WO2015119267A1

WO2015119267A1 - Uttered sentence collection apparatus and method

Info

Publication number: WO2015119267A1
Application number: PCT/JP2015/053448
Authority: WO
Inventors: Kenji Iwata
Original assignee: Kabushiki Kaisha Toshiba
Priority date: 2014-02-05
Filing date: 2015-02-03
Publication date: 2015-08-13
Also published as: JP6235360B2; JP2015148910A

Abstract

According to an embodiment, an uttered sentence collection apparatus includes following units. The input unit inputs an uttered sentence from a user. The extraction unit extracts, from the uttered sentence, at least one of an intention of the uttered sentence and a slot being information supplementing the intention. The management unit accumulates data on the at least one of the intention and the slot. The bias determination unit determines whether or not the data have a bias and acquires at least one of an infrequent intention and an infrequent slot from the data on determining that the data have the bias. The request output unit requests input of an uttered sentence related to the at least one of the infrequent intention and the infrequent slot.

Description

D E S C R I P T I O N

UTTERED SENTENCE COLLECTION APPARATUS AND METHOD

Cross-Reference to Related Applications This application is based upon and claims the benefit of priority from Japanese Patent Application

No. 2014-020811, filed February 5, 2014, the entire contents of which are incorporated herein by reference.

Field

Embodiments described herein relate generally to an uttered sentence collection apparatus and method.

Background

In recent years, a natural language understanding method utilized for a dialogue system (for example, an interactive voice response system) has been changing from a rule-based technique to a statistical technique.

Introduction of the statistical technique needs learning data in order to create a statistical model. The learning data needs to include a sufficient number and variety of uttered sentences that may be input to a dialogue system to be constructed.

Brief Description of Drawings

FIG. 1 is a block diagram schematically illustrating an uttered sentence collection apparatus according to a first embodiment;

FIG. 2 is a flowchart schematically illustrating an example of an uttered sentence collection process of the uttered sentence collection apparatus illustrated in

FIG . 1;

FIG. 3 is a diagram illustrating an example of a method in which a request output unit illustrated in FIG. 1 generates a request sentence;

FIG. 4 is a diagram illustrating an operation example of the uttered sentence collection apparatus illustrated in FIG. 1 where data on intentions and slots have no bias;

FIG. 5 is a diagram illustrating an operation example of the uttered sentence collection apparatus illustrated in FIG. 1 where data on intentions have a bias;

FIG. 6 is a diagram illustrating an operation example of the uttered sentence collection apparatus illustrated in FIG. 1 where data on slots have a bias;

FIG. 7 is a block diagram schematically illustrating an uttered sentence collection apparatus according to a second embodiment;

FIG. 8 is a flowchart schematically illustrating an example of an uttered sentence collection process of the uttered sentence collection apparatus illustrated in

FIG . 7 ;

FIG. 9 is a diagram illustrating a part of a flow of dialogues according to a third embodiment ;

FIG. 10 is a block diagram schematically illustrating an uttered sentence collection apparatus according to the third embodiment; FIG. 11 is a flowchart schematically illustrating an example of an uttered sentence collection process of the uttered sentence collection apparatus illustrated in

FIG. 10;

FIG. 12 is a flowchart schematically illustrating another example of the uttered sentence collection process of the uttered sentence collection apparatus illustrated in FIG. 10;

FIG. 13 is a diagram illustrating a part of a flow of dialogues according to a fourth embodiment;

FIG. 14 is a block diagram schematically illustrating an uttered sentence collection apparatus according to the fourth embodiment ; and

FIG. 15 is a flowchart schematically illustrating an example of an uttered sentence collection process of the uttered sentence collection apparatus illustrated in

FIG. 14.

Detailed Description

In general, according to an embodiment, an uttered sentence collection apparatus includes an input unit, an extraction unit, a management unit, a bias determination unit, and a request output unit. The input unit is configured to input an uttered sentence from a user. The extraction unit is configured to extract, from the uttered sentence, at least one of an intention and a slot, the intention being an intention of the uttered sentence, the slot being information supplementing the intention. The management unit is configured to accumulate data on the at least one of the intention and the slot. The bias

determination unit is configured to determine whether or not the data have a bias and to acquire at least one of an infrequent intention and an infrequent slot from the data on determining that the data have the bias . The request output unit is configured to request input of an uttered sentence related to the at least one of the infrequent intention and the infrequent slot acquired by the bias determination unit.

Hereinafter, embodiments will be described with reference to the drawings. The embodiments relate to uttered sentence collection apparatuses and methods which allow a variety of uttered sentences required for

construction of a dialogue system to be collected.

(First Embodiment)

FIG. 1 schematically illustrates an uttered sentence collection apparatus 100 according to a first embodiment. As illustrated in FIG. 1, the uttered sentence collection apparatus 100 includes an uttered sentence input unit 101, an intention and slot extraction unit 102, an intention and slot management unit 103, an intention and slot bias determination unit 104, and a request output unit 105. The intention and slot extraction unit 102, the intention and slot management unit 103, and the intention and slot bias determination unit 104 are also referred to as the

extraction unit 102, the management unit 103, and the bias determination unit 104, respectively.

The uttered sentence input unit 101 receives an input of an uttered sentence from a user. A method for inputting an uttered sentence may be text input using a keyboard or the like or speech input. For speech input, the uttered sentence input unit 101 receives a speech uttered by the user, through, for example, a microphone. Subsequently, the uttered sentence input unit 101 performs speech

recognition on the received speech to obtain, as an uttered sentence, a natural language text that is a result of the speech recognition. The uttered sentence input unit 101 saves (accumulates) the uttered sentence in a storage unit (not illustrated in the drawings) for collection. The uttered sentence input unit 101 further transmits the uttered sentence to the extraction unit 102.

The extraction unit 102 extracts at least one of an intention and a slot from the uttered sentence received from the uttered sentence input unit 101 and transmits the at least one of the intention and slot extracted to the management unit 103. The description of the present embodiment assumes that the extraction unit 102 extracts both an intention and a slot. The extraction unit 102 may extract an intention without extracting any slot or may extract a slot without extracting any intention.

The intention is information indicative of a content that the uttered sentence is intended to roughly

communicate. The intention can be expressed in a tag format. For example, a tag "weather search + request" is extracted from each of uttered sentences "Show me

tomorrow's weather" and "Show me the weather in Shibuya" . A tag "map search + desire" is extracted from "I want to see the locations of neighborhood convenience stores". The tag "request" and the tag "desire" may be unwanted

depending on the functional specifications of the dialogue system. A method for extracting an intention may be a rule-based method of manually predetermining keywords useful for the extraction and performing extraction using a database in which the keywords are stored in association with the intention tag or a statistical method of creating a learning model based on data on uttered sentences each associated with the intention tag and performing extraction using a support vector machine (SVM) or a maximum entropy model .

