JP5578571B2 - Multimodal dialogue program, system and method considering input / output device information - Google Patents

Description

  The present invention relates to a technique of a multimodal interaction system.

  The “multimodal dialogue system” refers to a technology that leads to an optimum result by advancing a scenario interactively with a user using not only input information by characters but also voice processing and image processing techniques. According to this technology, it is possible to optimally control the target device for each user through natural dialogue. “Multi-modal” means “multi-mode” and means that different media such as audio and video can be used in addition to characters.

  2. Description of the Related Art Conventionally, there is a technology of a dialogue device that realizes a natural dialogue by shortening the waiting time for dialogue with a human (see, for example, Patent Document 1). According to this technique, the terminal is a PDA (Personal Digital Assistance) or a personal computer, and includes a microphone unit that inputs voice from the user and a video camera unit that captures the movement of the user's body. Even if the voice of the user is not detected by the microphone unit, the end (response) of the dialogue can be predicted by detecting the body movement taken by the video camera unit, for example, “coincidence”. By detecting the end of the dialogue in the user, it is possible to proceed to the next dialogue scenario, and to shorten the waiting time for the response to the human.

  There is also a technique in which a multimodal dialogue system is applied to online shopping (see, for example, Non-Patent Document 1). According to this technology, it is composed of a document server module for storing documents used for shopping, a dialog control module for managing the progress of dialog, and a front-end module for managing input / output terminals. The document stored in the document server module includes, for example, a dialogue scenario, customer / product data (XML (Extensible Markup Language) format), display style, and the like. According to this system, a personal computer is assumed as a terminal on which a front end module is mounted. A shopping dialogue can be performed using a voice input unit, a pointing input unit, a synthesized voice output unit, an anthropomorphic agent output unit, and the like in a personal computer.

  Furthermore, there is a technology of interactive psychological learning using a partially-observed Markov decision process POMDP (Partially-Observed Markov Decision Process) (see, for example, Patent Document 2). The partial observation Markov determination process (hereinafter referred to as “POMDP”) refers to a process of determining the action a in the state s so that the reward Vt is maximized using the reward expectation value r (s, a). Hereinafter, POMDP will be described in detail.

  FIG. 1 is a system configuration diagram in which POMDP is applied to an interactive system.

  According to FIG. 1, the dialogue server module program 1 and the terminal module program 2 are configured. For example, the interactive server module program 1 is executed by a server, and the terminal module program 2 is executed by a terminal via a network.

  The dialogue server module program 1 includes a reward expected value accumulation unit 11, a POMDP unit 12, a dialogue scenario accumulation unit 13, an answer information reception unit 14, and a dialogue scenario transmission unit 15.

  The expected reward value accumulation unit 11 accumulates an expected reward value r (s, a) obtained when the action a is executed in the state s. The “reward expected value” represents the “expected” of the reward obtained when the action a is executed in the state s. The expected reward value r (s, a) is generally set by an experience value.

  The dialogue scenario storage unit 13 stores a dialogue scenario to be transmitted to the terminal for each action a. A “dialogue scenario” refers to a data structure expressed in a tree shape composed of questions and answers. For example, depending on the answers “red”, “blue”, “white”, etc. to the question “what color model do you like?”, The process moves to the next question (or information provision) that differs for each color.

  The partial observation Markov decision process POMDP unit 12 uses the expected reward value r (s, a) to determine the action a in the state s so that the reward Vt is maximized. According to FIG. 1, the POMDP unit 12 includes an action sequence accumulation unit 121, a probability distribution calculation unit 122, a reward calculation unit 123, and an action determination unit 124.

