JP7308550B2 - Utterance generation device, utterance generation method, and computer program - Google Patents

Description

The present invention relates to a speech generation device, a speech generation method, and a computer program.

2. Description of the Related Art In recent years, the development of dialogue devices that realize dialogue with a user by analyzing a user's utterance and returning a response according to the content of the utterance has been promoted. Such interactive devices are used, for example, in car navigation devices, guidance devices in public facilities and the like.

JP 2004-110524 A

However, since the above-described interactive devices are task-oriented and aimed at answering questions, they cannot be expected to have the effect of arousing changes in the user's mind, such as healing, treatment, advice, and idea support.

The present invention has been made in view of such circumstances, and provides an utterance generation device, a utterance generation method, and a computer program that can be expected to have the effect of arousing a change in the user's mind, such as healing, treatment, advice, and idea support. intended to provide

An utterance generation device according to an aspect of the present application is a utterance generation device that generates a system utterance to be output in response to an input user utterance, comprising: a recognition unit that recognizes a user's intention from the user utterance; Acquisition of extracting a noun included in an utterance and acquiring a plurality of associated words associated with the extracted noun and a plurality of co-occurring expressions including terms co-occurring with the noun and the plurality of associated words a classification unit that classifies the noun and the plurality of associative words into a plurality of clusters based on the amount of co-occurrence expressions commonly used between the noun and the plurality of associative words; and a selection unit that selects one noun or associative word from one of the clusters and selects one co-occurrence expression that co-occurs with any noun or associative word included in the one cluster; connects a conversion unit that converts the recognized user's intention into the intention to be added to the system utterance, the noun or associative word selected by the selection unit, and the co-occurrence expression selected by the selection unit; and an utterance generation unit that generates a system utterance by connecting the intention after conversion by the conversion unit to the co-occurrence expression .

An utterance generation method according to an aspect of the present application is an utterance generation method that uses a computer to generate a system utterance to be output in response to an input user utterance, wherein the computer extracts a user's intention from the user utterance. and extracting nouns included in the user utterance, and a plurality of co-occurrences including a plurality of associated words associated with the extracted nouns and terms co-occurring with each of the nouns and the plurality of associated words. and classifying the noun and the plurality of associative words into a plurality of clusters based on the amount of co-occurrence expressions commonly used between the noun and the plurality of associative words, and classifying the plurality of clusters. Select one noun or associative word from one of the clusters, select one co-occurring expression that co-occurs with any noun or associative word contained in the one cluster, and recognize the user's intention converting the system utterance into an intention to be imparted, connecting the selected noun or associative word with the selected co-occurring expression, and connecting the converted intention to the co-occurring expression to generate the system utterance; do.

A computer program according to an aspect of the present application is a computer program for causing a computer to execute processing for generating a system utterance to be output in response to an input user utterance, wherein the computer instructs the computer to generate a user utterance from the user utterance. extracting nouns included in the user utterance; and extracting a plurality of associated words associated with the extracted nouns; classifying the noun and the plurality of associative words into a plurality of clusters based on the amount of co-occurring expressions commonly used between the noun and the plurality of associative words; selecting one noun or associative word from one of the clusters, selecting one co-occurring expression co-occurring with any noun or associative word contained in the one cluster, and Converting the intention into the intention to be given to the system utterance, connecting the selected noun or associated word with the selected co-occurrence expression, and connecting the converted intention to the co-occurrence expression to generate the system utterance. is a computer program for executing the process of generating

According to the present application, effects such as comfort, treatment, advice, idea support, and the like that arouse a change in the user's mental state can be expected.

1 is a block diagram illustrating the overall configuration of a dialogue system according to Embodiment 1; FIG. 1 is a schematic diagram showing an example of dialogue in a dialogue system; FIG. It is a block diagram explaining the internal structure of a speech production|generation apparatus. It is a block diagram explaining the internal structure of a terminal device. FIG. 4 is an explanatory diagram for explaining an outline of speech generation processing; It is a figure which shows the extraction example of the noun by Word2Vec. It is an explanatory view explaining a clustering method in this embodiment. FIG. 10 is an explanatory diagram illustrating selection processing of nouns and co-occurrence expressions; FIG. 4 is a conceptual diagram showing an example of an intention conversion table; FIG. 4 is a conceptual diagram showing an example of an intention conversion table; 4 is a flowchart for explaining the procedure of processing executed by the speech generation device; FIG. 10 is an explanatory diagram illustrating a system utterance generation method according to Embodiment 2;