The slot is information that supplements the intention extracted from the uttered sentence. A combination of the intention and the slot enables what the uttered sentence is intended to communicate to be accurately determined. The slot includes a set of a word included in the uttered sentence and the attribute of the word. For example, a slot "attribute: date, value: tomorrow" is extracted from the uttered sentence "Show me tomorrow's weather". A slot "attribute: location, value: Shibuya" is extracted from the uttered sentence "Show me a map of Shibuya". No slot or a plurality of slots may be extracted depending on the uttered sentence. A method for extracting a slot may be, like the intention extraction method, a rule-based method of performing extraction using a database in which a keyword to be extracted as a slot is held in association with the intention tag or a statistical method of

performing extraction using conditional random fields (CRFs) .

The management unit 103 accumulates data on the intention and slot extracted by the extraction unit 102. Data on intentions and slots may be separately accumulated for each user or for each operation. Alternatively, data on intentions and slots acquired for all operations and for all users may be collectively accumulated. The latter method is more desirable because of more intentions and slots collectively accumulated.

In the management unit 103, data on the intention and slot extracted by the extraction unit 102 may be

accumulated in any format. By way of example, the

management unit 103 is provided with a database and adds the intention and slot extracted by the extraction unit 102 to the database. For example, when the intention tag

"weather search + request" and the slot "attribute: date, value: tomorrow" are extracted, the intention tag "weather search + request" is stored in an intention field in the database, and the slot "attribute: date, value: tomorrow" is stored in a slot field in the database. The database may include a user ID field that stores user information (for example, a user ID) allowing identification of the user having input the uttered sentence. The management unit 103 of the present embodiment manages the data on the intention and the slot in accordance with the above- described method.

In another example, the management unit 103 holds intention aggregate data that are a record of the frequency of each of types of intentions and slot aggregate data that are a record of the frequency of each of types of slots. For example, when the intention tag "weather search + request" is extracted, the management unit 103 increments, by one, the frequency of an intention type corresponding to the intention tag "weather search + request" in the

intention aggregate data. This method allows omission of an aggregate process of the bias determination unit 104 described below. The frequency indicates the number of times that information (for example, an intention) has been extracted, that is, the number of times that an uttered sentence related to the information has been input.

The bias determination unit 104 determines whether or not the data on the intention and slot accumulated in the management unit 103 have a bias. Specifically, the bias determination unit 104 aggregates the data accumulated in the management unit 103 for each of the intention and slot layers, and executes a determination process described below to determine whether or not the data have a bias . Upon determining that the data have a bias, the bias determination unit 104 extracts at least one of an

infrequent intention and an infrequent slot from the data accumulated in the management unit 103, and then transmits it to the request output unit 105.

When the bias determination unit 104 determines that the data have a bias, the request output unit 105 generates a request sentence requesting the user to input an uttered sentence related to the at least one of the intention and slot extracted by the bias determination unit 104. The request output unit 105 then presents the request sentence to the user. A method for generating a request sentence will be described below in detail . When the bias

determination unit 104 determines that the data have no bias, the request output unit 105 makes no request.

Alternatively, the request output unit 105 may generate and present a request sentence requesting the user to be free to input an uttered sentence, to the user. A presentation method may be, for example, a method of displaying a corresponding text on a screen viewed by the user or a method of audibly communicating the request to the user using a speech synthesis technique or a recorded speech.

Now, operation of the uttered sentence collection apparatus 100 will be described with reference to FIG. 1 and FIG. 2.

FIG. 2 schematically illustrates a procedure for a process in which the uttered sentence collection apparatus 100 collects uttered sentences. In step S201 in FIG. 2, the bias determination unit 104 determines whether or not the data (at least one of the data on the intention and the data on the slot) accumulated in the management unit 103 have a bias. When the bias determination unit 104

determines that the data have a bias,_. the process proceeds to step S202.

In step S202, the bias determination unit 104 extracts at least one of an infrequent intention and an infrequent slot. In step S203, the request output unit 105 generates a request sentence requesting the user to input an uttered sentence related to the at least one of the intention and slot extracted by the bias determination unit 104. The request output unit 105 then presents the request sentence to the user. Various uttered sentences with no bias can be sufficiently collected by requesting the user to input an uttered sentence related to at least one of an infrequent intention and an infrequent slot as described above.

On the other hand, when the bias determination unit 104 determines that the data have no bias, the process proceeds to step S204. In step S204, the request output unit 105 avoids limiting the content of an input uttered sentence and generates and presents a request sentence requesting the user to be free to input an uttered

sentence, to the user. The request output unit 105 may omit presenting a request sentence to the user.

The user inputs an uttered sentence in response to the presented request sentence which is generated in step S203 or step S204. The uttered sentence input unit 101 acquires the uttered sentence. The uttered sentence is accumulated in the storage unit (not illustrated in the drawings) . In step S205, the extraction unit 102 extracts an intention and a slot from the uttered sentence. In step S206, the extracted intention and slot are accumulated in the

management unit 103. The accumulation affects the content of a request sentence for the next uttered sentence input .

A method for determining whether or not the data have a bias will be described. A determination process is executed on the intention and on the slot. Here, a method for determining whether or not the data have a bias in connection with the intention will be mainly described. A determination process similar to the process described below may also be executed on the slot.

By way of example, the bias determination unit 104 determines that the data have a bias if any intention has a relative frequency equal to or less than a threshold when the total frequency of intentions accumulated in the management unit 103 reaches a specified value or more. The relative frequency is indicative of the percentage (%) of a certain intention in the total frequency of the intentions accumulated in the management unit 103. The specified value and the threshold may be manually predetermined. The threshold may be changed in accordance with the total frequency of the accumulated intentions. For example, the threshold decreases with increasing total frequency of the accumulated intentions. This is because a high total frequency is likely to mean a sufficient amount of learning data collected in spite of a low relative frequency. In an embodiment where both intentions and slots are accumulated in the management unit 103, the bias determination unit 104 may determine that the data have a bias in at least one of two cases: the case where any intention has a relative frequency equal to or lower than a first threshold when the total frequency of^" the intentions accumulated in the management unit 103 reaches a specified value or more, and the case where any slot has a relative frequency equal to or lower than a second threshold when the total frequency of the slots accumulated in the management unit 103 reaches a specified value or more. The first threshold and the second threshold may have the same value or different values. The first threshold and the second threshold may be manually predetermined. Furthermore, the first

threshold may be changed in accordance with the total frequency of the accumulated intentions. The second threshold may be changed in accordance with the total frequency of the accumulated slots.

In another example, the bias determination unit 104 determines that the data have a bias when the difference in frequency between the most frequent intention and the least frequent intention reaches a threshold or more. This threshold may be manually predetermined. In an embodiment where both intentions and slots are accumulated in the management unit 103, the bias determination unit 104 may- determine that the data have a bias in at least one of two cases: the case where the difference in frequency between the most frequent intention and the least frequent

intention is equal to or smaller than a first threshold and the case where the difference in frequency between the most frequent slot and the least frequent slot is equal to or smaller than a second threshold. The first threshold and the second threshold may have the same value or different values .

In yet another example, the bias determination unit 104 may determine that the data have a bias in at least one of two cases: the case where the rate, in the total frequency of the accumulated intentions, of the difference in frequency between the most frequent intention and the least frequent intention is equal to or higher than a first threshold and the case where the rate, in the total

frequency of the accumulated slots, of the difference in frequency between the most frequent slot and the least frequent slot is equal to or higher than a second

threshold. The first threshold and the second threshold may be manually predetermined. Furthermore, the first threshold may vary in accordance with the total frequency of the accumulated intentions. The second threshold may vary in accordance with the total frequency of the

accumulated slots.