The action sequence accumulation unit 121 accumulates the following probabilities.
・ Observation expectation probability P (o ′ | s ′, a):
Probability of receiving observation o 'for state s' after execution of action a-State transition probability P (s' | s, a):
Probability of transition from state s to state s ′ after executing action a Here, the definition is as follows.
S: finite set of “states s”, {s}
O: a set of finite “observations o”, {o}
A: A finite set of “actions a”, {a}

  The observation expectation probability P (o ′ | s ′, a) and the state transition probability P (s ′ | s, a) are determined based on the action sequence data (learning data) accumulated by a large amount of dialog simulation experiments. Reinforcement learning (for example, Q-learning or TD-learning) in the learning field

The “state s” is divided into the following four.
“S0”: Initial state “s1”: No input of answer information from the user “s2”: Input of answer information from the user within the range of the dialogue scenario “s3”: No input of answer from the user is required , Automatic transition to the next action “s4”: Other states

“Action a” is divided into three as follows.
“Ain”: Accepting input “aout”: Providing information “await”: Waiting Here, each action a is associated with each dialogue scenario stored in the dialogue scenario storage unit 13. For example, when “selecting the model of the mobile phone” is a task, there is a dialogue scenario of action a as follows.
"Ain": Accept input
"Enter cell phone color", "Enter age", ...
“Aout”: Provides information
"Introduction of interactive device", "Question about color of mobile phone", ...

η：はｂtの総和を１にするための正規化定数 The probability distribution calculation unit 122 calculates the probability distribution bt (s) at time t using the observation expected probability P (o ′ | s ′, a) and the state transition probability P (s ′ | s, a). . The probability distribution represents the probability of occurrence for each state s (random variable). In order to determine a policy function that outputs an optimal action sequence according to the conversation state, the state probability distribution bt (s ′) at time t is the state probability distribution bt−1 (s) at the previous time. And is calculated by the following equation.

η: is a normalization constant to set the sum of bt to 1

γ：１より小さい正定数 The reward calculation unit 123 calculates the reward Vt using the expected reward value r (s, a) and the probability distribution bt (s). When the action sequence aτ + t is given, the reward Vt that can be acquired after time t is expressed by the following equation.

γ: positive constant less than 1

  The action determination unit 124 determines the action a in the state s using the expected reward value r (s, a) so that the reward Vt is maximized. Here, a dialogue scenario corresponding to the determined action a is acquired from the dialogue scenario storage unit 13. The acquired dialogue scenario is presented to the user.

  The answer information receiving unit 14 receives answer information based on the input device from the terminal module program 2. The received answer information is input to the action determination unit 124 of the POMDP unit 12.

  The dialogue scenario transmission unit 15 transmits the dialogue scenario acquired by the behavior determination unit 124 of the POMDP unit 12 to the terminal module program 2.

  The terminal module program 2 has an input device and an output device for the user. When the input device is a microphone, the input voice processing unit converts the voice signal acquired by the microphone into answer information and transmits the answer information to the dialog module program 1. When the output device is a speaker, the output sound processing unit converts the answer information received from the dialogue module program 1 into a sound signal, and the sound signal is output from the speaker.

JP 2003-228449 A JP 2010-129033 A

Takenori Kobayashi, Yusaku Nakamura, Koichi Katsuta, Hirofumi Yamada, Kei Kobayashi, Tsuneo Nitta, “Development of Online Shopping System Using MMI Description Language XISL”, [online], [Search February 6, 2011], Internet <URL: http://ci.nii.ac.jp/naid/110002917299>

  In recent years, terminals that can be operated by a user are not limited to personal computers, but have spread to various devices such as car navigation terminals, digital signage terminals, digital photo frames, mobile phones, and smartphones. It is expected that convenience of operation for various devices is improved by providing an interactive interface to these various terminals.

  However, the dialogue scenario is generally described corresponding to an input / output device mounted on the terminal. According to Patent Document 1 and Non-Patent Document 1 described above, it is assumed that the terminal is preliminarily equipped with a microphone unit and a video camera unit, and the dialogue scenario also uses the microphone unit and the video camera unit. Pre-designed. Then, a car navigation terminal, a digital signage terminal, or the like that is not mounted with a microphone unit and a video camera unit in advance cannot be applied to this interactive system. That is, the dialogue scenario must be designed or adjusted for each terminal having different input / output devices.

