JP3776341B2

JP3776341B2 - Voice interaction method, apparatus for implementing the method, program, and recording medium recording the program

Info

Publication number: JP3776341B2
Application number: JP2001295672A
Authority: JP
Inventors: 宜仁安田; 浩二堂坂; 清明相川
Original assignee: Nippon Telegraph and Telephone Corp; NTT Inc USA
Current assignee: NTT Inc; NTT Inc USA
Priority date: 2001-09-27
Filing date: 2001-09-27
Publication date: 2006-05-17
Anticipated expiration: 2021-09-27
Also published as: JP2003108190A

Description

【０００１】
【発明の属する技術分野】
この発明は、音声対話方法および、この方法を実施する装置、プログラム、このプログラムを記録した記録媒体に関し、特に、より少ない対話のやり取り回数でユーザの要求内容を把握する音声対話方法および、この方法を実施する装置、プログラム、このプログラムを記録した記録媒体に関する。
【０００２】
【従来の技術】
ユーザとの間の対話により所定の動作を実行する装置（以下、音声対話装置、と呼ぶ）が知られている。この音声対話装置は、ユーザが発生した音声を文字列に変換する音声認識部を有し、この認識結果とこれまでの対話の履歴を利用して音声対話装置の次の発話内容を制御し、この発話内容を音声出力装置により音声として出力する。音声対話装置は、音声認識によりユーザの発話内容を必ず正しく認識するとは限らない。音声対話装置は、認識に誤りがあった場合でも、ユーザが求めている内容に応じた動作を行う必要があるので、通常、音声認識結果の確認をユーザに対して行う。また、認識結果のみに依っては実行すべき動作が定まらない場合がある。この場合、ユーザの要求内容を知るために、音声対話装置は、通常必要と思われる情報をユーザに対して質問する。この確認或いは質問に関して、音声対話装置とユーザとの間の対話のやり取り回数が多いと、ユーザの手間が増えて煩わしく、でき得る限りこのやり取り回数は少ない方が良い。
【０００３】
音声対話装置を実用に供することを想定した場合、処理可能なユーザの要求種類は１通りではなく、複数種類の要求を処理することができることが望ましい。例えば、スケジュール管理を行う音声対話装置を想定した場合、少なくとも「予定の照会」および「予定の削除」程度の要求種類を取り扱える必要があると思われる。ところが、この内の１種類の要求しか取り扱えないものとすると、この音声対話装置は実用的な装置であるとは言えない。複数の要求種類を受け付ける従来の音声対話装置における確認内容決定部においては、要求種類別に適切な確認手順とやり取り回数を計算する部分と、要求種類の確率分布を計算する部分を有し、これらのやり取り回数と確率分布に基づいて、すべての要求種類の確認順序の順列について期待やり取り回数を推定し、その期待やり取り回数が最も小さくなる順序を要求種類を単位として決定していた。
【０００４】
【発明が解決しようとする課題】
上述した通り、音声対話装置は、認識に誤りがある場合でも、ユーザが求めている内容に応じた動作を行う必要があるところから、通常、音声認識結果の確認をユーザに対して行い、正確にユーザの要求内容に応じた動作をする。この確認或いは質問は少ないことが望ましい。以上の従来技術には改善すべき点がある。即ち、相異なる要求種類間で共通に必要な情報を考慮していないということである。例えば、要求種類として「予定の追加」および「予定の照会」が含まれるスケジュール管理を行う音声対話装置の場合、何れの要求種類においても「日付」と「時間」は必須の項目と思われる。要求種類が定まらない場合に、他の項目を質問するより、先ず、これらの項目を質問する方が手間を削減することができる筈である。しかし、この従来技術の如く、要求種類を単位として計算する場合、仮に、「予定の追加」を仮定した場合のやり取り回数推定部６からの手順の出力が、先ず、「「予定の内容」の質問」であって、要求種類毎の確認順序が「予定の追加」→「予定の照会」である場合は、結果として「「予定の内容」の質問」をすることになる。
【０００５】
そこで、この発明は、要求種類間で共通の項目がある場合において、少ないやり取り回数によりユーザの要求の内容を決定する音声対話方法および、この方法を実施する装置、プログラム、このプログラムを記録した記録媒体を提供するものである。
【０００６】
【課題を解決するための手段】
請求項１：ユーザの音声を入力としてユーザの要求に応じた処理を実行して音声により応答内容を出力する音声対話方法において、音声信号を生成し、生成した音声信号に基づいてこれを文字列に変換し、文字列を解析して項目名と値の表より成る理解状態を作成し、現在処理することができるタスクの仕様を記述したタスク仕様データベース４を具備し、理解状態に基づいてユーザの要求内容種類の確率分布を推定し、理解状態と特定の要求種類名を入力としてユーザの要求している要求種類を仮定した場合にユーザの要求している内容を把握するまでに必要なやり取り回数とその際の手順を決定し、理解状態に基づいて次の行動の種類を決定し、要求種類を仮定していった確認順序の中で、仮定が誤っている場合に次の要求種類を仮定した場合に必要なやり取りの回数の推定を、前回までに仮定した要求種類で必要な項目と次に仮定した要求種類で必要な項目との差分から推定してユーザに対して行う確認および質問の内容を決定し、ユーザの要求に応じた行動を行い、確認内容或は応答内容を音声として出力する音声対話方法を構成した。
【０００７】
