JP3194719B2 - Dialogue system - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a dialog system for interacting with a human.

[0002]

2. Description of the Related Art Various models for dialogue understanding have been proposed so far, and most of them are intended not for chat but for purposeful dialogue. In the models proposed so far, dialogue is often considered as a plan, and each utterance is associated with an action to achieve a goal.

[0003] For example, the prior art document "Koichi Yamada et al.,"
Generation of User Utterance Model and Collaborative Response in Question Answering System ", Transactions of Information Processing Society of Japan, Vol. 35, N
o. 11, pp. 2265-2275, 1994 11
The month is characterized in that various cooperative responses can be enabled by inferring the intention of the user's question and generating a response according to the intention. In this prior art document, a user utterance model representing a relationship between a user's utterance and its intention is introduced, a cooperative response often seen in daily conversation is classified, and a relationship between the user's intention and the response is classified for each type of response. And a method of inferring the intention from the user's utterance based on the user's utterance model is disclosed. This method discloses a method of generating a cooperative response that has been classified using the inferred intentions, using independent intention inference rules in the domain and knowledge of the topic object.

[0004]

However, in the question answering system of the above-mentioned prior art document, there is a problem that the range of dialogue to be applied is narrow because models differ depending on the roles of the user and the system. . Also, it is conceivable to use a so-called slot insertion method in which the system side prepares a slot in advance and fills the slot by utterance, but the prepared slot is limited, and processing outside the standard can not be performed, There was a problem of lack of flexibility.

[0005] An object of the present invention is to solve the above problems and to provide a flexible dialog system capable of expanding a range of dialogs to be applied as compared with the prior art, and of being flexible.

[0006]

A dialogue system according to the present invention recognizes a voice of an uttered voice sentence as a character string, and responds to the voice-recognized character string to collect information. After generating a string of the utterance content of the response in,
A dialogue system for synthesizing and outputting a character string of utterance content, wherein an internal state indicating a dialogue state that changes with the progress of the dialogue is represented by: (a) a first state indicating that α knows φ Modal operator K (α, φ); (b) a second modal operator B (α, φ) indicating that α believes φ; and (c) α needs to know φ. And an internal state description storage device that expresses and stores a third propositional operator N (α, φ) using a modal operator, and (A) a general axiom expressing the properties of standard propositional modal logic using a modality operator And (B) inference rules representing the rules of necessity; and (C) background knowledge representing that if a plurality of facts are ANDed, certain facts are established; and utterances indicating restrictions on the order of utterances. A knowledge description storage device for storing task-dependent knowledge having a generation rule; An utterance generation rule storage device for storing an utterance generation rule indicating a utterance, a voice recognition unit for recognizing and outputting a voice of an uttered voice sentence as a character string, and displaying the character string as an information request or information provision content With reference to a conversion pattern model for converting to an intermediate language, a pre-processing means for converting the character string recognized by the voice recognition means to an intermediate language representing the content of the information request or information provision and outputting the same, From the intermediate language representing the content of the information request or information provision output from the preprocessing means, referring to the internal state in the internal state description storage device, and transitioning the internal state to update the internal state, State transition processing means for outputting the internal state after the transition; and internal state output from the state transition processing means, the knowledge in the knowledge description storage device and the utterance generation rule storage device Referring to the speech generation rule,
Inference processing means for inferring the content to be uttered in response to the dialogue in the current state, generating and outputting the intermediate language, and the inference processing with reference to the internal state in the internal state description storage device Data output processing means for converting the intermediate language of the content to be spoken in response to the information output from the means into output data of an intermediate language for requesting or providing information, and outputting the data; and an intermediate representing the content of the information request or providing information. With reference to an inverse conversion pattern model for inversely converting a language into a character string, the output data of the intermediate language representing the content of the information request or information provision output from the data output processing means corresponds to the intermediate language. Post-processing means for inverting the character string and outputting the same, and voice synthesizing means for synthesizing the character string output from the post-processing means and outputting the corresponding voice. To.

[0007]

Embodiments of the present invention will be described below with reference to the drawings.

FIG. 1 is a block diagram showing the configuration of a dialogue system according to an embodiment of the present invention. The dialogue system of the present embodiment recognizes the voice of the uttered voice sentence to be uttered into a character string, and responds to the voice-recognized character string in response to the utterance content in the dialogue for information collection. Is generated, and then a speech string of the utterance content is synthesized and output, and (a) an internal state description memory 31 for storing an internal state indicating the state of the dialogue that changes as the dialogue progresses; (B) Knowledge description memory 33 for storing knowledge including general axioms, inference rules, and task-dependent knowledge
(C) an utterance generation rule memory 32 for storing utterance generation rules indicating restrictions on the order of utterances, and (d) voice recognition for recognizing and outputting a digital voice signal of an uttered voice sentence to be uttered as a character string. Unit 3 and (e) converting the character string speech-recognized by the speech recognition unit 3 into an intermediate language representing the contents of the information request or information provision with reference to a predetermined conversion pattern model in the pattern model memory 11. (F) referring to the internal state in the internal state description memory 31 based on the information request or the intermediate language of information provision output from the preprocessing section 4 and A state transition processing unit 21 that changes the state to update the internal state and outputs the internal state after the transition, and (g) the knowledge description based on the internal state output from the state transition processing unit 21 With reference to the knowledge in the memory 33 and the utterance generation rules in the utterance generation rule memory 32, in consideration of the current state, the contents to be uttered in response to the dialogue are inferred to generate the intermediate language. (H) referring to the internal state in the internal state description memory 31 based on the intermediate language of the content to be uttered in response to the content output from the inference processing section 22; A data output processing unit 23 that converts the data into output data of an information request or information provision intermediate language and outputs the converted data; and (i) based on the information request or information provision intermediate language output data output from the data output processing unit 23 A post-processing unit 6 that refers to a predetermined inverse conversion pattern model in the pattern model memory 12 and inversely converts the character string corresponding to the intermediate language and outputs the character string;
(J) a voice synthesizing unit 7 for voice synthesizing the character string output from the post-processing unit 6 and outputting a voice corresponding thereto.