  On the other hand, in recent years, a plurality of various input / output devices are mounted for each terminal. Various input / output devices such as cameras, touch panels, displays (touch panel displays, multi-touch screens, three-dimensional displays), microphones, speakers (earphones), acceleration sensors, geomagnetic sensors, temperature sensors, barometric sensors, vibrators, etc. There is something. Even in the same type of input / output device, different characteristics (for example, display size) may be mounted depending on the terminal type. Designing or adjusting a dialogue scenario for each terminal equipped with such various input / output devices results in high costs.

  Therefore, the present invention provides a multimodal dialogue program, system and method capable of causing a different dialogue scenario to progress for each terminal equipped with various input / output devices by designing only one dialogue scenario. Objective.

According to the present invention, there is provided a multimodal dialogue server program that allows a computer to function to advance a scenario interactively with the user between terminals based on a user operation,
Dialog scenario storage means for storing a dialog scenario to be transmitted to the terminal for each action a;
Remuneration expected value accumulation means for accumulating an expected reward value r (s0, a0) obtained when the action a is executed in the state s;
Device action probability accumulation means for accumulating device action probability r (sd0, ad0) in each device action ain / aout available in state s for each of a plurality of different input / output devices;
Device type acquisition means for acquiring the presence or absence of one or more device types installed in the terminal;
Device available reward value calculating means for calculating the device available reward value r (sd, ad) by multiplying the presence / absence of the device type of the terminal and the device action probability r (sd0, ad0) for each input / output device; ,
Reward expected value calculating means for calculating a reward expected value r (s, a) obtained by weighting a device available reward value r (sd, ad) with respect to a reward expected value r (s0, a0) of the reward expected value accumulating means; ,
A partial observation Markov decision POMDP (Partially-Observed Markov Decision Process) processing means for determining an action a in the state s so that the reward Vt is maximized using the expected reward value r (s, a);
A dialog scenario based on the determined action a is acquired from the dialog scenario storage means, and the computer is caused to function as dialog scenario transmission means for transmitting the dialog scenario to a terminal.

According to another embodiment of the multimodal dialog server program of the present invention,
The partial observation Markov decision processing means
Observation expectation probability P (o ′ | s ′, a) for receiving observation o ′ for state s ′ after execution of action a, and state transition probability P for transitioning from state s to state s ′ after execution of action a action sequence accumulating means for accumulating (s ′ | s, a);
A probability distribution calculating means for calculating a probability distribution bt (s) at time t using the observation expected probability P (o ′ | s ′, a) and the state transition probability P (s ′ | s, a);
A reward calculating means for calculating a reward Vt using an expected reward value r (s, a) and a probability distribution bt (s);
It may include action determining means for determining the action a in the state s so that the reward Vt is maximized using the expected reward value r (s, a).

According to another embodiment of the multimodal dialog server program of the present invention,
Regarding the device action probability accumulation means, the device action probability r (sd0, ad0) indicates whether or not the device can be used for the dialogue scenario, ranging from 1 (should be used) to 0 to -1 (not to be used). The probability expressed by
Regarding the device available reward value calculation means, the presence / absence of one or more device types mounted on the terminal may be represented by 1 (valid) / 0 (invalid).

According to the present invention, a dialog server that causes the above-described multimodal dialog server program to function on a computer;
A user-operable terminal having one or more input / output devices;
A multimodal dialogue system comprising:
The terminal
Device information transmission means for transmitting the validity / invalidity for each device type to the dialogue server
An output device control means for presenting a dialog scenario received from the dialog server to the user by an output device;
It has an input device control means which transmits the reply information with respect to the input device input by user operation to a dialogue server.

According to another embodiment of the multimodal interaction system of the present invention,
The system further includes a user observation information analysis server,
The terminal further includes an observation feature value transmission means for transmitting the observation feature value detected for each input device to the user observation analysis server,
The user observation information analysis server may derive answer information for the user based on the observation feature quantity received from the terminal, and transmit the answer information to the dialogue server.