そして、請求項２：ユーザの音声を入力としてユーザの要求に応じた処理を実行して音声により応答内容を出力する音声対話装置において、音声を音声信号に変換する音声入力部１を具備し、音声入力部１により得られた音声信号を文字列に変換する音声認識部２を具備し、音声認識部２によって得られた文字列を解析し、項目名と値の表より成る理解状態を作成する言語理解部３を具備し、音声対話装置が現在処理することができるタスクの仕様を記述したタスク仕様データベース４を具備し、理解状態を入力としてユーザの要求内容種類の確率分布を推定する要求内容種類確率分布推定部５を具備し、理解状態と特定の要求種類名を入力としてユーザの要求している要求種類を仮定した場合に、ユーザの要求している内容を音声対話装置が把握するまでに必要なやり取り回数とその際の手順を決定するやり取り回数推定部６を具備し、理解状態を入力として音声対話装置の次の行動の種類を決定する行動決定部７を具備し、要求種類を仮定していった確認順序の中で、仮定が誤っている場合に次の要求種類を仮定した場合に必要なやり取りの回数の推定を、前回までに仮定した要求種類で必要な項目と次に仮定した要求種類で必要な項目との差分から推定して、装置がユーザに対して行う確認および質問の内容を決定する確認内容決定部８を具備し、ユーザの要求に応じた行動を行う応答内容決定部９を具備し、確認内容決定部８或いは応答内容決定部９により決定された確認内容或いは応答内容を音声として出力する音声出力部１０を具備する音声対話装置を構成した。
【０００８】
また、請求項３：コンピュータに対して、音声認識部２が音声入力部１を介して入力された音声信号を文字列に変換すべき指令をし、言語理解部３が音声認識部２により得られた文字列を解析して項目名と値の表より成る理解状態を作成すべき指令をし、要求内容種類確率分布推定部５が理解状態を入力としてユーザの要求内容種類の確率分布を推定すべき指令をし、やり取り回数推定部６が理解状態と特定の要求種類名を入力としてユーザの要求している要求種類を仮定した場合に、ユーザの要求している内容を音声対話装置が把握する迄に必要なやり取り回数とその際の手順を決定すべき指令をし、行動決定部７が理解状態を入力として音声対話装置の次の行動の種類を決定すべき指令をし、確認内容決定部８が、要求種類を仮定していった確認順序の中で、仮定が誤っている場合に次の要求種類を仮定した場合に必要なやり取りの回数の推定を、前回までに仮定した要求種類で必要な項目と次に仮定した要求種類で必要な項目との差分から推定して、装置がユーザに対して行う確認および質問の内容を決定すべき指令をし、応答内容決定部９がユーザの要求に応じた行動を行うべき指令をし、音声出力部１０が確認内容決定部８或いは応答内容決定部９により決定された確認内容或いは応答内容を音声として出力すべき指令をする音声対話プログラムを構成した。
【０００９】
更に、請求項４：コンピュータに対して、音声認識部２が音声入力部１を介して入力された音声信号を文字列に変換すべき指令をし、言語理解部３が音声認識部２により得られた文字列を解析して項目名と値の表より成る理解状態を作成すべき指令をし、要求内容種類確率分布推定部５が理解状態を入力としてユーザの要求内容種類の確率分布を推定すべき指令をし、やり取り回数推定部６が理解状態と特定の要求種類名を入力としてユーザの要求している要求種類を仮定した場合にユーザの要求している内容を音声対話装置が把握するまでに必要なやり取り回数とその際の手順を決定すべき指令をし、行動決定部７が理解状態を入力として音声対話装置の次の行動の種類を決定すべき指令をし、確認内容決定部８が、要求種類を仮定していった確認順序の中で、仮定が誤っている場合に次の要求種類を仮定した場合に必要なやり取りの回数の推定を、前回までに仮定した要求種類で必要な項目と次に仮定した要求種類で必要な項目との差分から推定して、装置がユーザに対して行う確認および質問の内容を決定すべき指令をし、応答内容決定部９がユーザの要求に応じた行動を行うべき指令をし、音声出力部１０が確認内容決定部８或いは応答内容決定部９により決定された確認内容或いは応答内容を音声として出力すべき指令をする音声対話プログラムを記録した記録媒体を構成した。
【００１０】
【発明の実施の形態】
この発明の実施の形態を図１の実施例を参照して説明する。
この発明による音声対話装置の実施例は、音声入力部１、音声認識部２、言語理解部３、タスク仕様データベース４、要求内容種類確率分布推定部５、やり取り回数推定部６、行動決定部７、確認内容決定部８、応答内容決定部９、および音声出力部１０により構成される。ここで、音声対話装置の各部位は、図示されてはいないが、コンピュータとこれに対して各部位の動作指令をするプログラムとにより構成される。
【００１１】
音声入力部１は、ユーザが発声した音声［１］を、ディジタル化された音声信号［２］に変換する。これには一般のマイク、電話機とＡＤコンバータの組み合わせを使うことができる。
音声認識部２は、音声入力部１からの音声信号［２］を受け、認識結果と認識結果の確からしさの対の文字列［３］を算出する。これは一般の隠れマルコフモデルＨＭＭを使った既存の音声認識器を使うことができる。
言語理解部３は、音声認識部２からの文字列［３］を解析して、理解状態を作成する。この理解状態としては項目名と値と値の確からしさと確認済みか否か［４］の表などを使う。項目名と値の対応を得るには、既存の構文解析器などを使うことができる。また、値と確からしさの対応を得るには、ＨＭＭを使った音声認識器の音響スコアなどを使うことができる。装置からユーザに発せられた確認の後に、肯定に相当する文に解析できる入力があった場合は、確認済みか否かを確認済みに変更する。この処理を経ていない項目の確認済みか否かはすべて未確認（否）である。
【００１２】
タスク仕様データベース４には、要求種類毎に必要な項目名とその項目に入りうる値［５］が記述してある表である。たとえば、スケジュール管理であれば、「予定の追加」という要求種類には「予定の内容」と「日付」と「時間が必要でり、「予定の内容」として可能な値としては、「会議」「打合せ」「発表」などがあり、「日付」として可能な値として年間の日付や曜日などがあり、「時間」として可能な値としては２４時間中の何れかがある。
要求内容種類確率分布推定部５は、言語理解部３から項目名と値と値の確からしさ［４］を入力し、現在の理解状態から要求内容それぞれの確率を推定し、ユーザ要求種類名および確率より成る対の集合［８］を出力する。このために、ｖi を項目ｓi の値、ｃi をその確からしさ、Ｎj を要求種類Ｇj において必要なスロット数、Ｖj をＧj で認められる値の集合、Ｍi をｖi を値域として認めるユーザの要求種類の個数としたとき、理解状態Ｓと要求種類Ｇj との間の関連度を
【００１３】
【数１】

