JP6306447B2 - Terminal, program, and system for reproducing response sentence using a plurality of different dialogue control units simultaneously - Google Patents

Description

  The present invention relates to a technology for an interactive system that outputs a response sentence in response to an utterance sentence from a user.

  A dialogue system that realizes natural dialogue with human beings is becoming more popular, especially on terminals such as smartphones and tablets. As a dialog system, there is a technique in which only an appropriate vocabulary according to the content of a dialog with a user is input to a dialog control process as a recognition target vocabulary (see, for example, Patent Document 1). According to this technique, only a minimum necessary vocabulary is stored as a recognition target vocabulary. Unregistered vocabulary is extracted from the user's utterance and the vocabulary is additionally registered. As a result, natural conversation processing can be realized while reducing the memory capacity and speeding up the vocabulary search.

  In addition, there is a technique for mounting a dialog control unit in which a dialog domain is limited in a mobile terminal (for example, see Non-Patent Document 1). According to this technique, the intention estimation process for extracting the intention from the user's utterance and the dialog control process for determining the dialog response to the extracted intention can be executed in the portable terminal.

  Furthermore, there is a technique using a plurality of voice interactive devices (see, for example, Patent Document 2). According to this technique, when the language understanding of the user's spoken sentence fails, the first voice interactive apparatus transmits a language understanding failure to the second voice interactive apparatus that has processed the speech sentence immediately before the user. When the second spoken dialogue device receives the language understanding failure, the second spoken dialogue device transmits, to the first spoken dialogue device, the utterance content obtained by reverse-analyzing the language understanding rules in the already-understood understanding state. As described above, the first voice interactive device and the second voice interactive device can cooperate with each other to analyze the user's utterance and proceed with the dialogue.

  Furthermore, there is a technique using four types of dialog control systems, that is, an information providing type, an information acquiring type, a question answering type, and an information receiving type (for example, see Patent Document 3). According to this technique, the interactive type is classified according to the user's non-input time and whether the content of the input utterance is a request or a question. By using a plurality of dialog control systems of a type suitable for the classification, an appropriate response sentence is output to the user even in a relatively complicated and wide range of dialog domains, and the user satisfaction of the dialog is improved. Can do.

JP 2002-149645 A JP 2004-240225 A JP 2009-198614 A

KDDI R & D Laboratories, “Development of Dialogue Platform for Smartphones with Multi-Device Collaboration”, [online], [Search May 20, 2014], Internet <http://www.kddilabs.jp/press/2013/ 1010.html>

However, according to the techniques described in Patent Document 1 and Non-Patent Document 1, although the response time of the dialogue can be shortened, the number of vocabularies and tasks to be recognized is limited, and the reliability of the response content is low. I must.
On the other hand, according to the techniques described in Patent Literature 2 and Patent Literature 3, since a plurality of dialogue control devices are used, the dialogue response time has to be long.
Thus, according to the dialogue system, there is a trade-off relationship between the response time in the dialogue and the reliability of the response content. Regardless of whether the response time is long or the reliability of the response content is low, the user is stressed by dialogue.

  Accordingly, an object of the present invention is to provide a terminal, a program, and a system for reproducing a response sentence in consideration of a trade-off between a response time and a reliability of the response content in a dialog with a user. To do.

According to the present invention, in a terminal having a user interface capable of voice interaction,
A first dialog control means for outputting a first response sentence after inputting the user's utterance sentence;
Connection word storage means storing "connection word";
A second dialogue control means for outputting a second response sentence having a longer response time and higher reliability of the response content than the first dialogue control means after the user's utterance sentence is input;
An utterance sentence input means for inputting the user's utterance sentence to both the first dialog control means and the second dialog control means;
When the second response text is output from the second dialog control means during the playback time in which the first response text output from the first dialog control means is played back to the user by voice, Response sentence similarity calculating means for calculating the similarity between the first response sentence and the second response sentence;
Response sentence reproduction means for reproducing a connected word and starting reproduction of a second response sentence continuously immediately after the reproduction of the first response sentence is finished when the similarity is not more than a first threshold value It is characterized by having.