According to the present invention, there is provided a multimodal dialogue method in a system having a terminal based on a user operation and a dialogue server that interactively progresses a scenario with the user,
The conversation server
A dialogue scenario storage unit that stores dialogue scenarios to be transmitted to the terminal for each action a;
An expected reward value accumulating unit that accumulates an expected reward value r (s0, a0) obtained when the action a is executed in the state s;
A device action probability accumulation unit that accumulates device action probabilities r (sd0, ad0) in each device action ain / aout available in the state s for each of a plurality of different input / output devices;
The conversation server
A first step of acquiring presence / absence of one or more device types installed in the terminal;
A second step of multiplying the input / output device by the presence / absence of the device type of the terminal and the device action probability r (sd0, ad0) to calculate a device available reward value r (sd, ad);
A third step of calculating an expected reward value r (s, a) obtained by weighting the device available reward value r (sd, ad) with respect to the expected reward value r (s0, a0) of the expected reward value accumulation unit;
A fourth step of the partial observation Markov decision process POMDP for determining the action a in the state s so that the reward Vt is maximized using the reward expectation value r (s, a);
And a fifth step of acquiring a dialogue scenario based on the determined action a from the dialogue scenario storage unit and transmitting the dialogue scenario to the terminal.

According to another embodiment of the multimodal interaction method of the present invention,
The dialogue server transitions from the state s to the state s ′ after executing the action a, and the observation expected probability P (o ′ | s ′, a) that receives the observation o ′ for the state s ′ after the action a is executed. state transition probability P (s' | s, a ) and further comprising the accumulated action sequence storing section,
For the fourth step of the interactive server,
Calculating a probability distribution bt (s) at time t using the observation expected probability P (o ′ | s ′, a) and the state transition probability P (s ′ | s, a);
Calculating the reward Vt using the expected reward value r (s, a) and the probability distribution bt (s);
A step of determining an action a in the state s so that the reward Vt is maximized using the expected reward value r (s, a).

According to another embodiment of the multimodal interaction method of the present invention,
Regarding the device action probability storage unit of the dialogue server , the device action probability r (sd0, ad0) indicates whether the device can be used for the dialogue scenario from 1 (should be used) to 0-1 (not to be used). ) In the range of
Regarding the second step of the dialogue server, the presence / absence of one or more device types mounted on the terminal may be represented by 1 (valid) / 0 (invalid).

  According to the multimodal dialogue program, system, and method of the present invention, different dialogue scenarios can be advanced for each terminal equipped with various input / output devices by designing only one dialogue scenario.

It is a system block diagram which applied the partial observation Markov decision process POMDP to the dialogue system. It is a block diagram of the multimodal dialog system in this invention. It is a functional block diagram of the multimodal dialogue program in the present invention. It is explanatory drawing showing the flow of the dialogue scenario in this invention. It is a system configuration diagram including a user trend analysis server.

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

  FIG. 2 is a configuration diagram of a multimodal dialogue system according to the present invention.

  According to FIG. 2, the dialogue server 1 and various types of terminals 2 are connected via a network. The terminal 2 includes a digital signage terminal, a digital photo frame, a mobile phone, a smartphone, and the like, and each terminal is equipped with various input / output devices such as a camera, a speaker, a display, and a sensor. The input / output device of the terminal 2 is controlled based on, for example, FLASH Player (registered trademark) or OpenGL (registered trademark) plugged into the web browser. FLASH Player by Adobe Systems, Inc. is software that reproduces Web content that combines voice, video, and vector graphics animation, and provides an application execution environment that allows user input operations. OpenGL by Silicon Graphics provides a programming interface for graphics processing. For example, the user can interact with an avatar displayed by FLASH or OpenGL.

  Note that, according to FIG. 2, the dialogue server 1 may acquire content based on the dialogue scenario to be transmitted to the terminal 2 from the content server 4. Specifically, the content server 4 stores content such as a still image, a moving image, and synthesized speech based on the dialogue scenario.