という式で求め、この関連度のすべての和を１にする正規化した場合の、Ｒeι（Ｓ、Ｇj ）を要求種類Ｇj の確率とする。
やり取り回数推定部６は、言語理解部３から項目名と値と値の確からしさと確認済みか否かの集合［４］を入力すると共に、タスク仕様データベース４から要求種類毎に必要な項目名とその項目に入りうる値［６］を入力し、或る要求種類を仮定した場合に、現在の理解状態からその要求種類で必要な項目をユーザとの対話により装置が得るまでに、効率的に確認を行なったとしてどの程度のやり取り回数が必要か推定し、このときの効率的な確認の手順であるユーザ要求種類名とターン数より成る対の集合［７］を出力する。出力は、例えば、（「日付」の確認→「時間」と「内容」の質問）７ターンという確認手順と数字（ターン数）の組である。
【００１４】
項目毎に受け付け可能な語彙の数を使って、或る項目を認識する場合の認識誤り率を語彙数の平方根に比例して求める。これにより得られた認識率（＝１−認識誤り率）をｒとしたときに、確認の際に必要なやり取り回数は１／ｒと見積もり、また、質問の際に必要なやり取り回数を１／ｒ＋１と見積もる。確認か質問かは［４］中で、項目に対応する値があれば、確認、値が空であれば、質問である。仮定した要求種類で必要な項目のすべての確認或は質問の組合せの中から、以上の計算により求めたやり取り回数はすべて計算可能であるので、この最も小さい場合のやり取り回数とそのやり取りの回数となる場合の確認手順とそのときの回数をこのやり取り回数推定部６の出力とする。
【００１５】
行動決定部７は、言語理解部３から項目名と値と値の確からしさ［４］を入力すると共に、タスク仕様データベース４から要求種類毎に必要な項目名とその項目に入りうる値［５］を入力し、装置がユーザの要求内容に対して応答可能かそれとも確認を継続するかを表す情報応答可能フラグ出力する。両者を比較して現在の理解状態が何れかの要求種類にとって必要な項目をすべて満たしていて、すべて確認済みであれば、既に要求内容を把握することができているとして応答可能フラグを応答可能として出力し、そうでなければ応答可能フラグを応答不可能として出力する。
【００１６】
確認内容決定部８は、要求内容種類確率分布推定部５の出力の確率分布、やり取り回数推定部６の出力のやり取り回数とその際の手順、行動決定部７から出力される「応答可能フラグ」を入力し、応答可能フラグが応答可能であれば何も出力を行わず、応答可能フラグが応答不可能であれば、各時点において装置は次にどういう項目について確認或いは質問すれば良いかを決定する。出力されるものは、項目名と確認或いは項目名と要求の対の集合［８］’である。例えば、（日付確認）（曜日確認）という情報を出力する。入力としては、要求種類の確率分布推定部５で推定した確率分布と、やり取り回数推定部６で推定した要求種類を仮定した場合のやり取り回数と、その際の手順を入力とする。
【００１７】
音声対話装置がどのユーザ要求について確認をすればよいかを考える。実際にユーザが要求している要求種類がＧi である確率をｐＧi 、その要求種類を仮定した場合の期待やり取り回数をｔGi、と表す。
装置が或る要求種類ＧS を仮定して確認を行なっていったけれども、実際要求種は異なるものだった場合があり得る。このときに次に要求種類を仮定して確認を行うために確認しなければならない項目は、前回仮定した要求種類で必要な項目と、今仮定した要求種類で必要な項目との差分とする。即ち、最初に「日付」と「時間」という項目が必要な要求種類を仮定し、次に「日付」と「時間」と「予定の内容」という項目が必要な要求種類を仮定する場合は、最初に仮定した要求種類が実は誤りだった場合に、次に仮定した要求種類で必要な項目は「予定の内容」のみとする。
【００１８】
この差分の項目以外の必要な項目は既に値が定まっている状態を理解状態として、再びやり取り回数推定部６に入力し、この差分についての確認を終えるために必要なやり取りの回数の推定結果を得る。この差分のやり取り回数をＧa(i)までの確認に失敗した場合に、Ｇa(i+1)の確認に必要なやり取り回数の推定結果を
【００１９】
【数２】

が最小となるような要求種類の選択順ａ(1)、ａ(2)、.......、ａ(n)での最初の要求種類Ｇa(1)を取り、この要求種類を仮定した場合のやり取り回数推定部６が出力結果のひとつである確認手順を得て、この確認手順の最初の内容を確認内容決定部８の出力［８］'とする。
応答内容決定部９は言語理解部からの理解状態と行動決定部７からの情報応答可能フラグを入力として受け取る。情報応答可能フラグが応答不可能であれば何も出力を行わず、情報応答可能フラグが応答可能であれば、理解状態［７］'と、タスク仕様データベース４から入力される要求種類毎に必要な項目名とその項目に入り得る値［５］とを比較することにより要求種類とユーザの要求内容を得て、その要求種類に対して事前にプログラムされた応答内容を決定し、文字列［９］を出力する。
【００２０】
音声出力部１０は、確認内容決定部８から出力される項目名と確認或いは項目名と要求の対の集合［８］'、或いは応答内容決定部９から出力される応答内容の文字列［９］を入力として受け取る。既存のテンプレートベースの言語生成器と、既存の音声合成器の組み合わせにより実現することができる。
実施例の音声対話のデータの流れは図１において丸付き数字により示される。
以下において、確認内容決定部８の具体例を説明する。
部署としては会社の秘書タスクを想定し、ユーザの要求種類は、「社内便の配送受付」および「社員の呼び出し」を扱えるものとする。
【００２１】
要求種類それぞれで必要な情報とその値域は、
・「社内便の配送受付」について、
商品名（社内にあるもの）
部署名（全部署）
人名（全員）
数量（１−１０）
・「社員の呼び出し」について、
部署名（関連部署）
人名（関連人員）
日を表す数（１−３０）
数量（１−５０）
であるものとする。
【００２２】
それぞれの要求を仮定した場合に確認終了迄に必要なターン数の推定結果は、
・「社内便の配送受付」について、
ｔ（ａ）ターン
・「社員の呼び出し」について、
ｔ（ｂ）ターン
であるものとする。
それぞれの要求の現在の理解状態における確率推定の結果の確率は、
・「社内便の配送受付」について、
ｐ（ａ）
・「社員の呼び出し」について、
ｐ（ｂ）
であるものとする。
【００２３】
この発明の実施例において、確認内容決定部８は、すべての要求種類の順列について、期待されるターン数を計算する。この具体例においては、「社内便の配送受付」および「社員の呼び出し」の２種類の要求種類があるものと仮定しているので、可能な順列は、