BEST MODE FOR CARRYING OUT THE INVENTION Hereinafter, the present invention will be specifically described based on the drawings showing the embodiments thereof.
(Embodiment 1)
FIG. 1 is a block diagram for explaining the overall configuration of a dialogue system according to Embodiment 1. As shown in FIG. A dialogue system according to the present embodiment includes a speech generation device 10 and a terminal device 20 that are connected to each other via a communication network N so as to be able to communicate with each other. The terminal device 20 is an information processing terminal used by a user, such as a personal computer, a smart phone, an AR (Augmented Reality) device, etc., and an application program for accessing the speech generation device 10 is installed. For example, when an access from the terminal device 20 is accepted, the speech generation device 10 performs user authentication, and provides an interactive service to the terminal device 20 when the user authentication is successful.

FIG. 2 is a schematic diagram showing an example of dialogue in the dialogue system. FIG. 2 shows an example of a dialogue displayed on the display screen 20A of the terminal device 20. As shown in FIG. The dialogue consists of user utterances input by the user using the terminal device 20 and system utterances that are the responses of the utterance generation device 10 to the user utterances. In the present embodiment, user utterances and system utterances are described as character information, but it goes without saying that voice information may also be used.

When the terminal device 20 receives the input of the user utterance U01, the terminal device 20 transmits the received user utterance U01 to the utterance generation device 10. FIG. When receiving a user utterance U01 transmitted from the terminal device 20, the utterance generation device 10 generates a system utterance S01 as a response, and transmits the generated system utterance S01 to the terminal device 20. FIG. After that, until the user utterance U06 of "bye-bye" in the example shown in FIG. Each time utterances U02, U03, . . . are received, system utterances S02, S03, .

Input user utterances U01, U02, U03, . . . and system utterances S01, S02, S03, . In the example of FIG. 2, on the right side of the display screen 20A, user utterances U01 to U05 are displayed in chronological order from the top to the bottom of the screen, and on the left side of the display screen 20A, utterances U01 to U05 are displayed from the top to the bottom of the screen. It shows how system utterances S01 to S05 are displayed in chronological order.

A method of generating the system utterance by the utterance generation device 10 will be described in detail later.

FIG. 3 is a block diagram for explaining the internal configuration of the speech generator 10. As shown in FIG. The speech generation device 10 is, for example, a server device, and includes a control section 11 , a storage section 12 , a communication section 13 , a display section 14 and an operation section 15 .

The control unit 11 includes, for example, a CPU (Central Processing Unit), a ROM (Read Only Memory), a RAM (Random Access Memory), and the like. The CPU provided in the control unit 11 expands various computer programs stored in the ROM or the storage unit 12 on the RAM and executes them, thereby causing the entire device to function as the speech generation device of the present application.

Note that the control unit 11 is not limited to the configuration described above, and may be any processing circuit including one or more CPUs, a multi-core CPU, a microcomputer, or the like. Further, the control unit 11 may have a function such as a timer for measuring the elapsed time from when the measurement start instruction is given until when the measurement end instruction is given, and a counter for counting the number.

The storage unit 12 is configured by a storage device such as a hard disk device, and stores various computer programs and various data. Here, the computer programs stored in the storage unit 12 include a computer program (speech generation program 120) for causing the speech generation device 10 to execute processing for generating system utterances from input user utterances.

The program stored in the storage unit 12 may be provided by a non-temporary recording medium M1 on which the program is readable. The recording medium M1 is, for example, a portable memory such as a CD-ROM, USB memory, SD (Secure Digital) card, micro SD card, compact flash (registered trademark). In this case, the control unit 11 uses a reading device (not shown) to read various programs from the recording medium M<b>1 and installs the read various programs in the storage unit 12 . Also, the programs stored in the storage unit 12 may be provided by communication via the communication unit 13 . In this case, the control unit 11 acquires various programs through the communication unit 13 and installs the acquired various programs in the storage unit 12 .

Various data stored in the storage unit 12 include an intention conversion table 121 and Word2Vec data 122, which will be described later. Details of these data will be described later.