According to another embodiment of the terminal of the present invention,
The connected word storage means is divided into a plurality of predetermined ranges in which the degree of similarity is not more than a first threshold and is higher to lower, and a “connected word” is set for each predetermined range of the degree of similarity. Stored in association with each other,
It is also preferable that the response sentence reproduction means selects a connection word corresponding to the similarity using the connection word storage means.

According to another embodiment of the terminal of the present invention,
The connected word storage means is divided into a predetermined range of three levels from the highest to the lowest, with the similarity being equal to or less than the first threshold Th1, and the connected words are stored in association with each other as follows. Similarity> first threshold Th1: the second response sentence is not reproduced First threshold Th1 ≧ similarity> second threshold Th1: cumulative connection word second threshold Th2 ≧ similarity> third threshold Th3: Reverse connection word Third threshold Th3 ≧ Similarity: A conversion connection word is also preferable.

According to another embodiment of the terminal of the present invention,
The response sentence similarity calculation means is as follows: (1) Until the reproduction of the first response sentence by voice is completed (2) After the reproduction of the first response sentence by voice, the utterance sentence from the user is detected It is also preferable that

According to another embodiment of the terminal of the present invention,
The first dialog control means and the second dialog control means are preferably scenario-type or statistical-type dialog control functions.

According to another embodiment of the terminal of the present invention,
Each of the first dialogue control means and the second dialogue control means outputs a reliability together with a response sentence.
The response sentence similarity calculating means outputs the first response sentence output from the first dialog control means to the user from the plurality of second dialog control means during the playback time during which the user is playing back by voice. When the second response sentence is output, the similarity between the second response sentence output from the second dialogue control unit with the highest reliability of the response content and the first response sentence is calculated. It is also preferable.

According to another embodiment of the terminal of the present invention,
Confidence in the first dialogue control means and the second interaction control device, it is also preferable that is calculated based on the average conversation correct answer rate P and real-time control confidence score C.

According to another embodiment of the terminal of the present invention,
The response sentence similarity calculating means extracts a plurality of words by morphological analysis for the first response sentence and the second response sentence, and the part of speech between the first response sentence word and the second response sentence word. Alternatively, it is also preferable to calculate a vector whose meaning has been analyzed and to calculate the cosine similarity of these vectors.

According to the present invention, in a program for causing a computer mounted on a terminal having a user interface capable of voice interaction to function,
A first dialog control means for outputting a first response sentence after inputting the user's utterance sentence;
Connection word storage means storing "connection word";
A second dialogue control means for outputting a second response sentence having a longer response time and higher reliability of the response content than the first dialogue control means after the user's utterance sentence is input;
An utterance sentence input means for inputting the user's utterance sentence to both the first dialog control means and the second dialog control means;
When the second response text is output from the second dialog control means during the playback time in which the first response text output from the first dialog control means is played back to the user by voice, Response sentence similarity calculating means for calculating the similarity between the first response sentence and the second response sentence;
When the similarity is not more than the first threshold value, as response sentence reproduction means for reproducing the connected word and starting reproduction of the second response sentence immediately after the reproduction of the first response sentence is completed It is characterized by making a computer function.

According to the present invention, in a system in which a terminal having a user interface capable of voice interaction and a dialogue control server are connected via a network,
The terminal
After input of the utterance Yu chromatography The, the first dialogue control means for outputting a first answering sentence,
Connection word storage means storing "connection word";
Have
The server has second dialogue control means for outputting a second response sentence having a longer response time and higher reliability of the response content than the first dialogue control means after inputting the user's utterance sentence. ,
The terminal
An utterance sentence input means for inputting the user's utterance sentence to the first dialog control means and transmitting it to the second dialog control means of the server;
When the second response text is received from the second dialog control means during the playback time in which the first response text output from the first dialog control means is played back to the user by voice, Response sentence similarity calculating means for calculating the similarity between the first response sentence and the second response sentence;
Response sentence reproduction means for reproducing a connected word and starting reproduction of a second response sentence continuously immediately after the reproduction of the first response sentence is finished when the similarity is not more than a first threshold value It is characterized by having.

