JP3883066B2 - Voice dialogue system and method, voice dialogue program and recording medium therefor - Google Patents

Description

【００６０】
今取り上げている例に戻る。現在のシステム理解状態（３）が正しい確率は、場所属性の認識精度０.６０と日属性の認識精度０.９５の積で与えられるとする。したがって、ｐ＝０.６０×０.９５＝０.５７でシステム理解状態（３）は正しい。ここでは、誤った応答を行った後の対話で正しいシステム応答が生成される確率ｑは、Pと等しいと仮定して、ｑ＝ｐ＝０.５７とする。誤った応答「明日の香川県は晴れです」に含まれる自立語は３個であるので、Ｌ１＝３である。誤った応答の後に続く各対話でやり取りされる自立語の期待値Ｌ２については、事前に収集された対話データの中で天気の問い合わせを行う対話に含まれる自立語の平均数をとり、Ｌ２＝１０となったと仮定する。（５）式より、対話手順Dのコストは、３＋０.４３・１０／０.５７＝３＋７.５４＝１０.５４となる。
【００６１】
次に、試行的な確認型の対話手順Ｂの対話コストについて説明する。対話手順Bは、ここでは、場所属性を確認し、その後で、確定した場所における明日の天気を応答するという手順である。まず、場所属性の値を確定するまでにやり取りされる自立語の期待数は、（４）式で、ｍ＝１、ｒ＝０.６５であるから、２・１／０.６５−１＋１＝３.０８となる。
【００６２】
場所属性が確定した後の状況では、日属性の値のみが未確定である。日属性の認識精度は０.９５であると仮定しているので、この時点でシステム理解状態が正しい確率は０.９５となる。この状況は、日属性の値が正しいと仮定してシステム応答を行うという試行的な即時応答型の対話手順を踏んでいるのと同じ状況である。したがって、場所属性の値が確定した後の対話に含まれる自立語の期待数は、（５）式において、ｐ＝ｑ＝０.９５、Ｌ１＝３、Ｌ２＝１０として、３＋０.０５・１０／０.９５＝３＋０.５３＝３.５３となる。結局、対話手順Bの対話コストは、３.０８＋３.５３＝６.６１となる。
【００６３】
以上をまとめると、
(１)確定的な確認型の対話手順Ａの対話コスト：９.０２
(２)試行的な確認型の対話手順Bの対話コスト：６.６１
(３)確定的な即時応答型の対話手順Cの対話コスト：６００
(４)試行的な即時応答型の対話手順Dの対話コスト：１０.５４
となる。
【００６４】
このように、対話手順コスト計算部１４０が各対話手順のコストを計算した後、確認侯補生成部１５０が、現在のシステム理解状態に含まれる属性のうち、確定済みでない属性のすべての組み合わせを確認候補として生成する（ステップ２０８）。先のシステム理解状態（３）から、組み合わせは、（１）場所を確認、（２）日を確認、（３）場所と日を確認の３通りである。ここでは、簡単のため、確認候補として、場所属性のみを確認するという確認候補、場所属性と日属性を同時に確認するという確認候補が生成されたとする。
【００６５】
次に、確認候補コスト生成手段１６０は、確認候補の各々に関して、確認侯補を含むような対話手順の中で、最小のコストをもつ対話手順を選び、提供情報種別の確率を考慮して、コストの期待値を計算し、その期待値を確認候補のコストとして生成する（ステップ２０９）。ここでは、可能な提供情報種別は天気のみであり、その確率は１である。
場所属性のみの確認を含む対話手順は、試行的な確認型の対話手順Bのみである。したがって、場所属性のみを確認するという確認候補のコストは「６.６１」となる。
場所属性と日属性を同時に確認することを含む対話手順は、確定的な確認型の対話手順Ａのみである。したがって、場所属性と日属性を同時に確認するという確認候補のコストは「９.０２」となる。
【００６６】
続いて、出力部１８０が、コスト判定し（ステップ２１０）、その判定結果に基づいて確定文又は応答文を生成して出力する（ステップ２１１）。詳しくは、まず、すべての確認候補（ここでは、場所のみ確認、場所と日を確認）のコストと、すべての即時応答型の対話手順（ここでは、対話手順ＣとＤ）のコストを比較する（ステップ２１１０）。そして、即時応答型の対話手順のコストの方が小さかったならば、即時応答型の対話手順の中から、最小のコストをもつ即時応答型（確定的な即時応答型又は試行的な即時応答型）を決定し（ステップ２１１１）、その対話手順にしたがって情報を提供するための応答文を生成する（ステップ２１１２）。一方、確認候補のコストの方が小さかったならば、確認候補の中から最小のコストをもつものを決定し（ステップ２１１３）、その確認候補の確認（ここでは、場所を確認又は場所と日を確認）を行うための確認文を生成する（ステップ２１１４）。
【００６７】
現在の例の場合、結局、最小のコストは、場所属性のみを確認するという確認侯補のコスト６.６１である。出力部１７０は、現在のシステム理解状態（３）にしたがって、場所属性の値のみを確認するための確認文「香川県ですか？」を出力する。
【００６８】
この後、ユーザが「神奈川県です」という訂正文を入力し、システムはその訂正文を正しく認識し、次に、「神奈川県ですか」という場所属性の値を確認するための確認文を出力し、ユーザが「はい」という承認文を入力し、システムはその承認文が正しく認識したとする。このとき、システム理解状態は次のようになる。
｛＜場所、神奈川県、済＞、＜日、明日、未＞、＜情報種別、天気、済＞｝（６）
日属性のみが未確定である。したがって、この場合もステップ２０５以降の処理に進む。
【００６９】
ここで、対話手順生成部１３０は、再び可能な対話手順をすべて生成する（ステップ２０６）。ここでは、簡単のため、次の２つの対話手順を取り上げて、説明を続ける。
(５) 確定的な確認型の対話手順Ａ：
日属性の値を確定した後で、確定した日における神奈川県の天気を応答するという手順。
(６) 試行的な即時応答型の対話手順Ｄ：
確認を行わずに、神奈川県の明日の天気を応答するという手順。ここでは、明日という値は確定済みではないが、正しい値であると仮定して応答が出力される。
【００７０】
次に、対話手順コスト計算部１４０が、各対話手順のコストを計算する（ステップ２０７）。今、可能な提供情報種別は天気のみであり、その確率は１である。
まず、(５)の対話手順Ａの対話コストについて説明する。日属性を確定するまでに必要な自立語の期待数は、日属性の認識精度は０.９５であるので、式（４）で、ｍ＝１、ｒ＝０.９５として、２・１／０.９５−１＋１＝２.１１となる。日属性が確定した後に生成される応答に含まれる自立語は３である。したがって、対話手順Ｅの対話コストは、２.１１＋３＝５.１１となる。
次に、(６)の対話手順Ｄの対話コストについて説明する。日属性の認識精度は０.９５であるので、現在のシステム理解状態（６）が正しい確率は０.９５である。また、天気についての問い合わせの対話は、平均して１０個の自立語をやり取りするとしているので、対話手順の対話コストは、式（５）で、ｐ＝ｑ＝０.９５、Ｌ１＝３、Ｌ２＝１０とおいて、３＋０.０５・１０／０.９５＝３＋０.５３＝３.５３となる。
【００７１】
まとめると、
確定的な確認型の対話手順Ａのコスト：５.１１
試行的な即時応答型の対話手順Ｄのコスト：３.５３
となる。
【００７２】
次に、確認候補コスト生成部１５０が、ここでは日属性のみを確認するという確認候補のみを生成する（ステップ２０８）。この確認候補のコストは、対話手順Ｅのコストと等しく、５.１１となる。
【００７３】
文出力部１８０は、日属性のみを確認するという確認候補のコスト５.１１と、試行的な即時応答型の対話手順Ｄのコスト３.５３を比較し、小さいコストを与える対話手順Ｄを選択する（ステップ２１０、２１１）。この選択にしたがって、日属性を確認するための確認文を出力することなしに、「明日の神奈川県は晴れです」という応答文が出力されることになる。
【００７４】
以上の対話をまとめると次のようになる。
（文１） ユーザ ：「神奈川県の明日の天気を教えてください」
（文２） システム：「天気ですか？」
（文３） ユーザ ：「はい」
（文４） システム：「香川県ですか？」
（文５） ユーザ ：「神奈川県です」
（文６） システム：「神奈川県ですか？」
（文７） ユーザ ：「はい」
（文８） システム：「神奈川県の明日の天気は晴れです」
ここでの説明は、文（４）以降のシステムの行動に焦点を絞った説明である。
【００７５】
なお、図３の処理フローにおいて、ある時点で、ステップ２０４にて、システム理解状態に未確定の属性がなくなったことが判定されれば、出力部１７０では、当該システム理解状態に基づいて直ちに応答文を生成し出力することになる。
【００７６】
本音声対話システム１０では、認識精度の低い場所属性の値は確認するが、認識精度の高い日属性の値は確認しないという振る舞いを示す。どの属性の値を確認するかという判断は、対話手順のコストという客観的な基準によって決定されている。この方法により、ユーザは短い対話で必要な情報を手に入れることが可能となる。
【００７７】
なお、図２で示した音声対話システム１０における各部の一部もしくは全部の処理機能をコンピュータのプログラムで構成し、そのプログラムをコンピュータを用いて実行して本発明を実現することができること、あるいは、図３、図４で示した処理手順をコンピュータのプログラムで構成し、そのプログラムをコンピュータに実行させることができることは言うまでもない。また、コンピュータでその処理機能を実現するためのプログラム、あるいは、コンピュータにその処理手順を実行させるためのプログラムを、そのコンピュータが読み取り可能な記録媒体、例えば、ＦＤ、ＭＯ、ＲＯＭ、メモリカード、ＣＤ、ＤＶＤ、リムーバブルディスクなどに記録して、保存したり、提供したりすることができるとともに、インターネット等のネットワークを通してそのプログラムを配布したりすることが可能である。
【００７８】
【発明の効果】
以上説明したように、本発明の音声対話システムによれば、対話全体の長さを増大させることがないと判断される場合には、ユーザ問い合わせ内容の一部或いは全てに関して、確認を省くことができ、確認文の出力数の増大に伴うユーザ不満足度の増大を避けることができる。また、システムが誤認識のため誤った応答を行ったときには、ユーザが拒否発話を行うということを前提としていないので、ユーザに拒否発話を行うことを強制する必要がなくなり、ユーザ満足度が増す。
【図面の簡単な説明】
【図１】本発明のシステム環境を示す全体構成図である。
【図２】本発明による音声対話システムの一実施例の機能ブロック図である。
【図３】本発明の一実施例の処理フローチャートである。
【図４】図３におけるコスト判定の詳細処理フローチャートである。
【符号の説明】
１０ 音声対話システム
２０ データベース
３０ メモリ装置
４０ 通信回線
５０ 利用者端末
１００ 入力部
１１０ 文理解部
１２０ 提供情報種別決定部
１３０ 対話手順生成部
１４０ 対話手順コスト計算部
１５０ 確認候補生成部
１６０ 確認候補コスト生成部
１７０ 出力部
１８０ 制御部[0001]
BACKGROUND OF THE INVENTION