The communication unit 13 has an interface for communicating with the terminal device 20 through the communication network N. FIG. When information to be transmitted to the terminal device 20 is input from the control unit 11, the communication unit 13 transmits the input information to the terminal device 20 and controls information received from the terminal device 20 through the communication network N. Output to unit 11 .

The display unit 14 has a display device such as a liquid crystal display or an organic EL display, and displays information to be notified to the administrator of the speech generation device 10 . Further, the operation unit 15 includes a touch panel and various buttons, receives operations by the administrator of the speech generation device 10 , and outputs received operation information to the control unit 11 .

In this embodiment, for the sake of simplification, the utterance generation device 10 is described as one device, but it may be configured by a plurality of server devices, and may be configured by one or a plurality of virtual machines. There may be.

Further, in this embodiment, for the sake of simplification, it is assumed that the storage unit 12 of the utterance generation device 10 includes the intention conversion table 121 and the Word2Vec data 122. may be prepared, and necessary data may be obtained by accessing an external server as necessary.

FIG. 4 is a block diagram for explaining the internal configuration of the terminal device 20. As shown in FIG. The terminal device 20 is, for example, a server device, an information processing terminal such as a personal computer or a smartphone, and includes a control unit 21 , a storage unit 22 , a communication unit 23 , a display unit 24 and an operation unit 25 .

The control unit 21 is composed of, for example, a CPU, a ROM, a RAM, and the like. The CPU provided in the control unit 21 controls the operation of the entire apparatus by deploying various computer programs stored in the ROM or the storage unit 22 on the RAM and executing them.

The storage unit 22 is composed of a nonvolatile memory such as an EEPROM (Electronically Erasable Programmable Read Only Memory), and stores various computer programs and various data. Here, the computer programs stored in the storage unit 22 include application programs for accessing the speech generation device 10 .

The communication unit 23 has an interface for communicating with the speech generation device 10 through the communication network N. FIG. When information to be transmitted to the utterance generation device 10 is input from the control unit 21, the communication unit 23 transmits the input information to the utterance generation device 10. Information is output to the control unit 21 .

The display unit 24 includes a display device such as a liquid crystal display and an organic EL display, and displays information (for example, dialogue sentences) to be notified to the user of the terminal device 20 . The operation unit 25 also has a touch panel and various buttons, receives operations by the user of the terminal device 20 , and outputs received operation information to the control unit 21 .

In the present embodiment, the terminal device 20 is configured to include the display unit 24 , but may be configured to include an audio output device in place of the display unit 24 or in addition to the display unit 24 . In this case, the terminal device 20 can output information to be notified to the user (for example, dialogue sentences) as voice from the voice output device.

The speech generation process performed by the speech generation device 10 will be described below.
FIG. 5 is an explanatory diagram for explaining the outline of the utterance generation process. When receiving a user's speech from the terminal device 20, the speech generation device 10 first extracts the content and the user's intention from the received user's speech. Therefore, the utterance generation device 10 uses a morphological analysis technique to decompose the user's utterance into a plurality of morphemes. In the example of FIG. 5, when the user utterance received from the terminal device 20 is "I want to study at school", the morphological analysis is performed to divide it into four morphemes: "school", "de", "learn", and "want". It shows disassembled. Also, when the user's utterance received from the terminal device 20 is "technical management?", it is decomposed into two morphemes, "technical management" and "?".

The utterance generation device 10 acquires the main part of speech based on the result of the morphological analysis. One or a plurality of main parts of speech are acquired from the user's utterance in the order of verbs, nouns, adjectives, or shape words starting from the terminal verb. The speech generation device 10 recognizes morphemes corresponding to nouns, verbs, and adjectives among the acquired parts of speech as content. In addition, the utterance generation device 10 recognizes morphemes mainly corresponding to particles and auxiliary verbs, excluding nouns and verbs, as intentions. In the example of FIG. 5, of the four morphemes "school", "de", "learn", and "want", "school" and "learn" are recognized as content, and "want" is the user's intention (this indicates that it was recognized as a desire). Also, of the two morphemes of "technical management" and "?", "technical management" was recognized as content, and "?" was recognized as the user's intention (question in this case). The user's intention extracted by the utterance generation device 10 may include information about the user's personality, emotion, and dialogue situation.

