JP6914094B2 - Utterance generator, utterance generation method and utterance generation program - Google Patents

Description

The present invention relates to devices, methods and programs for generating utterances of robots.

Conventionally, communication robots in which humans interact with computers have been developed. For example, a conversation bot called a chatbot has been developed as a robot that interacts while learning dialogue with humans.
Chatbots employ a method of mechanically generating plausible response sentences from human remarks. As a method of implementing chatbots, a dictionary type that prepares remarks and response sentences as conversation rules, and a log type that searches for the closest to the user's remarks from a large number of conversation logs and uses the response log. , There is a Markov sentence generation type that creates a Markov chain from a large number of sentences and generates a sentence (see, for example, Non-Patent Documents 1 and 2).

By the way, a software program or a communication robot incorporating an artificial brain is new at the beginning of use, but there is a problem that the user gets tired of using it for a while and does not use it. This is due to inadequate conversation patterns and chatbot responses to human utterances in many cases.
In addition, chatbots respond by actions (utterances) from humans, but humans rarely actively talk to robots in daily conversations, and the fact that they cannot think of the content of the utterance is also a cause of not being used for a long time. Conceivable.

Further, instead of a one-to-one conversation between a person and a robot, a technique has been proposed aimed at making it more enjoyable to watch than watching alone by watching TV together with the person and the robot.
For example, Patent Document 1 proposes a technique of determining a robot's behavior from a user's preference and a robot's preference and supporting them together while watching a program. Patent Document 2 proposes a technique for determining an emotional state from video impression information and user emotion information to express emotions that the robot sympathizes with.

Japanese Patent No. 4368316 Japanese Unexamined Patent Publication No. 2016-36883

Ryota Tomisaka, Takashi Suzuki, "Chatbot", Journal of the Institute of Image Information and Television Engineers, Vol. 64, No. 1, pp. 64-66 (2010) Tomotoshi Akiyama "Love Program-Artificial Brainless Made with Ruby", Mainichi Communications, 2005

According to the technologies of Patent Document 1 and Patent Document 2, the robot operates even if the human does not talk to the robot by interposing the television, so that the robot is used for a long time as compared with the one-to-one dialogue between the human and the robot. obtain.
However, these are technologies aimed at robots that enjoy watching TV while sympathizing with people, and are intended for people who enjoy watching TV. For example, many solitary people watch TV alone, and often leave the TV on to distract them from being too quiet, even though they don't particularly want to watch TV.

An object of the present invention is to provide an utterance generation device, an utterance generation method, and an utterance generation program capable of uttering words expressing emotions in accordance with a television program so that the robot has an independent emotion.

The utterance generator according to the present invention has an utterance template database that stores utterance sentence templates in association with representative words that indicate utterance features, a feature vector database that stores keyword feature vectors, and input keywords. An utterance sentence generation that generates an utterance sentence by combining the input keyword with respect to the distance calculation unit that calculates the distance between the feature vector and the feature vector of the representative word and the template selected based on the distance. It has a part and.

The utterance generation device includes a representative word ranking unit that ranks the representative words based on the distance, and the utterance sentence generation unit gives priority to the representative word having a higher ranking by the ranking. The template corresponding to the higher-level representative word may be selected.

The utterance generator may include a keyword extraction unit that extracts keywords registered in the keyword dictionary from the input sentence.

When the input sentence includes a keyword registered in the keyword dictionary and the representative word or another expression of the representative word, the utterance generator inputs the input as a template corresponding to the representative word. Among the sentences, a template registration unit for registering a sentence generalizing the keyword may be provided.

The utterance generator has a morphological analysis unit that analyzes the input sentence and converts it into a word-separated sentence, and when the representative word or another expression of the representative word is divided into a plurality of words in the word-separated sentence. An over-division correction unit that concatenates the plurality of words and restores them to the representative word or the different expression, a feature vector calculation unit that calculates the feature vector of each word in the word-separated sentence, and updates the feature vector database. May be provided.

The utterance generator includes an expression unification unit that replaces the other expression with the representative word in the word-separated sentence, and the feature vector calculation unit uses the feature vector after the other expression is replaced with the representative word. You may calculate.