In this embodiment, a dialogue model focusing on information exchange is disclosed for a dialogue for information collection, and what kind of information is conveyed by the dialogue, how it is influenced by what, and how it is influenced A dialogue system is constructed focusing on whether utterances occur. Here, in order to explain the occurrence of the utterance, a modal operator of "must know" is introduced, and the utterance generation rule is described using the operator. The utterance is either generated as information provision to the request of the other party, or is generated by an utterance generation rule based on the obtained information. With this framework, the utterance of the information requesting side and the utterance of the information providing side can be uniformly explained.

In this embodiment, the following three modal operators are introduced to represent the internal state. (A) K (α, φ) -α knows φ. (B) B (α, φ) -α believes φ. (C) N (α, φ) -α needs to know φ. Here, K and B are operators commonly used to indicate a mental situation, and it is known that axiomatic systems have modal logic axiomatic systems S4 and KD45, respectively. In addition, in this embodiment, an operator N is introduced. N
Describes the cause of the utterance, and corresponds to KD as an axiomatic system, and will be described in detail below. In the present embodiment, an expression that does not include the modality operator is called a fact.
Distinguish from mental expressions that include modal operators.

There are the following properties of the standard propositional modal logic, and the system differs depending on which property has an axiom. Here, O represents a modal operator.

Kc: O (α, φ → ψ) → (O (α, φ) → O
(Α, ψ))

Tc: O (α, φ) → φ

Dc: O (α, φ) → ¬O (α, ¬φ)

## EQU4 ## 4: O (α, φ) → O (α, O (α, φ))

5: ¬O (α, ¬φ) → O (α, ¬O (α, ¬φ))

Here, ¬ indicates negation, and so on. The property Kc indicates that O is closed under the implication (→). Further, the property Tc represents the reflection rule. Further, the property Dc represents a chain rule. In addition, the property 4 represents a transition law. Further, the property 5 represents the Euclidean rule. A system having properties Kc, Dc, and 4 as axioms is S4, property K
KD45 or weak-S for c, Dc, 4, 5
5. Properties Kc and Dc are called KD. Further, the inference rules can be expressed as:

[0013]

NEC: If | = φ then O (α, φ)

The above inference rules (NEC) represent inevitable rules, and the inference rules (NEC) are common to all systems.

Next, a description language of the conversation model will be described. Utterances are two types: request and inform.
Describe by type. First, it is assumed that all utterances are correctly (at least literally) transmitted to the other party. Below,
It is assumed that the dialogue is being conducted between the dialog participants α and β. However, in the dialogue system of the present embodiment, one of the dialogue participants is a machine such as a digital computer.

(1) Information request—request (α,
β, P) (a) Precondition: B (α, N (α, P)) ∧¬K (α,
P) (b) Post-condition: K (β, N (α, P)) This information request—request (α, β, P) is used even if the dialog participant α needs to know the fact P. If not currently known, α requests information about P from the other conversation participant β, so that β indicates that α knows that P needs to know P.

(2) Information provision-inform (α, β,
P) (a) Precondition: B (α, N (β, P)) ∧K (α, P) (b) Postcondition: K (β, K (α, P)) Provide this information—inform (α) , Β, P), the dialogue participant α thinks that the dialogue participant β needs to know the fact P, and if he knows the information, α informs β of information about P. , And consequently, β indicates that α knows P. At this time, note that K (β, P) also holds according to the axiom.

Next, the task-dependent knowledge will be described. Each participant has background knowledge and utterance generation rules as task-dependent knowledge in addition to general axioms and inference rules.

The background knowledge is expressed by the following equation.

[Equation 7] _{P 1 ∧ ... ∧P n → Q} 1 ∧ ... ∧Q m _{However, P 1, ..., P n} , Q 1 ... Q m is a fact. For example, in a hotel reservation, the background knowledge that the number of people, the schedule, and the type of room are determined, and if there is a room that satisfies the conditions, the reservation is achieved is expressed as follows.
Example 1: Background knowledge

[Equation 8] number @ datas @ roomtype
∧available → reserved

Next, the utterance generation rules will be described.
In general, even if the interlocutor has the information to be given to the other party from the beginning, it does not mean that the information can be given at any time. For example, in the case of a hotel reservation, the customer would first tell the will of the reservation and then give information about the room type and schedule, and the hotel would only ask about the payment method and arrival time when the desired room was taken. That would be normal. Such restrictions on the order of utterances are described as utterance generation rules as follows.

[0021]

K (α, P ₁ ) ∧... ∧K (α, P _n ) → B (α,
_{N (α 1, Q 1)} ) ∧ ... ∧B (α, N (α m, Q m)) _{However, P 1, ..., P n} , Q 1, ..., Q m fact, α _{i (i}
= 1,..., M) is α or β.

If the dialog participant α knows P ₁ ,..., P _n , then the dialog participant α _i (i = 1 _,.
Indicates that you need to know Each utterance generation rule is applied at most once for one dialog.

For example, a hotel front (represented by c)
The knowledge that it is necessary to know the desired number of people, the schedule, and the type of room if the accommodation request (DesRes) is known is expressed as follows.