（１）「社内便の配送受付」→「社員の呼び出し」
（２）「社員の呼び出し」→「社内便の配送受付」
の２通りである。
【００２４】
（１）の場合において、仮に最初に確認を試みた要求推定が誤っていた場合の推定ターン数を考える場合に、従来は必要なターン数は
ｔ（ａ）＋ｔ（ｂ）
としていた。この発明の実施例においては、最初に確認を試みた要求推定が誤っていた場合においても、そこで必要とされた情報の一部は次に仮定する要求種類の役に立つであろうという観点から、「社内便の配送受付」を仮定して失敗した後で「社員の呼び出し」を確定するには、「社員の呼び出し」で必要とされる情報であるけれども「社内便の配送受付」においては必要とされない情報である「日を表す数」のみを確定する必要があると仮定してターン数の推定を行う。ここで推定されたターン数をｔ'（ｂ）（これはｔ（ｂ）より小さい筈である）とすれば、この発明においては
ｔ（ａ）＋ｔ'（ｂ）
をこの場合の期待ターン数とする。これにより、より正確なターン数推定を実行することができる。
【００２５】
（２）の場合においても、同様に、「社員の呼び出し」の確認を試みて失敗した場合に、その後に「社内便の配送受付」を確定するに必要なターン数の推定結果をｔ'（ａ）とすれば、
ｔ'（ａ）＋ｔ（ｂ）
をこの場合の期待ターン数とする。これらにより、この発明においては、（１）の場合の
ｔ（ａ）＋（１−ｐ（ａ））（ｔ（ａ）＋ｔ'（ｂ））
と、（２）の場合の
ｔ（ｂ）＋（１−ｐ（ｂ））（ｔ'（ａ）＋ｔ（ｂ））
の内の小さい値を返した方の最小にあるユーザの要求種類の確認を最初に試みることにする。
【００２６】
【発明の効果】
以上の通りであって、この発明によれば、より少ない対話のやり取り回数でユーザの要求内容を把握することができ、ユーザの手間を軽減することができる。
【図面の簡単な説明】
【図１】実施例を説明するブロック図。
【符号の説明】
１音声入力部２音声認識部
３言語理解部４タスク仕様データベース
５要求内容種類確率分布推定部６やり取り回数推定部
７行動決定部８確認内容決定部
９応答内容決定部１０音声出力部[0001]
BACKGROUND OF THE INVENTION
The present invention relates to a voice interaction method, an apparatus for implementing the method, a program, and a recording medium recording the program, and more particularly, a voice interaction method for grasping a user's request contents with a smaller number of conversations and the method. The present invention relates to an apparatus, a program, and a recording medium on which the program is recorded.
[0002]
[Prior art]
2. Description of the Related Art A device that performs a predetermined operation by a dialogue with a user (hereinafter referred to as a voice dialogue device) is known. This voice interactive device has a voice recognition unit that converts a voice generated by a user into a character string, and controls the next utterance content of the voice interactive device using the recognition result and the history of previous dialogue, The utterance content is output as voice by the voice output device. The voice interactive apparatus does not always recognize the user's utterance content correctly by voice recognition. Since the voice interactive apparatus needs to perform an operation according to the content requested by the user even when there is an error in recognition, the voice interactive device is normally confirmed for the user. In addition, the operation to be executed may not be determined depending on the recognition result alone. In this case, in order to know the content of the user's request, the voice interaction apparatus asks the user about information that is normally considered necessary. Regarding this confirmation or question, if the number of conversations between the voice interaction apparatus and the user is large, the user's trouble increases and is troublesome, and it is better that the number of exchanges is as small as possible.