The utterance generator may include a representative word dictionary update unit that extracts words within a predetermined distance from the feature vector of the representative word from the feature vector database and registers them in the dictionary as the alternative expression.

In the utterance generation method according to the present invention, a computer that stores a template of an utterance sentence in association with a representative word indicating the characteristic of the utterance and stores the feature vector of the keyword causes the input keyword feature vector and the representative word. A distance calculation step for calculating the distance from the feature vector of the above, and an utterance sentence generation step for generating an utterance sentence by combining the input keywords with respect to the template selected based on the distance are executed. ..

The utterance generation program according to the present invention stores the template of the utterance sentence in association with the representative word indicating the characteristic of the utterance, and the feature vector of the keyword input to the computer storing the feature vector of the keyword and the representative word. A distance calculation step for calculating the distance from the feature vector of the above, and an utterance sentence generation step for generating an utterance sentence by combining the input keywords with respect to the template selected based on the distance are executed. Is for.

According to the present invention, the robot can speak emotional words according to the TV program being watched as if it had independent emotions, and the surroundings of a small child being happy to watch TV. It can be expected to have the effect of making an adult look at it with a smile.

It is a block diagram which shows the functional structure of the utterance generator which concerns on embodiment. It is a figure which shows the registration example of the representative word dictionary which concerns on embodiment. It is a figure which shows the registration example of the utterance template database which concerns on embodiment. It is a figure which shows the registration example of the feature vector database which concerns on embodiment. It is a figure which shows the conversion example of the input sentence which concerns on embodiment. It is a flowchart which shows the template registration process which concerns on embodiment. It is a figure which shows the operation image of the related word selection which concerns on embodiment. It is a flowchart which shows the update process of the representative word dictionary which concerns on embodiment.

Hereinafter, an example of the embodiment of the present invention will be described.
The utterance generator 1 according to the present embodiment includes words and phrases including verbs expressing desires such as "want to eat", "want to talk", and "want to go", and words such as "beautiful", "delicious", and "big". For representative words including adjectives, prepare template sentences including these representative words and feature vectors of many keywords including the representative words.

The utterance generator 1 compares the feature vector of this keyword with the feature vector of the representative word for the keyword of the program being viewed, and combines the target keyword with the template sentence for the representative word that is close to the keyword. Generate utterance sentences.
As a result, the utterance generator 1 selects an appropriate verb or adjective representative word for the keyword, and for example, for the keyword "ramen", "want to eat", "delicious", etc. as the representative word. You can select and generate utterances such as "I want to eat ramen" and "It looks delicious ramen" from the template.

FIG. 1 is a block diagram showing a functional configuration of the utterance generation device 1 according to the present embodiment.
The utterance generation device 1 is an information processing device provided with various interfaces in addition to the control unit and the storage unit, and the control unit executes software (utterance generation program) stored in the storage unit to execute the software (utterance generation program) of the present embodiment. Various functions are realized.

The control unit of the utterance generation device 1 includes a morphological analysis unit 11, an overdivision correction unit 12, a template registration unit 13, an expression unification unit 14, a feature vector calculation unit 15, a keyword extraction unit 16, and a vector distance calculation. A unit 17, a representative word ranking unit 18, an utterance sentence generation unit 19, and a representative word dictionary update unit 20 are provided.
Further, the storage unit of the utterance generation device 1 includes a keyword dictionary 21, a representative word dictionary 22, an utterance template database (DB) 23, and a feature vector database 24.

In the present embodiment, the utterance generator 1 is incorporated inside a communication robot placed on the side of the television.
The utterance generation device 1 will be described as having a subtitle sentence (closed caption) added to a television broadcast program input as an input sentence I. As a method of extracting subtitles from a broadcast program, for example, a robot interlocks with a television receiver to acquire subtitle information, or as a broadcast interlocking service, subtitle information distributed using communication equipment such as the Internet is used. A method of receiving is conceivable.
Further, the input sentence I is not limited to the subtitle sentence, and an open caption may be obtained by image processing from a television image using a camera or a microphone equipped on the robot, or by voice recognition from voice. The statement may be extracted.

The keyword dictionary 21 is a database in which keywords are registered. Here, the keyword is mainly a person's name, a place name, a proper noun such as a facility, or a general noun. In the present embodiment, it is assumed that the keyword is registered in the keyword dictionary 21 in advance.