The present invention relates to an interactive information providing system. More specifically, when a user inputs a query about the contents of a database using voice, the system performs an exchange for confirming the contents of the query of the user, if necessary. The present invention relates to a voice dialogue system and method, a voice dialogue program, and a recording medium for the same, which provide a user with the contents of a database by voice according to the contents of a query that is recognized.
[0002]
[Prior art]
For example, when there is a database that stores various information such as weather information, TV program schedules, schedules, and transportation times, a voice dialogue system allows a user to recognize a sentence meaning a query about the contents of the database. When the voice is input through the device, the inquiry content recognized by the system is expressed as a system understanding state expressed as a set of triples indicating the attribute, the value of the attribute, and the value of the attribute. Hold.
[0003]
In such a voice interaction system, the result of recognizing the contents of the user inquiry is not always correct due to a voice recognition error. Therefore, if necessary, a confirmation sentence is output as a voice to the user in order to confirm all or part of the system understanding state. When the user inputs an approval sentence (for example, “Yes”) indicating confirmation of confirmation, the system records the confirmed contents of the system understanding state as confirmed. When the user inputs a correction sentence meaning correction for confirmation, the system updates the system understanding state according to the content of the correction sentence. After the exchange for confirmation, the system determines the contents of the user inquiry according to the system understanding state, and outputs a response sentence for providing information to the user.
[0004]
On the other hand, if no confirmation is made, immediately after the user inputs the inquiry, the system determines the contents of the user inquiry according to the system understanding state at that time, and outputs a response sentence for providing information to the user. It becomes.
[0005]
The advantage of confirmation is that if the user enters an approval sentence for the confirmation sentence for confirming the attribute value, the probability that the attribute value is correct will increase, so the system will convey the information required by the user more accurately. It is possible to do that. However, since the length of the dialogue is increased by performing the confirmation, the user's satisfaction may be reduced.
[0006]
Conventionally, as a technology for confirming the contents of user inquiries by the spoken dialogue system, the first technology for generating a response sentence according to only the attribute values that have been confirmed by the approval sentence from the user, the confirmation is completely Second technique not to perform, record the reliability of the speech recognition result obtained when recognizing the value of each attribute in the system understanding state, if the reliability is greater than a certain threshold, Assume that the third technique (for example, see Non-Patent Document 1) that does not check the attribute value, and that not only the attribute value that has been confirmed but also the value of the attribute that has not been confirmed are correct values. When a response sentence is generated, if the response sentence is incorrect, it is assumed that there is a rejection utterance from the user, and if a rejection utterance is received from the user, the next is the value of the confirmed attribute Is positive As a have value, the fourth technique of generating a response text (e.g., Non-Patent Documents 2 and 3, reference) is.
[0007]
[Non-Patent Document 1]
Komaya Kazunori and Kawahara Tatsuya, “Dialogue Management to Deal with Speech Recognition Errors in Mixed-initiated Dialogue”,
Proc. Of the 6th Annual Conference of the Language Processing Society, 336-339 (2000)
[Non-Patent Document 2]
Koji Dosaka, Yoshihito Yasuda, Kiyoaki Aikawa, “Efficient Spoken Dialogue Control under System Knowledge Restriction”,
Natural Language Processing, Vo1.9, No. 1 (2002) 43-63
[Non-Patent Document 3]
Koji Dosaka, Nobuhito Yasuda, Kiyoaki Aikawa, “Spoken Dialogue Control Based on Maximizing Information Transmission Efficiency”,
Proc. Of the 8th Annual Conference of the Language Processing Society, 260-263 (2002)
[0008]
[Problems to be solved by the invention]
In the voice interactive system, the result of recognizing the contents of the user inquiry is not always correct due to a voice recognition error. Therefore, in the speech dialogue system, a technique for responding as much as possible to the content of the user's inquiry has been developed even if a speech recognition error exists.