[0003]
When it is assumed that the voice interactive apparatus is put into practical use, it is desirable that the number of types of user requests that can be processed is not one, and that a plurality of types of requests can be processed. For example, when a voice interactive apparatus that performs schedule management is assumed, it may be necessary to handle at least the request types of “schedule inquiry” and “schedule deletion”. However, if only one type of request among them can be handled, this spoken dialogue device cannot be said to be a practical device. In the confirmation content determination unit in the conventional voice interactive apparatus that accepts a plurality of request types, there are a part for calculating an appropriate confirmation procedure and the number of exchanges for each request type, and a part for calculating the probability distribution of the request type. Based on the number of exchanges and the probability distribution, the expected number of exchanges is estimated for the permutation of the confirmation order of all the request types, and the order in which the expected number of exchanges is the smallest is determined for each request type.
[0004]
[Problems to be solved by the invention]
As described above, the voice interaction device normally performs confirmation of the voice recognition result to the user because it is necessary to perform an operation according to the content requested by the user even when there is an error in recognition. It operates according to the user's request contents. It is desirable that there are few confirmations or questions. There is a point which should be improved in the above prior art. That is, it does not consider information that is commonly required between different request types. For example, in the case of a voice interactive apparatus that performs schedule management including “addition of schedule” and “query inquiry” as request types, “date” and “time” are considered to be indispensable items in any request type. If the request type is not determined, it is possible to save time by asking these items first, rather than asking other items. However, when the calculation is performed with the request type as a unit as in this prior art, the procedure output from the number-of-interactions estimation unit 6 assuming that “schedule addition” is assumed is first “the contents of“ schedule ”. If it is a “question” and the confirmation order for each request type is “schedule addition” → “schedule inquiry”, a “question of“ schedule content ”question” is obtained as a result.
[0005]
Accordingly, the present invention provides a voice interaction method for determining the content of a user request with a small number of exchanges when there are common items between request types, a device for implementing this method, a program, and a record in which this program is recorded A medium is provided.
[0006]
[Means for Solving the Problems]
[Claim 1] In a voice interaction method for executing a process according to a user's request by inputting a user's voice and outputting a response content by voice, a voice signal is generated, and this is converted into a character string based on the generated voice signal. The task specification database 4 describing the specifications of tasks that can be processed at present is prepared, and the user is based on the understanding state. Necessary exchanges for grasping the contents requested by the user when estimating the probability distribution of the requested contents type and assuming the requested kind requested by the user with the understanding state and the specific request type name as input The number of times and the procedure at that time are determined, the next action type is determined based on the understanding state, and if the assumption is incorrect in the confirmation order assuming the request type, the next request type is determined. Assumed Confirmation and questions to be made to the user by estimating the number of necessary exchanges from the previous item by using the difference between the items required for the previously assumed request type and the items required for the next assumed request type The voice dialogue method is configured to output the confirmation content or the response content as voice by performing the action according to the user's request.
[0007]
According to a second aspect of the present invention, there is provided a voice interaction apparatus for executing a process according to a user's request by inputting a user's voice and outputting a response content by voice, comprising a voice input unit 1 for converting voice into a voice signal, It has a speech recognition unit 2 that converts a speech signal obtained by the speech input unit 1 into a character string, analyzes the character string obtained by the speech recognition unit 2, and creates an understanding state consisting of a table of item names and values A task comprehension unit 3, a task specification database 4 describing task specifications that can be processed by the voice interaction device, and a request for estimating a probability distribution of a user's requested content type using an understanding state as an input When the content type probability distribution estimation unit 5 is provided and the request type requested by the user is assumed by inputting the understanding state and the specific request type name, the content requested by the user is expressed as a voice dialogue apparatus. An exchange number estimation unit 6 for determining the number of exchanges necessary for grasping and a procedure at that time, and an action determination unit 7 for determining the next type of action of the voice interaction device with the understanding state as an input, In the confirmation order assuming the request type, if the assumption is incorrect, the estimated number of exchanges required if the next request type is assumed And a confirmation content determination unit 8 that determines the contents of confirmation and questions that the device performs on the user, estimated from the difference between the required items in the request type assumed next and the action type according to the user's request The voice interaction device is configured to include the response content determination unit 9 that performs the above and the voice output unit 10 that outputs the confirmation content or the response content determined by the confirmation content determination unit 8 or the response content determination unit 9 as a voice.
[0008]
Further, the voice recognition unit 2 instructs the computer to convert the voice signal input via the voice input unit 1 into a character string, and the language understanding unit 3 obtains the command from the voice recognition unit 2. The requested character type probability distribution estimation unit 5 estimates the probability distribution of the user's requested content type using the understood state as an instruction to generate an understanding state consisting of a table of item names and values. The voice interaction device grasps the contents requested by the user when the exchange number estimation unit 6 assumes the request type requested by the user with the understanding state and the specific request type name as input. A command to determine the number of necessary exchanges and the procedure at that time, and a command to determine the type of next action of the voice interactive device by inputting the understanding state as an action determination unit 7 Part 8 assumes the request type In the confirmation order, if the assumption is incorrect and the next request type is assumed, the number of required exchanges is estimated. The command to determine the contents of the confirmation and question that the device performs to the user is estimated from the difference from the necessary items in the response, and the response content determination unit 9 issues a command to perform the action according to the user's request Then, the voice output program that instructs the voice output unit 10 to output the confirmation content or the response content determined by the confirmation content determination unit 8 or the response content determination unit 9 as a voice is configured.