The representative word dictionary 22 is a database in which another expression having a similar meaning to the representative word is registered in association with the representative word indicating the characteristics of the utterance.
FIG. 2 is a diagram showing a registration example of the representative word dictionary 22 according to the present embodiment.
In this example, the representative word dictionary 22 stores data of three fields of type, representative word, and different expression as a list.

The type represents the type (part of speech) of the representative word, for example, v represents a verb and a represents an adjective.
The representative word is a word including a verb or an adjective expressing emotion. For example, "want to speak", "want to hear", and "want to eat" are registered as representative words of the verb, and "pretty" is registered as the representative word of the adjective. , "Interesting", and "Like" are registered.

In the field of another expression, a different expression having the same or similar meaning as the representative word is registered. For example, "I want to talk", "I want to talk", "I want to talk", "I want to say", "I want to say" are different expressions of "I want to talk", but "I want to hear" is another expression of "Listen". "I want to eat", "I want to hear", and "I want to listen" are registered as other expressions of "I want to eat", "I want to eat", "I want to eat", and "I want to eat".
Furthermore, "beautiful" is registered as another expression of "beautiful", "interesting" is registered as another expression of "interesting", and "love" is registered as another expression of "like".

Here, the hiragana notation and the kanji notation are registered as different words in the representative word dictionary 22. For example, the subtitle text may be matched to any notation, although the notation used differs depending on the broadcasting station or program.
In this embodiment, it is assumed that the representative word and another expression are registered in advance in the representative word dictionary 22.

The utterance template database 23 is a database in which templates of utterance sentences generated by the utterance generator 1 are registered.
FIG. 3 is a diagram showing a registration example of the utterance template database 23 according to the present embodiment.

In the utterance template database 23, utterance sentence templates are registered for each representative word. The representative word of the utterance template database 23 is the same as the word registered in the representative word dictionary 22.

For example, in the utterance template database 23, for the representative word "I want to talk", "I want to talk with% key", "I want to talk with% key", "I want to talk with% key", Template sentences such as "I have something to say to% key" are registered.
Here, "% key" indicates a part that can be replaced with some keyword.

Similarly, in the example of the figure, template sentences corresponding to the representative words of "want to hear", "want to eat", "beautiful", "interesting", and "like" are registered.
The data of the utterance template database 23 may be prepared in advance, or a new template may be added by the template registration unit 13 by using the input sentence I as described later.

The feature vector database 24 is a database in which feature vectors of each word including representative words and other keywords are registered.
FIG. 4 is a diagram showing a registration example of the feature vector database 24 according to the present embodiment.
A word and a corresponding feature vector are registered in the feature vector database 24.

In the present embodiment, the feature vector calculated by word2vec, which is known software, is used as the feature vector. Word2vec uses a three-layer neural network, and it is empirically known that when a large amount of divided sentences are input, the calculated feature vector becomes a vector in which similar words are close in distance.
The method of calculating the feature vector is not limited to word2vec, and any calculation method may be used as long as the feature vectors of words that are semantically close to each other are distributed close to each other in the vector space.

In the example of the figure, vc1 is registered as the feature vector of "ramen", and below, as the feature vector of "o", "want to eat", "○ river x child", "ha", "very", "beautiful". Vec2 to vc7 are registered, respectively, and a total of N feature vectors are registered.
In the feature vector database 24, a feature vector learned in advance by inputting a large number of sentences may be registered, or as described later, the feature vector calculation unit 15 may update the feature vector using the input sentence I.

[Database update process]
Of the above-mentioned databases, the utterance template database 23 and the feature vector database 24 are subjected to a series of processes of the morphological analysis unit 11, the overdivision correction unit 12, the template registration unit 13, the expression unification unit 14, and the feature vector calculation unit 15. Will be updated.

When the input sentence I is input to the utterance generation device 1, the morphological analysis unit 11 performs morphological analysis of the input sentence I and converts the input sentence I into a word-separated sentence.
Here, the word-separated sentence is a sentence in which a delimiter (for example, a blank character) is inserted between words divided into morphemes.

