatus 20 according to the second embodiment includes a phoneme analysis unit 21 that analyzes the phoneme of the user's voice information, and a conflict of the user's voice information. It is characterized in that it further includes an interaction generation unit 22 that generates a response.

  The phoneme analysis unit 21 analyzes the phoneme of the user's voice information based on the user's voice information acquired by the microphone 6. The phoneme analysis unit 21 is a specific example of phoneme analysis means. For example, the phonological analysis unit 21 estimates a break in a user's voice by detecting a change in volume level or frequency (such as a fundamental frequency) of voice information. The phoneme analysis unit 21 outputs the phoneme analysis result to the conflict generation unit 22.

  Based on the phonological analysis result output from the phonological analysis unit 21, the harmonious generation unit 22 generates a response to the user's voice (hereinafter referred to as a “compatibility response”). The interaction generating unit 22 is a specific example of the interaction generating unit. For example, when the volume level of the voice information is equal to or lower than the threshold, the conflict generation unit 22 searches the standard response database 23 in which the conflict pattern is stored. Then, the conflict generation unit 22 randomly selects a conflict response from the standard response database 23. The standard response database 23 stores a plurality of patterns used for consideration such as “Yes. The fixed response database 23 is constructed in the memory 1b and the like. The interaction generating unit 22 outputs the generated interaction response to the response output unit 4.

  The response output unit 4 causes the speaker 7 to output the conflict response generated by the conflict generation unit 22 before the repeated response sentence generated by the repeat generation unit 5. Note that the phonological analysis unit 21 performs phonological analysis using feature quantities with low processing costs. For this reason, the generation time of the conflict response is shorter than the generation time of the repeated response sentence, and the processing cost is lower.

  Therefore, it is possible to output a conflict response with a lower processing cost before outputting the repeated response sentence. Thereby, the connection between dialogs becomes smoother, and the uncomfortable feeling of dialog can be eased more. Further, more responses and response sentences having different processing costs are generated in parallel and output in the order of generation. As a result, it is possible to maintain a smoother continuity of dialogue and realize a more natural dialogue that does not impair the sense of tempo.

  Note that the conflict generation unit 22 routinely generates a conflict response, and the repeat generation unit 5 generates a repeat response sentence by only performing surface interpretation of the speech recognition result. Therefore, it is assumed that the response output unit 4 generates an optional response candidate similar to the interaction response generated by the interaction generation unit 22 and the repetition response generated by the repetition generation unit 5.

  On the other hand, the response output unit 4 excludes from the voluntary response candidates the voluntary response candidates that overlap with the conflict response generated by the conflict generation unit 22 and the repeat response generated by the repeat generation unit 5. And the response output part 4 selects an optimal candidate from the excluded voluntary response candidates, and makes it an arbitrary response sentence. This eliminates redundant useless words and realizes a more natural dialogue.

  For example, in response to the user's utterance “Today is hot”, the conflict generation unit 22 generates a conflict response “Yes”. Subsequently, the repeat generation unit 5 generates a repeat response sentence “hot”. On the other hand, the response output unit 4 generates a voluntary response candidate “I don't like it”, “How long will it be hot?”, “It ’s hot”, “I think so”, and so on. The response output unit 4 eliminates “hot” overlapping with the repeated response sentence generated by the repeated generation unit 5 from the generated voluntary response candidates. Then, the response output unit 4 selects, for example, “How long will it be hot?” From the excluded optional response candidates, and uses it as an optional response sentence.

  Note that, in the response generation device 20 according to the second embodiment, the same parts as those in the response generation device 1 according to the first embodiment are denoted by the same reference numerals, and detailed description thereof is omitted.

Hereinafter, an example of the interaction between the response generation device 20 and the user will be shown. In the following example, M is a response sentence and a response of the response generation device 20, and U is a user's utterance.
M (topic provided): What did you eat at lunch?
U: I ate Tonkatsu.
M (conformity response): Yeah. Yup.
M (repeat response): Tonkatsu. (Add "ka")
M (judgment processing): It is judged that the first word “tonkatsu” and the second word “pork” are related. M (optional response sentence): Tonkatsu is pork?
U: (Positive answer) That's right.
M (Aiso response): That's right.
M (voluntary response): Where did you eat it?
...
U: I ate Tonkatsu.
M (conformity response): I see.
M (repeat response): Pork, pork. (Add “Dane” and replace “Tonkatsu” with “Pork”)
M (voluntary response): I ate it.

  As shown in the example of the above dialogue, when the user utters, the response response of the response generation device 20, the repeated response text, and the voluntary response text continue to the utterance at a fast pace, and the connection between the dialogs is smoother. I understand that Further, it is understood that the related word information can be automatically updated as appropriate (the related word “pork” corresponding to the keyword “tonkatsu” is added), and no human load is applied. Furthermore, it is understood that the naturalness of the dialogue is further improved by adding the additional word to the verb or the noun and replacing it with the related word to give the repeated response sentence diversity.