The utterance generation device 10 executes a process of extracting associated words associated with recognized content. In this embodiment, a configuration for extracting nouns as associated words associated with content will be described, but the associated words may be verbs or adjectives. The utterance generation device 10 synthesizes the main part of speech of the most recent user utterance, the noun, verb or adjective of the most recent system utterance, and the theme while holding the subject noun or initial noun as a theme, and refers to the Word2Vec data 122, Extract only the set number of nouns. For example, when associating a noun from the content included in the second user utterance, "Technology management?" It is possible to associate nouns by adding "school" and "learning"). After extracting a plurality of nouns associated with the content, the utterance generation device 10 selects one noun to be included in the system utterance from the group of extracted nouns. In addition, co-occurring expressions including predicates co-occurring with the selected noun are extracted as necessary. Note that when the associated words extracted from the content are verbs or adjectives, co-occurring expressions including nominal words co-occurring with the verbs or adjectives may be extracted.

In this embodiment, the Word2Vec data 122 is used when extracting nouns associated with recognized content. Of course, it is also possible to extract nouns associated with the recognized content using the converted data.

Further, in the present embodiment, an association word is extracted from recognized content, but the association may be intentionally biased by adding a set term. For example, by adding the term "physics" it is possible to bias the association towards concepts close to physics.

Note that, in the present embodiment, it is allowed to select (parrot) the same content as the content included in the user's utterance. The example of FIG. 5 shows that the parrot has selected the content of the system utterance for the content "learning" included in the first user utterance "I want to learn at school".

The utterance generation device 10 uses the intention conversion table 121 to execute processing for converting the user's intention recognized from the user's utterance into the intention to be added to the system utterance. Note that, in the present embodiment, the intention after conversion is allowed to be the same as the intention of the user recognized from the original user utterance. In the example shown in FIG. 5, the user's intention (desire) contained in the first user utterance is converted into the system utterance's intention (desire), and the user's intention (question) contained in the second user utterance is converted into It shows that it has been converted into the intention (desire) of the system utterance.

The utterance generation device 10 generates system utterances including the extracted nouns and the converted intentions. In the example of FIG. 5, for the first user utterance "I want to learn at school", the content selected by the parrot ("learn") and the intention after conversion ("I want to") It shows that a system utterance such as "I want to learn" was generated. Also, for the second user utterance, the noun (content) extracted by the association is “introduction”, the co-occurrence expression “to take” including the predicate that co-occurs with this noun, and the converted It shows that a system utterance such as "I want to take a general discussion" containing the intention ("tai") of the was generated.

The utterance generation device 10 generates a system utterance each time it receives a user utterance from the terminal device 20 and returns it to the terminal device 20 each time. The terminal device 20 causes the display unit 24 to display the input user speech and the system speech received from the speech generation device 10 in chronological order.

Details of the process of extracting nouns associated with content included in user utterances will be described below.

FIG. 6 is a diagram showing an example of noun extraction by Word2Vec. The speech generation device 10 acquires nouns associated with the content from the Word2Vec data 122 based on the content recognized from the user's speech. Nouns associated with the content are extracted from the Word2Vec data 122 as neighboring vocabularies of the content. Neighboring vocabularies are vocabularies located within a predetermined distance range from the target content in the language space, and are determined based on the cosine distance between the content and each lexical. The example of FIG. 6 shows part of the neighborhood vocabulary acquired from the Word2Vec data 122 based on the content (in this example, "meaning") included in the user's utterance such as "meaning?". It can be seen that the acquired neighboring vocabulary includes nouns such as "imi", "wake", "intention", "true intention", "nuance", and so on, and verbs such as "understand". In this way, the neighborhood vocabulary obtained from the Word2Vec data 122 may include not only nouns but also verbs (or adjectives) and the like. Excluded from extraction targets.