[0009]
Furthermore, the voice recognition unit 2 instructs the computer to convert the voice signal input via the voice input unit 1 into a character string, and the language understanding unit 3 obtains the command from the voice recognition unit 2. The requested character type probability distribution estimation unit 5 estimates the probability distribution of the user's requested content type using the understood state as an instruction to generate an understanding state consisting of a table of item names and values. The voice interaction device grasps the content requested by the user when the number of exchanges is assumed and the request type requested by the user is assumed with the understanding state and the specific request type name as input. A command to determine the number of exchanges necessary and the procedure at that time, and the behavior determination unit 7 issues a command to determine the type of the next action of the voice interaction device with the understanding state as an input. 8 assumes the request type In the confirmation order, if the assumption is incorrect and the next request type is assumed, the number of required exchanges is estimated. The command to determine the contents of the confirmation and question that the device performs to the user is estimated from the difference from the necessary items in the response, and the response content determination unit 9 issues a command to perform the action according to the user's request Then, a recording medium on which a voice dialogue program that instructs the voice output unit 10 to output the confirmation content or the response content determined by the confirmation content determination unit 8 or the response content determination unit 9 as a voice is configured.
[0010]
DETAILED DESCRIPTION OF THE INVENTION
An embodiment of the present invention will be described with reference to the example of FIG.
The embodiment of the voice dialogue apparatus according to the present invention includes a voice input unit 1, a voice recognition unit 2, a language understanding unit 3, a task specification database 4, a request content type probability distribution estimation unit 5, an exchange count estimation unit 6, and an action determination unit 7. , A confirmation content determination unit 8, a response content determination unit 9, and a voice output unit 10. Here, although not shown in the drawings, each part of the voice interactive apparatus is constituted by a computer and a program for giving an operation command of each part to the computer.
[0011]
The voice input unit 1 converts the voice [1] uttered by the user into a digitized voice signal [2] . For this, a general microphone, a combination of a telephone and an AD converter can be used.
The voice recognition unit 2 receives the voice signal [2] from the voice input unit 1 and calculates a character string [3] of a pair of recognition result and probability of the recognition result. For this, an existing speech recognizer using a general hidden Markov model HMM can be used.
The language understanding unit 3 analyzes the character string [3] from the speech recognition unit 2 and creates an understanding state. As the understanding of state use, such as tables of whether or not the item name and certainty of value and values and verified [4]. To obtain the correspondence between item names and values, you can use an existing parser. Moreover, in order to obtain the correspondence between the value and the probability, the acoustic score of a speech recognizer using HMM can be used. After the confirmation issued from the device to the user, if there is an input that can be analyzed into a sentence corresponding to affirmation, whether or not confirmation has been confirmed is changed to confirmed. Whether or not an item that has not undergone this processing has been confirmed is unconfirmed (not confirmed).
[0012]
The task specification database 4 is a table in which item names necessary for each request type and values [5] that can be included in the items are described. For example, in the case of schedule management, the type of request “add schedule” requires “schedule content”, “date”, and “ time ”. Possible values for “schedule content” are “meeting”. “Meetings”, “Announcement”, etc., possible values for “Date” include date and day of the year, and possible values for “Time” are any of 24 hours.
The request content type probability distribution estimation unit 5 inputs the item name, value, and probability [4] of the value from the language understanding unit 3, estimates the probability of each request content from the current understanding state, and calculates the user request type name and A set of pairs [8] consisting of probabilities is output. For this purpose, vi is the value of the item si, ci is the certainty, Nj is the number of slots required in the request type Gj, Vj is a set of values allowed in Gj, and Mi is a range of values required in the range vi. When the number is the number, the degree of association between the understanding state S and the request type Gj is
[Expression 1]

Reι (S, Gj) when normalized by setting the sum of all the relevances to 1 as the probability of the request type Gj.
The number-of-interactions estimation unit 6 inputs a set [4] of the item name, the value, the probability of the value, and whether or not the value has been confirmed from the language understanding unit 3, and the necessary item name for each request type from the task specification database 4. and enter the possible values [6], which enters the item, in assuming a certain request types, so far device by user interaction required items in the request type from the current understanding states obtained efficiently If the confirmation is performed, the number of exchanges is estimated, and a set [7] of pairs consisting of the user request type name and the number of turns, which is an efficient confirmation procedure at this time, is output. The output is, for example, a combination of a confirmation procedure of 7 turns (number of turns) and confirmation of “date” → question of “time” and “content”.
[0014]
Using the number of vocabularies that can be accepted for each item, the recognition error rate when recognizing a certain item is found in proportion to the square root of the number of vocabularies. Assuming that the recognition rate (= 1−recognition error rate) obtained by this is r, the number of exchanges necessary for confirmation is estimated as 1 / r, and the number of exchanges necessary for questioning is 1 / r. Estimate r + 1. Whether it is confirmation or question is a question in [4] , if there is a value corresponding to the item, it is confirmation, and if the value is empty, it is a question. Since the number of exchanges obtained by the above calculation can be calculated from all confirmations or combinations of questions required for the assumed requirement type, the smallest number of exchanges and the number of exchanges The confirmation procedure and the number of times at that time are used as the output of the number-of-exchanges estimation unit 6.
[0015]
The action determination unit 7 inputs the item name, the value, and the probability [4] of the value from the language understanding unit 3, and the item name necessary for each request type and the value [5] that can be entered in the item from the task specification database 4. ] Is output, and an information response enable flag indicating whether the apparatus can respond to the user's request content or whether the confirmation is continued is output. Comparing the two, if the current understanding state satisfies all the necessary items for any request type, and if all have been confirmed, it is possible to respond with a response flag as having already grasped the request content Otherwise, the response enable flag is output as a response impossible.