According to the present invention, in a system in which a terminal having a user interface capable of voice dialogue and a plurality of dialogue control servers are connected via a network,
A first dialog control server that returns a first response after receiving the user's utterance;
A second dialog control server that outputs a second response sentence having a longer response time and higher reliability of the response content than the first dialog control server after receiving the user's utterance sentence;
The terminal
Connection word storage means storing "connection word";
An utterance sentence input means for transmitting the user's utterance sentence to both the first dialog control server and the second dialog control server;
When the second response text is received from the second dialog control server during the playback time in which the first response text received from the first dialog control server is being played back by voice to the user, A response sentence similarity calculating means for calculating a similarity between the first response sentence and the second response sentence;
Response sentence reproduction means for reproducing a connected word and starting reproduction of a second response sentence continuously immediately after the reproduction of the first response sentence is finished when the similarity is not more than a first threshold value It is characterized by having.

  According to the terminal, the program, and the system of the present invention, it is possible to reproduce the response sentence in consideration of the trade-off between the response time and the reliability of the response content in the dialog with the user.

It is a function block diagram of the terminal in this invention. It is explanatory drawing showing the 1st specific reproduction | regeneration timing of a response sentence. It is explanatory drawing showing the 2nd concrete reproduction | regeneration timing of a response sentence. It is explanatory drawing showing the 3rd concrete reproduction | regeneration timing of a response sentence. It is a 1st system block diagram in this invention. It is a 2nd system block diagram in this invention.

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

  FIG. 1 is a functional configuration diagram of a terminal in the present invention.

  According to FIG. 1, the terminal 1 is like a smartphone or a tablet, and has a user interface that allows voice interaction. When the input device is a microphone, the input voice processing unit 101 converts a voice signal acquired by the microphone into an utterance sentence, and inputs the utterance sentence to the dialogue control unit. When the output device is a speaker, the output voice processing unit 102 converts the response sentence output from the dialogue control unit into a voice signal, and outputs the voice signal to the speaker.

  According to FIG. 1, the terminal 1 includes an utterance sentence input unit 111, a response sentence similarity calculation unit 112, a response sentence reproduction unit 113, a connected word storage unit 114, a first dialog control unit 121, Two dialog control units 122. These functional components are realized by executing a program that causes a computer installed in the terminal to function.

[First Dialog Control Unit 121 / Second Dialog Control Unit 122]
The dialogue control unit establishes a dialogue between the person and the system, similar to the dialogue between the people. The dialogue control unit has dialogue logic that recognizes a natural language utterance and outputs an appropriate response sentence. There are two types of dialogue control functions as follows, for example.
Scenario-type dialog control function Statistical-type dialog control function

  The scenario type dialogue control function proceeds with dialogue in a fixed scenario described in advance by hand. It is task-oriented and suitable for applications where the task (purpose) to be achieved is clear. The fixed scenario is designed in advance based on prior dialogue experience and system know-how, and the transition between dialogue nodes (or node groups) is fixed.

  The statistical dialog control function advances a natural dialog by accumulating a large number of dialog nodes and transitioning to the next dialog node having the maximum transition probability with respect to the current dialog node. Without a specific task, the transition to another dialog node is determined by the preceding and following transition probabilities. This transition probability is automatically and sequentially updated by machine learning.

