JP3770099B2 - Expert device, reasoning method, and expert system construction method - Google Patents

Abstract

PROBLEM TO BE SOLVED: To easily input the certainty factor of a proposition in a knowledge base or the like. SOLUTION: When constructing a knowledge base in an exert system or making an inquiry to the knowledge base, an inputted proposition (a sentence) is analyzed, certainty factor previously imparted to words used in the sentence is used to calculate the certainty factor of the sentence as a whole. In the case of 'nearly', a numerical value such as a certainly degree of 0.6 is assigned to a word, and if a plurality of words are used, calculation is done by using a calculating method (for example, a Comb method, a multiplying method, a mean value method, and the like) designated to each word. When making an inquiry to the knowledge base, a user has to reply to a question by voice, and the certainty factor is obtained from information such as the strength of a tone of the voice or whether a specific vocal sound is extended or not.

Description

となる。こうして計算した全体の確信度は、データベース構築部Ｍ５により集められ、命題データベースＳＤＢに、命題自身と共に記憶される。こうして、エキスパートシステムの基礎となる命題データベースＳＤＢが用意される。
【００２５】
（２）第２の実施の形態：
本発明の第２の実施の形態は、図２に示すように、第１の実施の形態の構成要素との比較により説明すると、文入力部Ｍ１に代えて音声入力部Ｍ１１が、形態素解析部Ｍ２に代えて音声分析部Ｍ１２が、字句解析部Ｍ３に代えて音声解析部Ｍ１３が、それぞれ設けられている。その余の構成は、第１の実施の形態とほぼ同一である。ここで、音声入力部Ｍ１１は、音声の認識までは行なっておらず、音の強さと長さを取り込むマイクやアンプからなる回路である。また音声分析部Ｍ２は、音声の強さを５段階に、長さを３段階に分けて認識する回路である。音声解析部Ｍ１３は、認識した音声の強さＶと長さＬとから、確信度辞書ＤＤＢを参照して、命題に関する入力者の確信度を求める回路である。このとき、単語は音声認識により入力しても良いし、第１の実施の形態と同様、キーボード等から入力するものとしても良い。単語毎に、音声の強さＶや長さＬから確信度ｘが得られたなら、これを集計してデータベースとして蓄積する点は、第１の実施の形態と同様である。
【００２６】
かかる構成によっても、命題についての入力者の入力に基づいて、そこに存在する特徴（この実施の形態では、音声の強さと長さ）により確信度を抽出することができ、データベースにこれを蓄積していくとができる。
【００２７】
上述した両実施の形態ともデータベースへの確信度と命題の記憶までしか含んでいないが、一旦これらの情報がデータベースとして蓄積されれば、このデータベースに蓄積された各命題に構造を与えることにより、これをエキスパートシステムとして構築することは容易である。
【００２８】
以上説明した本発明の実施の形態を一層明らかにするために、以下実施例について詳しく説明する。
（３）実施例における装置構成：