  Moreover, according to FIG. 2, the user trend analysis server 3 is further connected to the network. The user trend analysis server 3 receives the observed feature quantity in the input device from the terminal 2 and analyzes the feature to generate response information for the dialogue server 1. The answer information is transmitted to the dialogue server 1.

  According to FIG. 2, the dialogue server module program is executed by the computer mounted on the dialogue server 1. Further, the terminal module program 2 is executed by a computer mounted on the terminal 2. Of course, both module programs may be executed integrally by one computer of a single device. According to the present invention, the target device on which the server module program and the terminal module program are to be executed is not specified.

  FIG. 3 is a functional configuration diagram of the multimodal dialogue program in the present invention.

  According to FIG. 3, the dialogue server module program 1 in the multimodal dialogue program causes the computer to function in such a manner as to proceed with a scenario interactively with the user with the terminal 2 based on the user operation. According to FIG. 3, the dialog server module program 1 is similar to FIG. 1 in that the reward expected value storage unit 11, the partial observation Markov decision process POMDP unit 12, the dialog scenario storage unit 13, and the answer information receiving unit 14 And a dialogue scenario transmission unit 15. The processing contents of these functional components are basically the same as those in FIG. 1 described above as the prior art.

  The dialog server module program 1 of the present invention further includes a device action probability accumulation unit 16, a device type acquisition unit 17, a device available reward value calculation unit 18, and a reward expected value calculation unit 19.

[Device action probability accumulation unit 16]
The device action probability accumulation unit 16 accumulates the device action probability r (sd0, ad0) in each device action ain / aout available in the state s for each of a plurality of different input / output devices. The device action probability r (sd0, ad0) is a probability representing whether or not the device can be used for the dialogue scenario in a range of 1 (which should be used) to 0 to -1 (which should not be used).

A set A of device actions a is A {ain, aout, await} and is expressed as follows, for example.
ain1: “Input request by camera”
ain2: "Input request via touch panel"
ain3: “Input request by microphone”
aout1: “Providing information through speakers” (voice production)
aout2: “Providing information on the display” (displaying characters by speech balloons)
aout3: “Providing related information on the display” (displaying related images)

Further, for example, assume the following input / output device. Of course, whether or not the following input / output devices are installed differs depending on the terminal based on the user operation.
[Input device]
SM: Microphone status
SC: Camera status
ST: Touch panel status [Output device]
SS: Speaker status
SD: Display status

A table of specific device action probabilities r (sd0, ad0) is expressed as follows, for example.

  For example, since the action “ain1” requires a camera input action, the action probability value of the device “camera SC” is “1” (which should be used). Here, the action probability value of the device “microphone SM” is “−1” (not to be used), and the action probability value of the device “touch panel ST” is also “−1” (not to be used).

  Also, according to FIG. 3, for example, the action “aout1” requires an action of sound output, and therefore the action probability value of the device “speaker SS” is “1” (should be used). Here, the action probability value of the device “display SD” is “0.5”. This means that the probability of using a display as the output device is somewhat higher even in the case of the voice output action “aout1” (0.5 is 0.5 according to the range from 1 to −1). Equivalent to the top 1/4). The device action probability table value is set by an experience value or a statistical value based on a large amount of user evaluation data.

[Device Type Acquisition Unit 17]
The device type acquisition unit 17 acquires the presence / absence of one or more device types mounted on the terminal. For example, when the terminal is a photo frame, the terminal transmits the following device type information to the dialogue server 1.
Sd_t = “SM, SC, ST, SS, SD” = “0,0,1,1,1”
0: Invalid 1: Valid According to this device type information, the terminal is equipped with a touch panel, a speaker, and a display, but is not equipped with a microphone and a camera. In addition, the terminal transmits device type information to the dialogue server 1 in real time when the presence / absence of the input / output device changes.