Upon determining that the data on the intentions have a bias, the bias determination unit 104 extracts an

infrequent intention. When a plurality of infrequent intentions are present, the bias determination unit 104 may extract any of the infrequent intentions. An extraction method may be a method of extracting the most infrequent intention or a method of extracting an intention the user has not input before. In the latter case, the management unit 103 needs to manage the intention along with user information such as the user ID which is used to identify the user so that the bias determination unit 104 extracts the intention with reference to the user information. A method for acquiring the user information may be, for example, a method of receiving the user information from the user when the uttered sentence collection apparatus 100 is activated and holding the information in the uttered sentence collection apparatus 100 while the uttered

sentence collection apparatus 100 is active or a method in which the uttered sentence input unit 101 receives the user information simultaneously with each input of an uttered sentence .

When a plurality of infrequent intentions are present, the request output unit 105, instead of the bias

determination unit 104, may determine which intention is to be adopted when the request output unit 105 outputs a request. In this case, the bias determination unit 104 transmits information indicative of a plurality of

infrequent intentions to the request output unit 105. The request output unit 105 can determine an adopted intention in accordance with a method similar to the above-described determination method of the bias determination unit 104. The request output unit 105 may generate a request sentence including all of the plurality of infrequent intentions and request the user to input an uttered sentence related to one of the infrequent intentions .

A method in which the request output unit 105

generates a request sentence may be a rule-based method, a statistical method, or the like. The statistical method may be a method of preparing a large number of sentences for each content of request and generating a request sentence using a model resulting from learning of the data. The rule-based method may be, for example, a method of preparing request sentences for each intention or for each slot and selecting one of the request sentences in

accordance with an extracted intention or slot, or a method of preparing request sentences each designed such that a partial sentence can be added to the request sentence for the corresponding intention, selecting one of the request sentences according to an extracted intention, and then adding a sentence indicative of the attribute of an

extracted slot to the request sentence to generate one request sentence. The latter method is illustrated in FIG. 3. In the example in FIG. 3, the request output unit 105 selects a request sentence "Enter an uttered sentence requesting a search for the weather at XX" associated with the extracted intention "weather search" and uses a

location that is the attribute of the extracted slot to generate a request sentence "Enter an uttered sentence requesting a search for the weather at a certain location" . When the attribute of the slot is converted into a text corresponding to a part of the request sentence, the expressed attribute itself may be used but texts may be prepared for the respective attributes of slots as

illustrated in FIG. 3. A plurality of candidates for the request sentence may be generated for one intention or slot so that one of the candidates can be selected randomly or according to the user's preference, and various request sentences may be output so as to increase the variety of expressions of an uttered sentence that may be affected by the request sentence. The above-described methods are used when a request sentence is generated based on rules. In adopting a statistical technique, the request output unit 105 stochastically selects a request sentence from

resultant N-best candidates.

In the above description, the bias is determined and a request sentence is created using both the intention and the slot. However, the bias is determined and a request sentence is generated using only the intention or the slot.

Now, operation examples of the uttered sentence collection apparatus 100 will be described with reference to FIGS . 4 to 6 to indicate that the uttered sentence collection apparatus 100 can sufficiently collect a variety of uttered sentences that may actually be input to the dialogue system. In examples described below, a terminal including the uttered sentence collection apparatus 100 is in front of the user, and the user inputs an uttered sentence to the terminal in a text format . A request to the user is displayed on a screen of the terminal. A dialogue system to be actually constructed can accept the intentions "weather search" and "map search" and accept the slots "attribute: date" and "attribute: location" for the "weather search" and the slot "attribute: location" for the "map search" .

FIG. 4 illustrates an operation example of the uttered sentence collection apparatus 100 where data on the

accumulated intentions have no bias and data on the

accumulated slots have no bias. In the example in FIG. 4, the terminal requests the user to be free to input an uttered sentence. Specifically, the terminal displays a request sentence "Be free to enter an uttered sentence" . In response to the request, the user inputs an uttered sentence "I want to know the weather forecast for

tomorrow" . The terminal extracts an intention and a slot from the uttered sentence, resulting in accumulation of the intention "weather search" and the slot "attribute: date, value: tomorrow".

FIG. 5 illustrates an operation example of the uttered sentence collection apparatus where data on the intentions have a bias. In the example in FIG. 5, as a result of the advanced accumulation of intentions and slots, the

frequency of the intention "weather search" is much higher than the frequency of the intention "map search" . In this state, the uttered sentence collection apparatus 100 determines that the data on the intentions have a bias. In this case, the uttered sentence collection apparatus 100 extracts the infrequent intention "map search" and

generates a request sentence "Enter an uttered sentence requesting a map search" based on the extracted intention to present the request sentence to the user. In response to the request sentence, the user inputs a sentence "Show me a map of Tokyo" including the "map search" as an

intention. In this manner, the uttered sentence

corresponding to the infrequent intention is input and accumulated.

FIG. 6 illustrates an operation example of the uttered sentence collection apparatus 100 where data on the slots have a bias. In the example in FIG. 6, as a result of advanced accumulation of intentions and slots, the slot "attribute: date", included in the slots accumulated in association with the intention "weather search", has a much higher frequency than the three other types of slots. The three types of slots are no slot, the slot "attribute:

location", and a set of the slot "attribute: date" and the slot "attribute: location". In this state, the uttered sentence collection apparatus 100 determines that the data on the accumulated slots have a bias. In this case, the uttered sentence collection apparatus 100 selects the slot "attribute: location" from the three types of infrequent slots, and creates a request sentence "Enter an uttered sentence requesting a search for the weather at a certain location" to present the request sentence to the user. In response to the request sentence, the user inputs the uttered sentence "Show me the weather in Tokyo" including the intention "weather search" and including the slot "attribute: location". In this manner, the uttered

sentence corresponding to the infrequent slot is input and accumulated.

In the example in FIG. 6, the uttered sentence

collection apparatus 100 determines the presence or absence of a bias for the slots accumulated in association with the intention "weather search" . However, the uttered sentence collection apparatus 100 may determine the presence or absence of a bias for all of the accumulated slots.

As described above, the uttered sentence collection apparatus according to the first embodiment requests the user to input an uttered sentence related to an infrequent intention and/or slot to enable sufficient collection of a variety of uttered sentences that may be input to the dialogue system.

(Second Embodiment)

According to the first embodiment, the user inputs an uttered sentence in accordance with a request sentence presented by the request output unit. In this case, the user does not actually have a dialogue with the dialogue system, thus uttered sentences that can be collected may be limited to uttered sentences that are input in a dialogue state where the user can be free to input uttered

sentences; for example, as in the case of an initial state in which the user starts a dialogue with the dialogue system. According to a second embodiment, a response sentence actually output by the dialogue system is

presented to the user along with a request sentence, allowing collection of uttered sentences provided in various dialogue states while the dialogue system is in use .