[0009]
In any of the first technique, the second technique, the third technique, and the fourth technique, the user's inquiry contents recognized by the system at each point of the dialogue are attribute and attribute It is represented as a set of triples of values and flags that indicate whether the values have been finalized. This triplet set is called a system understanding state.
[0010]
In the first technique, the system outputs a confirmation message for confirming the value of the recognized attribute to the user, and confirms when the approval sentence from the user, for example, the sentence “Yes” is recognized. Record in the system understanding state that the value of the attribute has been finalized. The system searches the database according to only the attribute values that have been confirmed, and outputs a response sentence for transmitting the search result to the user. Since this method outputs a response sentence according to only the value approved by the user, there is an advantage that the system is less likely to make an incorrect response according to the result of misrecognizing the user's inquiry content. While the output of confirmation sentences is effective to increase the reliability of the response sentence of the system, the number of confirmations increases the time required for the user to obtain the necessary information, which increases the degree of user dissatisfaction. It also causes the problem.
[0011]
The first conventional technique checks all the attribute values necessary to create a response sentence, and has not been devised to reduce the number of confirmation sentences, so there is a problem that the degree of user dissatisfaction increases. is there.
[0012]
The second conventional technique is a method in which no confirmation is performed. Unlike the first conventional technique, this method does not increase the degree of user dissatisfaction in that no confirmation text is output. However, since no confirmation is performed at all, the system may output a response sentence based on the erroneous recognition result of the user inquiry content. As a result, the user makes an inquiry again, and there is a problem that the length of the dialog until the user obtains the necessary information increases.
[0013]
The conventional third technique records the reliability of the speech recognition result of the attribute value, does not check the attribute value having a reliability greater than a certain threshold, and the reliability of the speech recognition result is low. This is sometimes done by checking. In this method, the threshold value is determined on the basis of improving the understanding rate of the user inquiry content. There is a problem that the standard of improving the understanding rate of the user's inquiry content does not necessarily result in shortening the length of dialogue required to obtain the information required by the user as much as possible.
[0014]
In the conventional fourth technique, at each point in the dialogue, the possible procedures for the dialogue until the response sentence is output are listed, and the dialogue is performed according to the dialogue procedure that minimizes the length of the dialogue. It is a method of determining the behavior of the system at each time point. As a procedure of dialogue, after confirmation, a definite confirmation type dialogue procedure that outputs a response sentence only according to the confirmed system understanding state, without understanding, understanding the system that has been confirmed at present A deterministic immediate response type interactive procedure that outputs a response sentence only according to the state, a response sentence is output according to the system understanding state including attribute values that have not been confirmed at present, and the user rejects the response sentence Consider a confirmation-needs-search type dialogue procedure in which a dialogue is retried when an error occurs, and a response sentence is output only according to the confirmed system understanding state in the redo dialogue. In the fourth technique, for each of these dialogue procedures, the dialogue length is estimated, and the dialogue procedure with the shortest dialogue length is selected. The next action of the system is determined according to the selected interaction procedure.
[0015]
In the fourth technology, if it is determined that the confirmation-needs search type interaction procedure can obtain necessary information in a shorter interaction than other types of interaction procedures, this confirmation requirement is required. A sex search type interactive procedure is selected, and confirmation of some attribute values can be omitted. At this time, it is also guaranteed that the total length of the dialogue does not increase compared to other dialogue procedures.
[0016]
The conventional fourth technique has the advantage that some confirmations can be omitted compared to the conventional first technique, and the erroneous response sentence in the conventional second technique is output many times. As a result, the problem that the length of the dialogue increases can be overcome. In the third technology, the criterion for selecting a threshold is a criterion that improves the understanding rate of user queries. In the fourth technology, the length of the dialogue is shortened when following each dialogue procedure. Standards are used.
[0017]
However, the fourth technique is based on the premise that the user will make a refusal utterance when the system gives an incorrect response, and there is a problem that it cannot be applied if this premise is not satisfied. In an actual dialogue, the user may give a refusal utterance such as “No” when the system gives an incorrect response, but without making an explicit refusal utterance, the user can ask again. It may be possible to repeat the dialogue repeatedly. Also, forcing the user to make a rejection utterance when the system gives an incorrect response puts a burden on the user and increases the degree of user dissatisfaction.
[0018]
In this way, the fourth technology has the advantage of overcoming the problems of the first, second, and third technologies, but on the premise that the user makes a refusal utterance in response to an incorrect response of the system. However, there is a problem that this assumption does not necessarily hold in an actual dialogue.
[0019]
The present invention has been made in view of the above-described problems of the prior art in a voice interaction system, and is based on the premise that a user makes a rejection utterance when the system makes a wrong response due to a recognition error. The purpose is to reduce the output of the confirmation text for confirming the system understanding state as much as possible, and to suppress the increase in the length of the dialog due to the re-execution of the dialog.
[0020]
[Means for Solving the Problems]
In the present invention, the confirmation necessity search type interactive procedure considered in the conventional fourth technique is stopped. This is because the confirmation necessity search type dialogue procedure is a dialogue procedure based on the user's rejection utterance. Instead, a confirmation text for confirming all or part of the attribute values included in the system understanding state is output, and after the user inputs an approval text, not only the confirmed attribute values but also the confirmed values are not confirmed. Assuming that the value of the attribute is also correct, a trial confirmation type interactive procedure of outputting a response sentence for providing information, and not confirming to the user, not only the confirmed attribute value, Assuming that the value of the attribute that has not been finalized is also correct, consider a trial immediate response type interactive procedure in which a response sentence for immediately providing information to the user is output. In the trial confirmation type interactive procedure and the trial immediate response type interactive procedure, the rejection utterance from the user is not assumed for the erroneous response of the system.
[0021]
In addition, in the present invention, in addition to the above-described trial confirmation type interactive procedure and trial immediate response type interactive procedure, it is possible to confirm all or part of the attribute values included in the system understanding state. A definite confirmation type interactive procedure that outputs a confirmation sentence, and after the user inputs an approval sentence, outputs a response sentence for providing information to the user according to the value of the confirmed attribute, and confirmation to the user Considering a deterministic immediate response type interactive procedure in which a response sentence for promptly providing information to the user is output according to the value of the determined attribute.
[0022]
At each point in the dialogue, there are four types of dialogue procedures: deterministic confirmation type interaction procedure, trial confirmation type interaction procedure, deterministic immediate response type interaction procedure, and trial immediate response type interaction procedure. Consider all of them, and select the dialogue procedure that will minimize the length of the dialogue among the possible dialogue procedures. If the length of the dialogue when the dialogue is performed according to the dialogue procedure is called the cost of the dialogue procedure, the dialogue procedure that minimizes the dialogue procedure cost may be selected.
[0023]
The interaction cost of each interaction procedure can be calculated as follows.