ｒ(ｓd,ａd)は、各テーブルの値を加算したものである。
ｒ(ｓd,ａd)＝
（ｒ(SM,ainα）＋ｒ(SC,ainα）＋ｒ(ST,ainα）＋
ｒ(SS,aoutβ)＋ｒ(SD,aoutβ)) [Device Usable Reward Value Calculation Unit 18]
The device available reward value calculation unit 18 multiplies the presence / absence of the device type of the terminal by the device action probability r (sd0, ad0) for each input / output device, and obtains the device available reward value r (sd, ad). calculate. For example, when the device type information Sd_t = [SM, SC, ST, SS, SD] = [0, 0, 1, 1, 1] is multiplied by the device action probability r (sd0, ad0) in Table 1, A table of probabilities of r (SM, ain α), r (SS, aout β), etc. is calculated. α and β are subscripts of ain and aout, respectively.

r (sd, ad) is the sum of the values in each table.
r (sd, ad) =
(R (SM, ainα) + r (SC, ainα) + r (ST, ainα) +
r (SS, aoutβ) + r (SD, aoutβ))

[Reward Expectation Value Calculation Unit 19]
The expected reward value calculation unit 19 weights the expected reward value r (s0, a0) obtained when the action a is executed on the state s with the device available reward value r (sd, ad). r (s, a) is calculated. This can lead the existing expected reward value r (s0, a0) to the most appropriate action a by the device available reward value r (sd, ad).

s0及びa0は、対話状態と対話装置行動に関するパラメータである。
sd及びadは、デバイス可用報酬値と対話装置動作タイプに関するパラメータである。 Thus, the partial observation Markov decision process POMDP unit 12 determines the action a in the state s so that the reward Vt is maximized by using the expected reward value r (s, a). The POMDP unit 12 is defined as follows for the reward r for the action a in Equation 2 described above.

s0 and a0 are parameters relating to the dialog state and dialog device behavior.
sd and ad are parameters related to the device available reward value and the interaction device operation type.

  FIG. 4 is an explanatory diagram showing the flow of a dialogue scenario in the present invention.

  According to FIG. 4, when the terminal 2 is a photo frame, the photo frame has a touch panel ST as an input device and a speaker SS and a display SD as output devices.

[Time t = 0] (State S0)
For the initial state, the operation a that maximizes r (s0, A) is determined. At this time, it is assumed that the following operation a is determined.
Action a = “aout”
Device action = (aout1, aout3)
The dialogue scenario here is as follows, for example.
“Aout1”: From the speaker, say “This is a recommended service for mobile phones”.
“Aout3”: As a related information, a balloon of the latest model or a character of campaign information is displayed from the mouth of the character on the display.

[Time t = 1]
Next, the state is shifted to the state s3 by the above-described Expression 2 and Expression 3 (POMDP processing). Then, an operation a (device action) that maximizes the reward Vt is determined. For example, assume that the following is determined.
Action a (device action) = (aout1, ain2)
“Aout1”: Say “What color do you like” from the speaker.
“Ain2”: A color selection button is displayed on the touch panel to accept a user's touch.

[Time t = 2]
In contrast, it is assumed that the user touches the “red” button on the touch panel. As a result, it is assumed that the response information red @ touch is obtained. Next, the state s2 is transferred by the above-described Expression 2 and Expression 3 (POMDP processing). Then, an operation a (device action) that maximizes the reward Vt is determined. For example, assume that the following is determined.
Action a (device action) = (aout1, aout3)
“Aout1”: From the speaker, “Recommended red cell phones are S005 and CA006”.
“Aout3”: Display related information such as function information and word-of-mouth information of S005 and CA006 on the display.

  Here, when the time t = 2, it is assumed that the user has muted the speaker of the terminal 2 (or the speaker may be broken). At this time, the terminal 2 transmits device type information indicating that the speaker SS has become invalid (0) to the dialog server 1. According to the present invention, the terminal detects valid / invalid device type information in real time, and the information is immediately transmitted to the dialogue server 1.

At this time, it is changed as follows.
Sd_t = “SM, SC, ST, SS, SD” = “0,0,1,0,1”
As a result, the device available reward value r (sd, ad) is also changed as follows.