FIG. 7 schematically shows an uttered sentence

collection apparatus 700 according to the second

embodiment. As illustrated in FIG. 7, the uttered sentence collection apparatus 700 includes an uttered sentence input unit 101, an intention and slot extraction unit 102, an intention and slot management unit 703, an intention and slot bias determination unit 704, a request output unit 105, and a response output unit 706. The intention and slot extraction unit 102, the intention and slot management unit 703, and the intention and slot bias determination unit 704 are also referred to as the extraction unit 102, the management unit 703, and the bias determination unit 704, respectively. Elements in FIG. 7 which are identical to the corresponding elements in FIG. 1 are denoted by the same reference numerals in FIG. 1 and duplicate descriptions are omitted. Furthermore, the management unit 703 and bias determination unit 704 illustrated in FIG. 7 basically perform operations similar to the operations of the management unit 103 and bias determination unit 104 illustrated in FIG. 1. Thus, for the management unit 703 and the bias determination unit 704, operations different from the operations of the management unit 103 and the bias determination unit 104 will be described.

The response output unit 706 presents a response sentence to be output by the actual dialogue system, to the user. At this time, the response output unit 706 transmits a response ID indicative of the content of the response sentence to the management unit 703 and the bias

determination unit 704. The management unit 703 stores an intention and a slot extracted by the extraction unit 102 in association with the response ID received from the response output unit 706.

The bias determination unit 704 determines whether or not a bias is present in data accumulated in the management unit 703 and for which the response ID matches the response ID received from the response output unit 706. Upon determining that the data have a bias, the bias

determination unit 704 extracts an infrequent intention or slot from the data associated with the response ID, and transmits the infrequent intention or slot to the request output unit 105.

Now, operations of the uttered sentence collection apparatus 700 will be described with reference to FIG. 7 and FIG. 8.

FIG. 8 schematically shows a procedure for a process in which the uttered sentence collection apparatus 700 collects uttered sentences. Processing in steps S804, S805, and S806 illustrated in FIG. 7 is similar to the processing in steps S203, S204, and S205 illustrated in FIG. 2. Thus, the processing in steps S804, S805, and S806 will not be described below in detail.

In step S801 illustrated in FIG. 8, the response output unit 706 presents the response sentence to the user and determines a response ID indicative of the content of the response ID. The presentation of the response sentence may be performed before or after or simultaneously with the determination of the response ID. The response output unit 706 may present exclusively a response sentence to be actually output by the dialogue system in any dialogue state or may present, along with the response sentence to be presented, dialogues generated in a pseudo manner and corresponding to all dialogues exchanged until the response sentence appears . A method for selecting a response sentence and a response ID may be, for example, a method of acquiring a response sentence and a response ID from a database in which response sentences and response IDs are described. Which of the response sentences in the database and which of the response IDs in the database are to be selected may be determined in accordance with a method of random selection, a method of selecting from response sentences in dialogue states close to the initial state, a method of selecting a response ID for which the accumulated intentions or slots have a bias, or the like. In any of these cases, the response output unit 706 is configured to reference the management unit 703 or the bias determination unit 704 is configured to determine the presence or absence of a bias for each response ID to notify the response output unit 706 of a response ID involving a bias.

In step S802, the bias determination unit 704 extracts data for which the response ID matches the response ID . determined in step S801, from the data accumulated in the management unit 703. The bias determination unit 704 then determines whether or not the extracted data (at least one of the intention data and the slot data) have a bias. When the data are determined to have a bias, the process

proceeds to step S803.

In step S803, the bias determination unit 704 extracts at least one of an infrequent intention and an infrequent slot from the data for which the response ID matches the response ID determined in step S801. In step S804, the request output unit 105 generates a request sentence requesting input of an uttered sentence related to the at least one of the intention and slot extracted by the bias determination unit 704, and presents the request sentence to the user.

When the data are determined not to have a bias, the process proceeds to step S805. In step S805, the request output unit 105 creates and presents a request sentence requesting free input of an uttered sentence, to the user.

The processing in step S804 or step S805 results in presentation of both the response sentence and the request sentence to the user. In this state, the user inputs an uttered sentence. The input uttered sentence is

accumulated in the storage unit (not illustrated in the drawings) and transmitted to the extraction unit 702. In step S806, the extraction unit 702 extracts an intention and a slot from the uttered sentence input by the user. In step S807, the management unit 703 stores the intention and slot extracted in step S806 in association with the

response ID determined in step S801.

The second embodiment uses the response ID as

information indicative of the content of the response sentence. The information indicative of the content of the response sentence may be any information such as the response sentence itself or a dialogue state in the actual dialogue system provided that the information allows determination of which response sentence has been output.

As described above, the uttered sentence collection apparatus according to the second embodiment presents the user with the response sentence actually output by the dialogue system and the request sentence requesting input of an uttered sentence related to an infrequent intention or slot, to enable collection of uttered sentences provided in various dialogue states while the dialogue system is in use .

(Third Embodiment)

The second embodiment presents a response sentence output by the dialogue system. According to a third embodiment, the actual dialogue system is operated to allow uttered sentences to be collected. Thus, the user inputs an uttered sentence in accordance with the flow of

dialogues, allowing collection of uttered sentences with the flow of dialogues taken into account, that is,

collection of uttered sentences that are likely to be input while the system is in operation.

FIG. 9 illustrates a part of dialogues designed for a dialogue system that searches for restaurants. A response sentence is set for each node. At each node, a response sentence is output, the user inputs an uttered sentence, and to which of the nodes the dialogue transitions varies depending on the content of the uttered sentence . In the third embodiment, the node is referred to as a dialogue state, and a transition to the node is referred to as a dialogue transition.

For example, a response sentence "What kind of a restaurant do you want to go?" is output in a dialogue state 901. When, in response, the user inputs an uttered sentence "I want to go to a Japanese restaurant", a

dialogue transition to a dialogue state 902 occurs.

Specifically, the dialogue system extracts an intention and a slot from the uttered sentence "I want to go to a

Japanese restaurant", and the state of the dialogue system transitions to the dialogue state 902 based on the

extracted intention and slot. When the user inputs an uttered sentence "Inexpensive restaurant" in the dialogue state 901, a dialogue transition to a dialogue state 903 occurs. In FIG. 9, two state transition destinations from the dialogue state 901 are illustrated. However, more state transition destinations may actually be set.

A problem may occur when a bias occurs in an intention or a slot extracted from an uttered sentence input in a dialogue state (for example, a dialogue state 902) that is not the initial state (dialogue state 901) and the uttered sentence in the dialogue state needs to be collected. For example, a bias may occur when an intention or slot included in an uttered sentence "Limit the restaurants to those where alcohol is served" , which is input in response to a response sentence "X restaurants have been found.

Which do you like best?" in a dialogue state 904, has a smaller frequency than the other intentions or slots. In the third embodiment, a request for input of such an uttered sentence as allows a dialogue state with a bias to be reached is issued to the user at the point in time of the initial state, enabling collection of an uttered sentence for a dialogue state with a bias in intention or slot. For example, in a dialogue state 903 where the dialogue system responds with a sentence "Do you prefer a single-friendly restaurant?", the user is requested to input an uttered sentence including an intention

corresponding to "Yes". Then, a dialogue state 904 with a bias in intention or slot can be reached.