Dialog cost of a deterministic confirmation-type dialog procedure: Expected value of the number of independent words that the system and the user exchange when performing a dialog according to the dialog procedure.
Dialog cost of a deterministic immediate-response dialog procedure: Expected value of the number of independent words included in a system response sentence when a dialog is performed according to the dialog procedure.
Dialogue cost of a trial confirmation-type dialogue procedure: When dialogues are conducted according to the dialogue procedure, when the expected value of the independent word exchanged between the system and the user and the system response sentence included in the dialogue procedure are incorrect The sum of the expected number of free words that the system and the user exchange in the subsequent dialogue.
Dialogue cost of a trial procedure of immediate response type: When performing dialogue according to the dialogue procedure, the expected value of the number of free words included in the system response sentence and the subsequent dialogue if the system response is incorrect The sum of the expected number of independent words that the system and user exchange in.
[0024]
According to the present invention, all of these four types of dialogue procedures are considered, and the dialogue procedure having the smallest dialogue cost is selected under the objective criterion of dialogue cost. If the user does the refusal, the output of the confirmation sentence for confirming the system understanding state is reduced as much as possible, and at the same time, the increase in the length of the conversation due to the re-conversation is suppressed. Can do.
[0025]
DETAILED DESCRIPTION OF THE INVENTION
Hereinafter, embodiments of the present invention will be described in detail with reference to the drawings.
FIG. 1 shows a system environment of a voice dialogue system according to the present invention. The voice interactive system 10 includes a database 20 and a memory device 30 and is connected to a large number of user terminals (computer terminals) 50 through a communication line 40 or the like. When the user uses the user terminal 50 to input a sentence meaning an information inquiry by voice, the voice interaction system 10 performs an inquiry after confirming all or part of the inquiry content with the user. Information held in the database 20 according to the contents is provided to the user by voice through the communication line 40 and the user terminal 50. The entity of the voice interaction system 10 is a computer system. The memory device 30 is a working memory for the voice interaction system 10 and stores a system understanding state, an interaction procedure, an interaction cost, a confirmation candidate, a confirmation candidate cost, and the like at each point in the interaction, as will be described later.
[0026]
FIG. 2 is a functional block showing a configuration example of the voice interaction system 10 according to the present invention. In the figure, an input unit 100 and an output unit 170 are connected to a user terminal 50 through a communication line 40 or the like.
[0027]
The input unit 100 approves a sentence that indicates a query of information or a sentence that indicates correction of a query content or a confirmation that is input as a voice (voice signal) by a user via a voice recognition device of the user terminal 50 as a computer terminal. The sentence which means is taken in.
[0028]
The sentence understanding unit 110 recognizes the content of a user inquiry input by voice (voice signal) from the input unit 100, and holds the inquiry content recognized by the system as a system understanding state. Actually, the system understanding state is held in the memory device 30. The system understanding state is expressed as a set of triples of an attribute, an attribute value, and a flag indicating whether the attribute value has been confirmed. When the sentence understanding unit 110 recognizes that a sentence meaning an information inquiry is input from the user, the sentence understanding unit 110 generates a system understanding state from the recognition result of the inquiry sentence. When the correction text of the inquiry content is input from the user, the current system understanding state is updated according to the result of recognizing the content of the correction text. If the system recognizes that an approval statement has been input by the user after outputting a confirmation message to confirm to the user all or part of the attribute values included in the system understanding state, the confirmed value It is recorded in the current system understanding state that has been confirmed.
[0029]
The provided information type determination unit 120 determines all the types of information that can be provided to the user as the provided information type based on the system understanding state generated / updated by the sentence understanding unit 110, and provides the provided information In addition to the type, the probability of the provided information type is calculated. Although omitted in FIG. 1, the provision information and the probability thereof are also held in the memory device 30.
[0030]
The dialogue procedure generating unit 130 has confirmed after exchanging for confirming all or part of the attribute values included in the system understanding state for each of the provided information types determined by the provided information type determining unit 120. Assuming only the value of the attribute is correct, the definite confirmation type interactive procedure (A) of outputting a response sentence for providing information to the user, and the value of the attribute included in the system understanding state To provide information to users on the assumption that not only the value of the attribute that has been confirmed but also the value of the attribute that has not been confirmed is the correct value after the exchange to confirm all or part of it Assuming that only the value of the confirmed attribute is the correct value without prompting the user, the information is immediately sent to the user. Provide Deterministic immediate response type interactive procedure (C) of outputting a response sentence for confirmation and a value not only for a confirmed attribute but also for a non-confirmed attribute without confirming to the user As a result, a trial immediate response type interactive procedure (D) is generated that immediately outputs a response sentence for providing information to the user. Each generated dialog procedure (dialog plan) is held in the memory device 30.
[0031]
The dialog procedure cost calculation unit 140 calculates the cost (dialog cost) of each dialog procedure generated by the dialog procedure generation unit 130. Regarding the deterministic confirmation type interactive procedure (A), the expected value of the number of independent words exchanged between the system and the user when performing dialogue according to the interactive procedure is used as the cost of the definite confirmation type interactive procedure. calculate. With regard to the trial confirmation type interactive procedure (B), when the expected value of the independent word exchanged between the system and the user when performing the conversation according to the interactive procedure and the system response sentence included in the interactive procedure are incorrect The sum of the expected values of the number of freestanding words exchanged between the system and the user in the subsequent dialogue is calculated as the cost of the trial confirmation type dialogue procedure. With respect to the deterministic immediate response type interactive procedure (C), when the system responds according to the interactive procedure, the expected value of the number of free words included in the system response sentence is set as the definite immediate response type interactive procedure. Calculate as cost. Regarding the interactive procedure (D) of the trial immediate response type, when the system responds according to the interactive procedure, the expected value of the number of free words included in the system response sentence is followed, and the system response is incorrect. The sum of expected values of the number of independent words exchanged between the system and the user in the dialogue is calculated as the trial immediate response type dialogue procedure. The cost of each generated dialog procedure is held in the memory device 30 in association with the dialog procedure.
[0032]
The confirmation candidate generation unit 150 generates all combinations of attributes that have not been confirmed among the attributes included in the system understanding state as confirmation candidates. Each confirmation candidate generated is held in the memory device 30.
[0033]
The confirmation candidate cost generation unit 160 selects, for each of the confirmation candidates, an interaction procedure having the minimum cost among the interaction procedures including confirmation of the confirmation candidate for each provision information type, and determines the probability of the provision information type. Considering this, the expected value of the cost is calculated, and the expected value is generated as the cost of the confirmation candidate. The cost of each confirmation candidate generated is stored in the memory device 30 in association with the confirmation candidate.
[0034]
The output unit 170 compares the costs of all confirmation candidates with the costs of all immediate response type interactive procedures. If the cost of the immediate response type interactive procedure is smaller as a result of the comparison, a response sentence for providing information according to the minimum immediate response type interactive procedure is generated and output, and the cost of the confirmation candidate If is smaller, a confirmation sentence for confirming a confirmation candidate with the smallest cost is generated and output.
[0035]
The control unit 180 controls the operation of each unit such that the units 100 to 170 perform the cooperative operation as described above.
[0036]
FIG. 3 shows a process flowchart of the voice dialogue system 10. FIG. 4 is a detailed process flowchart of the cost determination step 210 in FIG.