Example 2: Utterance generation rule

## EQU10 ## K (c, DesRes) → B (c, N (c,
number)) {B (c, N (c, dates))}
B (c, N (c, roomType))

Further, the internal state in the state transition will be described. Each conversation participant has an internal state that changes as the conversation progresses. The internal state is a set of knowledge and beliefs at that time, and each element is K (α, φ), B
(Α, φ). Here, α is a dialog participant, and φ is an expression including a fact or modality operator.
The internal state transits from S _{i -1} to S _{i according} to the utterance u _i (i = 1,..., N). Then, the dialogue is equivalent to the following finite sequence of internal states.

[0026]

[Number 11] _{u 1 u 2 u n S 0} ⇒S 1 ⇒ ... ⇒S n

The initial state is represented by the following equation.

S ₀ = ｛K (a, P ₁ ), ..., K (a, P _n ), K (b, Q ₁ ), ..., K (b, Q _m ), K (a, N ( α, P))｝

Here, a is a participant in the dialogue with the intention of collecting information, b is its partner, and P ₁ ,..., P _n , Q ₁ ,.
Q _m and P are facts. α is a or b. That is, the dialogue starts from an internal state that describes all the knowledge of a and b, and further describes that a knows that α needs to know P. The first utterance is inform
(A, b, P) or request (a, b, P).

The goal in the dialogue (that is, the expression that should be satisfied in the final state due to the purpose of collecting information) is K
(A, P). Here, a is a dialogue participant having an information collection intention, and P is information to be acquired. In the present embodiment, for certain α ₁ ,..., Α _m ∈S _n , ¬ (α ₁ ,..., Α _m → false) If α ₁ ∧… ∧α _m → K (a, P), define the dialogue as “successful”

Further, the mechanism of the state transition will be described. Utterance is inform (α, β, P) or request
It has the form st (α, β, P). Upon receiving the utterance in the state S _i-1 is the interactive participants as an input, state transition to S _i occurs according to the following procedure. (A) Transition by request (α, β, P)

(Equation 13) (B) Transition by inform (α, β, P)

Equation 14] if B (α, N (β , P)), K (α, P) ∈S i-1 then S i = S i-1 - {B (α, N (β, P))} ∪ ｛K (β,
K (α, P)) {K (β, P)}

In S _i , the inference based on the background knowledge and the inference based on the utterance generation rule are subsequently performed, and are output as the next utterance. The above is the framework of the utterance generation process to be responded in the present embodiment.

Next, the configuration and operation of the interactive system according to the present embodiment will be described with reference to FIG. In FIG. 1, a voice uttered by a dialog participant (human) 100 is input to a microphone 1 and is converted into an analog voice signal (electric signal).
After that, the digital audio signal is A / D converted by the A / D converter 2 and input to the voice recognition unit 3. The voice recognition unit 3 converts the input digital voice signal into, for example, LP
After performing speech analysis by the C method and extracting acoustic feature parameters, a known speech recognition method is performed by referring to a hidden Markov model (HMM) or a statistical language model based on the extracted acoustic feature parameters. Then, the speech is recognized, converted into text data of a character string, and output to the preprocessing unit 4.

The pre-processing unit 4 is connected to a pattern model memory 11 for storing in advance a pattern model for converting a character string into an intermediate language for requesting information or providing information. Referring to the pattern model stored in the pattern model memory 11 based on the column,
It is converted into an information request or an information provision intermediate language described in the above description language and output to the utterance generation processing unit 5. The following table shows an example of a pattern model in the pattern model memory 11 for the preprocessing unit 4. In this example of the pattern model, the character string of the conversion source, the intermediate language after the conversion, and the example number described in the present embodiment (only when there is an example) are shown.

[0034]

[Table 1] Example of a pattern model in the pattern model memory 11 for the preprocessing unit 4 ――――――――――――――――――――――――――――― ――― (a) Yes, how many people are there? Request (c, g, nbr); (Example 3) ―――――――――――――――――――――――――――――――――― (b )Understood. So can you give me your name? Request (c, g, name) ―――――――――――――――――――――――――――――――――― (c) September of September I would like one night on one Sunday. ; Inform (g, c, dates) ―――――――――――――――――――――――――――――――――― (d) Yes, so No thank you. ; Inform (g, c, fix); (Example 4) ―――――――――――――――――――――――――――――――― (e ) Please take three days from Saturday, August 13, to the 15th. ; Form (g, c, dates); (Example 5) ―――――――――――――――――――――――――――――――――― (f ) I am staying at Hotel New Otani Los Angeles Room 602. The telephone number is 213,443,1700. Inform (g, c, ad), inform (g, c, tel); (Example 6) ――――――――――――――――――――――――――― ―――――――

The utterance generation processing unit 5 includes a state transition processing unit 21
, An inference processing unit 22, and a data output processing unit 23, and the state transition processing unit 21 and the data output processing unit 23.
Is connected to an internal state description memory 31, and the inference processing unit 22 includes an utterance generation rule memory 32 and a knowledge description memory 3.
3 are connected.

Here, the internal state description memory 31 stores a set of knowledge and beliefs at that time, and stores, for each dialogue participant, an internal state indicating the state of the dialogue that changes as the dialogue progresses. I do. The following table shows a description example of the internal state in the internal state description memory 31. In the description example of the internal state, the above-described operators K and B are used, and the numbers of the examples described in the present embodiment are shown at the end of each example.