In this embodiment, clustering is performed on the extracted nouns in order to alleviate restrictions on speech performance due to biased language resources. For example, if there are only linguistic resources "carbonated water - drink" and "water - drink, water - is delicious", the particle and verb "to drink" are unique for "carbonated water". will be connected. That is, if the noun associated with the user's utterance is "carbonated water", only an utterance such as "drink carbonated water" can be generated as the system utterance. On the other hand, if “carbonated water” and “water” can be clustered as the same cluster in terms of “drinking”, either “drink” or “is delicious” is selected for “carbonated water”. Since it becomes possible to do so, it is possible to generate system utterances such as "carbonated water is delicious" in addition to system utterances such as "drink carbonated water". In this way, clustering can expand the possibilities of particles and verbs (adjectives) that can be naturally connected as utterances.

FIG. 7 is an explanatory diagram for explaining the clustering method in this embodiment. In this embodiment, nouns obtained from the Word2Vec data 122 are clustered using an undirected graph and Kmeans++. As an example, "meaning", which is content recognized from user utterances, and nouns "imi", "why", "intention", "real intention", and "nuance" obtained from the Word2Vec data 122 as neighboring vocabularies, A process of clustering a total of 10 nouns of "truth", "context", "mechanism", and "reason" will be described. The utterance generation device 10 generates an undirected graph based on the acquired ten nouns and co-occurrence expressions including predicates co-occurring with each noun. Co-occurring expressions include, for example, verbs (or adjectives) that co-occur with each noun, and particles that connect nouns and verbs (or adjectives). Each noun is distributed and arranged on the undirected graph according to the number of co-occurrence expressions connected to each noun.

The utterance generation device 10 clusters 10 nouns placed on the undirected graph using Kmeans++. In clustering by Kmeans++, the center of gravity is placed as many times as the number of clusters to be divided, and the process of associating each data (noun in this example) with the closest center of gravity and the process of moving the center of gravity to the average value of the associated data are repeated. Determine clusters. In Kmeans++, the number of clusters must be set in advance, but the elbow method may be used to automatically set the number of clusters to be divided.

In the example shown in FIG. 7, the three nouns "meaning", "nuance", and "context" are in the first cluster, and the three nouns "truth", "true meaning", and "intention" are in the second cluster. , “wake”, “imi”, “mechanism”, and “reason” are clustered into the third cluster.

In this embodiment, for the sake of simplification, a configuration in which 10 nouns are clustered into 3 clusters has been described. It is not limited. For example, the number of nouns acquired from content may be about 50, and in this case, about 8 to 16 clusters may be generated by clustering.

The utterance generation device 10 refers to the clustering results to select nouns and co-occurrence expressions to be included in the system utterance.

FIG. 8 is an explanatory diagram illustrating selection processing of nouns and co-occurrence expressions. FIG. 8 schematically shows the flow of processing from extracting a plurality of nouns from content recognized based on user utterances to selecting nouns and co-occurrence expressions to be included in system utterances. When the speech generation device 10 recognizes content based on user speech, the speech generation device 10 extracts nouns associated with the content from the Word2Vec data 122 as described above. In addition, the utterance generation device 10 also reads co-occurrence expressions including predicates that co-occur with each of the extracted nouns. Co-occurring expressions include, for example, particles and verbs, or particles and adjectives. Note that the noun extracted from the Word2Vec data 122 may be the same (parrot) as the content recognized from the original user utterance. Also, the co-occurring expression read together with the noun may be the same (parrot) as the co-occurring expression included in the original user utterance. In FIG. 8, nouns and co-occurrence expressions extracted as parrots are shown with hatching.

When the nouns extracted from the Word2Vec data 122 and the co-occurrence expressions read together are obtained, the utterance generation device 10 generates an undirected graph as described above and performs clustering using the Kmeans++ method. The example of FIG. 8 shows that n clusters are generated, the first cluster contains three nouns, and the n-th cluster contains four nouns including Parrot. The co-occurring expressions associated with each cluster are a set of co-occurring expressions containing predicates that co-occur with each of the nouns contained within each cluster. That is, even if the co-occurrence expression of "carbonated water" does not include "ga delicious", the co-occurrence expression of "water" included in the same cluster as "carbonated water" includes "ga delicious". If it is included, this cluster will include "ga delicious".