[0037]
Hereinafter, detailed operations of the embodiment of the present invention will be described together with specific examples.
Here, as an example, consider a spoken dialogue system that provides weather information guidance. Assume that there are two types of provided information, weather and warning. It is assumed that the content of the user's inquiry is represented by three attributes: place, date, and information type. Assume that the place attribute has values of prefecture names and city names such as Kanagawa Prefecture and Kagawa Prefecture, the day attribute has values of today and tomorrow, and the information type has values of weather and warning. Further, 100 locations are registered in the database 20, and the weather category predicted for each location and the type of alarm announced are stored.
[0038]
It is assumed that the input unit 100 first inputs a sentence meaning the following information inquiry from the user (step 201).
“Tell me about tomorrow's weather in Kanagawa” (1)
[0039]
The sentence understanding unit 110 recognizes the content of the inquiry sentence from the user and generates a system understanding state (step 202). The system understanding state is expressed as a set of three sets of attributes, attribute values, and flags indicating whether the attribute values have been confirmed. It is assumed that the system understanding state generated by the sentence understanding unit 110 is as follows.
[0040]
{<Place, Kagawa, not yet>, <day, tomorrow, not yet><information type, weather, not yet}} (2)
In the above expression, the triple <attribute, value, not> indicates that the attribute value has not been confirmed. In the case of the example of (1), the value of the place attribute is Kanagawa Prefecture, the value of the day attribute is tomorrow, the value of the information type attribute is weather, and neither has been confirmed. Further, in the generated system understanding state (2), the value of the location attribute is erroneously recognized as “Kagawa Prefecture”. The values of the day attribute and information type attribute are recognized correctly.
[0041]
The control unit 180 determines whether there is an unconfirmed attribute value for the system understanding state (step 204). If it exists, the process proceeds to step 205, and if it does not exist, the process proceeds to step 212. At this point, since there is no confirmed attribute, the process proceeds to step 205, but the first steps 205 to 211 are omitted.
[0042]
Thereafter, the voice dialogue system 10 outputs a confirmation sentence such as “weather?” For confirming that the value of the information type is weather, and the user inputs an approval sentence such as “Yes” to the confirmation sentence. Assume the situation.
[0043]
When the sentence understanding unit 110 recognizes the approval sentence from the user, the sentence understanding unit 110 records in the system understanding state that the confirmed value has been confirmed (step 203). Here, the contents of the system understanding state are expressed as follows.
{<Place, Kagawa, not yet>, <day, tomorrow, not>, <information type, weather, finished>} (3)
In the above expression, the triplet <attribute, value, done> indicates that the value of the attribute has been confirmed. The value of the information type has been confirmed to be weather. Even at this time, since the values of the place and date attributes are unconfirmed, the process proceeds to step 205.
[0044]
The provided information type determination unit 120 determines the types of all information that can be provided to the user based on the system understanding state, and calculates the respective probabilities (step 205). Here, the weather is generated as the provided information type based on the system understanding state (3). Since there is only one possible provided information type, the probability of the information type of weather is 1. When there are a plurality of provided information types, for example, there is a method of calculating the probability of each provided information type assuming that the probabilities of each provided information type are equal probabilities.
[0045]
The dialog procedure generator 130 generates a dialog procedure (dialog plan) that is possible under the current system understanding state (3) (step 206). A comprehensive description of possible dialogue procedures is as follows.
[0046]
(1) Definite confirmation type dialogue procedure A
(I) A procedure of confirming only the value of the location attribute and then responding with the weather for all the days (today and tomorrow) of the confirmed location.
(Ii) A procedure of confirming only the value of the day attribute and then responding to the weather of all places (100 places) on the confirmed day.
(Iii) A procedure of confirming the value of the location attribute, then confirming the value of the day attribute, and then responding with the weather of the confirmed location and day.
(Iv) A procedure of confirming the value of the day attribute, then confirming the value of the location attribute, and then responding with the weather of the confirmed day and location.
(V) A procedure of confirming the value of the place attribute and the day attribute at the same time, and then responding with the confirmed place and day weather.
[0047]
(2) Trial confirmation type dialogue procedure B
(I) A procedure of confirming the location attribute and then responding tomorrow's weather at the determined location. In this case, the value of tomorrow is not finalized, but a response is generated assuming that it is a correct value.
(Ii) A procedure of confirming the day attribute and then responding with the weather of the place “Kagawa Prefecture” on the confirmed day. In this case, the value of the place “Kagawa Prefecture” is not fixed, but a response is generated assuming that the value is correct.
[0048]
(3) Definite immediate response type interactive procedure C
The procedure of responding to the weather of all locations (100 locations) and all days (today, tomorrow) without confirmation.
[0049]
(4) Trial immediate response type interactive procedure D
(I) A procedure of responding to the weather on all days (today, tomorrow) in Kagawa Prefecture without performing confirmation. In this case, the response is generated assuming that the value of the location attribute “Kagawa Prefecture” has not been confirmed or is a correct value.
(Ii) The procedure of responding to the weather of all tomorrow's places (100 places) without confirmation. In this case, the value of the day attribute “tomorrow” is not finalized, but a response is generated on the assumption that it is a correct value.
(Iii) The procedure of responding to Kagawa Prefecture and tomorrow's weather without confirmation. In this case, although the values of the place attribute “Kagawa Prefecture” and the date attribute “Tomorrow” are not finalized, a response is generated assuming that the values are correct.
[0050]
Here, simply, (1) in the definite confirmation type interactive procedure A, the procedure of confirming the location and date in the case of (v) at the same time, and then responding with the confirmed location and date weather, (2) In the trial confirmation type interactive procedure B, the procedure of confirming the location in the case of (i) and then responding tomorrow's weather at the confirmed location; (3) Deterministic immediate response type In the dialogue procedure C, as described above, the procedure of responding to the weather of all locations and all days, (4) In the trial procedure of immediate response type D, the weather of Kagawa Prefecture, tomorrow Assume that four interactive procedures are generated:
[0051]
The dialog procedure cost calculation unit 140 calculates the cost of each dialog procedure (step 207). Now, the provided information type is one of weather, and its probability is 1.
First, the dialogue cost of the deterministic immediate response type dialogue procedure C will be described. The dialogue procedure C is a dialogue procedure in which a response is made only in accordance with the currently determined value of weather. Since the unfixed values of the place attribute and the day attribute are not used, the weather of all places and days is responded. In the situation assumed here, 100 locations are registered in the database 140, and the value of the day attribute is two values of today and tomorrow. In order to respond to the weather in one place and one day, three independent words “Tomorrow”, “Kanagawa Prefecture”, and “Sunny” are necessary, like “Tomorrow's Kanagawa Prefecture is sunny”. And Since the interactive procedure responds to today's and tomorrow's weather at 100 locations, it will output a response sentence containing 600 independent words. Therefore, the cost of the dialog procedure C is “600”.
[0052]
Next, the cost of the definite confirmation type dialogue procedure A will be described. In the dialogue procedure A, first, the values of two attributes of place and date are confirmed. It is assumed that the expected value of the number of independent words exchanged in the confirmation dialogue is calculated by the method used in the fourth conventional method (see Non-Patent Document 2). Here, when the recognition accuracy of the attribute to be confirmed is r and the number of the attributes to be confirmed is m, the number of independent words exchanged until the values of those attributes are determined is represented by the following expression. .