FIG. 10 schematically illustrates an uttered sentence collection apparatus 1000 according to the third

embodiment. The uttered sentence collection apparatus 1000 comprises an uttered sentence input unit 101, an intention and slot extraction unit 102, an intention and slot

management unit 1003, an intention and slot bias

determination unit 1004, a request output unit 1005, a response output unit 1006, and a dialogue management unit 1007. The intention and slot extraction unit 102, the intention and slot management unit 1003, and the intention and slot bias determination unit 1004 are also referred to as the extraction unit 102, the management unit 1003, and the bias determination unit 1004, respectively. Elements in FIG. 10 which are identical to the corresponding

elements in FIG. 1 are denoted by the same reference numerals in FIG. 1 and duplicate descriptions are omitted. Furthermore, the management unit 1003, bias determination unit 1004, and request output unit 1005 illustrated in FIG. 10 perform operations basically similar to the

operations of the management unit 103, bias determination unit 104, and request output unit 105 illustrated in

FIG. 1. Thus, for the management unit 1003, bias

determination unit 1004, and request output unit 1005, operations different from the operations of the management unit 103 and the bias determination unit 104 will be described.

The dialogue management unit 1007 has functions of a dialogue system to perform dialogue processing based on an intention and a slot extracted by the extraction unit 102. The dialogue processing includes a search using a received slot, determination of a response ID indicative of a content to be output as a response, and a transition of the dialogue state. The determined response ID and the result of the search are transmitted to the response output unit 1006. Moreover, a dialogue state present before a

transition (that is, the dialogue state present when an uttered sentence is input) is transmitted to the management unit 1003 as a pre-transition dialogue state. A dialogue state succeeding the transition is transmitted to the bias determination unit 1004 as a post-transition dialogue state. In some implementations, the post-transition dialogue state is transmitted to the request output unit 1005. The information indicative of the content to be output as a response is not limited to the response ID but may be any information that allows a response sentence itself or a response sentence for the post-transition dialogue state or the like to be uniquely determined.

The response output unit 1006 creates a response sentence in accordance with the response ID and the result of the search received from the dialogue management unit 1007 and presents the response sentence to the user. A method for creating a response sentence may be, like the method of creating a request sentence executed by the request output unit 105, an existing method such as a rule- based creation method, a method of embedding the result of the search into a part of the sentence to create a response sentence, or a statistical generation method.

The management unit 1003 stores the intention and slot extracted by the extraction unit 102 in association with the pre-transition dialogue state received from the

dialogue management unit 1007. The extraction unit 102 may transmit the extracted intention and slot to the management unit 1003 directly or via the dialogue management unit 1007.

The bias determination unit 1004 determines whether or not any route is present which reaches a dialogue state with a bias in at least one of the intention data and slot data accumulated in the management unit 1003. That is, the bias determination unit 1004 determines whether or not a dialogue state with a bias in at least one of the intention data and slot data accumulated in the management unit 1003 is present up ahead on a route from the pre-transition dialogue state received from the dialogue management unit 1007. The bias determination unit 1004 also determines that such a route is present when the dialogue state with a bias matches the post-transition dialogue state. Upon determining that such a route is present, the bias determination unit 1004 extracts an infrequent intention or slot and a corresponding dialogue state to the request output unit 1005. The bias determination unit 1004 may also transmit a dialogue state succeeding execution of dialogue processing by the dialogue management unit 1007, to the request output unit 1005.

The request output unit 1005 matches the post- transition dialogue state received from the dialogue management unit 1007 against the dialogue state with a bias. When the dialogue states are the same, the request output unit 105 creates a request sentence requesting the user to input an uttered sentence related to an infrequent intention or an infrequent slot associated with the

dialogue state with a bias and presents the request sentence to the user. When the dialogue states are different, the request output unit 1005 extracts an

intention or a slot needed to transition from the post- transition dialogue state to the dialogue state with a bias. The request output unit 1005 creates a request sentence requesting the user to input an uttered sentence related to the intention or slot to present the request sentence to the user.

Now, operations of the uttered sentence collection apparatus 1000 will be described with reference to FIG. 10 and FIG. 11.

FIG. 11 schematically illustrates a procedure for a process in which the uttered sentence collection apparatus 1000 collects uttered sentences. Processing in steps S1105, S1107, and S1108 illustrated in FIG. 11 is similar to the processing in steps S203, S204, and S205 illustrated in FIG. 2. Thus, the processing in steps S1105, S1107, and S1108 will not be described below in detail.

When the uttered sentence collection apparatus 100 is activated, the dialogue state of the dialogue management unit 1007 starts at the initial state. In step S1101 in FIG. 11, the response output unit 1006 receives a response ID for the initial state and outputs a response sentence according to the response ID.

In step S1102, the bias determination unit 1004 determines whether any dialogue state with a bias in at least one of the intention data and slot data is present and whether any route from the current dialogue state is present which reaches a dialogue state with a bias. The bias determination unit 1004 also determines that such a route is present when the current dialogue state is a dialogue state with a bias. When the bias determination unit 1004 determines that no such route is present, the process proceeds to step S1107. In step S1107, the request output unit 105 creates and presents a request sentence requesting the user to be free to input an uttered

sentence, to the user.

On the other hand, when the bias determination unit 1004 determines that such a route is present, the process proceeds to step S1103. In step S1103, the bias determination unit 1004 extracts at least one of an

infrequent intention and an infrequent slot along with a corresponding dialogue state. In step S1104, the request output unit 1005 determines whether or not the dialogue state extracted by the bias determination unit 1004 matches the current dialogue state of the dialogue management unit 1007. When these dialogue states match each other, the process proceeds to step S1105. In step S1105, the request output unit 1005 creates a request sentence requesting the user to input an uttered sentence related to an infrequent intention or slot in the dialogue state with a bias, and presents the request sentence to the user.

When the dialogue states are determined to fail to match each other, the process proceeds to step S1106. In step S1106, the request output unit 1005 extracts an intention or a slot needed to transition from the dialogue state of the dialogue management unit 1007 to the dialogue state with a bias. The request output unit 1005 creates a request sentence requesting the user to input an uttered sentence related to the intention or slot to present the request sentence to the user.

The processing in step S1105 or step S1106 or step S1107 results in presentation of both the response sentence and the request sentence to the user. In this state, the user inputs an uttered sentence. The input uttered

sentence is accumulated in the storage unit (not

illustrated in the drawings) and transmitted to the extraction unit 102. In step S1108, the extraction unit 102 extracts an intention and a slot from the uttered sentence input by the user. In step S1109, the management unit 1003 stores the intention and slot extracted in step S1108 in association with the current dialogue state.

In step S1110, the dialogue management unit 1007 performs dialogue processing based on the intention and slot extracted by the extraction unit 102. In step Sllll, the response output unit 1006 outputs a response sentence corresponding to a response ID resulting from the dialogue processing. In step S1112, the dialogue management unit 1007 determines whether or not the dialogues are complete. When the dialogues are complete, the process is ended.