FIG. 5 is a diagram showing a conversion example of the input sentence I according to the present embodiment.
When the sentence "I want to eat this ramen" is input as the input sentence I, the analysis result of the morphological analysis unit 11 is divided and "This / ramen / is / eat / want / want / is / is /". Ne /. "Is output. Here, "/" is used instead of a space as a delimiter for the word-separated sentence.

The overdivision correction unit 12 connects words that have become too fine as a result of analysis by the morphological analysis unit 11. Specifically, the over-division correction unit 12 includes a keyword registered in the keyword dictionary 21 or a representative word or another expression registered in the representative word dictionary 22 in the input sentence I, and these are morphological elements. When the word is divided into a plurality of words by the analysis unit 11, these divided words are reconnected and restored to the original word.

In the example of FIG. 5, in the output of the morphological analysis unit 11, the word "want to eat" registered as another expression of "want to eat" in the representative word dictionary 22 is three words "eat / te / like". It is divided into. Therefore, the over-division correction unit 12 connects these to one word of "I want to eat", and as a result, outputs a word-separated sentence "This / ramen / is / wants to eat / is / ne /.".

The template registration unit 13 creates a template sentence as necessary from the word-separated sentence output from the overdivision correction unit 12, and registers it in the utterance template database 23.
When the input sentence I includes a keyword registered in the keyword dictionary 21 and a representative word or another expression of the representative word dictionary 22, the template registration unit 13 uses the input sentence as a template corresponding to the representative word. Of I, register a sentence that generalizes the keyword.

FIG. 6 is a flowchart showing a template registration process according to the present embodiment.
In step S1, the template registration unit 13 determines whether or not the divided input sentence I includes a representative word or another expression registered in the representative word dictionary 22. If this determination is YES, the process proceeds to step S2, and if the determination is NO, the process ends.

In step S2, the template registration unit 13 determines whether or not the input sentence I includes a keyword registered in the keyword dictionary 21. If this determination is YES, the process proceeds to step S3, and if the determination is NO, the process proceeds to step S4.

In step S3, the template registration unit 13 replaces the keyword portion included in the input sentence I with the character string “% key”.
In step S4, the template registration unit 13 deletes the space that is the delimiter in order to return the word-separated sentence to a normal sentence.

In step S5, the template registration unit 13 determines whether or not the same sentence as the result of step S4 is registered as the template of the representative word in the utterance template database 23. If this determination is YES, the process ends, and if the determination is NO, the process proceeds to step S6.
In step S6, the template registration unit 13 registers the result of step S4 in the utterance template database 23 as a new template.

By this process, the template registration unit 13 sends a sentence "I want to eat this% key" in response to the input "I want to eat this / ramen / / is / ne /." For example. , Register as a template for the representative word "I want to eat" in the utterance template database 23.

When the output sentence of the overdivision correction unit 12 includes another expression of the representative word registered in the representative word dictionary 22, the expression unification unit 14 replaces this other expression with the representative word.
For example, as shown in FIG. 5, when the output sentence of the overdivision correction unit 12 is "this / ramen / is / is / ne /.", The representative word dictionary 22 is representative of "want to eat". Since the word "want to eat", the expression unification department 14 outputs the sentence "this / ramen / is / wants to eat / is / ne /." By replacing "want to eat" with "want to eat".

The feature vector calculation unit 15 gives the sentence output by the expression unification unit 14 as the input of word2vec, calculates the feature vector of each word in the word-separated sentence in which another expression is replaced with the representative word, and stores it in the feature vector database 24. Update the feature vector.

In this way, the template registration unit 13 and the feature vector calculation unit 15 register the template sentence including the representative word or another expression for the representative word in the utterance template database 23, and the representative in the feature vector database 24. The feature vector of the word unified to the word is registered.
As a result, a word that rarely appears as another expression of the representative word (for example, "I want to eat" as another expression of "I want to eat") is used for learning the feature vector of "I want to eat". This alternative expression will be registered as a template, and the variation of the utterance sentence will expand. Further, by summarizing them into representative words, the amount of calculation of the vector distance calculation unit 17 described later can be reduced.

[Utterance generation process]
The utterance generation device 1 outputs an output sentence O (utterance sentence) to the input sentence I by a series of processes of the keyword extraction unit 16, the vector distance calculation unit 17, the representative word ranking unit 18, and the utterance sentence generation unit 19. do.