2m / r-m + 1 (4)
[0053]
Assume that the location attribute recognition accuracy is 0.60 and the day attribute recognition accuracy is 0.95. The recognition accuracy for correctly recognizing the place attribute and the day attribute at the same time is 0.60 × 0.95 = 0.57, if expressed by the product of the recognition accuracy of each attribute. Therefore, the expected number of independent words exchanged in the dialogue until the values of the place attribute and the day attribute are determined simultaneously is (2) where m = 2 and F = 0.57. 0.52-2 + 1 = 6.02.
[0054]
In the dialogue procedure A, after the values of the place attribute and the day attribute are confirmed, the weather in the confirmed place and day is returned. In this response, it is assumed that three independent words are output, such as “Today's Kanagawa Prefecture is sunny”. 6.02 independent words are required for the confirmation, and 3 independent words are required for the response after the confirmation, and a total of “9.02” independent words are required. As a result, the dialogue cost of the dialogue procedure A is 9.02.
[0055]
Next, the cost of a trial immediate response type interactive procedure D will be described. The dialogue procedure D is a dialogue procedure in which a response is made according to the values of Kagawa Prefecture and tomorrow and the value of confirmed weather that have not been confirmed in the current system understanding state (3). Here, the system responds with “Tomorrow's Kagawa Prefecture is sunny”. Of course, since Kagawa Prefecture is a misrecognized value, the information required by the user is not given by this response.
[0056]
Here, the cost of a trial immediate response type interactive procedure will be described in a general form. The description will be divided into a case where the system understanding state is correct and a case where the system understanding state is not correct, where p is the probability that the current system understanding state is correct. First, consider the case where the current system understanding state is correct. If the system understanding state is correct, the system response is consistent with the user's query and the user can obtain the necessary information, and the dialog ends here. Let L1 be the number of free words included in the system response when the system understanding state is correct.
[0057]
Next, consider a case where the current system understanding state is incorrect. The probability that the system understanding state is incorrect is 1-p. At this time, since the system response is incorrect, the user cannot obtain necessary information. In order to obtain the necessary information, the user starts over the current inquiry from the beginning, and after several dialogues, obtains the necessary information.
[0058]
Here, in order to facilitate the calculation of costs, even in the dialogue that follows the output of an incorrect response, the attributes that have been confirmed in the current system understanding state are confirmed, and the attributes that are not confirmed are not confirmed. Assume that a system response is output in the above state. Let L2 be the expected value of an independent word exchanged in one continuous conversation. With respect to the expected value L2, the average value of the number of free words included in the dialog for the currently inquired information type is calculated using the dialog data collected in advance, and this average value is set to L2. Assuming that a correct response is output with probability q in each cascaded dialog, the expected value of the number of dialogs required to perform a correct response is 1 / q.
[0059]
To summarize the above considerations, when the current system understanding state is correct with probability p, the system responds correctly and the user succeeds in obtaining the necessary information. At this time, the number of independent words included in the system response is L1. Next, if the current system understanding state is incorrect with probability 1-p, an incorrect system response containing L1 independent words is output, followed by 1 / q dialogues containing L2 independent words, The user succeeds in obtaining the necessary information. Therefore, the interactive cost of the trial immediate response type interactive procedure is as follows. This is defined in advance.

[0060]
Let's go back to the example we're taking now. The probability that the current system understanding state (3) is correct is given by the product of the location attribute recognition accuracy 0.60 and the day attribute recognition accuracy 0.95. Therefore, the system understanding state (3) is correct when p = 0.60 × 0.95 = 0.57. Here, it is assumed that q = p = 0.57, assuming that the probability q that a correct system response is generated in a dialog after an erroneous response is made is equal to P. Since there are three independent words included in the erroneous response “Tomorrow's Kagawa is sunny”, L1 = 3. For the expected value L2 of the independent word exchanged in each dialogue following the erroneous response, the average number of independent words included in the dialogue for inquiring the weather among the dialogue data collected in advance is taken, and L2 = Assume that 10 is reached. From the equation (5), the cost of the dialogue procedure D is 3 + 0.43 · 10 / 0.57 = 3 + 7.54 = 10.54.
[0061]
Next, the dialogue cost of the trial confirmation type dialogue procedure B will be described. Here, the dialogue procedure B is a procedure of confirming the place attribute and then responding tomorrow's weather at the decided place. First, since the expected number of independent words exchanged until the value of the location attribute is determined is m = 1 and r = 0.65 in the equation (4), 2.1 / 0.65-1-1 = 3.08.
[0062]
In the situation after the place attribute is confirmed, only the value of the day attribute is not confirmed. Since the recognition accuracy of the day attribute is assumed to be 0.95, the probability that the system understanding state is correct at this time is 0.95. This situation is the same as when a trial immediate response type interactive procedure is performed in which a system response is made assuming that the value of the day attribute is correct. Therefore, the expected number of free words included in the dialogue after the value of the place attribute is determined is 3 + 0.05 · 10 in the equation (5), where p = q = 0.95, L1 = 3, and L2 = 10. /0.95=3+0.53=3.53. Eventually, the dialogue cost of the dialogue procedure B is 3.08 + 3.53 = 6.61.
[0063]
In summary,
(1) Dialogue cost of definite confirmation type dialogue procedure A: 9.02
(2) Dialogue cost of trial confirmation type dialogue procedure B: 6.61
(3) Dialog cost of deterministic immediate response type dialog procedure C: 600
(4) Dialog cost of trial immediate response type dialog procedure D: 10.54
It becomes.
[0064]
As described above, after the dialog procedure cost calculation unit 140 calculates the cost of each dialog procedure, the confirmation supplement generation unit 150 selects all combinations of attributes that are not determined among the attributes included in the current system understanding state. A confirmation candidate is generated (step 208). From the previous system understanding state (3), there are three combinations: (1) Confirm location, (2) Confirm date, and (3) Confirm location and date. Here, for simplicity, it is assumed that a confirmation candidate for confirming only the location attribute and a confirmation candidate for confirming the location attribute and the day attribute at the same time are generated as confirmation candidates.
[0065]
Next, for each of the confirmation candidates, the confirmation candidate cost generating means 160 selects an interaction procedure having the minimum cost among the interaction procedures including the confirmation supplement, and considers the probability of the provided information type, The expected value of cost is calculated, and the expected value is generated as the cost of the confirmation candidate (step 209). Here, the only possible provision information type is weather, and its probability is 1.
The interactive procedure including confirmation of only the location attribute is only a trial confirmation type interactive procedure B. Therefore, the cost of the confirmation candidate for confirming only the location attribute is “6.61”.
The only interactive procedure that includes confirming the place attribute and the day attribute at the same time is the definite confirmation type interactive procedure A. Therefore, the cost of the confirmation candidate for confirming the place attribute and the day attribute at the same time is “9.02”.