[0016]
The confirmation content determination unit 8 outputs the probability distribution of the output of the request content type probability distribution estimation unit 5, the number of exchanges of the output of the exchange number estimation unit 6, the procedure at that time, and the “response possible flag” output from the behavior determination unit 7. enter a, or response flag is without applying any output if response available-response flag is responsive not possible, it may be confirmed or interrogating device will now what items at each time point To decide. What is output is a set [8] ′ of item name and confirmation or item name and request pairs. For example, the information (date confirmation) (day confirmation) is output. As inputs, the probability distribution estimated by the request type probability distribution estimation unit 5, the number of exchanges when the request type estimated by the exchange number estimation unit 6 is assumed, and the procedure at that time are input.
[0017]
Consider which user request the voice interaction device should confirm. The probability that the request type actually requested by the user is Gi is expressed as pGi, and the expected number of exchanges when the request type is assumed is expressed as tGi.
Although the apparatus performs the confirmation assuming a certain request type GS, the actual request type may be different. At this time, the items that must be confirmed in order to confirm the next request type are the differences between the items required for the previously assumed request type and the items required for the currently assumed request type. In other words, assuming the required types of items "date" and "time" are required first, and then assuming the required types of items "date", "time" and "planned contents" If the first assumed requirement type is actually incorrect, the only item required for the next assumed requirement type is “scheduled contents”.
[0018]
Necessary items other than this difference item are input to the number-of-interactions estimation unit 6 again in a state where the values have already been determined, and the estimation result of the number of exchanges necessary to complete the confirmation of this difference is obtained. obtain. If the confirmation of the difference number of exchanges up to Ga (i) fails, the estimated number of exchanges necessary for confirmation of Ga (i + 1) is expressed as follows.
[Expression 2]

The first request type Ga (1) in the selection order a (1), a (2),. When the assumption is made, the number-of-interactions estimation unit 6 obtains a confirmation procedure that is one of the output results, and the initial content of this confirmation procedure is the output [8] ′ of the confirmation content determination unit 8.
The response content determination unit 9 receives the understanding state from the language understanding unit and the information response enable flag from the behavior determination unit 7 as inputs. If the information response possible flag is impossible, no output is performed. If the information response possible flag is possible, it is necessary for each understanding type [7] 'and each request type input from the task specification database 4. The request type and the user's request content are obtained by comparing the item name and the value [5] that can be entered in the item, the response content programmed in advance for the request type is determined, and the character string [ 9] is output.
[0020]
The voice output unit 10 sets the item name and confirmation or item name / request pair set [8] ′ output from the confirmation content determination unit 8 or the response content character string [9 ] output from the response content determination unit 9. ] As input. It can be realized by a combination of an existing template-based language generator and an existing speech synthesizer.
The data flow of the voice dialogue in the embodiment is indicated by circled numbers in FIG.
Below, the specific example of the confirmation content determination part 8 is demonstrated.
As a department, a secretary task of the company is assumed, and the user request types can handle “delivery receipt of in-house flights” and “call of employees”.
[0021]
Information required for each request type and its value range are:
・ About "delivery receipt for internal flights"
Product name (in-house)
Department name (all departments)
Person name (all members)
Quantity (1-10)
・ About "call of employee"
Department name (related department)
Person name (related personnel)
Number representing day (1-30)
Quantity (1-50)
Suppose that
[0022]
Assuming each request, the estimated number of turns required until the end of confirmation is
・ About "delivery receipt for internal flights"
t (a) turn ・ About "calling employees"
Let t (b) turn.
The probability of the result of the probability estimation in the current understanding state of each request is
・ About "delivery receipt for internal flights"
p (a)
・ About "call of employee"
p (b)
Suppose that
[0023]
In the embodiment of the present invention, the confirmation content determination unit 8 calculates the expected number of turns for all permutations of request types. In this specific example, it is assumed that there are two types of request, “acceptance of in-house delivery” and “call of employees”, so possible permutations are:
(1) “In-house delivery delivery” → “Employee call”
(2) “Calling employees” → “In-house delivery”
There are two ways.
[0024]
In the case of (1), when considering the estimated number of turns when the request estimation first attempted to be confirmed is incorrect, the number of turns conventionally required is t (a) + t (b)
I was trying. In the embodiment of the present invention, even if the request estimation first attempted to be verified is incorrect, some of the information required there will be useful for the next assumed type of request. In order to confirm “calling of employees” after failure assuming “delivery of in-house flights”, information required for “calling of employees” is necessary for “delivery of in-house flights” The number of turns is estimated on the assumption that it is necessary to determine only the “number representing the day”, which is information that is not performed. If the number of turns estimated here is t ′ (b) (which should be smaller than t (b)), in the present invention, t (a) + t ′ (b)
Is the expected number of turns in this case. Thereby, more accurate turn number estimation can be performed.
[0025]
Similarly, in the case of (2), if an attempt to confirm “call of employee” fails, the estimation result of the number of turns necessary for confirming “delivery receipt of in-house flight” thereafter is t ′ ( a)
t ′ (a) + t (b)
Is the expected number of turns in this case. Accordingly, in the present invention, t (a) + (1-p (a)) (t (a) + t ′ (b)) in the case of (1)
And t (b) + (1−p (b)) (t ′ (a) + t (b)) in the case of (2)
Let's first try to identify the user request type that is the smallest of the ones that returned the smaller value.
[0026]
【The invention's effect】
As described above, according to the present invention, the user's request content can be grasped with a smaller number of conversations and the user's labor can be reduced.
[Brief description of the drawings]
FIG. 1 is a block diagram illustrating an embodiment.