The keyword extraction unit 16 determines whether or not the input sentence I includes a keyword registered in the keyword dictionary 21, and extracts the included keyword. If the input sentence I does not include a keyword, the utterance generator 1 outputs nothing as an output sentence O to the input sentence I.

The vector distance calculation unit 17 calculates the vector distance between the keyword extracted by the keyword extraction unit 16 and each of the representative words registered in the representative word dictionary 22.
Here, when the feature vector of the keyword k is vector and the feature vector of the representative word t is vector, the cosine similarity cos (vac, vector) of Equation 1 can be used as the distance between the vectors.
(Equation 1) cos (veck, vect) = (veck ・ vect) / (| veck || vect |)

The representative word ranking unit 18 ranks by sorting the distance between the feature vector of the keyword and the feature vector of each representative word in descending order of this value, and the ranking strongly related to the keyword is the highest representative. Select a word.

FIG. 7 is a diagram showing an operation image of related word selection by the vector distance calculation unit 17 and the representative word ranking unit 18 according to the present embodiment.
In this example, the keywords are "ramen" (feature vector vc1) and "Kyoto" (feature vector vecy), and the representative words are "want to go" (feature vector vcx) and "want to eat" (feature vector vc3). It is defined.

As for the feature vector of the present embodiment, for example, the feature vectors of words having similar tendencies are close to each other so that "udon", "curry", "yakisoba" and the like are distributed near "ramen". At the same time, the feature vectors of related words are also close. That is, the feature vectors of foods such as "ramen" are distributed at positions closer to the feature vector of "want to eat" than the feature vector of "want to go". This means that the word "ramen" is closely related to "want to eat" as a representative word, and the word "Kyoto" is closely related to "want to go" as a representative word.

The utterance sentence generation unit 19 randomly selects from the templates registered in the utterance template database 23 for the representative words ranked by the representative word ranking unit 18, and "% key" in the selected template. By replacing the part with a keyword, an utterance sentence is generated and output as an output sentence O.

For example, when the keyword is "ramen" and the representative word is "want to eat", if the third template registered for the representative word "want to eat" is selected from the utterance templates in Fig. 3, the utterance sentence will be changed. "I want to eat this ramen because it's okay once."
The utterance sentence generation unit 19 has described that a plurality of registered templates are randomly selected for the representative word, but the present invention is not limited to this. The utterance generation device 1 may, for example, weight a specific template to give a bias to the selection probability, or update the selection probability as appropriate in relation to the viewing program to give the utterance sentence individuality. You may let me.

[Update process of representative word dictionary 22]
The representative word dictionary 22 may be automatically updated based on the feature vector database 24.
The representative word dictionary update unit 20 additionally registers words with different expressions for each representative word in the representative word dictionary 22.
As described above, the feature vectors registered in the feature vector database 24 are close to words having similar meanings. Utilizing this, the representative word dictionary update unit 20 registers an unregistered word as another expression of the representative word among the feature vectors whose distance from the feature vector of the representative word is within a predetermined value.

FIG. 8 is a flowchart showing an update process of the representative word dictionary 22 according to the present embodiment.
In step S11, the representative word dictionary update unit 20 sets the variable n for loop processing to 1.
In step S12, the representative word dictionary update unit 20 acquires the nth representative word among the representative words registered in the representative word dictionary 22.
In step S13, the representative word dictionary update unit 20 acquires the upper M of the terms of the feature vector close to the feature vector of the nth representative word acquired in step S12 from the feature vector database 24.

In step S14, the representative word dictionary update unit 20 sets the variable m for loop processing for the acquired M terms to 1.
In step S15, the representative word dictionary update unit 20 determines whether or not the mth term among the M terms acquired in step S13 is registered as another expression of the nth representative word in the representative word dictionary 22. judge. If this determination is YES, the process proceeds to step S17, and if the determination is NO, the process proceeds to step S16.

In step S16, the representative word dictionary update unit 20 registers the m-th term in the representative word dictionary 22 as another expression of the nth representative word.
In step S17, the representative word dictionary update unit 20 counts up the variable m and repeats the loop process until m = M.
In step S18, the representative word dictionary update unit 20 counts up the variable n and sets n = N in order to perform the above processing for all the number N of the representative words registered in the representative word dictionary 22. Repeat the loop process until.