[0037]

[Table 2] Description example of internal state in internal state description memory 31 ―――――――――――――――――――――――――――――――――― a) B (c, N (c, nbr)) (Example 3) ――――――――――――――――――――――――――――――― ― (B) K (g, N (c, nbr)) (Example 3) ――――――――――――――――――――――――――――――― ――― (c) K (c, K (g, nbr)) (Example 3) ――――――――――――――――――――――――――― ――――― (d) B (g, N (c, fix)), K (g, fix) (Example 4) ――――――――――――――――――― ―――――――――――――― (e) K (c, K (g, fix)), K (c, fix), B (c, N (c, name)), B (C, N (c, pay)), K (g, fix) (Example 4) ――――――――――――――――――――――――――――― ――――― (f) B (g, N (c, ad)), B (g, ad), B (g, N (c, tel)), B (g, tel) (Example 6) ― ――――――――――――――――――――――――――――――― (g) B (c, B (g, ad)), B ( c, ad), B (c, B (g, tel ')), B (c, tel'), B (c, N (g, ad)), B (g, ad), B (c, N) (G, tel ')), B (g, tel) (Example 6) ―――――――――――――――――――――――――――――――― -(H) B (g, B (c, ad)), B (g, ad), B (c, ad), B (g, B (c, tel ")), B (g, tel) , B (g, tel ") (Example 6) ----------------------------------

The knowledge description memory 33 describes and stores task-dependent knowledge in addition to general axioms and inference rules. The following table shows an example of knowledge description in the knowledge description memory 33. Here, an explanation will be added for each example.

[0039]

[Table 3] Example of knowledge description in knowledge description memory 33 ―――――――――――――――――――――――――――――――― (a) B (Α, φ → ψ) → (B (α, φ) → B (α, ψ)) (Axiom of B) (Explanation) B is closed under implication. ―――――――――――――――――――――――――――――――――― (b) N (α, φ → ψ) → (N (α , Φ) → N (α, ψ)) (Axiom of N) (Explanation) N is closed under implication. ―――――――――――――――――――――――――――――――――― (c) K (α, φ) → φ (K axiom) (Explanation) We only know that α is true. ―――――――――――――――――――――――――――――――――― (d) number∧datas∧roomtype∧available → reserved (Example 1) (Description) See Background Knowledge above. When booking a hotel, the reservation is achieved if the number of people, the schedule, the type of room are determined, and there are rooms that meet the conditions. ―――――――――――――――――――――――――――――――――― (e) tree @ check → extraBed (Explanation) Cheap room with 3 people Then put in an extra bed. ―――――――――――――――――――――――――――――――――― (f) single → double → false (Description) Room type is single And never double. ――――――――――――――――――――――――――――――――――

The utterance generation rule memory 32 stores in advance utterance generation rules indicating restrictions on the order of utterances.
Here, "if you know a fact (s), you must know a particular fact (s)", or
It is described in the form of "if one knows one or more facts, the other party has to know a particular fact (s)".
However, in the case of a weak model, "know" becomes "believe". Hereinafter, (f) is an example in a weak model. Here, the weak model is a model for a dialogue in which the utterance may not be transmitted correctly. The following table shows an example of an utterance generation rule in the utterance generation rule memory 32. Here, an explanation will be added for each example.

[0041]

[Table 4] Example of utterance generation rules in utterance generation rule memory 32 ―――――――――――――――――――――――――――――――― ) K (c, DesRes) → B (c, N (c, number)) ∧B (c, N (c, dates)) ∧B (c, N (c, roomType)) (Explanation) (Example 2) The hotel front desk (represented by c) needs to know the desired number of people, the schedule, and the type of room once the accommodation request (DesRes) is known. ―――――――――――――――――――――――――――――――― (b) K (c, fix) → B (c, N ( (c, name)) @ B (c, N (c, pay)) (Explanation) (Example 4) When the reservation is confirmed, the hotel asks for the name and payment method of the partner. ―――――――――――――――――――――――――――――――――― (c) K (g, breakfast) → B (g, N ( c, bfType)) (Explanation) If the guest (represented by g) knows that breakfast is included, I think that the hotel should know the type of breakfast desired. ―――――――――――――――――――――――――――――――――― (d) K (c, roomType) ∧K (c, notAvailable) → B (c, N (g, notAvailable)) ∧B (c, N (g, roomType ')) (Description) If the hotel knows the type of room desired and there is no vacancy in that type, the customer I need to know that and know that there is another type of room. ―――――――――――――――――――――――――――――――――― (e) K (c, dates) → B (c, N ( g, dates)) (Explanation) (Example 4) If the hotel knows the schedule, the guest should know it (for confirmation). ―――――――――――――――――――――――――――――――――― (f) B (c, ad) → B (c, N ( g, ad)) (Explanation) (Example 6) If the hotel side knows the contact information (if they believe it), the guest should know it (for confirmation). ――――――――――――――――――――――――――――――――――

Accordingly, in the utterance generation processing unit 5, the state transition processing unit 21 firstly determines the internal state in the internal state description memory 31 based on the information request or the intermediate language of information provision input from the preprocessing unit 4. , The internal state is transited to update the internal state in the internal state description memory 31, and the internal state after the transition is output to the inference processing unit 22.
Next, the inference processing unit 22 refers to the knowledge in the knowledge description memory 33 and the utterance generation rule in the utterance generation rule memory 32 based on the input internal state, and considers the dialogue in consideration of the current state. In response, the contents to be uttered are inferred, the intermediate language is generated, and output to the data output processing unit 23. Further, in response to this, the data output processing unit 2
3 refers to the internal state in the internal state description memory 31 based on the intermediate language of the content to be uttered,
The data is converted into output data in an intermediate language for information request or information provision and output to the post-processing unit 6.