[0066]
Subsequently, the output unit 180 determines the cost (step 210), generates a definite sentence or a response sentence based on the determination result, and outputs it (step 211). Specifically, first, the costs of all the confirmation candidates (here, only the location is confirmed, the location and the date are confirmed) are compared with the costs of all the immediate response type interactive procedures (here, the interactive procedures C and D). (Step 2110). If the cost of the immediate response type interactive procedure is smaller, an immediate response type with a minimum cost (deterministic immediate response type or trial immediate response type) is selected from the immediate response type interactive procedures. ) Is determined (step 2111), and a response sentence for providing information is generated according to the interactive procedure (step 2112). On the other hand, if the cost of the confirmation candidate is smaller, the confirmation candidate having the smallest cost is determined (step 2113), and confirmation of the confirmation candidate (in this case, confirm the place or place and date). A confirmation sentence for performing (confirmation) is generated (step 2114).
[0067]
In the case of the present example, the minimum cost after all is the confirmation compensation cost 6.61 of confirming only the location attribute. The output unit 170 outputs a confirmation sentence “Is it Kagawa Prefecture?” For confirming only the value of the location attribute according to the current system understanding state (3).
[0068]
After this, the user inputs a correction sentence “This is Kanagawa Prefecture”, the system recognizes the correction sentence correctly, and then outputs a confirmation sentence to confirm the value of the location attribute “Is Kanagawa Prefecture?” Assume that the user inputs an approval sentence “Yes” and the system correctly recognizes the approval sentence. At this time, the system understanding state is as follows.
{<Place, Kanagawa Prefecture, Finished>, <Sunday, Tomorrow, Not>, <Information Type, Weather, Finished}} (6)
Only the date attribute is indeterminate. Therefore, also in this case, the process proceeds to step 205 and subsequent steps.
[0069]
Here, the dialog procedure generator 130 generates all possible dialog procedures again (step 206). Here, for the sake of simplicity, the following two interactive procedures will be taken up and the explanation will be continued.
(5) Definite confirmation type dialogue procedure A:
After confirming the value of the day attribute, the procedure of responding the weather in Kanagawa Prefecture on the confirmed date.
(6) Trial immediate response type interactive procedure D:
The procedure of responding to tomorrow's weather in Kanagawa without checking. Here, the value of tomorrow is not finalized, but a response is output assuming that it is a correct value.
[0070]
Next, the dialog procedure cost calculation unit 140 calculates the cost of each dialog procedure (step 207). At present, the only possible provision information type is weather, and its probability is 1.
First, the dialogue cost of the dialogue procedure A in (5) will be described. The expected number of independent words required until the date attribute is determined is that the recognition accuracy of the day attribute is 0.95. Therefore, in Equation (4), m = 1 and r = 0.95, 2 · 1 / 0.95-1 + 1 = 2.11. The number of free words included in the response generated after the date attribute is determined is 3. Therefore, the dialogue cost of the dialogue procedure E is 2.11 + 3 = 5.11.
Next, the dialogue cost of the dialogue procedure D in (6) will be described. Since the recognition accuracy of the day attribute is 0.95, the probability that the current system understanding state (6) is correct is 0.95. In addition, since the dialogue for inquiries about the weather exchanges 10 independent words on average, the dialogue cost of the dialogue procedure is expressed by equation (5), where p = q = 0.95, L1 = 3, When L2 = 10, 3 + 0.05 · 10 / 0.95 = 3 + 0.53 = 3.53.
[0071]
Summary,
Cost of deterministic confirmation type dialogue procedure A: 5.11
Cost of trial interactive response procedure D: 3.53
It becomes.
[0072]
Next, the confirmation candidate cost generation unit 150 generates only a confirmation candidate that confirms only the date attribute here (step 208). The cost of this confirmation candidate is equal to the cost of the dialogue procedure E and is 5.11.
[0073]
The sentence output unit 180 compares the cost 5.11 of the confirmation candidate for confirming only the date attribute with the cost 3.53 of the trial immediate response type interactive procedure D, and selects the interactive procedure D that gives a small cost. (Steps 210 and 211). According to this selection, a response sentence “Tomorrow's Kanagawa Prefecture is fine” is outputted without outputting a confirmation sentence for confirming the day attribute.
[0074]
The above dialogue is summarized as follows.
(Sentence 1) User: “Please tell me the weather tomorrow in Kanagawa Prefecture”
(Sentence 2) System: “Weather?”
(Sentence 3) User: “Yes”
(Sentence 4) System: “Is it Kagawa Prefecture?”
(Sentence 5) User: “It is Kanagawa Prefecture”
(Sentence 6) System: “Is it Kanagawa Prefecture?”
(Sentence 7) User: “Yes”
(Sentence 8) System: “Tomorrow's weather in Kanagawa Prefecture is fine”
The explanation here focuses on the behavior of the system after sentence (4).
[0075]
In the processing flow of FIG. 3, if it is determined in step 204 that there are no more uncertain attributes in the system understanding state, the output unit 170 immediately responds based on the system understanding state. A sentence is generated and output.
[0076]
In the present spoken dialogue system 10, the value of the place attribute with low recognition accuracy is confirmed, but the value of the day attribute with high recognition accuracy is not confirmed. The determination of which attribute value to check is determined by an objective criterion of the cost of the interactive procedure. This method allows the user to obtain necessary information in a short dialogue.
[0077]
The processing functions of some or all of the components in the voice interaction system 10 shown in FIG. 2 can be configured by a computer program, and the program can be executed using the computer to realize the present invention, or It goes without saying that the processing procedures shown in FIGS. 3 and 4 can be configured by a computer program and the program can be executed by the computer. In addition, a computer-readable recording medium such as an FD, MO, ROM, memory card, CD, or the like is stored in the computer. In addition, the program can be recorded and stored on a DVD, a removable disk, etc., and the program can be distributed through a network such as the Internet.
[0078]
【The invention's effect】
As described above, according to the voice dialogue system of the present invention, when it is determined that the length of the whole dialogue is not increased, confirmation of part or all of the user inquiry content may be omitted. It is possible to avoid an increase in the degree of user dissatisfaction associated with an increase in the number of output confirmation sentences. In addition, when the system makes an erroneous response due to misrecognition, it is not premised that the user makes a refusal utterance, so it is not necessary to force the user to make a refusal utterance, and user satisfaction is increased.
[Brief description of the drawings]
FIG. 1 is an overall configuration diagram showing a system environment of the present invention.
FIG. 2 is a functional block diagram of an embodiment of a voice interaction system according to the present invention.
FIG. 3 is a process flowchart of an embodiment of the present invention.