In the present embodiment, in step S16, all the unregistered words are registered in the representative word dictionary 22, but the present invention is not limited to this, and for example, the user accepts a selection input as to whether or not to register. May be good.

According to the present embodiment, the utterance generator 1 generates a representative word strongly related to the keyword based on the distance between the feature vector of the keyword in the television program and the feature vector of each of the plurality of representative words expressing emotions and the like. When selected, an utterance sentence is generated by combining a template sentence corresponding to a representative word and a keyword.
Therefore, the utterance generation device 1 can spontaneously generate an utterance sentence indicating an emotion from a keyword related to the program. As a result, as if the robot incorporating the utterance generator 1 has independent emotions, it is possible to speak words expressing emotions according to the TV program being watched, and a small child is pleased to watch TV. You can expect the effect of watching the situation with a smile from the surrounding adults.

By incorporating the utterance generator 1 into the robot, the robot responds to the television and mutters words including the robot's own emotions, so that the user does not have to actively speak, and the user exchanges utterances between the television and the robot. Can be enjoyed from a bird's-eye view. In addition, since the robot speaks spontaneously, the effect of encouraging the user to speak can be expected. Furthermore, even if the user is not interested in the TV program itself being broadcast, the effect of being interested in the program can be expected as a result of the robot tweeting.

The utterance generation device 1 ranks representative words based on the distance of the feature vector, and generates an utterance sentence by giving priority to the representative word having a higher ranking, that is, a representative word strongly related to the keyword of the program. Therefore, the utterance generation device 1 can generate an utterance sentence by using an appropriate representative word for the input keyword.

The utterance generation device 1 can efficiently generate an utterance sentence for a keyword for which a feature vector is registered by extracting a keyword registered in the keyword dictionary 21 from the input sentence I.

The utterance generator 1 registers an unregistered template and updates the feature vector based on the input sentence I. Therefore, the utterance generator 1 can learn popular words, change the registered template sentences and feature vectors according to the TV program viewed, and cause variations in the utterance content depending on the TV viewing tendency. can.

When the input sentence I includes another expression of the representative word, the utterance generator 1 calculates the feature vector after replacing the other expression with the representative word. Therefore, the utterance generator 1 can collectively process similar words including different expressions, and can share the feature vector for the representative word regardless of the appearance frequency of each different expression. As a result, the feature vector of the representative word can be appropriately updated, and multiple expressions of the template can be treated equally, and the variation of the utterance sentence can be increased.

The utterance generator 1 uses the feature vector database 24 to register a term whose feature vector distance is close to that of the representative word in the representative word dictionary 22 as a separate expression. Can be increased.

Although the embodiments of the present invention have been described above, the present invention is not limited to the above-described embodiments. Further, the effects described in the present embodiment merely list the most preferable effects arising from the present invention, and the effects according to the present invention are not limited to those described in the present embodiment.

The utterance sentence generation unit 19 may generate not only one sentence but also a plurality of sentences and output them. The utterance sentence generation unit 19 may generate utterance sentences for the upper n representative words of the representative word ranking unit 18, for example, or the utterance sentence for the representative words whose cosine similarity value is equal to or greater than the threshold value. May be generated. Further, the utterance sentence generation unit 19 may perform ranking for each type of representative word (verb or adjective) and generate an utterance sentence using each higher-level representative word.
By generating a plurality of utterance sentences, the utterance sentence generation unit 19 can provide a plurality of options for the utterance sentence selection process that can be provided as a subsequent process, and can increase the variation of the utterance content. ..
Further, at this time, the utterance sentence generation unit 19 may output the distance value of the feature vector together with the utterance sentence for the subsequent processing.

In the present embodiment, the utterance template database 23 and the feature vector database 24 have been described as being updated every time the input sentence I is input, but if sufficient data is accumulated, the morphological analysis unit 11 will be overloaded. The database update process by the division correction unit 12, the template registration unit 13, the expression unification unit 14, and the feature vector calculation unit 15 may be omitted. Further, in this case, as the input to the utterance generation device 1, a keyword may be simply input instead of the input sentence I. In this case, the keyword dictionary 21 and the keyword extraction unit 16 are also unnecessary.