次に、本発明の実施例について説明する。図３は、本発明の第１実施例である確信度入力を行なう。図３に示した例では、本発明の確信度入力を具現化するプログラムが、インターネットのようなネットワーク１０に接続されたデータベースサーバ２０にインストールされている例を示している。データベースサーバ２０は、それ自身スタンドアロンのコンピュータとして使用可能であるが、以下で説明するように、サーバとして他のクライアントから利用することが可能である。すなわち、ネットワーク１０に接続された多数のコンピュータ３０，４０の利用者が、各人の知識をルール（命題）として取り出し、これをネットワーク１０を介してデータベースサーバ２０に入力することで、短期間に大規模なエキスパートシステムを、データベースサーバ２０に構築しようとするものである。入力部分については、データベースサーバ２０，コンピュータ３０，４０共に同じなので、以下、データベースサーバ２０を例として説明を行なう。なお、エキスパートシステムができあがれば、コンピュータ３０，４０からの問い合わせをネットワーク１０を介して受けたデータベースサーバ２０が、内部に構築されたエキスパートシステムを用いて推論を行ない、その推論果を、ネットワーク１０を介して、クライアントであるコンピュータ３０などに返すといった形態で利用される。エキスパートシステムとしては、どのようなものでも適用可能であるが、以下の説明では、病気の診断用エキスパートシステムを例として説明する。
【００２９】
データベースサーバ２０の内部構成を図３に基づいて説明する。データベースサーバ２０は、モデムやルータ１８を介してネットワーク１０とのデータのやり取りを制御するネットワークインタフェース（ＮＴ−Ｉ／Ｆ）２１、処理を行なうＣＰＵ２２、処理プログラムや固定的なデータを記憶するＲＯＭ２３、ワークエリアとしてのＲＡＭ２４、時間を管理するタイマ２５、モニタ２９への表示を司る表示回路２６、テキストデータをデータベースとして蓄積しているハードディスク（ＨＤ）２７、キーボード１１，マウス１２，マイク１３とのインタフェースを司る入力インタフェース（Ｉ／Ｆ）２８等を備える。なお、ハードディスク２７は、固定式のものとして記載したが、着脱式のものでも良いし、着脱式の記憶装置（例えばＣＤ−ＲＯＭ、ＣＤ−Ｒ、ＣＤ−ＲＷ、ＤＶＤ−ＲＯＭ、ＤＶＤ−ＲＡＭ、フレキシブルディスクなど）を併用することも可能である。また、この実施例では、データベースサーバ２０の処理プログラムは、ＲＯＭ２３内に記憶されているものとしたが、ハードディスク２７に記憶しておき、起動時にＲＡＭ２４上に展開して実行するものとしても良い。あるいは、上述した着脱式の記録媒体から読み込むものとしても良い。更には、ネットワーク１０を介して、他のサーバから読み込んで実行するものとしても良い。
【００３０】
ハードディスク２７には、仮名漢字変換辞書ＤＩＣが記憶されている。この仮名漢字変換辞書は、キーボード１１から入力されたかな文字列を解析して、仮名漢字文字列に変換する際に参照されるものである。この辞書ＤＩＣには、単語の読み、表記、文法情報の他に、図４に例示するように、単語の一部については、確信度のデータＣＮＦおよび確信度の集計方法を示す演算方法が記憶されている。もとよりこれらのデータは、全ての単語について付与されている訳ではなく、確信度に影響を与える可能性のある単語についてのみ付与されている。
【００３１】
図３に示したデータベースサーバ２０では、ハードディスク２７にデータベースが構築される。このデータベースＤＢは、基本的にはテキストデータからなる命題の集合として構築される。この命題は、キーボード１１から入力した文として入力される。なお、このデータベースＤＢを構築する際、キーボードから入力された文をそのまま登録しても良いが、一定の標準化を行なって登録しても良い。標準化を行なうことにより、文の作成者が異なる場合などに生じやすい用語のばらつきや表記のばらつきなどを統一することができ、データベースＤＢを用いてエキスパートシステムを検索する際のルール（構造）を簡略化することができる。
【００３２】
この文の標準化について簡略に説明する。データベースサーバ２０は、後述する文をハードディスク２７に登録する際、図５に示すように、まず文を形態素解析し、文が、文法情報を備えた単語から成り立つものとして把握する。その後、これらの単語から構成された文に対して、次の４つのレベルの標準化を行なう。（Ａ）文字の標準化処理（予め定めた文字に置き換える文字の標準化）、
文字レベルの表記のゆれであり、カタカナ語尾の長音の有無、カタカナ全角／半角といった全く概念に変化のない表記のゆれを統一するのである。
（Ｂ）表記の統一処理（表記のゆれを予め定めた表記に統一する処理）、
単語レベルの表記のゆれであり、ほとんど意味の異ならない複数漢字表記などを統一するのである。
（Ｃ）自立語処理（自立語を、予め定めた置き換えの基準に従って、他の自立語に置き換える処理）、
自立語の表記のゆれであり、わずかに意味の異なることがある単語間の入れ替えを行なって、使用する自立語を統一するのである。
【００３３】