When selecting nouns and co-occurring expressions to be included in the system utterance, the utterance generation device 10 first randomly selects one cluster from the generated clusters. Next, the utterance generation device 10 randomly selects one noun from the nouns included in the selected cluster, and randomly selects one co-occurrence expression from the set of co-occurrence expressions associated with the cluster. Choose an expression. In the example shown in FIG. 8, the n-th cluster is selected from the first to n-th clusters, one noun is selected from the selected n-th cluster, and the co-occurrence expression associated with the n-th cluster is selected. This indicates that one co-occurrence expression is selected from the set of .
When parroting back the user's utterance, nouns and co-occurrence expressions included as parrots are selected.

Through the above processing, the utterance generation device 10 can extract nouns and co-occurrence expressions to be included in system utterances.

Next, the utterance generation device 10 executes a process of converting the intention recognized from the user's utterance into the intention to be added to the system utterance. In this embodiment, intention conversion is performed using an intention conversion table 121 that defines the relationship between the intention in user utterances and the intention in system utterances.

9 and 10 are conceptual diagrams showing an example of the intention conversion table 121. FIG. The intention conversion table 121 is composed of a first table 121A shown in FIG. 9 and a second table 121B shown in FIG.

The first column of the first table 121A shown in FIG. 9 represents tags that classify the intention of the user's utterance, and the second to tenth columns represent candidates selected as the intention of the system utterance. . In the present embodiment, user utterances are [present/affirmative/regular], [past/affirmative/regular], [*1/negative/regular], [present/affirmative/interrogative], [past/affirmative] · Interrogative sentence], [*1 negative interrogative sentence], [present affirmative 7W2H], [past affirmative 7W2H], and [*1 negative 7W2H].

For example, if the user utterance is "I wanted to study", it is classified into [past/affirmative/normal sentence], so in the example of the first table 121A, the row of "0x00000002" is a candidate for selection. That is, the intention of the system utterance is "0x00000000" to "0xFFFFC000", and the utterance generation device 10 randomly selects the candidate of the system utterance with a preset weight.

When selecting, for example, "0x00000000" as a system utterance candidate, the utterance generation device 10 refers to the page number (page 10 in this example) registered in the first table 121A, and selects the second table 121B of the corresponding page. is read out from the storage unit 12 .

FIG. 10 shows the second table 121B of the tenth page read out from the storage unit 12. As shown in FIG. As an input to the second table 121B, only intention tags of user utterances are used. Here, since the above-described user utterance "I wanted to study" represents "subject/desire", the row of "subject/desire" in the second table 121B becomes a selection candidate. That is, there are six intention candidates in the system utterance: "Possible/Normal", "Possible/Desire", "Subject/Desire", "Subject/Desire", "Obligation", and "Proposal". The utterance generation device 10 randomly selects an intention candidate in the system utterance using preset weights. For example, when "suggestion" is selected by weighted randomness, the intention (subject/desire) of the user's utterance is converted into the intention (proposal) to be added to the system utterance.

In the column where the row of "Subject/Desire" and the column of "Proposal" intersect, there are six S's, which indicate whether or not noun association and verb (adjective) association are possible or not. is a controller for controlling If any S is selected, it is determined whether it associates a noun/not associates a noun and associates a verb (adjective)/does not associate a verb (adjective).

Also, when recognizing content from user utterances, content may be recognized in consideration of intention information of user utterances.

In this embodiment, the intention conversion table 121 is used to convert the intention recognized from the user utterance into the intention to be given to the system utterance. It may be selected, and the configuration may be such that the selected parameter is linked with other information and dynamically changed. Furthermore, the intention conversion table 121 may be machine-learned using dialogue data.

In addition, the intention recognized from the user's utterance may be converted into the intention to be added to the system utterance using parameters such as user's attribute, emotion, personality, character setting, conversation state, and conversation flow.

The procedure of processing executed by the speech generation device 10 will be described below.
FIG. 11 is a flowchart for explaining the procedure of processing executed by the utterance generation device 10. As shown in FIG. The control unit 11 of the speech generation device 10 receives an input of user speech transmitted from the terminal device 20 through the communication unit 13 (step S101). The control unit 11 determines whether or not the input user utterance is a standard sentence (step S102), and if it determines that it is a standard sentence (S102: YES), outputs the standard sentence as a system utterance (step S102). S103). A system utterance output from the control unit 11 is transmitted to the terminal device 20 through the communication unit 13 .