When the dialogues are not complete, the process returns to step S1102.

When whether or not the intentions or slots

accumulated in the management unit 1003 have a bias is determined or an infrequent intention or slot is extracted, the bias determination unit 1004 may determine the presence or absence of a bias and perform the extraction for each of the dialogue states or for all of the intentions or slots in all of the states.

To determine whether or not a dialogue state with a bias in intention or slot is reached by advancing the dialogues from the post-transition dialogue state, the bias determination unit 1004 references a transition rule for the dialogue state. Furthermore, to extract an intention or a slot needed to transition to a dialogue state with a bias, the request output unit 1005 references the

transition rule for the dialogue state. The transition rule for the dialogue state describes information

indicating to which of the dialogue states a certain dialogue state transitions when a certain intention or slot is extracted in the dialogue state. In general, the dialogue management unit 1007 is provided with the

transition rule for the dialogue state. The bias

determination unit 1004 and the request output unit 1005 references the transition rule for the dialogue state. The transition rule for the dialogue state may be prepared independently for the bias determination unit 1004 and the request output unit 1005. Processing executed by the bias determination unit 1004 may be implemented by searching for a route by following dialogue states that can be

transitioned to and from one another, from the post- transition dialogue state to a target dialogue state.

Processing executed by the request output unit 1005 may be implemented by extracting an intention or a slot that allows implementation of the first transition on the route of the dialogue transition resulting from the processing executed by the bias determination unit 1004. The

transition rule for the dialogue state may be manually created or may be statistically created based on learning data. A statistically created transition rule may in principle enable transition from all the dialogue states to all the dialogue states. In that case, enabling the transition to all the dialogue states can be avoided by- providing a threshold for a transition probability value and utilizing only information on dialogue state

transitions with a transition probability value equal to or larger than the threshold in following the dialogue states. The threshold may be fixed or may vary depending on the dialogue state.

A plurality of routes may be present which leads to a dialogue state with a bias in intention or slot, as in the example in FIG. 9. In that case, a method for selecting such a route may be a method of randomly selecting a route, a method of selecting a route with a smaller number of procedural steps needed until the dialogue state is

reached, a method of selecting a route for which intentions or slots have a low input frequency, or the like. The input frequency of intentions or slots may be the

cumulative frequency of intentions or slots measured from a post- transition dialogue state until the target dialogue state is reached or the frequency of the intention or slot needed for the post- transition dialogue state to transition to the next dialogue state .

Furthermore, the uttered sentence collection apparatus 100 is likely to collect uttered sentences for dialogue states at which any dialogues arrive (for example, the initial state) . Consequently, as a whole, the intentions or slots are likely to have a bias. Thus, uttered sentences corresponding to an intention or a slot for a dialogue state with a low arrival frequency can be

efficiently collected by requesting, when a dialogue state with a bias in intention or slot is reached, the user to input an uttered sentence for that dialogue state a plurality of consecutive times.

FIG. 12 illustrates an operation procedure for the uttered sentence collection apparatus 1000 designed such that the user inputs an uttered sentence in a dialogue state with a bias in intention or slot a plurality of times. Processing in steps S1201 to S1209 and S1212 to S1214 illustrated in FIG. 12 is identical to the processing in steps S1101 to S1112 illustrated in FIG. 11. Thus, description of the processing in steps S1201 to S1209 and S1212 to S1214 is omitted.

When the user inputs an uttered sentence, the input uttered sentence is accumulated in the storage unit (not illustrated in the drawings) and transmitted to the

extraction unit 102. In step S1208, the extraction unit 102 extracts an intention and a slot from the uttered sentence input by the user. In step S1209, the management unit 1003 accumulates the intention and slot extracted in step S1108 in association with the current dialogue state.

In step S1210, the bias determination unit 1004 determines whether or not an uttered sentence with the same content needs to be re-input. This determination method may be a method of predetermining the number of times that an uttered sentence for the same dialogue state is

collected, a method of collecting as many uttered sentences as possible the number of which is approximately equivalent to a certain percentage of the total frequency of

intentions or slots collected in other dialogue states, or a method of collecting uttered sentences until a bias in intentions or slots accumulated in the management unit 1003 is corrected to some degree. When the bias determination unit 1004 determines that the uttered sentence needs to be re-input, the process proceeds to step S1211. In step S1211, the request output unit 1005 requests the user to input an uttered sentence with the same intention or slot. Subsequently, the uttered sentence input unit 101 waits to receive an uttered sentence input by the user. When the bias determination unit 1004 determines that no uttered sentence needs to be re-input, the process proceeds to step S1212 where the dialogue management unit 1007 executes dialogue processing.

When the method illustrated in FIG. 12 is implemented, whether a plurality of input uttered sentences has

different expressions is desirably checked. Determination of whether the uttered sentences have different expressions may be, for example, a method of checking whether the uttered sentences are exactly the same.

As described above, the uttered sentence collection apparatus 1000 according to the third embodiment operates the actual dialogue system to collect uttered sentences. This enables collection of uttered sentences that are likely to be collected while the dialogue system is in operation with the flow of dialogues taken into account.

(Fourth Embodiment)

In the third embodiment, an infrequent intention or slot is extracted to allow control of a dialogue state to be reached. When dialogue data with various dialogue transitions are to be collected, the arrival frequencies of dialogue states themselves may be accumulated and a request for an uttered sentence may be made to the user based on a bias in the arrival frequencies of the dialogue states. Collection of such dialogue data enables a statistical dialogue management model to be created.

For example, in the flow of dialogues illustrated in FIG. 13, a dialogue state 903 where the system responds with the sentence "Do you prefer a single- friendly

restaurant?" has a much lower arrival frequency than a dialogue state 902 where the system responds with a

sentence "What kind of a Japanese food do you like?". In this case, according to the fourth embodiment, a request to input an uttered sentence with the intention or slot corresponding to "Inexpensive restaurant" is issued in a dialogue state 901 with the sentence "What kind of a restaurant do you want to go?".

FIG. 14 schematically illustrates an uttered sentence collection apparatus 1400 according to the fourth

embodiment. As illustrated in FIG. 14, the uttered sentence collection apparatus 1400 includes an uttered sentence input unit 101, an intention and slot extraction unit 102, a request output unit 1405, a response output unit 1006, a dialogue management unit 1407, a dialogue state management unit 1408, and a dialogue state bias determination unit 1409. The intention and slot extraction unit 102, the dialogue state management unit 1408, and the dialogue state bias determination unit 1409 are also referred to as the extraction unit 102, the management unit 1408, and the bias determination unit 1409, respectively. Elements in FIG. 14 which are identical to the

corresponding elements in FIG. 1 and FIG. 10 are denoted by the same reference numerals in FIG. 1 and FIG. 10 and duplicate descriptions are omitted. Furthermore, the dialogue management unit 1407 illustrated in FIG. 14 performs operations basically similar to the operations of the dialogue- management unit 1007 illustrated in FIG. 10. The dialogue management unit 1407 is different from the dialogue management unit 1107 in that the dialogue

management unit 1407 transmits a dialogue state preceding a dialogue transition resulting from dialogue processing to the management unit 1408 as a pre-transition dialogue state and transmits a dialogue state succeeding the dialogue transition to the bias determination unit 1409 as a post- transition dialogue state. In some implementations, the dialogue management unit 1407 transmits the post-transition dialogue state to the request output unit 1405. The management unit 1408 accumulates data on the pre- transition dialogue state received from the dialogue management unit 1407. The bias determination unit 1409 determines whether or not the data accumulated in the management unit 1408 have a bias. Upon determining that the data have a bias, the bias determination unit 1409 determines whether or not any route from the current dialogue state is present which reaches a dialogue state with a low arrival frequency. That is, the bias

determination unit 1409 determines whether or not any dialogue state with a low arrival frequency is present up ahead on a route from the current dialogue state. When a dialogue state with a bias matches the current dialogue state, the bias determination unit 1409 determines that no such route is present, unlike in the third embodiment.