[Explanation of symbols]
DESCRIPTION OF SYMBOLS 1 Speech input part 2 Speech recognition part 3 Language understanding part 4 Task specification database 5 Request content type probability distribution estimation part 6 Exchange frequency estimation part 7 Action determination part 8 Confirmation content determination part 9 Response content determination part 10 Voice output part

Claims

In a voice interaction method in which a user's voice is input and processing according to a user's request is executed and a response content is output by voice.
Generate an audio signal,
Convert this into a string based on the generated audio signal,
Parse the string to create an understanding state consisting of a table of item names and values,
It has a task specification database that describes the specifications of tasks that can currently be processed,
Estimate probability distribution of user request content type based on understanding state,
When the understanding type and specific request type name are input and the request type requested by the user is assumed, the number of exchanges required to grasp the content requested by the user and the procedure at that time are determined.
Determine the next action type based on your understanding,
In the confirmation order assuming the request type, if the assumption is incorrect, the estimated number of exchanges required if the next request type is assumed Is estimated from the difference between the required item and the next requirement type,
Obtain the sum of the number of exchanges for the estimated difference and the number of exchanges up to the previous time for each of the assumed next request types, and calculate the product of the sum and the probability that the previous understanding was wrong , Determine the sum of the product and the previous number of interactions, determine the type of request and confirmation to the user for the request type that minimizes the sum,
Take action according to user requirements,
A voice dialogue method characterized by outputting confirmation contents or response contents as voice.

In a voice interaction device that performs processing according to a user's request with the user's voice as input and outputs the response content by voice,
An audio input unit for converting audio into an audio signal;
A voice recognition unit that converts a voice signal obtained by the voice input unit into a character string;
Analyzing the character string obtained by the speech recognition unit, comprising a language understanding unit that creates an understanding state consisting of a table of item names and values,
A task specification database that describes the specifications of tasks that the voice interaction device can currently process;
A request content type probability distribution estimation unit that estimates a probability distribution of a user's request content type with an understanding state as an input;
Assuming the request type requested by the user with the understanding state and the specific request type name as input, the number of exchanges required for the voice interactive device to grasp the content requested by the user and the procedure at that time An exchange number estimation unit for determining
An action determining unit for determining the next action type of the voice interaction device with the understanding state as an input;
In the confirmation order that assumed the request type, if the assumption is incorrect, an estimate of the number of interactions required when the next request type is assumed And the difference between necessary items in the next assumed request type, and for each assumed next request type, find the sum of the number of exchanges for the estimated difference and the number of exchanges up to the previous time. And the probability that the previous understanding was wrong, the sum of the product and the number of exchanges up to the previous time is obtained, and the device performs confirmation to the user about the request type that minimizes the sum And a confirmation content determination unit for determining the content of the question,
A response content determination unit that performs an action according to a user request,
A voice dialogue apparatus comprising: a voice output unit that outputs the confirmation content or the response content determined by the confirmation content determination unit or the response content determination unit as a voice.

Against the computer,
The voice recognition unit instructs to convert the voice signal input via the voice input unit into a character string,
The language understanding unit analyzes the character string obtained by the speech recognition unit and gives an instruction to create an understanding state consisting of a table of item names and values,
The request content type probability distribution estimation unit gives an instruction to estimate the probability distribution of the user's request content type with the understanding state as input,
The number of exchanges required until the spoken dialogue device grasps the contents requested by the user, assuming that the request type requested by the user is input with the understanding state and a specific request type name as input. And a directive to determine the procedure at that time,
The action determining unit gives an instruction to determine the next action type of the voice interactive device with the understanding state as an input,
In the confirmation order where the confirmation content decision section assumed the request type, if the assumption was incorrect, the estimation of the number of exchanges required when the next request type was assumed was assumed up to the previous time. Estimated from the difference between the items required in the request type and the items required in the next assumed request type, the number of exchanges for the estimated difference and the number of exchanges up to the previous time for each of the assumed next request types Find the sum , find the product of the sum and the probability that the previous understanding was wrong, find the sum of the product and the number of exchanges until the previous time, and for the request type that minimizes the sum, the device Command the user to determine what to check and ask
The response content determination unit gives an instruction to perform an action according to the user's request,
A voice dialogue program characterized in that the voice output unit gives a command to output the confirmation content or the response content determined by the confirmation content determination unit or the response content determination unit as a voice.

Against the computer,
The voice recognition unit instructs to convert the voice signal input via the voice input unit into a character string,
The language understanding unit analyzes the character string obtained by the speech recognition unit and gives an instruction to create an understanding state consisting of a table of item names and values,
The request content type probability distribution estimation unit gives an instruction to estimate the probability distribution of the user's request content type with the understanding state as input,
The number of exchanges required until the spoken dialogue device grasps the contents requested by the user, assuming that the request type requested by the user is input with the understanding state and a specific request type name as input. And a directive to determine the procedure at that time,
The action determining unit gives an instruction to determine the next action type of the voice interactive device with the understanding state as an input,
In the confirmation order where the confirmation content decision section assumed the request type, if the assumption was incorrect, the estimation of the number of exchanges required when the next request type was assumed was assumed up to the previous time. Estimated from the difference between the items required in the request type and the items required in the next assumed request type, and for each assumed next request type, the number of exchanges for the estimated difference and the number of exchanges up to the previous time Find the sum , find the product of the sum and the probability that the previous understanding was wrong, find the sum of the product and the number of exchanges until the previous time, and for the request type that minimizes the sum, the device Command the user to determine what to check and ask
The response content determination unit gives an instruction to perform an action according to the user's request,
A recording medium in which an audio output program is recorded in which a voice output unit instructs to output the confirmation content or the response content determined by the confirmation content determination unit or the response content determination unit as a voice.