Embodiment 3
In the first and second embodiments, when the user's answer to the question of the user question unit 11 is affirmative, the memory determination unit 12 uses the first word as a keyword and the second word as an associative word. The word information is registered sequentially. When the storage determination unit 12 registers keywords and associative words in the related word information in the memory 1b without limitation, it is also assumed that there is no room in the storage capacity of the memory 1b.

  Therefore, the storage determination unit 12 according to the third embodiment limits the number of keywords of related word information stored in the memory 1b under a predetermined condition. In the related word information, for example, the keyword, the related word of the keyword, the number of keywords used, and the keyword registration time are associated as follows.

Keyword Related Words Number of Uses Registration Time “Tonkatsu” “Pork” 1 YY: MM: DD
“Steak” “Beef” 0 YY: MM: DD
・ ・ ・ ・ ・ ・ ・ ・ ・ ・ ・ ・ ・ ・ ・ ・ ・ ・ ・ ・ ・ ・

For example, in the related word information stored in the memory 1b, the storage determination unit 12 leaves a predetermined number (such as 50 words) of keywords and related words in the order of new registration times of keywords and related words, and deletes others. May be. The storage determination unit 12 may leave a predetermined number of keywords and related words in the order of descending use of keywords and related words in the related word information stored in the memory 1b, and delete others. In the related word information stored in the memory 1b, the storage determination unit 12 deletes keywords and related words that have passed a predetermined period (N days, etc.) from the registration time and whose usage number is less than a predetermined number (such as once). Also good.
The storage determination unit 12 may delete the keyword and the related word in the related word information stored in the memory 1b based on the associative strength and the user profile information (user preference, preference pattern, etc.).

  The related word information may be stored and updated in the network device 101 such as a server or database on the network 100 (FIG. 5). In this case, for example, a plurality of response generation apparatuses 1 are connected to the network 100.

  Each response generation device 1 transmits the keyword of the first word and the associative word of the second word to the network device 101 on the network 100, and updates the related word information stored in the network device 101. Furthermore, the network device 101 limits the keywords of the related word information and the number of related words under the predetermined condition as described above. Then, the network device 101 transmits the related word information limited by the predetermined condition to each response generation device 1. Each response generation device 1 stores the related word information transmitted from the network device 101 in the memory 1b. As described above, the related word information is collectively updated by the network device based on the information from the plurality of response generation devices 1, thereby making it possible to provide diversity by repeated response sentences. In the third embodiment, the same parts as those in the first and second embodiments are denoted by the same reference numerals, and detailed description thereof is omitted.

Note that the present invention is not limited to the above-described embodiment, and can be changed as appropriate without departing from the spirit of the present invention.
In the above embodiment, the repetition generation unit 5 adds the additional word of the additional information in the memory 1b to the noun or verb output from the part-of-speech extraction unit 9, and then associates the noun or verb with the related word information in the memory 1b. Although it is replaced with a word, it is not limited to this. The repetition generation unit 5 replaces the noun or verb output from the part-of-speech extraction unit 9 with the related word of the related word information in the memory 1b, and then adds the additional word of the additional information in the memory 1b to the replaced related word. It may be added.

  For example, the repetition generation unit 5 uses a related word that matches the noun or verb output from the part of speech extraction unit 9 based on the noun or verb output from the part of speech extraction unit 9 and the related word information in the memory 1b. Select information keywords. The repetition generation unit 5 replaces the noun or verb output from the part-of-speech extraction unit 9 with a related word corresponding to the selected keyword. Based on the noun or verb output from the part-of-speech extraction unit 9 and the additional information in the memory 1b, the repetition generation unit 5 selects the keyword of the additional information that matches the noun or verb output from the part-of-speech extraction unit 9. select. The repeat generation unit 5 generates a repeat response sentence by adding an additional word corresponding to the selected keyword to the replaced related word.

  In the above-described embodiment, the response output unit 4 outputs the conflict response generated by the conflict generation unit 22 from the speaker 7, but is not limited thereto. The response output unit 4 may perform an arbitrary response with a low processing load based on the conflict response generated by the conflict generation unit 22. For example, the response output unit 4 may perform vibrations of the vibration device, lighting / flashing of the light device, display of the display device, operation of each part such as a robot's limbs, head, and trunk, and any combination thereof. You may go.

  In the above embodiment, the response output unit 4 outputs the repeated response text generated by the repeated generation unit 5 from the speaker 7, but is not limited thereto. The response output unit 4 may output an arbitrary repeated response sentence with a low processing load based on the repeated response sentence generated by the repeated generation unit 5. For example, the response output unit 4 may output a repetitive response sentence using a display on the display device or the like, or may output it by arbitrarily combining means. In this case, for example, the output mode of the response output unit 4 may be settings such as character size, brightness, and shape.

In addition, the present invention can realize the processing shown in FIG. 3 by causing the CPU 1a to execute a computer program, for example.
The program may be stored using various types of non-transitory computer readable media and supplied to a computer. Non-transitory computer readable media include various types of tangible storage media. Examples of non-transitory computer-readable media include magnetic recording media (for example, flexible disks, magnetic tapes, hard disk drives), magneto-optical recording media (for example, magneto-optical disks), CD-ROMs (Read Only Memory), CD-Rs, CD-R / W and semiconductor memory (for example, mask ROM, PROM (Programmable ROM), EPROM (Erasable PROM), flash ROM, RAM (random access memory)) are included.