As a result, the operation a (device action) that maximizes the reward Vt is different. For example, assume that the following is determined.
Action a (device action) = (aout2, aout3)
“Aout2”: “Recommended red cell phone is S005 and CA006” is displayed in a balloon character on the display character's mouth.
“Aout3”: Display related information such as function information and word-of-mouth information of S005 and CA006 on the display.
As a result, a dialogue scenario can be advanced using a display instead of a speaker.

  FIG. 5 is a system configuration diagram including a user trend analysis server.

  According to FIG. 5, the terminal 2 acquires an observation feature amount from an input device (for example, a camera, a microphone, a touch panel), and transmits the observation feature amount to the user trend analysis server 3. Examples of observed feature quantities include the following.

"User's face image observation feature" O_face: Observed feature value obtained from camera Face face information detected Face trajectory of face angle Number of user's face matching Low-level Gabor feature extracted directly from face image (still image) Face image feature value (Haar-Like feature value of face image)
“User's voice observation feature” O_voice: Observation feature obtained from microphone Audio power / pitch average / standard deviation / maximum / maximum-minimum difference Voice feature (voice power information, MFCC information, pitch information, etc.)
“User touch operation observation feature” O_touch: Feature obtained from touch panel Touched panel coordinates Touched event information (link information, etc.)

  The terminal 2 transmits the data acquired from the input device to the dialogue server 1 and transmits the observed feature quantity detected by the input device to the user trend analysis server 3. The user trend analysis server 3 transmits response information (recognition result) based on the observation feature quantity to the dialogue server 1. Thereby, the dialogue server 1 can advance the dialogue scenario in consideration of the answer information.

  As described above in detail, according to the multimodal dialogue program, system and method of the present invention, it is possible to create different dialogue scenarios for each terminal equipped with various input / output devices by designing only one dialogue scenario. Can be advanced. In particular, the dialogue scenario can be advanced in accordance with the configuration of the input / output device mounted on the terminal.

  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 Multimodal dialogue program, dialogue server 11 Remuneration expectation value accumulation | storage part 12 Partial observation Markov decision process POMDP part 121 Behavior series accumulation part 122 Probability distribution calculation part 123 Compensation calculation part 124 Action decision part 13 Dialogue scenario accumulation part 14 Answer information receiving part DESCRIPTION OF SYMBOLS 15 Dialogue scenario transmission part 16 Device action probability storage part 17 Device type acquisition part 18 Device usable reward value calculation part 19 Reward expected value calculation part 2 Terminal 3 User trend analysis server 4 Content server

Claims (8)