Upon determining that such a route is present, the bias determination unit 1409 extracts and transmits a dialogue state with a low arrival frequency to the request output unit 1405. The bias determination unit 1409 may

simultaneously transmit a dialogue state succeeding

execution of dialogue management by the dialogue management unit 1407, to the request output unit 1405.

The request output unit 1405 utilizes the post- transition dialogue state and the dialogue state with a low arrival frequency both received from the dialogue

management unit 1407 to extract at least one of an

intention and a slot needed to transition to the dialogue state with a low arrival frequency. The request output unit 1405 creates a request sentence requesting the user to input an uttered sentence related to at least one of the intention and slot and presents the request sentence to the user.

Now, operations of the uttered sentence collection apparatus 1400 will be described with reference to FIG. 14 and FIG. 15.

FIG. 15 illustrates a procedure example for a process in which the uttered sentence collection apparatus 1400 collects uttered sentences. Processing in steps S1505 and S1506 illustrated in FIG. 15 is similar to the processing in steps S204 to S205 illustrated in FIG. 2. Processing in steps S1501, S1504, and S1508 to S1510 illustrated in

FIG. 15 is similar to the processing in steps S1101, S1106, and S1110 to S1112 illustrated in FIG. 11. Thus,

description of the processing in steps S1501, S1504, S1505, S1506, and S1508 and S1510 is omitted.

In step S1501 in FIG. 15, the response output unit 1006 presents a response sentence to the user. In step S1502, the bias determination unit 1409 determines whether or not a dialogue state with a low arrival frequency- included in dialogue states, accumulated in the management unit 1408 is present up ahead of the current dialogue state. At this time, when the current dialogue state is the dialogue state with a low arrival frequency, the bias determination unit 1409 determines that the dialogue state with a low arrival frequency is not present up ahead of the current dialogue state. When the management unit 1408 determines that the dialogue state with a low arrival frequency is present up ahead of the current dialogue state, the process proceeds to step S1503. Otherwise, the process proceeds to step S1505 where the request output unit 1405 presents the user with a request sentence

requesting the user to be free to input an uttered

sentence .

In step S1503, the bias determination unit 1409 extracts the dialogue state with a low arrival frequency from the management unit 1408. In step S1504, the request output unit 1405 presents a request sentence requesting arrival at the dialogue state with a low arrival frequency. A method for determining a dialogue state with a low arrival frequency may be similar to the method in which the intention and slot management unit extracts an infrequent intention or slot in the first to third embodiments.

The processing in step S1504 or step S1505 results in presentation of both the response sentence and the request sentence to the user. In this state, the user inputs an uttered sentence. The input uttered sentence is

accumulated in the storage unit (not illustrated in the drawings) and transmitted to the extraction unit 102. In step S1506, the extraction unit 102 extracts an intention and a slot from the uttered sentence input by the user.

In step S1507, the management unit 1408 accumulates data on the pre-transition dialogue state held by the dialogue management unit 1407. The accumulated dialogue state is a dialogue state preceding execution of a dialogue transition based on the intention or slot extracted from the input uttered sentence .

In step S1508, the dialogue management unit 1407 executes dialogue processing utilizing the intention and slot extracted by the extraction unit 102. In step S1509, the response output unit 1006 outputs a response sentence corresponding to a response ID resulting from the dialogue processing. In step S1510, the dialogue management unit 1407 determines whether or not the dialogues are complete. When the dialogues are complete, the process is ended.

When the dialogues are not complete, the process returns to step S1502.

In the above-described example, the management unit 1408 accumulates data on dialogue states. Alternatively, the management unit 1408 may accumulate data on transitions between the dialogue states. Accumulation of dialogue transitions enables the manner of transition of dialogues between the user and the dialogue system to be more

appropriately understood, allowing infrequently followed routes to be accurately determined. In this case, the dialogue management unit 1407 saves, to the management unit 1408, dialogue transitions generated when dialogue

processing is executed based on the intention and slot extracted from the input uttered sentence. Upon outputting a request sentence, the dialogue management unit 1407 extracts a dialogue transition with a low frequency which is utilized before the bias determination unit 1409 can reach the target on a route from the post-transition dialogue state. Subsequently, the^' request output unit 1405 extracts an intention or a slot needed to advance the dialogue so as to generate the above-described dialogue transition in the post-transition dialogue state, and requests the user to input an uttered sentence with the intention or slot.

Furthermore, the management unit 1408 may accumulate a dialogue state succeeding a dialogue transition instead of the dialogue state preceding a dialogue transition. This can be achieved by replacing step S1508 with step S1509 in FIG. 15.

As described above, the uttered sentence collection apparatus according to the fourth embodiment can collect dialogue data for various dialogue state transitions by requesting the user to provide an uttered sentence needed to reach a dialogue state with a low arrival frequency when data on the dialogue state have a bias.

According to at least one of the embodiments described above, there is provided an uttered sentence collection apparatus and method which allow a sufficient number and variety of uttered sentences to be collected.

The instructions provided in the process procedure illustrated in the above-described embodiments can be executed based on a program that is software. Effects similar to the effects of the uttered sentence collection apparatuses in the above-described embodiments can be exerted by using a general-purpose computer system to prestore and read in this program. The instructions described above in the embodiments are stored in a magnetic disk (a flexible disk, a hard disk, or the like) , an optical disc (CD-ROM, CD-R, CD-RW, DVD-ROM, DVD+R, DVD+RW, or the like) , a semiconductor memory, or a similar

recording medium as a program that can be executed by a computer. The recording medium may be in any storage form provided that the recording medium can be read by a

computer or an embedded system. Effects similar to the effects of the uttered sentence collection apparatuses in the above-described embodiments can be exerted by using the computer to read in the program from the recording medium and to allow a CPU to carry out the instructions described in the program based on the program. Of course, in

acquiring or reading in the program, the computer may acquire or read in the program through a network.

Furthermore, the processes implementing the

embodiments may partly be executed by an OS (Operating System) operating on the computer based on the instructions in the program installed from the recording medium into the computer or the embedded system, or MW (Middle Ware) such as database management software or a network.

Moreover, the recording medium according to the embodiments is not limited to a medium independent of the computer or the embedded system but includes a recording medium in which the program communicated through a LAN, the Internet, or the like is downloaded and stored or

temporarily stored.