If it is determined that the sentence is not a fixed phrase (S102: NO), the control unit 11 recognizes the content and intention from the user's utterance (step S104). That is, the control unit 11 acquires main parts of speech by performing morphological analysis on user utterances, recognizes morphemes corresponding to nouns among the acquired parts of speech as content, and recognizes morphemes excluding nouns, verbs, and adjectives based on morphemes excluding nouns, verbs, and adjectives. Recognize user intent.

Next, the control unit 11 refers to the Word2Vec data 122 and extracts nouns associated with the recognized content (step S105). At this time, the control unit 11 reads co-occurrence expressions connected to each of the extracted nouns.

Next, the control unit 11 generates an undirected graph based on the extracted nouns and co-occurrence expressions (step S106), and clusters the nouns arranged on the generated undirected graph (step S107). For clustering, for example, a method based on Kmeans++ is used. Moreover, the number of clusters to be divided may be automatically set by using the elbow method as necessary.

Next, the control unit 11 randomly selects one cluster from the clusters generated in step S107 (step S108). Further, the control unit 11 randomly selects one noun from the nouns included in the selected cluster, and randomly selects one co-occurring expression from the set of co-occurring expressions associated with the cluster. Select (step S109).

Next, the control unit 11 refers to the intention conversion table 121 and converts the intention of the user's utterance into the intention to be added to the system utterance (step S110). In the present embodiment, the procedure for extracting nouns associated with content included in user utterances, selecting nouns and co-occurrence expressions to be included in system utterances, and then performing intention conversion is described. After the execution, extract nouns associated with the content included in the user utterance, select nouns and co-occurrence expressions to be included in the system utterance, and then perform intention conversion. You can go there on purpose.

Next, the control unit 11 generates system utterances using the nouns and co-occurrence expressions selected in step S109 and the intention of the system utterances converted in step S110 (step S111). At this time, the control unit 11 appropriately converts the conjugation of the verbs or adjectives so that the verbs or adjectives included in the co-occurrence expressions are appropriately connected to the intention of the system utterance. A system utterance generated by the control unit 11 is transmitted to the terminal device 20 through the communication unit 13 .

As described above, in the present embodiment, after recognizing the contents and intentions included in the user's utterance, a plurality of nouns associated with the contents and co-occurrence expressions including predicates co-occurring with these nouns are used. is used to generate system utterances, so dialogue can proceed without depending on a predetermined pattern such as a scenario. As a result, it is possible to provide the user with an interactive environment aimed at arousing mental changes such as healing, treatment, advice, and idea support.

In the present embodiment, a system utterance is generated based on a user utterance. information may be used to generate system utterances.

Furthermore, information on the user and information on the character that utters the system utterance may be stored in the storage unit 12, and the system utterance may be generated in consideration of these pieces of information.

(Embodiment 2)
In Embodiment 2, a configuration will be described in which mental vocabulary associated with content is evaluated and system utterances are generated based on the evaluation results.
Note that the overall configuration of the system and the internal configurations of the speech generation device 10 and the terminal device 20 are the same as those of the first embodiment, so description thereof will be omitted.

FIG. 12 is an explanatory diagram for explaining a system utterance generation method according to the second embodiment. The second embodiment aims to generate a system utterance such as "it's painful" in response to a user utterance such as "I have a hangover". When a user's utterance is input to the utterance generation device 10, it is decomposed into main parts of speech using morphological analysis to recognize the content, as in the first embodiment. For example, from a user utterance such as "I have a hangover", "hangover" is recognized as content.

Next, the utterance generation device 10 uses the Word2Vec data 122 to calculate the cosine distance between "Hangover" and mental vocabularies such as (happy, sad), (relieved, worried), and (fun, painful). , and the cosine distance between the two points exceeds a threshold value, Sens analysis is performed to determine which of the two points is closer to the intention. The example of FIG. 12 indicates that the cosine distance to "pain" is close to 1 and exceeds the threshold.

The utterance generation device 10 acquires the system utterance ("painful" in the example of FIG. 12) stored in association with the evaluation result "pain" from the utterance template, and generates "painful" as the system utterance. .

Generation of system utterances by Sens analysis in Embodiment 2 need not always be performed, and may be performed randomly.

The embodiments disclosed this time should be considered as examples in all respects and not as restrictive. The scope of the present invention is indicated by the scope of the claims rather than the meaning described above, and is intended to include all modifications within the scope and meaning equivalent to the scope of the claims.