The post-processing unit 6 stores in advance a pattern model for inversely converting an intermediate language for information request or information provision, which is input from the data output processing unit 23 of the utterance generation processing unit 5, into text data of a character string. The pattern model memory 12 to be stored is connected, and the post-processing unit 6 refers to the pattern model stored in the pattern model memory 12 based on an input information request or an intermediate language for providing information, and The described information request or the intermediate language of information provision is inversely converted into text data of a character string corresponding thereto and output to the speech synthesis unit 7. The following table shows an example of a pattern model in the pattern model memory 12 for the post-processing unit 6. In this example of the pattern model, a conversion source intermediate language, a character string after conversion, and an example number (if any) described in the present embodiment are shown.

[0044]

[Table 5] Example of pattern model in pattern model memory 12 for post-processing unit 6 ――――――――――――――――――――――――――――― --- (a) I would like two adults. Inform (g, c, nbr); (Example 3) ―――――――――――――――――――――――――――――――― (b Yes, my name is Amy Harris. Inform (g, c, name) ―――――――――――――――――――――――――――――――――― (c) Yes, how many people Is it? Request (c, g, nbr) ―――――――――――――――――――――――――――――――― (d) I was smart. So can you give me your name? Request (c, g, name) (Example 4) ―――――――――――――――――――――――――――――――――― (e) Yes, you will be checked in on August 13 and staying for two nights. ; Inform (c, g, dates) (Example 5) ―――――――――――――――――――――――――――――――― (f) Yes, Mr. Suzuki is staying in Room 602 of the New Otani Hotel. And your phone number is 714,443,1700. ; Inform (c, g, ad '), inform (c, g, tel') (Example 6) ―――――――――――――――――――――――――― ――――――――

The speech synthesizer 7 includes a pulse generator, a noise generator, a variable gain amplifier, and a filter, and based on voiced and unvoiced, a pulse signal from the pulse generator and a signal from the noise generator. After switching between the noise and the noise, the text data of the input character string is changed using a known speech synthesis method in which the gain of the amplifier is changed according to the amplitude of the synthesized speech and the filter coefficient is changed according to the synthesized speech. Is converted into a digital audio signal and output to the D / A converter 8.
In response, the D / A converter 8 D / A converts the input digital audio signal into an analog audio signal and outputs it as a synthesized audio via the speaker 9.

The interactive system configured as described above includes an A / D converter 2, a speech recognition unit 3, a preprocessing unit 4, a state transition processing unit 21, an inference processing unit 22, and a data output processing unit 23. The utterance generation processing unit 5, the post-processing unit 6, the speech synthesis unit 7, and the D / A converter 8 are constituted by, for example, digital computers, and include pattern model memories 11, 12, an internal state description memory 31, and an utterance generation rule memory 32. , And the knowledge description memory 33 are configured by a storage device such as a hard disk memory.

The following table shows an example of the system operation used in this embodiment. In the following example, (f) is an operation example for a weak model. The inputs and outputs in each example correspond to the examples of the pre-processing and post-processing described above, respectively.

[0048]

[Table 6] System operation example ―――――――――――――――――――――――――――――――― (a) Q & A (Example 3) request (c, g, nbr) ↓ State transition K (g, N (c, nbr)) ↓ Inference B (g, N (c, nbr)) ∧ K (g, nbr) ↓ Output inform (g, c, (nbr) ―――――――――――――――――――――――――――――――― (b) Q & A (dialogue immediately after Example 4) request (C, g, name) ↓ State transition K (g, N (c, name)) ↓ Inference B (g, N (c, name)) ∧ K (g, name) ↓ Output inform (g, c, name) ) ―――――――――――――――――――――――――――――――――― (c) Topic transition (dialogue just before Example 3) (G , C, data) ↓ State transition K (c, K (g, data)) ↓ Utterance generation B (c, N (c, nbr)) ↓ Inference B (c, N (c, nbr)) ∧¬ K ( c, nbr) ↓ Output request (c, g, nbr) ―――――――――――――――――――――――――――――― (d ) Topic transition (Example 4) inform (g, c, fix) ↓ State transition K (c, K (g, fix)) ↓ Inference K (c, fix) ↓ Utterance generation B (c, N (c, name) )) ∧B (c, N (c, pay)) ↓ Inference B (c, N (c, pay)) ∧B (c, N (c, pay)) ∧ ¬K (c, name) ∧¬K (C, pay) ↓ Output request (c, g, name) ―――――――――――――――――――――――――――――――― ( e) Confirmation Example 5) inform (g, c, data) ↓ state transition K (c, K (g, data)) ↓ inference K (g, data) ∧ K (c, data) ↓ utterance generation B (c, N (g) , Dates)) ∧K (c, dates) ↓ Output inform (c, g, dates) ――――――――――――――――――――――――――――― ――――― (f) Confirmation (Example 6) inform (g, c, ad), inform (g, c, tel) ↓ state transition B (c, B (g, ad)) ∧ B (c, ad) ) ∧B (c, B (g, tel ′)) ∧B (c, tel ′) ↓ Utterance generation B (c, N (g, ad)) ∧B (c, ad) ∧B (c, N ( g, tel ')) ∧ B (c, tel') ↓ Output inform (c, g, ad), inform (c, g, tel ') ――― ------------------------------

Further, an example of the operation of the utterance generation processing section 5 of the interactive system configured as described above will be described in detail. In the utterance generation processing unit 5, three processes of state transition, inference, and output data are executed along the processing flow of the utterance generation processing unit 5 based on the input information request or the intermediate language of information provision. An intermediate language for requesting or providing information corresponding to words to be spoken in a dialog is generated.