  The program may also be supplied to the computer by various types of transitory computer readable media. Examples of transitory computer readable media include electrical signals, optical signals, and electromagnetic waves. The temporary computer-readable medium can supply the program to the computer via a wired communication path such as an electric wire and an optical fiber, or a wireless communication path.

  DESCRIPTION OF SYMBOLS 1 Response production | generation apparatus, 2 Speech recognition part, 3 Structure analysis part, 4 Response output part, 5 Repeat production | generation part, 6 Microphone, 7 Speaker, 8 Incomplete dictionary database, 9 Part of speech extraction part, 10 Related word extraction part, 11 User question part, 12 memory determination part, 21 phonological analysis part, 22 phase generation part, 23 routine response database

Claims (6)

Recognizing the user's voice;
Analyzing the structure of the recognized speech;
Storing information in which a plurality of keywords are associated with an associative word and an additional word related to each keyword,
Extracting a noun or verb from the recognized speech;
Select a keyword of the stored information that matches the extracted noun or verb, add an additional word corresponding to the selected keyword to the extracted noun or verb, and select the noun or verb Generating a repetitive response sentence for repeating the user's voice by replacing with a related word corresponding to the keyword,
Generating an optional response sentence for the user's voice based on the analyzed voice structure, outputting the repeated response sentence, and then outputting the optional response sentence;
A response generation method including:
Extracting a first noun or verb and a second noun or verb associated with the first noun or verb from predetermined sentence set information;
Asking the user whether the extracted first noun or verb and the second noun or verb are the same;
Storing the first noun or verb as the keyword and storing the second noun or verb as the associative word when a user's answer to the question is affirmative;
A response generation method comprising: The response generation method according to claim 1, comprising:
Extracting the first noun or verb and the second noun or verb from the predetermined sentence set information existing on the network,
The degree of similarity between the first noun or verb and the second noun or verb is calculated using the co-occurrence frequency of the extracted first noun or verb and the second noun or verb, and the calculation The first noun or verb extracted and the second noun or verb are identical to the user when the similarity is equal to or greater than a threshold,
When the user's answer to the question is positive, the first noun or verb is used as the keyword, and the second noun or verb is stored as the associative word. The response generation method according to claim 1 or 2,
The response generation apparatus further comprising a step of limiting a number of associative words and additional words associated with the stored keyword under a predetermined condition. The response generation method according to any one of claims 1 to 3,
Analyzing the phoneme of the user's voice;
Generating a reciprocal response to the user's voice based on the analyzed phonological analysis results; and
Before outputting the generated repeated response sentence, the generated response of the response is output. Voice recognition means for recognizing the user's voice;
Structure analysis means for analyzing the structure of the voice recognized by the voice recognition means;
Storage means for storing information in which a plurality of keywords are associated with an associative word and an additional word related to each keyword,
Part of speech extraction means for extracting a noun or verb from the speech recognized by the speech recognition means;
Selecting a keyword of the storage means that matches the noun or verb extracted by the part-of-speech extraction means, and adding an additional word corresponding to the selected keyword to the extracted noun or verb; Repetitive generating means for generating a repetitive response sentence for repeating the user's voice by replacing a verb with a related word corresponding to the selected keyword;
Based on the structure of the voice analyzed by the structure analysis means, an optional response sentence for the user's voice is generated, and after outputting the repeated response sentence, a response output means for outputting the optional response sentence;
A response generating device including:
Extraction means for extracting a first noun or verb and a second noun or verb related to the first noun or verb from predetermined sentence set information;
Question means for asking the user whether or not the first noun or verb extracted by the extraction means and the second noun or verb are the same;
When the user's answer to the question is affirmative, a determination unit that registers the first noun or verb as the keyword and the second noun or verb as the association word in the storage unit;
A response generation device comprising: Processing to recognize the user's voice,
Processing for analyzing the structure of the recognized speech;
Processing to extract nouns or verbs from the recognized speech;
Information in which an associative word and an additional word related to each keyword are associated with a plurality of keywords is stored, the keyword that matches the extracted noun or verb is selected, and an additional corresponding to the selected keyword is stored. By adding a word to the extracted noun or verb and replacing the noun or verb with a related word corresponding to the selected keyword, a repeated response sentence for repeating the user's voice is generated. Processing,
Based on the structure of the analyzed voice, generating an optional response sentence to the user's voice, outputting the repeated response sentence, and then outputting the optional response sentence;
A response generation program to be executed by a computer,
A process of extracting a first noun or verb and a second noun or verb related to the first noun or verb from predetermined sentence set information;
Processing for asking the user whether or not the extracted first noun or verb and the second noun or verb are the same;
When the user's answer to the question is positive, the first noun or verb as the keyword and the second noun or verb as the associative word; and
A response generation program characterized by causing a computer to execute.

Images