4 is a detailed process flowchart of cost determination in FIG. 3;
[Explanation of symbols]
10 Spoken dialogue system
20 database
30 Memory device
40 Communication line
50 User terminal
100 input section
110 sentence comprehension department
120 Provided information type determination unit
130 Dialog procedure generator
140 Dialog procedure cost calculator
150 Confirmation candidate generator
160 Confirmation candidate cost generator
170 Output unit
180 Control unit

Claims (7)

When the user inputs a sentence meaning an inquiry of information by voice, after exchanging to confirm all or part of the recognized inquiry content, the information held in the database is changed according to the inquiry content. In a spoken dialogue system that provides users with voice,
Recognized inquiry content is generated as a system understanding state expressed by a set of triples indicating an attribute, a value of the attribute, and whether or not the value has been confirmed, and the system understanding state at each time point of the dialogue A means of updating
After exchanging for confirming all or part of the attribute values included in the system understanding state, the user inquiry content is determined assuming that only the attribute values that have been confirmed are correct values. , A definite confirmation type interactive procedure (A) of outputting a response sentence for providing information to the user, and an exchange for confirming all or part of attribute values included in the system understanding state Later, it is assumed that not only the value of the attribute that has been confirmed but also the value of the attribute that has not been confirmed is a correct value, the user inquiry content is determined, and a response statement for providing information to the user is output. Trial confirmation type interactive procedure (B), without confirming to the user, assuming that only the value of the confirmed attribute is the correct value, determining the user inquiry contents, and providing information to the user A deterministic immediate response type interactive procedure (C) that immediately outputs a response sentence to be provided, and a value for a non-determined attribute as well as a definite attribute value without confirmation to the user A means for generating a trial immediate response type interactive procedure (D) of determining a user inquiry content on the assumption that the value is also a correct value and outputting a response sentence for providing information to the user;
The cost of the definite confirmation type interactive procedure (A), the cost of the definite immediate response type interactive procedure (B), the cost of the trial confirmation type interactive procedure (C), and the trial Means for calculating the cost of a typical immediate response interactive procedure (D);
Means for generating all or part of attribute values that have not been confirmed in the system understanding state as confirmation candidates;
Means for calculating the cost of the confirmation candidate based on the cost of the interactive procedure including confirmation of the confirmation candidate for each of the confirmation candidates;
Comparing the costs of all confirmation candidates with the costs of all immediate response interactive procedures, and if the cost of the immediate response interactive procedure is the lowest cost, follow the prompt response interactive procedure Generating a response sentence for providing the confirmation candidate, and if the cost of the confirmation candidate is the minimum cost, means for generating a confirmation sentence for confirming the confirmation candidate;
A spoken dialogue system comprising: The voice interaction system according to claim 1,
The means for calculating the cost of the dialogue procedure is to calculate the expected value of the number of independent words that the system and the user exchange when performing the dialogue according to the definite confirmation type dialogue procedure as the cost of the definite confirmation type dialogue procedure. Calculate and try the expected value of the number of free words contained in the system response sentence as the cost of the deterministic immediate response interactive procedure when the system responds according to the deterministic immediate response interactive procedure System and the user interact with each other in the subsequent conversation if the expected value of the independent word that the system and the user interact when interacting according to a typical confirmation-type interaction procedure and the system response text included in the interaction procedure is incorrect When the system responds by calculating the sum of the expected value of the number of free words to be used as the cost of a trial confirmation type interactive procedure The sum of the expected number of free words included in the system response sentence and the expected value of the number of free words exchanged between the system and the user in the subsequent dialogue when the system response is incorrect is a trial immediate response. A spoken dialogue system characterized by calculating the cost of a type dialogue procedure. The spoken dialogue system according to claim 1 or 2,
The means for calculating the cost of the confirmation candidate selects, for each of the confirmation candidates, the dialogue procedure having the minimum cost from among the dialogue procedures including confirmation of the confirmation candidate, and takes the probability of the provided information type into consideration. A spoken dialogue system characterized by calculating an expected value and using the expected value as a cost of a confirmation candidate. When the user inputs a sentence meaning an information inquiry by voice, after exchanging to confirm all or part of the recognized inquiry content, the information stored in the database is changed according to the inquiry content. In the voice interaction method provided to the user by voice,
Receiving an inquiry sentence meaning an inquiry of information input by the user, a correction sentence meaning correction of the content of the inquiry, and an approval sentence meaning approval of the confirmation;
When a query is input from the user, it is represented by a set of triples of an attribute, a value indicating the value of the attribute, and a flag indicating whether or not the value has been determined according to the result of recognizing the content of the query. When the system understanding state is generated and the correction text of the inquiry content is input from the user, the system understanding state is updated according to the result of recognizing the content of the correction text, and all the attribute values included in the system understanding state Alternatively, when the confirmation sentence is recognized by the user after the system outputs a confirmation sentence for confirming a part, a step of recording in the system understanding state that the confirmed value is confirmed; ,
Determining from the system understanding state, determining the type of all information that can be provided to the user as the provided information type, and calculating the respective probabilities of the provided information types;
For each provided information type, only the attribute values that have been confirmed after the exchange for confirming all or part of the attribute values included in the system understanding state are correct values. It is assumed that the user inquiry content is determined, and a definite confirmation type interactive procedure (A) of outputting a response sentence for providing information to the user, all or one of the attribute values included in the system understanding state After exchanging for confirmation, the user inquiry content is determined assuming that not only the value of the attribute that has been confirmed but also the value of the attribute that has not been confirmed is the correct value. Trial confirmation type interactive procedure (B) of outputting a response sentence to provide, without confirmation to the user, assuming that only the value of the confirmed attribute is the correct value A deterministic immediate response type interactive procedure (C) in which the contents of the inquiry are determined and a response sentence for providing information to the user is immediately output. In addition, it is assumed that the value of the attribute that has not been finalized is also a correct value, determines the content of the user inquiry, and outputs a response sentence for providing information to the user. Generating D);
For each of the provided information types, the cost of the definite confirmation type interactive procedure, the cost of the deterministic immediate response type interaction procedure, the cost of the trial confirmation type interactive procedure, the trial immediate response Calculating the cost of a type interaction procedure;
Generating all or part of attribute values that have not been confirmed in the system understanding state as confirmation candidates;
For each of the confirmation candidates, for each provided information type, calculating the cost of the confirmation candidate based on the cost of the interactive procedure including confirmation of the confirmation candidate;
Comparing the costs of all confirmation candidates with the costs of all immediate response interactive procedures, and if the cost of the immediate response interactive procedure is the lowest cost, follow the prompt response interactive procedure Generating a response sentence for providing the confirmation candidate, and generating and outputting a confirmation sentence for confirming the confirmation candidate if the cost of the confirmation candidate is the minimum cost;
A voice interaction method characterized by comprising: The voice interaction method according to claim 4.
The step of calculating the cost of the dialogue procedure is to calculate the expected value of the number of independent words exchanged between the system and the user as the cost of the definite confirmation type dialogue procedure. Calculate and try the expected value of the number of free words contained in the system response sentence as the cost of the deterministic immediate response interactive procedure when the system responds according to the deterministic immediate response interactive procedure The system and the user interact with each other in the subsequent conversation if the expected value of the independent word that the system interacts with the user when performing a conversation according to a typical confirmation-type conversation procedure and the system response text included in the conversation procedure is incorrect Calculates the sum of expected values for the number of free words to be used as the cost of a trial-confirmed interactive procedure, and the system responds according to a trial immediate-response interactive procedure In this case, the sum of the expected value of the number of independent words included in the system response sentence and the expected value of the number of independent words exchanged between the system and the user in the subsequent dialogue when the system response is incorrect Calculate as the cost of an instant response interactive procedure,
In the step of calculating the cost of the confirmation candidate, for each of the confirmation candidates, the dialogue procedure having the minimum cost is selected from among the dialogue procedures including confirmation of the confirmation candidate, and the probability of the provided information type is taken into consideration. Calculate the expected value and use the expected value as the cost of the confirmation candidate.
A voice dialogue method characterized by the above. 6. A program for causing a computer to execute the processing steps of the voice interaction method according to claim 4. A computer-readable recording medium recording a program for causing a computer to execute the processing steps of the voice interaction method according to claim 4.

Images