In the present embodiment, the configuration and operation of the utterance generator 1 have been mainly described, but the present invention is not limited to this, and the present invention includes each component and is configured as a method or a program for generating an utterance sentence of a robot. You may.

Further, the program for realizing the function of the utterance generator 1 may be recorded on a computer-readable recording medium, and the program recorded on the recording medium may be read by the computer system and executed. good.

The term "computer system" as used herein includes hardware such as an OS and peripheral devices. Further, the "computer-readable recording medium" refers to a portable medium such as a flexible disk, a magneto-optical disk, a ROM, or a CD-ROM, or a storage device such as a hard disk built in a computer system.

Furthermore, a "computer-readable recording medium" is a communication line that transmits a program via a network such as the Internet or a communication line such as a telephone line, and dynamically holds the program for a short period of time. It may also include a program that holds a program for a certain period of time, such as a volatile memory inside a computer system that serves as a server or a client in that case. Further, the above-mentioned program may be a program for realizing a part of the above-mentioned functions, and may be a program for realizing the above-mentioned functions in combination with a program already recorded in the computer system. ..

1 Utterance generator 11 Morphological analysis unit 12 Overdivision correction unit 13 Template registration unit 14 Expression unification unit 15 Feature vector calculation unit 16 Keyword extraction unit 17 Vector distance calculation unit 18 Representative word ranking unit 19 Speech generation unit 20 Representative word dictionary update Part 21 Keyword dictionary 22 Representative word dictionary 23 Speaking template database 24 Feature vector database

Claims (9)

An utterance template database that stores utterance template templates in association with representative words that include emotional verbs or adjectives.
A feature vector database that stores keyword feature vectors and
A distance calculation unit that calculates the distance between the input feature vector of the noun keyword and the feature vector of the representative word, and
An utterance generation device including a utterance sentence generation unit that generates an utterance sentence by combining the input noun keywords with respect to the template selected based on the distance. A representative word ranking unit that ranks the representative words based on the distance is provided.
The utterance generation device according to claim 1, wherein the utterance sentence generation unit gives priority to a representative word having a higher ranking based on the ranking and selects the template corresponding to the higher representative word. The utterance generator according to claim 1 or 2, further comprising a keyword extraction unit for extracting noun keywords registered in a keyword dictionary from an input sentence. When the input sentence includes a noun keyword registered in the keyword dictionary and the representative word or another expression of the representative word, the input sentence of the input sentence is used as a template corresponding to the representative word. The utterance generator according to claim 3, further comprising a template registration unit for registering sentences that generalize noun keywords. A morphological analysis unit that analyzes the input sentence and converts it into a word-separated sentence.
In the word-separated sentence, when the representative word or another expression of the representative word is divided into a plurality of words, the over-division correction unit that concatenates the plurality of words and restores the representative word or another expression to the representative word or the different expression.
The utterance generator according to claim 3 or 4, further comprising a feature vector calculation unit that calculates a feature vector of each word in the word-separated sentence and updates the feature vector database. In the word-separated sentence, the expression unification unit that replaces the other expression with the representative word is provided.
The utterance generation device according to claim 5, wherein the feature vector calculation unit calculates the feature vector after the alternative expression is replaced with the representative word. The present invention according to any one of claims 4 to 6, further comprising a representative word dictionary update unit that extracts words within a predetermined distance from the feature vector of the representative word from the feature vector database and registers them in the dictionary as the alternative expression. The utterance generator described. A computer that memorizes the template of the utterance sentence and memorizes the feature vector of the keyword in association with the representative word including the verb or adjective expressing emotion.
A distance calculation step for calculating the distance between the feature vector of the input noun keyword and the feature vector of the representative word, and
An utterance generation method for executing a utterance sentence generation step of combining the input noun keywords with respect to the template selected based on the distance to generate an utterance sentence. A computer that stores a template of an utterance sentence and stores a feature vector of a keyword in association with a representative word containing an emotional verb or adjective.
A distance calculation step for calculating the distance between the feature vector of the input noun keyword and the feature vector of the representative word, and
An utterance generation program for executing a utterance sentence generation step of combining the input noun keywords with respect to the template selected based on the distance to generate an utterance sentence.

Images