なお、これらの標準化の他に、（Ｄ）付属語処理（文節レベルの表記のゆれの処理）も考えられるが、付属語処理の標準語は、文末表現の統一など、確信度の影響を与える要素を調整することがある。そこで、本実施例では、付属語の標準化処理は行なっていない。
【００３４】
上記の標準化の処理は、通常は、ハードディスク２７内に備えられた標準化辞書を参照することにより行なわれる。標準化辞書には、上記の標準化処理に対応した辞書が含まれており、形態素解析を行なった後、各単語について辞書を参照し、各単語について複数の表記、複数の単語が登録されていれば、これを予め定めた一つの表記、一つの単語に置き換えるのである。
【００３５】
次に、本実施例において、確信度の入力を行なう際の処理について説明する。図６は、確信度の入力を行なう際の処理プログラムを示すフローチャートである。図示するように、このプログラムは、命題（文）の入力の際に実施されるものである。この処理が起動されると、まずキーボード１１から文を入力する処理を行なう（ステップＳ１００、１０５）。入力は文単位で、即ち句点が入力されるまでを単位として行なわれる。文が入力されると、文の標準化処理を行なう（ステップＳ１１０）。標準化処理は上述したか、用語を統一するための処理である。例えば、「熱がある」「ねつがある」「ネツがある」などを一つの標準的な表記「熱がある」に統一するといった処理である。この後、入力した文が命題としふさわしいか否かのチェックを行なう（ステップＳ１２０）。
【００３６】
ここで入力している文は、最終的にエキスパートシステムを構築し、病名の診断などに用いられる。従って、こうした目的から見て命題として無意味な文（例えば、「病気はやだな。」「病い時は病気だ」）や、命題として矛盾した文（例えば、「熱が高くて平熱だ」）などは、この段階で排除される。
【００３７】
入力した文が、命題として適切であると判断した場合には（ステップＳ１２０）、次に、入力文を形態素解析する処理を行なう（ステップＳ１３０）。形態素解析を用いれば、文を構成する単語を容易に切り出すことができる。こうした形態素解析は、例えば最小コスト法といった仮名漢字変換の技術として確立した技術を用いることができる。図７は、入力した文ｎが「結核だとだいたい微熱が出たりする。」である場合を示す。この文ｎを形態素解析すると、図示するように、「結核」「だと」「だいたい」「微熱」「が」「出」「たりする」といった単語を取り出すことができる。
【００３８】
次に、これらの単語について、辞書ＤＩＣを参照して、確信度のデータＣＮＦと演算方法を読み出す処理を行なう（ステップＳ１４０）。なお、本実施例では、確信度のデータは、仮名漢字変換用の辞書ＤＩＣに記憶したが、確信度ＣＮＦと演算方法だけを与えるように別の辞書として用意しても良い。図４に示した例では、「だいたい」について確信度「０．６」、演算方法としてＣｏｍｂ法が、「たりする」について確信度「０．８」、演算方法として乗算法が、それぞれ見い出された。こうした確信度の数値は、例えば「わずかに」「やや」「より」「だいたい」「普通」「多分」「かなり」「非常に」「極めて」といった程度を表わす語について、それぞれ「わずかに（０．１）」「やや（０．２）」「より（０．３）」「だいたい（０．４）」「普通（０．５）」「多分（０．６）」「かなり（０．７）」「非常に（０．８）」「極めて（０．９）」のように、何段階かに分けて登録されている。なお、この例では、０．１刻みにしたが、０．１５など、１／２０刻みや１／１００刻みで値を与えても良い。また、単語毎に使用する演算方法として、最適のものが指定されている。一般に確信度を上げる種類の単語（例えば「一層」「非常に」など）についてはＣｏｍｂ法が、確信度を著しく下げる種類の単語（例えば「聞いた」「らしい」など）については乗算法が、また確信度に対して他の語との関連が小さい種類の単語（例えば「平均的には」「通常は」など）については平均値法が、それぞれ指定される傾向にある。
【００３９】
次に、得られた確信度を用いて、文ｎの全体の確信度を演算する処理を行なう（ステップＳ１５０）。この実施例では、確信度の計算は、単語「たりする」について指定された演算方法、つまり乗算法を用いることになる。この手法の詳細は、既に説明したので繰り返さないが、この例では、文ｎの全体の確信度ｘｎは、
ｘｎ＝０．６＊０．８＝０．４８
となった。なお、文中に現われた単語がいずれも確信度を持っていないようはな場合には、その文の確信度は０．５に設定される。例えば「天気は崩れる。」といった命題文の場合には、確信度を調整するような単語は存在しないので、文全体の確信度は、値０．５とされるのである。次に、この文全体の確信度ｘｎを文ｎと組にして、ハードディスク２７に記憶する処理を行なう（ステップＳ１６０）。以上で、本実施例における確信度の入力処理を完了する。