Each dialog control function outputs a different response sentence even if the same utterance sentence is input according to the characteristic. The reliability Ps of the response content of the response sentence in the dialogue control unit is calculated as follows based on the average dialogue correct answer rate P and the real-time control reliability score C.
Ps (N) = P (N) ・ C (N)
P (N): Average dialogue correct answer rate of dialogue control N C (N): Real-time control reliability score of response sentence from dialogue control N

The average dialogue correct answer rate P (N) is a ratio calculated by inputting preliminarily prepared dialogue test data to the dialogue control unit and comparing the response sentence with a pre-prepared correct answer sentence.
Average dialogue correct answer rate P (N) = number of correct sentences / number of all response sentences

  The real-time control reliability score C (N) is a reliable measure calculated in real time for the response sentence output from the dialogue control function. For scenario-type dialog control, C (N) is the degree of intent understanding reliability for the user's dialog input. For example, when a statistical model of intent understanding is learned by Support Vector Machines (SVM), the reliability of the result of intent understanding is the distance from the separation plane of the SVM. In addition, regarding the statistical dialog control function, C (N) is the maximum transition probability of the selected response sentence based on a statistical model established from a large number of dialogue examples by machine learning techniques.

[Speech sentence input unit 111]
The utterance sentence input unit 111 inputs the user's utterance sentence output from the input voice processing unit 101 to both the first dialog control unit 121 and the second dialog control unit 122.

[Response sentence similarity calculation unit 112]
The response sentence similarity calculation unit 112 outputs the first response sentence output from the first dialog control unit 121 to the user during the “playback time” while playing back the first response sentence by voice. When the second response text is output from 122, the “similarity” between the first response text and the second response text is calculated. The response sentence similarity calculating unit 112 recognizes that the voice of the first response sentence is “during reproduction time” in cooperation with the response sentence reproduction unit 113.

Here, there are the following two patterns as “during playback time”.
(1) Until reproduction of the first response sentence by voice is completed (2) After reproduction of the first response sentence by voice, until an utterance sentence from the user is detected, that is, for the first response sentence Unless the user responds and speaks, a response sentence with the highest reliability is output.

  When the reliability Ps of the second dialog control unit 122 is lower than the reliability Ps of the first dialog control unit 121, the second dialog control unit during the playback time of the first response sentence. Even if the second response sentence is output from 122, the response sentence similarity calculation unit 112 ignores the similarity without calculating the similarity. This is because it is not necessary to intentionally reproduce the second response sentence having lower reliability than the first response sentence.

For the “similarity”, the response sentence similarity calculation unit 112 first extracts a plurality of words from the first response sentence and the second response sentence by morphological analysis. Then, the response sentence similarity calculation unit 112 calculates a vector obtained by analyzing the part of speech or the meaning between the words of the first response sentence and the words of the second response sentence, and the cosine similarity (Bag of words based). Of course, for each response sentence, a key word may be detected using a key word dictionary in which key words (nouns) having a high appearance frequency in the dialogue corpus are accumulated. The cosine distance is calculated according to the word extracted from each response sentence and its noun type. For example, it is calculated by the following conceptual formula.
Feature vector of word of first response sentence: D
Feature vector of word of second response sentence: E
Similarity between two sentences: sim (D, E)
sim (D, E) = cos θ = (D · E) / (| D || E |)
For the cosine distance, the similarity weight is set to 1 when the words are the same, and the similarity weight is set to 1 when the words are the same category. Of course, the similarity weight of the same word category can be set to a value between 0 and 1.

[Response sentence reproduction unit 113]
When the similarity is equal to or lower than the first threshold (when the similarity is low), the response sentence playback unit 113 starts playback of the second response sentence immediately after the end of playback of the first response sentence. (See FIG. 2 described later).
In addition, when the second response sentence is output from the second dialogue control unit 122 after the reproduction of the first response sentence by voice is completed, the response sentence reproducing unit 113 does not intentionally reproduce the second response sentence. (See FIG. 3 described later).
Furthermore, when the similarity is higher than the first threshold (when the similarity is high), the response sentence reproduction unit 113 does not reproduce the second response sentence (see FIG. 4 described later).

[Connected word storage unit 114]
The connection word storage unit 114 stores “connection words”. The response sentence reproduction unit 113 may reproduce the connected word and start reproduction of the second response sentence immediately after the reproduction of the first response sentence is completed. The connection word storage unit 114 stores the “connection word” associated with each predetermined range of the similarity.