A multi-modal dialog server program that allows a computer to function to proceed with a scenario interactively with a user based on a user operation,
Dialog scenario storage means for storing a dialog scenario to be transmitted to the terminal for each action a;
Remuneration expected value accumulation means for accumulating an expected reward value r (s0, a0) obtained when the action a is executed in the state s;
Device action probability accumulation means for accumulating device action probability r (sd0, ad0) in each device action ain / aout available in state s for each of a plurality of different input / output devices;
Device type acquisition means for acquiring presence or absence of one or more device types mounted on the terminal;
Device available reward value calculating means for calculating device available reward value r (sd, ad) by multiplying the input / output device by the presence / absence of the device type of the terminal and the device action probability r (sd0, ad0). When,
Expected reward value for calculating an expected reward value r (s, a) obtained by weighting the device available reward value r (sd, ad) with respect to the expected reward value r (s0, a0) of the expected reward value accumulation unit. A calculation means;
Partial observation Markov decision POMDP (Partially-Observed Markov Decision Process) processing means for determining the action a in the state s so that the reward Vt is maximized using the expected reward value r (s, a);
A multimodal dialog server program for causing a computer to function as dialog scenario transmission means for acquiring a dialog scenario based on the determined action a from the dialog scenario storage means and transmitting the dialog scenario to the terminal. The partial observation Markov decision processing means includes:
Observation expectation probability P (o ′ | s ′, a) for receiving observation o ′ for state s ′ after execution of action a, and state transition probability P for transitioning from state s to state s ′ after execution of action a action sequence accumulating means for accumulating (s ′ | s, a);
A probability distribution calculating means for calculating a probability distribution bt (s) at time t using the observed expected probability P (o ′ | s ′, a) and the state transition probability P (s ′ | s, a); ,
A reward calculating means for calculating a reward Vt using the expected reward value r (s, a) and the probability distribution bt (s);
2. The multimodal according to claim 1, further comprising behavior determining means for determining the behavior a in the state s so that the reward Vt is maximized using the reward expectation value r (s, a). Interactive server program. Regarding the device action probability accumulation means, the device action probability r (sd0, ad0) indicates whether the device can be used for the dialogue scenario, from 1 (use) to 0-1 (not use). The probability expressed in the range of
The device usable reward value calculating means is represented by 1 (valid) / 0 (invalid) indicating the presence or absence of one or more device types mounted on the terminal. The interactive server program described in 1. A dialogue server that causes the multimodal dialogue server program according to any one of claims 1 to 3 to function on a computer;
A user-operable terminal having one or more input / output devices;
A multimodal dialogue system comprising:
The terminal
Device information transmission means for transmitting validity / invalidity for each device type to the dialogue server;
Output device control means for presenting the dialog scenario received from the dialog server to a user by an output device;
A multimodal dialog system comprising: input device control means for transmitting response information to an input device input by a user operation to the dialog server. The system further includes a user observation information analysis server,
The terminal further includes an observation feature amount transmitting means for transmitting the observation feature amount detected for each input device to the user observation analysis server,
The said user observation information analysis server derives | leads-out the reply information in the said user based on the observation feature-value received from the said terminal, and transmits this reply information to the said dialogue server. Multimodal dialogue system. A multimodal interaction method in a system having a terminal based on a user operation and an interaction server that interactively progresses a scenario with the user,
The dialogue server
For each action a, a dialogue scenario storage unit that stores a dialogue scenario to be transmitted to the terminal;
An expected reward value accumulating unit that accumulates an expected reward value r (s0, a0) obtained when the action a is executed in the state s;
A device action probability accumulation unit that accumulates device action probabilities r (sd0, ad0) in each device action ain / aout available in the state s for each of a plurality of different input / output devices;
The dialogue server is
A first step of acquiring presence / absence of one or more device types mounted on the terminal;
A second step of calculating the device available reward value r (sd, ad) by multiplying the input / output device by the presence / absence of the device type of the terminal and the device action probability r (sd0, ad0);
A reward expectation value r (s, a) is calculated by weighting the device available reward value r (sd, ad) with respect to the reward expectation value r (s0, a0) of the reward expectation value storage unit. Steps,
Fourth step of partial observation Markov decision POMDP (Partially-Observed Markov Decision Process) process for determining action a in state s so that reward Vt is maximized using the expected reward value r (s, a) When,
And a fifth step of acquiring a dialogue scenario based on the determined action a from the dialogue scenario storage unit and transmitting the dialogue scenario to the terminal. The dialogue server transitions from the state s to the state s ′ after executing the action a, and the observation expected probability P (o ′ | s ′, a) that receives the observation o ′ for the state s ′ after the action a is executed. a state transition probability P (s' | s, a ) and further comprising the accumulated action sequence storing section,
About a fourth step of the interaction server,
Calculating a probability distribution bt (s) at time t using the observed expectation probability P (o ′ | s ′, a) and the state transition probability P (s ′ | s, a);
Calculating a reward Vt using the expected reward value r (s, a) and the probability distribution bt (s);
The multimodal interaction method according to claim 6, further comprising: determining an action a in the state s so that the reward Vt is maximized using the expected reward value r (s, a). . Regarding the device action probability accumulation unit of the dialog server, the device action probability r (sd0, ad0) indicates whether the device can be used for the dialog scenario, from 1 (should be used) to 0-1 (use Is a probability expressed in the range of
The second step of the dialogue server is expressed by 1 (valid) / 0 (invalid) indicating whether one or more device types are installed in the terminal. 8. The multimodal dialogue method according to 7.

Images