First, a typical question and answer will be described. request (α, β, P) gives K (β, N
(Α, P)) occurs. By inference, B (β, N (α,
P)) is derived. At that time the internal state becomes K (β, P)
If there is, the precondition of information provision is satisfied, and
(Β, α, P) is caused. The following example is a dialogue between a hotel front desk and a customer in a hotel reservation scene (all the examples used in this embodiment are taken from a corpus for travel dialogue owned by the present applicant. ), G represents a guest, and c represents a hotel front (clerk).
In hotel reservation dialogue, customers usually have knowledge of accommodation schedule, desired room type, price, etc., communicate them in turn with the front desk, and finally know that "booking was done under desired conditions" .

Example 3: Q & A c: Yes, how many people are there? g: I would like two adults. The first utterance is request
st (c, g, nbr), next utterance is inform
(G, c, nbr). The internal state immediately before this dialogue is S ₁ , the state transited by each utterance is S ₂ ,
When S _3, the transition of the internal state of this interaction is as follows.

[0052]

S ₁ : B (c, N (c, nbr)) ∧¬K (c, nbr) ↓ request (c, g, nbr) S ₂ : K (g, N (c, nbr)) → B (g, N (c, nbr)) ∧K (g, nbr) ↓ inform (g, c, nbr) S ₃ : K (c, K (g, nbr)) → K (c, nbr)

Since the number of customers is known from the beginning, the initial state includes K (g, nbr). Assume further from the previous dialog that state S ₁ contains B (g, N (c, nbr)) and does not contain K (c, nbr). Then, the precondition of the information request is satisfied and request (c, g, n
br) is issued. Thus occurs a state transition to state _{S 2 K (g, N (} c, nbr)) is occurring. Next, B (g, N (c, nbr)) is derived from K (g, N (c, nbr)) by the inference rule. On the other hand, K
(G, nbr) so is also stored in the state S _2, it holds the previous condition of providing information inform (g, c, nbr)
Is issued. As a result, it occurs state transition to S ₃ K
(C, K (g, nbr)) occurs. From the axiom of K to K
(C, nbr) holds. Therefore, the hotel reception finally obtains information on the number of people.

Next, in the example of the utterance order, a process in which a request or notification regarding new information occurs after obtaining certain information will be described. K (β, K (α, P)) is generated by inform (α, β, P). K by inference
(Β, P)). Here, it is assumed that there is an applicable utterance generation rule, and as a result, B (β, N (α, Q)) is generated. If there is no K (β, Q) in the internal state at that time, the precondition for the information request is satisfied and request (β, Q)
α, Q). The result may be inform (β, α, Q) depending on the utterance generation rule to be applied.

Example 4: Order of utterance g: Yes, that is fine. c: I was smart. So can you give me your name?

The first utterance conveys the intention of the guest to make a reservation at the hotel, and is described as inform (g, c, fix) and the next utterance is request (c, g, name). Assuming that the internal state immediately before the dialogue is S ₁ , and the states transited by each utterance are S ₂ and S ₃ , the internal state transitions due to the dialogue are as follows.

[0057]

S ₁ : B (g, N (c, fix)) ∧K (g, fix) ↓ inform (g, c, fix) S ₂ : K (c, K (g, fix)) → K (C, fix) → B (c, N (c, name)) ∧B (c, N (c, pay)) ¬K (c, name) ∧¬K (c, pay) ↓ request (c, g) , Name) S ₃ : K (g, N (c, name))

When the reservation is confirmed, the hotel asks for the name of the partner and the payment method. This order is described as an utterance generation rule as follows.

[0059]

K (c, fix) → B (c, N (c, nam
e)) ∧B (c, N (c, pay))

From the conversation so far, S ₁ is B (g, N
(C, fix)) and K (g, fix). Then, the pre-conditions for providing information are satisfied,
rm (g, c, fix) is emitted. As a result, a state transition to the state S ₂ occurs, and K (c, K (g, fix))
Occurs. Next, according to the inference rule, K (c, fix)
Is led. Then, according to the above utterance generation rules,

B (α, N (c, name)) ∧B (α, N
(C, pay)) occurs. On the other hand, in the state S ₂ , K (c, name) is also K
Since (c, pay) is not included, the precondition of the information request is satisfied, and request (c, g, name) or r
request (c, g, pay) is issued. Which occurs first is non-deterministic. The result of this speech, takes place the state transition to state _{S 3 K (c, N (} g, nam
e)) occurs. After this, if you continue inference,
m (c, g, name) is triggered to provide information. From the above, it can be understood that both the information request and the information provision can be explained by this framework.

Next, an example of the confirmation process will be described. inform (α, β, P) followed by inform
When (β, α, P) occurs, the latter corresponds to a confirmation rather than a mere provision of information. Hereinafter, this process will be described. The rule that causes the confirmation utterance can be described as one of the utterance generation rules as follows.

[0062]

[Equation 19] K (α, P ₁ ) ∧ ... ∧K (α, P _n ) → B
(Α, N (β, P))

This rule states that α is a specific fact (P ₁ ,...,
If P _n ) is known, it means that P needs to be confirmed to β.

[0064]

B (α, N (β, P)) ∧K (α, P) ∧B
If inform (α, β, P) occurs in the state of (α, K (β, P)), it is confirmed. This condition indicates that α believes that β needs to know P, and that α knows that he knows the information and β knows P. In the internal state that has transitioned as a result of confirmation,

The following holds: K (α, K (β, P)) ∧K (β, K (α, P)), and P is mutual knowledge between the dialog participants α and β.

Example 5: Confirmation g: I would like three days from Saturday, the 13th of August to the 15th. c: Yes, you check-in on August 13th and you will stay for two nights.