  FIG. 2 is an explanatory diagram showing the first specific reproduction timing of the response sentence.

It is assumed that the first dialogue control unit 121 and the second dialogue control unit 122 have the following trade-off relationship.
First dialogue control unit 121: Although the response time of the dialogue is short, the reliability of the response content is low. Second dialogue control unit 122: Although the response time of the dialogue is long, the reliability of the response content is high. For example, it is determined as follows.
First dialog control unit 121: average response time = 0.7 seconds
Average dialog correct answer rate P (1) = 0.65
System reliability C (1) = 0.70
Reliability Ps (1) = C (1) × P (1) = 0.455
Second dialog control unit 122: average response time = 1.0 second
Average dialog correct answer rate P (2) = 0.85
System reliability C (2) = 0.70
Reliability Ps (2) = C (2) × P (2) = 0.595
That is, it is assumed that the first dialog control unit has a shorter response time of the dialog compared to the second dialog control unit, but the reliability of the response content is low.

(S21) It is assumed that the user utters "How is the weather tomorrow?" On the other hand, the input voice processing unit 101 outputs the utterance sentence to the utterance sentence input unit 111 by voice recognition. On the other hand, the utterance sentence input unit 111 inputs the following utterance sentence to both the first dialog control unit 121 and the second dialog control unit 122.
"What's the forecast for tomorrow?"

(S221) On the other hand, it is assumed that the first dialog control unit 121 with a short response time first outputs the following response sentence to the response similarity calculation unit 112.
“No plans tomorrow”
It is assumed that the first dialog control unit 111 erroneously determines that the schedule is asked while the weather is asked because the reliability of the response content is relatively low although the response time is short.
(S222) On the other hand, the response similarity calculation unit 112 outputs the response sentence as it is to the response sentence reproduction unit 113.
“No plans tomorrow”
(S223) The response sentence reproduction unit 113 sequentially utters the output audio processing unit 102 as follows, and responds to the user from the speaker.
“Ah, you”, “Ah,” “Ah,” “Ah,” “Ah,” “Ah,” “You”
(This voice takes 1.6 seconds, for example)

(S231) Next, it is assumed that the second dialogue control unit 122 having a long response time outputs the following response sentence to the response similarity calculation unit 112. Here, it is assumed that this response sentence is within 1.6 seconds after the response reproduction unit 113 starts to utter the first response sentence in S223.
"The weather will be sunny tomorrow"

(S232) The response sentence similarity calculation unit 112 reproduces the first response sentence “There is no tomorrow scheduled” output from the first dialogue control unit 121 to the user during the playback time. I recognize that there is. If the second response sentence “Tomorrow's weather is sunny” is output from the second dialogue control unit 122 during this playback time, the similarity between the first response sentence and the second response sentence is calculated. calculate.
The first response "There is no plan tomorrow"
Words detected as vectors: “Tomorrow”, “plan”, “not”
Second response sentence “Tomorrow's weather is sunny”
Words detected as vectors: “Tomorrow” “Weather” “Sunny”
In this case, the cosine distance between the first response sentence and the second response sentence is 0.33.

  According to FIG. 2, it is assumed that the cosine distance is equal to or smaller than the first threshold Th1, and it is determined that the second response sentence should be reproduced. At this time, it is assumed that the response sentence reproduction unit 113 starts reproduction of the second response sentence immediately after the reproduction of the first response sentence is completed.

(S233) In the connected word storage unit 114, the similarity is equal to or lower than the first threshold value, and is divided into a plurality of predetermined ranges from higher to lower. The response sentence reproducing unit 113 selects and reproduces the connected word corresponding to the similarity range of the connected word storage unit 114.