【００４０】
確信度の演算方法について、補足する。辞書ＤＩＣに、次のように単語毎の確信度と演算方法が登録されているものとする。

【００４１】
この場合、次の各文の確信度は、以下のように演算される。
例文１：「天気は崩れる。」
確信度 ０．５（文中に、確信度を有する単語がないからである）。
例文２：「天気はちょっと崩れるらしい。」
確信度ｘ２＝０．４＊０．５＝０．２
例文３：「天気はちょっと崩れるらしいと聞いた。」
確信度ｘ３＝（０．４＊０．５）＊０．５＝０．１
例文４：「天気はかなり激しく崩れる。」
確信度ｘ２′＝０．８＋０．９−０．８＊０．９＝０．９８
例文５：「天気はかなり激しく崩れるらしい。」
確信度ｘ３′＝０．９８＊０．５＝０．４９５
【００４２】
以上説明した本実施例の確信度入力方法によれば、キーボード１１から入力した文に含まれる単語について予め辞書ＤＩＣに用意した確信度と演算方法とを求め、これに基づいて文全体の確信度を与えることができる。従って、入力者は、命題（文）毎にいちいち確信度を数値入力したり、メンバーシップ関数を定義するといった煩わしい処理を行なう必要がない。また、文の確信度を、通常使用している自然な日本語文を介して設定できるので、数値を入力に伴う個人差などが生じにくいという利点も得られる。更に、確信度を単語毎に与えるだけでなく、確信度を集計する演算方法も単語毎に与えているので、確信度の決定をきめ細かく行なうことができる。また、本実施例では、入力文について標準化の処理を行なっているので、入力者により命題である文が微妙に異なると言うことがない。
【００４３】
（４）エキスパートシステムの構築：
以上説明した手法により、命題と確信度とが、ハードディスク２７内にデータベースとして蓄積された。そこで、次に、これをエキスパートシステムとして構成する手法について説明する。図８は、実現しようとするエキスパートシステムの概略構成図ある。図示するように、このエキスパートシステムは、知識ベースＡＢ、推論機構ＡＮ、知識獲得支援モジュールＡＳ、推論過程説明モジュールＡＥ、ユーザインタフェースＵＩなどから構成されている。上述した第１の実施例において命題と確信度を入力した処理は、知識獲得支援モジュールＡＳに相当し、その結果、得られた命題と確信度が、知識ベースＡＢに格納される。本実施例では、基本的にプロダクションシステムを用いており、知識ベースＡＢは、「鼻水が出ていれば風邪かも知れない。」「咳と微熱が続けば、結核の疑いがある。」といった「○○ならば、××である」というタイプの知識として、知識ベースＡＢは構成されている。
【００４４】
推論機構ＡＮは、こうして蓄積された知識ベースを用いて推論を行なうものである。推論は、ユーザインターフェースＵＩを介して入力された条件を用いて、前向きまたは後ろ向きの推論を適宜使い分けしつつ、行なわれる。ユーザインタフェースは、後述する第２実施例において説明するように、推論機構ＡＮから出した質問に、ユーザが回答するものを受け付け、これを推論機構ＡＮに渡す。例えば、「質問：どんな症状ですか？」「答え：鼻水が出て熱があります。」「質問：鼻水が出ているのですか？」「答え：はい。」「質問：熱があるのですね？」「答え：すごーく高いのです。」といったやり取りを介して、ユーザインタフェースＵＩは、推論機構ＡＮに、推論の基礎となる命題「鼻水か出ている」「熱がある」とそれぞれについてのユーザの確信度を渡す。
【００４５】
推論機構ＡＮは、これらの命題および確信度を用いて推論を行ない、結論を推論過程説明モジュールＡＥに出力する。推論過程説明モジュールＡＥは、推論の結果を、ユーザインタフェースＵＩを介して、使用者に提示する。例えば、上記の質問と回答から、熱が高いという命題の確信度は高いと判断し、「あなたは、風邪症状ですが、インフルエンザの可能性があります。」といった推論結果と、「鼻水が出ていて、熱がかなりあるからです」といった推論過程の説明を、例えばディスプレイ上に表示するのである。
【００４６】
（４）第２実施例：
次に、本発明の第２実施例について説明する。第２実施例は、第１実施例の処理により蓄積されたデータベースを、上述した手法でエキスパートシステムにおける知識ベースＡＢとして構築したケースにおいて、ユーザインタフェースＵＩが、ユーザの回答を推論機構ＡＮに渡す部分に適用されたものである。即ち、第２実施例の確信度入力は、上述した第１実施例と基本的には同じハードウェア構成を用い、推論機構ＡＮが、「質問：熱はありますか？」に対して、ユーザが答えた回答の確信度を入力するものである。第２実施例では、ユーザの音声を用いて、確信度の入力を行なう。
【００４７】