REFERENCE SIGNS LIST 10 speech generation device 11 control unit 12 storage unit 120 speech generation program 121 intention conversion table 122 Word2Vec data 13 communication unit 14 display unit 15 operation unit 20 terminal device 21 control unit 22 storage unit 23 communication unit 24 display unit 25 operation unit

Claims (9)

An utterance generation device that generates a system utterance to be output in response to an input user utterance,
a recognition unit that recognizes a user's intention from the user's utterance;
extracting nouns included in the user utterance, and acquiring a plurality of associated words associated with the extracted nouns and a plurality of co-occurrence expressions including terms co-occurring with the nouns and the plurality of associated words, respectively; an acquisition unit that
a classification unit that classifies the noun and the plurality of associative words into a plurality of clusters based on the amount of co-occurrence expressions commonly used between the noun and the plurality of associative words;
A selection unit that selects one noun or associative word from one of the plurality of clusters and selects one co-occurrence expression that co-occurs with any noun or associative word included in the one cluster. and,
a converting unit that converts the user's intention recognized by the recognizing unit into an intention that should be added to the system utterance;
connecting the nouns or associative words selected by the selection unit and the co-occurrence expressions selected by the selection unit, and connecting the intention after conversion by the conversion unit to the co-occurrence expressions to generate a system utterance; and a speech generator. The speech generation device according to claim 1, wherein the selection unit randomly selects the one cluster, the one noun or associative word, and the one co-occurrence expression. 3. The speech generation device according to claim 1, wherein the acquisition unit acquires a plurality of associated words positioned within a predetermined distance range from the noun in language space. The speech generation device according to any one of claims 1 to 3, wherein the associative words are nouns, verbs, or adjectives. Equipped with an intention conversion table that defines the relationship between intentions recognized from user utterances and intentions to be given to system utterances,
The utterance according to any one of claims 1 to 4, wherein the conversion unit refers to the intention conversion table and converts an intention recognized from an input user utterance into an intention to be added to the system utterance. generator. the intention conversion table includes a plurality of candidates for intention to be added to the system utterance;
The speech generation device according to claim 5, wherein the conversion unit selects one intention to be added to the system speech from intention candidates included in the intention conversion table. 7. The system according to any one of claims 1 to 6, further comprising: when the user utterance is a fixed phrase, a fixed phrase output unit that outputs a system utterance including the fixed phrase instead of the system utterance generated by the utterance generation unit. A speech generator as described. An utterance generation method for generating a system utterance to be output in response to an input user utterance using a computer, comprising:
The computer is
recognizing the user's intention from the user utterance;
extracting nouns included in the user utterance;
Acquiring a plurality of associated words associated with the extracted noun and a plurality of co-occurrence expressions including terms co-occurring with the noun and the plurality of associated words,
classifying the noun and the plurality of associative words into a plurality of clusters based on the amount of co-occurrence expressions commonly used between the noun and the plurality of associative words;
Selecting one noun or associative word from one cluster of the plurality of clusters, and selecting one co-occurrence expression that co-occurs with any noun or associative word contained in the one cluster,
Converting the recognized user's intention to the intention to be attached to the system utterance,
An utterance generation method for generating a system utterance by connecting selected nouns or associated words and selected co-occurrence expressions, and connecting intention after conversion to the co-occurrence expressions. A computer program for causing a computer to execute processing for generating a system utterance to be output in response to an input user utterance,
to said computer;
recognizing the user's intention from the user utterance;
extracting nouns included in the user utterance;
Acquiring a plurality of associated words associated with the extracted noun and a plurality of co-occurrence expressions including terms co-occurring with the noun and the plurality of associated words,
classifying the noun and the plurality of associative words into a plurality of clusters based on the amount of co-occurrence expressions commonly used between the noun and the plurality of associative words;
Selecting one noun or associative word from one cluster of the plurality of clusters, and selecting one co-occurrence expression that co-occurs with any noun or associative word contained in the one cluster,
Converting the recognized user's intention to the intention to be attached to the system utterance,
A computer program for connecting a selected noun or associative word with a selected co-occurring expression, and connecting an intention after conversion to the co-occurring expression to generate a system utterance.

Images