The connected word storage unit 114 has a similarity degree equal to or lower than the first threshold Th1 and is divided into predetermined ranges of three levels from higher to lower, and associates connected words as follows: Remember.
Similarity> first threshold Th1: the second response sentence is not reproduced First threshold Th1 ≧ similarity> second threshold Th1: a conjunctive word “cumulative”
"Progress"->"More""Beyond"
Second threshold Th2 ≧ similarity> third threshold Th3: concatenation word of “reverse connection”
"Reverse connection"->"But""But""I want to say""That'swhy"
Third threshold Th3 ≧ similarity: conjunctive word for “conversion”
“Conversion”-> “By the way” “Well” “It ’s a different thing”
For example, Th1 = 0.8, Th2 = 0.5, and Th3 = 0.3 may be set. The lower the similarity, the more connected words that change the story are selected.
According to FIG. 2, when the cosine distance between the first response sentence and the second response sentence is, for example, 0.33, the response sentence reproduction unit 113 selects the reverse connection word “but”. .

(S234) The response sentence playback unit 113 plays back as follows.
“A”, “Y”, “T”, “A”, “R”, “M”, “N” (1.6 seconds)
“Ke” re ”do” mo ”(0.5 seconds)

(S235) The response sentence similarity calculating unit 112 outputs the second response sentence “Tomorrow's weather is sunny” to the response sentence reproducing unit 113.
(S236) The response sentence playback unit 113 plays back as follows.
“A”, “Y”, “T”, “A”, “R”, “M”, “N” (1.6 seconds)
“Ke” re ”do” mo ”(0.5 seconds)
"A""" A """""""""""""""""""""""""" (1.9 seconds) "

  As is clear from FIG. 2, the first response sentence output from the first dialog control unit with a short response time and low reliability is reproduced, and during the reproduction time, the response time is long and the reliability is high. When the second response sentence is output from the second dialog control unit, the second response sentence is reproduced in succession to the first response sentence. Further, when the similarity between the first response sentence and the second response sentence is lower than the first threshold, a connection word corresponding to the similarity is used as the first response sentence and the second response sentence. By inserting and connecting between the two, it is possible to respond so as not to make the user feel unnatural as much as possible.

  FIG. 3 is an explanatory diagram showing a second specific reproduction timing of the response sentence.

  According to FIG. 3, the response sentence reproduction unit 113 receives the second response sentence “Tomorrow's weather is sunny” from the second dialogue control unit 122 after the reproduction of the first response sentence “No tomorrow is scheduled” by voice. ”Is output, the second response sentence is not reproduced. Normally, the user tries to react immediately after playback of the response text is interrupted. This is because if the first response sentence “There is no plan tomorrow” is interrupted and then the second response sentence “Tomorrow's weather is sunny” is played, then there is a high possibility that it will collide with the user's utterance. .

  FIG. 4 is an explanatory diagram showing a third specific reproduction timing of the response sentence.

  According to FIG. 4, when the similarity is higher than the first threshold (when the similarity is high), the response sentence reproducing unit 113 does not reproduce the second response sentence. The high degree of similarity is because response sentences having the same meaning are reproduced twice in succession.

  FIG. 5 is a first system configuration diagram in the present invention.

  According to FIG. 5, the first dialog control unit 121 is provided in the terminal 1, and the second dialog control unit 122 is provided in an external server via a network. Since the terminal 1 generally has a small amount of memory and low arithmetic processing capability, the response time of the dialogue is short, but only a lightweight dialogue control unit with low reliability of the response content can be mounted. On the other hand, since the external server has a large amount of memory and a high arithmetic processing capacity, it is possible to mount a dialogue control unit with a high reliability of response contents although the response time of the dialogue is long.

  FIG. 6 is a second system configuration diagram according to the present invention.

  According to FIG. 6, both the first dialog control unit 121 and the second dialog control unit 122 are provided in an external server via a network. Since the terminal 1 does not have a dialog control function, the response time is relatively long, but various dialog control functions of an external server can be used.

<3 or more dialog control units>
According to the above-described embodiment, the two dialog control units have been described. Of course, three or more dialog control units may be mounted. For each dialog control unit, there is a difference in the response time of the dialog and the reliability of the response content. Specifically, a plurality of different types of second dialogue control units 122 may be mounted on the basis of the first dialogue control unit 121.