図９は、第２実施例における処理の概要を示すフローチャートである。この処理ルーチンは、エキスパートシステムが既にできあがっており、ユーザが、エキスパートシステムを利用する際に起動される。このルーチンが起動されると、まずモニタ２９に「症状を入力して下さい」という表示を行なう（ステップＳ２００）。ユーザは、この表示を見て、キーボード１１を用いて「熱があります」といった症状を入力する（ステップＳ２１０）。ユーザの入力を受け取ると、この文を解析し、推論機構ＡＮは、推論の条件が十分なものになるまで（ステップＳ２２０）、質問を繰り返す（ステップＳ２３０）。例えば、「どのくらい熱いのですか？」という質問を出力する。これに対してユーザは、キーボード１１から「すごく熱い」というように答えても良いが、本実施例では、マイク１３を用いて、回答を入力するように促される（ステップＳ２４０）。そこで、ユーザがマイク１３に向かって、「あつい」とか「あつーい」と返事をすると、これを入力し（ステップＳ２５０）、ユーザの発音に備わった特徴について解析する（ステップＳ２６０）。
【００４８】
ここで、音声の解析は、次の２点について行なわれる。一つは音声の強さであり、もう一つは音声の長さである。音声の強さは、簡便には、マイク１３から入力した音声の絶対値により認識することができるが、音声の絶対値はマイクと話者との距離や角度により変化しやすい。そこで、本実施例では、子音の部分と母音の部分との音声の強弱比を検出して音声の強さを判定している。いわゆる語気の強い言葉は、母音の部分に対する子音の部分の音が強いからである。母音の部分の検出は、音声に含まれるフォルマント構造を認識することにより検出することができる。音声に含まれるフォルマント構造を認識するためには、音声をフーリエ変換し、特定の周波数領域にピークが存在するかどうかを判定すればよい。こうして特定された母音部分の音の強さ（波高値の絶対値）と、この前後に存在する子音の部分（通常フーリエ変換により、含まれる周波数帯域が高い領域として認識できる）の音の強さとの比を採ることにより、子音の部分の強調を認識することができる。
【００４９】
一方、音声の長さは、次のように検出している。例えば「あつい」と「あつーい」を比べると、「つ」の継続期間が長いほど、話し手が確信を持っている、あるいは強調したいと考えていることが分かる。そこで、こうした音節が所定時間を超えて継続している部分があるかないかを検出するのである。一般に日本語の場合、継続できるのは母音の部分に限られるので、マイク１３から入力した音声をフーリエ変換し、フォルマント構造を見い出して母音の区間を特定し、この区間が通常の話し手の話のピッチから見て、２倍以上に及ぶ部分があれば、これを強調とみなすのである。
【００５０】
次に、解析した音声の特徴に基づいて、確信度を与える処理を行なう（ステップＳ２７０）。この場合、音声の強さと強調されている母音の長さを入力とするテーブルやマップを用意し、これを用いて、確信度を与えればよい。図１０に、音声の強さと特定の母音の長さに基づいて、確信度を与えるテーブルの一例を示す。ユーザインタフェースＵＩは、与えられた確信度ｘを、推論機構ＡＮに渡す。推論機構ＡＮは、入力された命題（「熱があります」）と、これについての確信度ｘとを組にして受け取り、知識ベースＡＢを参照して、推論を進める。推論が結論を導くことができる程度に進むまで、上記処理を繰り返す。
【００５１】
推論機構ＡＮが推論を完了したと判断すると（ステップＳ２２０）、この推論を結果をモニタ２９に表示し（ステップＳ２８０）、必要に応じて、推論の過程の説明を加え（ステップＳ２９０，Ｓ２９５）、本ルーチンを終了する。
【００５２】
以上説明した第２実施例によれば、使用者は、入力した命題に対する確信度をいちいち数値で入力する必要がなく、質問に対して、口頭でかつ普通の日本語で答えるだけで、確信度を入力することができる。従って、確信度の入力に煩瑣な作業を要求されることもなく、かつ数値入力における個人差や数値で入力するためについつい高い確信度を入れてしまう、といった弊害を生じることがない。エキスパートシステムは、「熱があります」といった単純な命題だけでなく、この命題に対する入力者の確信度も併せて入力することができるので、推論機構ＡＮを精密に運用して、精度の高い判断結果を示すことができる。
【００５３】
なお、上記実施例では、命題は、キーボード１１からの文の入力を基本としたが、音声認識を利用して命題を入力するようにしても差し支えない。更に、モニタ２９上にいくつかの選択肢を表示し、マウス１２を用いて選択することで入力するようにしても良い。例えば「熱があります」という入力に対して、「熱が高い」を異なる大きさの文字でいくつかボタン状に表示し、いずれかを入力者に選択させることで、確信度を受け付けるようにしても良い。あるいは文字の色を変えて、赤い文字を選択した場合には、確信度が高いといった関連づけを行なうことも可能である。いずれの手法にせよ、入力者の行なった入力に備わった特徴を解析して確信度を抽出することに変わりはない。上記の第２実施例において、確信度の判断を、音声の強弱や長短に代えて、あるいはこれらと共に、アクセントの強弱、基本周波数の高低などにより判断するものとしても良い。