The first utterance is inform (g, c, da
tes), the next utterance is inform (c, g, date)
s). It is assumed that K (g, dates) is given in an initial state, and B (g, N (c, dates)) is generated by the immediately preceding utterance. For simplicity, the utterance generation rule is represented by the following expression.

[0067]

## EQU22 ## K (c, dates) → B (c, N (g, d
ates)).

The state transition in this case is as follows.

B (g, N (c, dates)) ， K (g, dates) ↓ inform (g, c, dates) K (c, K (g, dates)) → K (g, dates) ∧ K (c, dates) → B (c, N (g, dates)) ∧K (c, dates) ↓ inform (c, g, dates) K (g, K (c, dates))

Here, it can be seen from the two underlined equations that mutual knowledge about dates was obtained between the hotel and the guest.

Next, an example in the case of a wrong listening will be described. Until now, it was assumed that information was transmitted correctly, so there was no failure due to confirmation. However, in an actual dialogue, information may not be transmitted correctly due to a mistake in listening, and confirmation and correction are often performed. In order to describe this process, request, inform
Weaken the condition.

(1) Information request—request (α,
β, P) (a) Precondition: B (α, N (α, P)) ∧¬B (α,
P) (b) Post-condition: B (β, N (α, P ′)) P ′ is true but does not always coincide with P.

(2) Information provision-inform (α, β,
P) (a) Precondition: B (α, N (β, P)) ∧B (α, P) (b) Postcondition: B (β, B (α, P ′)) ∧B (β,
P ')

In this model, the other party does not always obtain accurate information as a result of the utterance, but merely obtains the belief about the internal state of the utterer.

(24) B (α, N (β, P)) ＰB (α, P) ∧B
If inform (α, β, P) occurs in the state of (α, B (β, P)), it is confirmed.

Further, the utterance generation rule has a form in which K is replaced with B as follows.

## EQU25 ## B (α, P ₁ ) ∧... B (α, P _n ) → B
(Α, N (α ₁ , Q ₁ )) ∧ ... ∧B (α, N (α _m ,
Q _m )) In the internal state that transits as a result of the confirmation,

(26) B (α, P) ∧B (β, P ′) 'B (α, B
(Β, P)) ∧B (β, B (α, P ′)) holds.
here,

If P = P ', then P corresponds to the mutual belief between α and β.

## EQU28 ## If P ≠ P ’, it is a misunderstanding.

Example 6: Finding Mistakes in Listening g: Stay at Hotel New Otani Los Angeles 602
Issue room. The telephone numbers are 213, 443 and 1700. c: Yes, Mr. Suzuki is staying at room 602 of the New Otani Hotel. And your phone number is 7
14,443,1700.

This conversation is performed in the form (g, c, a
d), inform (g, c, tel), the next utterance is i
nform (c, g, ad), inform (c, g,
tel ′). B (g, ad), B (g, t
el) is given in the initial state, and B (g, N (c, a
d)) and B (g, N (c, tel)) are assumed to have been generated by the immediately preceding utterance. For simplicity, the utterance generation rule is

## EQU29 ## B (c, ad) → B (c, N (c, ad))

## EQU30 ## B (c, tel) → B (c, N (c, te)
l)) Then, the transition of the internal state is as follows.

[0077]

B (g, N (c, ad)) ∧B (g, ad) ∧B (g, N (c, tel)) ∧ [B (g, tel)] ↓ inform (g, c, ad), inform (g, c, tel) B (c, B (g, ad)) ∧ B (c, ad) Ｂ B (c, B (g, tel ')) ∧ B (c, tel') → B (c, N (g, ad ')) ∧B (c, ad) ∧B (c, N (g, tel')) ∧B (c, tel ') ↓ inform (c, g, ad') , Inform (c, g, tel ') B (g, B (c, ad)) {B (g, ad)} B (g, B (c, tel "))} [B (g, tel") ]

In this example, as shown by the underlined part, c believes in ad and g believes in ad. The same can be said for g's belief. Therefore,
Mutual beliefs hold for addresses. On the other hand, as shown in the box of [], g believes in tel and also believes in tel. ”(Tel @ tel”), so that the telephone number was not a mutual belief and a misunderstanding was found.

According to the present embodiment, a dialogue system using a dialogue model based on information exchange for a dialogue for information collection has been disclosed. In this model, both the request of the information collecting side and the notification of the information providing side can provide a unified explanation by introducing the modality operator who needs to know. Also, dynamically occurring phenomena, such as confirmations in actual conversations and discovery of listening mistakes, can be explained by the same framework.

As described above, the embodiment according to the present invention has the following effects. (A) It is possible to provide a dialogue system based on the exchange of information for a dialogue for information collection independent of tasks. (B) By adding utterance generation rules and knowledge, it is easy to extend the response function and incorporate the learning function, and it is possible to cope with various nonstandard specifications. Therefore, it is possible to broaden the range of dialogue to be applied as compared with the related art, and it is possible to provide a flexible cooperative dialogue system.