Furthermore, the recording medium according to the embodiments is not limited to a single recording medium but includes a plurality of media used to execute the

processing according to the embodiments. The media may have any configuration.

The computer or embedded system according to the embodiments executes the processes according to the

embodiments based on the program stored in the recording medium. The computer or embedded system according to the embodiments may have any configuration and may be, for example, a single apparatus such as a personal computer or a microcomputer, or a system with a plurality of

apparatuses connected together via a network.

Furthermore, the computer according to the embodiments is not limited to a personal computer but includes an arithmetic processing apparatus or a microcomputer included in information processing equipment. The computer

according to the embodiments is a general term for

equipment and apparatuses that can implement the functions according to the embodiments based on the program.

While certain embodiments have been described, these embodiments have been presented by way of example only, and are not intended to limit the scope of the inventions.

Indeed, the novel embodiments described herein may be embodied in a variety of other forms; furthermore, various omissions, substitutions and changes in the form of the embodiments described herein may be made without departing from the spirit of the inventions. The accompanying claims and their equivalents are intended to cover such forms or modifications as would fall within the scope and spirit of the inventions.

Claims

C L A I M S

1. An uttered sentence collection apparatus

comprising:

an input unit configured to input an uttered sentence from a user;

an extraction unit configured to extract, from the uttered sentence, at least one of an intention and a slot, the intention being an intention of the uttered sentence, the slot being information supplementing the intention; a management unit configured to accumulate data on the at least one of the intention and the slot;

a bias determination unit configured to determine whether or not the data have a bias and to acquire at least one of an infrequent intention and an infrequent slot from the data on determining that the data have the bias; and a request output unit configured to request input of an uttered sentence related to the at least one of the infrequent intention and the infrequent slot acquired by the bias determination unit.

2. The apparatus according to claim 1, wherein the bias determination unit determines that the data have the bias in at least one of a case where an intention with a relative frequency lower than a first threshold is present when a total frequency of intentions included in the data reaches a specified value or more and a case where a slot with a relative frequency lower than a second threshold is present when a total frequency of slots included in the data reaches a specified value or more.

3. The apparatus according to claim 1, wherein the bias determination unit determines that the data have the bias in at least one of a case where a difference in frequency between a most frequent intention and a least frequent intention is equal to or larger than a first threshold and a case where a difference in frequency between a most frequent slot and a least frequent slot is equal to or larger than a second threshold.

4. The apparatus according to claim 1, further comprising a response output unit configured to present a response sentence to be output by a dialogue system.

5. The apparatus according to claim 4, wherein the response output unit presents the response sentence to the user before the user inputs the uttered sentence.

6. The apparatus according to claim 4, wherein the response output unit further presents a content of

dialogues occurred until the response sentence appears.

7. The apparatus according to claim 1, further comprising :

a dialogue management unit configured to perform dialogue processing based on at least one of the intention and the slot extracted by the extraction unit; and

a response output unit configured to present a

response sentence associated with the dialogue processing, wherein the management unit stores the at least one of the intention and the slot extracted by the extraction unit in association with a dialogue state of the dialogue management unit which is present when the uttered sentence is input,

the bias determination unit determines whether or not the data have a bias for each dialogue state, determines whether or not a route is present which reaches a dialogue state with the bias on determining that the data have the bias, and acquires at least one of an infrequent intention and an infrequent slot associated with the dialogue state with the bias on determining that the route is present, and the request output unit presents a request sentence requesting input of an uttered sentence related to at least one of the intention and the slot acquired by the bias determination unit when the dialogue state with the bias matches a dialogue state succeeding execution of the dialogue processing by the dialogue management unit, and requests input of an uttered sentence related to at least one of an intention and a slot needed to reach the dialogue state with the bias when the dialogue state with the bias fails to match the dialogue state succeeding execution of the dialogue processing by the dialogue management unit.

8. The apparatus according to claim 7, wherein, when a plurality of candidates for at least one of the intention and slot needed to reach the dialogue state with the bias is present, the request output unit selects at least one of a lowest intention and a lowest slot from the candidates, and requests input of an uttered sentence related to the selected at least one of the intention and the slot.

9. The apparatus according to claim 7, wherein the bias determination unit determines whether or not a

plurality of uttered sentences need to be input in one dialogue state, and

the request output unit requests re-input of an uttered sentence related to the at least one of the

intention and the slot acquired by the bias determination unit when the bias determination unit determines that a plurality of uttered sentences need to be input in the one dialogue state.

10. The apparatus according to claim 1, wherein the intention is information indicative of a content that the uttered sentence is intended to roughly communicate.

11. An uttered sentence collection method comprising: inputting an uttered sentence from a user;

extracting, from the uttered sentence, at least one of an intention and a slot, the intention being an intention of the uttered sentence, the slot being information

supplementing the intention;

accumulating data on the at least one of the intention and the slot;

determining whether or not the data have a bias and acquiring at least one of an infrequent intention and an infrequent slot from the data on determining that the data have the bias ; and

requesting input of an uttered sentence related to the at least one of the infrequent intention and the infrequent slot acquired by the bias determination unit.

12. A non-transitory computer readable medium

including computer executable instructions, wherein the instructions, when executed by a processor, cause the processor to perform a method comprising:

inputting an uttered sentence from a user;

supplementing the intention;

accumulating data on the at least one of the intention and the slot;

13. An uttered sentence collection apparatus

comprising:

an input unit configured to input an uttered sentence from a user;

an extraction unit configured to extract, from the uttered sentence, at least one of an intention and a slot, the intention being an intention of the uttered sentence, the slot being information supplementing the intention; a dialogue management unit configured to perform dialogue processing based on the at least one of the intention and the slot;

a response output unit configured to present a

response sentence associated with the dialogue processing; a management unit configured to accumulate data on a dialogue state of the dialogue management unit;

a bias determination unit configured to determine whether or not the data have a bias and to acquire an infrequent dialogue state from the data on determining that the data have the bias ; and

a request output unit configured to request input of an uttered sentence related to at least one of an intention and a slot needed to reach the infrequent dialogue state.

14. The apparatus according to claim 13, wherein the management unit accumulates data on a transition between dialogue states resulting from the dialogue processing, and the bias determination unit determines whether or not the transition has a bias.

15. An uttered sentence collection method comprising: inputting an uttered sentence from a user;

supplementing the intention;

performing dialogue processing based on the at least one of the intention and the slot;

presenting a response sentence associated with the dialogue processing;

accumulating data on a dialogue state;

determining whether or not the data have a bias and acquiring an infrequent dialogue state from the data on determining that the data have the bias; and

requesting input of an uttered sentence related to at least one of an intention and a slot needed to reach the infrequent dialogue state .

16. A non-transitory computer readable medium

inputting an uttered sentence from a user;

supplementing the intention;

presenting a response sentence associated with the dialogue processing;

accumulating data on a dialogue state;

determining whether or not the data have a bias and acquiring an infrequent dialogue state from the data on determining that the data have the bias; and requesting input of an uttered sentence related to at least one of an intention and a slot needed to reach the infrequent dialogue state .