【００５４】
以上、本発明の実施例といくつかの変形例について説明したが、本発明は上記実施例や変形例に何ら限定されるものではなく、その要旨を逸脱しない範囲内において、種々の態様により具現化されることは勿論である。例えば、上記実施例ではデータベースを構築してエキスパートシステムを作るものとしたが、単に確信度を入力し、コンピュータと人間の対話に味付けをするという利用にとどめても良い。人間との会話ができるコンピュータやロボットにおいて、対話相手である人間の発言を理解することは極めて重要な事項であり、意味論的な解析に加えて、話者の確信度を評価して応答を変えることは、自然なマンマシンインタフェースを実現する上で、極めて効果的である。また、電話の自動応答システムなどに適用し、お客様サービスなどの電話番号に電話してくるユーザの声から、クレームの緊急度を判断して、オペレータに接続する構成、１１０番などの緊急電話番号に掛けてきた人の要件の緊急性などの判断に用いた構成（結果的にいたずら電話などは自動応答で対処できる）、など種々なる態様で実施することができる。
【図面の簡単な説明】
【図１】本発明の第１の実施の態様を示すブロック図である。
【図２】本発明の第２の実施の態様を示すブロック図である。
【図３】本発明の実施例のハードウェア構成を示す説明図である。
【図４】仮名漢字変換辞書ＤＩＣの登録内容を説明する説明図である。
【図５】標準化の処理の概要を示す説明図である。
【図６】第１実施例における確信度入力処理ルーチンを示すフローチャートである。
【図７】確信度入力の一例を示す説明図である。
【図８】実施例におけるエキスパートシステムの概略構成を示す説明図である。
【図９】第２実施例における確信度入力処理ルーチンを示すフローチャートである。
【図１０】音声の特徴と確信度との関係を例示する説明図である。
【図１１】メンバーシップ関数の一例を示す説明図である。
【符号の説明】
１０…ネットワーク
１１…キーボード
１２…マウス
１３…マイク
１８…ルータ
２０…データベースサーバ
２２…ＣＰＵ
２３…ＲＯＭ
２４…ＲＡＭ
２５…タイマ
２６…表示回路
２７…ハードディスク
２９…モニタ
３０，４０…コンピュータ
ＡＢ…知識ベース
ＡＥ…推論過程説明モジュール
ＡＮ…推論機構
ＡＳ…知識獲得支援モジュール[0001]
BACKGROUND OF THE INVENTION
The present invention relates to a technique for inputting a certainty of an input person with respect to a proposition and a technique for constructing an expert system using the input of the certainty.
[0002]
[Prior art]
Conventionally, an attempt has been made to construct an expert system that makes it possible for anyone to use expert knowledge in a database. Initially, these expert systems were expressed as a combination of simple rules that would result in a “YES” or “NO” answer, but it is understood that human knowledge and the decisions that accompany it are ambiguous. Efforts have been made to make it possible to express this on a computer. In other words, an attempt was made to build an expert system based on a rule that allowed ambiguity such as "If it is XX, maybe XX" for a simple proposition "XX, if it is XX". It has come.
[0003]