For example, it is assumed that the reliability is set as follows for each dialogue control unit as follows.
First dialogue control unit 121: reliability Ps (1) = 0.455
The first response "There is no plan tomorrow"
21st dialogue control unit 122: reliability Ps (2) = 0.595
The 21st response "Tomorrow's weather is fine"
Twenty-second dialog control unit 122: reliability Ps (3) = 0.720
The 22nd response sentence "The chance of precipitation tomorrow is 20%"

  The response sentence similarity calculation unit 112 is connected to the 21st dialog control unit 122 during the playback time during which the first response sentence output from the first dialog control unit 121 is played back to the user by voice. When the second response sentence is output from the twenty-second dialog control unit 123, the second response sentence and the first response output from the twenty-second dialog control unit 123 having the highest reliability Ps of the response content. Calculate similarity between sentences. Here, it is assumed that the cosine distance between the first response sentence and the twenty-second response sentence is 0.29, for example.

On the other hand, the response sentence reproduction unit 113 selects the reverse connection word “by the way”. Finally, the response sentence reproduction unit 113 reproduces as follows.
“A”, “Y”, “T”, “A”, “R”, “M”, “N” (1.6 seconds)
"To" and "Kororo""(0.5 seconds)
“A” su “ko” u “su” i ”or“ ku ”ri“ tsu ”is“ ju ”“ pa ”-“ se ”and“ su ”(2.7 seconds) )

  As described above in detail, according to the terminal, the program, and the system of the present invention, in the dialogue with the user, the response sentence is considered in consideration of the trade-off between the response time and the reliability of the response content. Can be played. In particular, the response time is shortened with respect to the user's utterance sentence, and the response sentence having the highest reliability can be reproduced in the series of response sentences.

  The present invention is suitable for a voice dialogue system in which a “character agent” is displayed on a display of a terminal based on a user operation, and a dialogue between a user and an agent is advanced.

  Various changes, modifications, and omissions of the above-described various embodiments of the present invention can be easily made by those skilled in the art. The above description is merely an example, and is not intended to be restrictive. The invention is limited only as defined in the following claims and the equivalents thereto.

DESCRIPTION OF SYMBOLS 1 Terminal 111 Utterance sentence input part 112 Response sentence similarity calculation part 113 Response sentence reproduction | regeneration part 114 Conjunction memory | storage part 121 1st dialog control part 122 2nd dialog control part 21, 22 Server

Claims (11)