[0081]

As described above in detail, according to the present invention, the voice of an uttered voice sentence to be uttered is recognized as a character string, and the dialogue for collecting information is performed in response to the recognized character string. A dialogue system that generates a character string of the utterance content of the response in and then synthesizes and outputs the character string of the utterance content, wherein the internal state indicating the status of the dialogue that changes with the progress of the dialogue is represented by (a) ) A first modal operator K (α, φ) indicating that α knows φ, and (b) a second modal operator B (α, φ) indicating that α believes φ. (C) an internal state description storage device which is represented and stored using a third modality operator N (α, φ) indicating that α needs to know φ, and (A) standard propositional modal logic. General axioms that express the properties of, using modal operators; (B) inference rules that represent the rules of necessity; and (C) multiple things. A knowledge description storage device for storing background knowledge indicating that a certain fact is satisfied if AND is established, and task-dependent knowledge having utterance generation rules indicating constraints on the order of utterances; An utterance generation rule storage device for storing utterance generation rules indicating restrictions on the order of speech, a voice recognition means for recognizing and outputting a voice of an uttered voice sentence as a character string, and requesting or providing information on the character string Before converting the character string recognized by the voice recognition unit into an intermediate language representing the content of the information request or information provision with reference to a conversion pattern model for converting the content into an intermediate language representing the content of The internal state is transited by referring to the internal state in the internal state description storage device from the processing means and the intermediate language representing the content of the information request or information provision output from the preprocessing means. State update processing means for updating the internal state and outputting the internal state after the transition, knowledge from the knowledge description storage device based on the internal state output from the state transition processing means, and generation of the utterance With reference to the utterance generation rule in the rule storage device,
Inference processing means for inferring the content to be uttered in response to the dialogue in the current state, generating and outputting the intermediate language, and the inference processing with reference to the internal state in the internal state description storage device Data output processing means for converting the intermediate language of the content to be spoken in response to the information output from the means into output data of an intermediate language for requesting or providing information, and outputting the data; and an intermediate representing the content of the information request or providing information. With reference to an inverse conversion pattern model for inversely converting a language into a character string, the output data of the intermediate language representing the content of the information request or information provision output from the data output processing means corresponds to the intermediate language. Post-processing means for inverting and outputting the character string, and voice synthesizing means for synthesizing the character string output from the post-processing means and outputting the corresponding voice.

Accordingly, the present invention has the following effects. (A) It is possible to provide a dialogue system based on the exchange of information for a dialogue for information collection independent of tasks. (B) By adding utterance generation rules and knowledge, it is easy to extend the response function and incorporate the learning function, and it is possible to cope with various nonstandard specifications. Therefore, it is possible to broaden the range of dialogue to be applied as compared with the related art, and it is possible to provide a flexible cooperative dialogue system.

[Brief description of the drawings]

FIG. 1 is a block diagram showing a configuration of a dialogue system according to an embodiment of the present invention.

[Explanation of symbols]

 DESCRIPTION OF SYMBOLS 1 ... Microphone, 2 ... A / D converter, 3 ... Voice recognition part, 4 ... Pre-processing part, 5 ... Speech generation processing part, 6 ... Post-processing part, 7 ... Voice synthesis part, 8 ... D / A converter Reference numeral 9: Speaker, 11: Pattern model memory, 12: Pattern model memory, 21: State transition processing unit, 22: Inference processing unit, 23: Data output processing unit, 31: Internal state description memory, 32: Utterance generation rule memory 33: knowledge description memory; 100: dialog participants.

──────────────────────────────────────────────────続 き Continuation of the front page (56) References JP-A-7-239694 (JP, A) JP-A-7-210391 (JP, A) "Belief, Awarenes, and Limited Reasoning" Artificial Intelligence, Vol. . 34, No. 1, December 1987, p. 39 (58) Fields investigated (Int. Cl. ⁷ , DB name) G06F 17/27-17/30 G10L 3/00-9/20 G10L 13/00-15/00 JICST file (JOIS)

Claims (1)

(57) [Claims] 1. A speech recognition system comprising: a speech recognition unit that recognizes a voice of an uttered voice sentence to be uttered into a character string and generates a character string of an utterance content of a response in a dialogue for information collection in response to the character recognition string; A dialogue system for speech-synthesizing and outputting a character string of the utterance content, wherein an internal state indicating the state of the dialogue changing with the progress of the dialogue is represented by: (a) a state indicating that α knows φ. Modal operator of 1
K (α, φ) and (b) a second modal operator indicating that α believes φ.
B (α, φ) and (c) the third aspect that α needs to know φ
An internal state description storage device that expresses and stores using an operator N (α, φ); and (A) expresses the properties of standard propositional modal logic using modal operators.
General axioms that, (B) if the inference rules that represent the necessity of regulations, (C) a plurality of facts is established in logical, facts
And the order of utterance
Task-dependent with utterance generation rules indicating constraints on
And knowledge description storage device that stores the knowledge, and the speech generation rule storage unit for storing a speech production rule indicating the constraints on the order of speech, speech to be output by the speech recognition speech utterance voice statement string uttered Recognition means and convert character strings into intermediate language to express information request or information provision
Refer to the conversion pattern model for conversion and
A pre-processing unit that converts the character string recognized by the voice recognition unit into an intermediate language representing the contents of the information request or information provision and outputs the intermediate request; and an information request or information provision output from the pre-processing unit.
State transition processing means for referring to the internal state in the internal state description storage device from the intermediate language representing the content, transiting the internal state and updating the internal state, and outputting the internal state after the transition; from the internal state is outputted from the state transition processing means, with reference and knowledge in the knowledge description storage unit, and a speech generation rule in the utterance generation rule storage unit, put the current state
That in response in the dialogue to infer the contents to be uttered and inference processing means for generating and outputting an intermediate language, with reference to the internal states within the internal state description storage device, upper
Content to be uttered in response to output from the inference processing means
Of the intermediate language, and a data output processing means for converting the output data of the intermediate language information request or information providing, the intermediate language into a character string representing the content of the information request or information providing
Refer to the inverse transformation pattern model for inverse transformation, and
Information request or information output from the data output processing means
Post-processing means for inverting the output data of the intermediate language representing the content of the provision into a character string corresponding to the intermediate language and outputting the converted character string; A dialogue system, comprising: voice synthesis means for outputting voice.