In a terminal having a user interface capable of voice interaction,
A first dialog control means for outputting a first response sentence after inputting the user's utterance sentence;
Connection word storage means storing "connection word";
A second dialogue control means for outputting a second response sentence having a longer response time and higher reliability of the response content than the first dialogue control means after the user's utterance sentence is input;
An utterance sentence input means for inputting the user's utterance sentence to both the first dialog control means and the second dialog control means;
When the second response text is output from the second dialog control means during the playback time in which the first response text output from the first dialog control means is played back to the user by voice, Response sentence similarity calculating means for calculating the similarity between the first response sentence and the second response sentence;
If the similarity is less than or equal to the first threshold and is not similar, response sentence playback that plays back the connected word and starts playback of the second response sentence immediately after the end of playback of the first response sentence And a terminal. The connection word storage means is divided into a plurality of predetermined ranges in which the similarity is equal to or less than a first threshold and is higher to lower. "Is stored in association with each other,
The terminal according to claim 1 , wherein the response sentence reproduction unit selects a connection word corresponding to the similarity using the connection word storage unit. The connection word storage means is divided into a predetermined range of three levels from the higher to the lower with the similarity being equal to or less than the first threshold Th1, and associating the connection words as follows: Similarity> First threshold Th1: Do not reproduce second response sentence First threshold Th1 ≧ similarity> Second threshold Th1: Cumulative connection word Second threshold Th2 ≧ Similarity> Third The terminal according to claim 2 , wherein the threshold value Th 3 is a reverse connection word, and the third threshold value Th 3 ≧ similarity is a conversion connection word. The response sentence similarity calculation means determines that during the reproduction time, (1) until reproduction of the first response sentence by voice is completed. (2) After reproduction of the first response sentence by voice, an utterance sentence from the user is received. The terminal according to any one of claims 1 to 3 , wherein the terminal is detected. The terminal according to any one of claims 1 to 4 , wherein the first dialog control means and the second dialog control means have a scenario type or statistical type dialog control function. Each of the first dialogue control means and the second dialogue control means outputs a reliability together with a response sentence.
A plurality of second dialogue control means are provided,
The response sentence similarity calculating means receives a first response sentence output from the first dialog control means from a plurality of second dialog control means during a playback time during which the user is playing back by voice. When the second response sentence is output, the similarity between the second response sentence output from the second dialogue control means with the highest reliability of the response content and the first response sentence is calculated. terminal according to claim 1, any one of 5, characterized in that it has. The confidence in the first dialogue control means and the second interaction control device, it is intended to be calculated on the basis of the average interaction correct answer rate P and real-time control confidence score C to claim 6, wherein The listed terminal. The response sentence similarity calculating means extracts a plurality of words by morphological analysis for the first response sentence and the second response sentence, and between the first response sentence word and the second response sentence word The terminal according to any one of claims 1 to 7 , wherein a vector obtained by analyzing part of speech or meaning is calculated, and a cosine similarity of these vectors is calculated.
In a program for causing a computer mounted on a terminal having a user interface capable of voice interaction to function,
A first dialog control means for outputting a first response sentence after inputting the user's utterance sentence;
Connection word storage means storing "connection word";
A second dialogue control means for outputting a second response sentence having a longer response time and higher reliability of the response content than the first dialogue control means after the user's utterance sentence is input;
An utterance sentence input means for inputting the user's utterance sentence to both the first dialog control means and the second dialog control means;
When the second response text is output from the second dialog control means during the playback time in which the first response text output from the first dialog control means is played back to the user by voice, Response sentence similarity calculating means for calculating the similarity between the first response sentence and the second response sentence;
If the similarity is less than or equal to the first threshold and is not similar, response sentence playback that plays back the connected word and starts playback of the second response sentence immediately after the end of playback of the first response sentence A program that causes a computer to function as means. In a system in which a terminal having a user interface capable of voice communication and a dialog control server are connected via a network,
The terminal
After input of the utterance Yu chromatography The, the first dialogue control means for outputting a first answering sentence,
Connection word storage means storing "connection words" , and
The server has second dialog control means for outputting a second response sentence having a longer response time and higher reliability of the response content than the first dialog control means after inputting the user's utterance sentence. And
The terminal
An utterance sentence input means for inputting the user's utterance sentence to the first dialog control means and transmitting to the second dialog control means of the server;
When the second response text is received from the second dialog control means during the playback time in which the first response text output from the first dialog control means is played back to the user by voice, Response sentence similarity calculating means for calculating the similarity between the first response sentence and the second response sentence;
If the similarity is less than or equal to the first threshold and is not similar, response sentence playback that plays back the connected word and starts playback of the second response sentence immediately after the end of playback of the first response sentence And a system. In a system in which a terminal having a user interface capable of voice interaction and a plurality of dialogue control servers are connected via a network,
A first dialog control server that returns a first response after receiving the user's utterance;
A second dialog control server that outputs a second response sentence having a longer response time and higher reliability of the response content than the first dialog control server after receiving the user's utterance sentence;
The terminal
Connection word storage means storing "connection word";
An utterance sentence input means for transmitting the user's utterance sentence to both the first dialog control server and the second dialog control server;
When the second response text is received from the second dialog control server during the playback time in which the first response text received from the first dialog control server is being played back by voice to the user, A response sentence similarity calculating means for calculating a similarity between the first response sentence and the second response sentence;
If the similarity is less than or equal to the first threshold and is not similar, response sentence playback that plays back the connected word and starts playback of the second response sentence immediately after the end of playback of the first response sentence And a system.

Images