In such a system, in order to express the probability of a proposition, a certainty factor is converted into a numerical value and attached to a rule, or a judgment (eg, “28 degrees C today”) is made for a condition (for example, “28 degrees C today”). “Today is hot”)) was defined as a membership function. In the former case, the certainty (hereinafter referred to as certainty factor) is expressed by a numerical value such as “0 to 1.0” or “−1.0 to 1.0”. For example, with respect to the proposition “the weather is good today”, a certainty factor is input as a certainty factor = [0.6] to define the certainty factor of each proposition.
[0004]
The latter, that is, a method using a membership function is a method taken in fuzzy theory, and gives a possibility y that a proposition holds for a variable x by a function as shown in FIG. For example, if the temperature T is 28 degrees C, the confidence factor 0.8 is obtained by using the membership function shown in FIG.
[0005]
[Problems to be solved by the invention]
However, there is a problem that such quantification of confidence and setting of membership function does not always work well in an expert system using human knowledge. For one thing, when inputting a huge number of rules (propositions), it is difficult to enter the certainty level numerically, and most rules are set to a certainty level of 0.5. There was a problem that it was easy to deal with only propositions by evaluating and giving numerical values. In this case, naturally, the accuracy and reliability of the system are lowered. Of course, it is possible to input certainty values for all propositions, but from the nature of the input person who quantifies the certainty, etc. There is a problem that when the system is viewed as a whole, the variation in judgment becomes large and the reliability is not necessarily increased.
[0006]
Membership functions are useful to some extent when conditions are easy to quantify (for example, “if body temperature is XX degrees”), but members that cannot be quantified (for example, “if you are sick”) A ship function cannot be defined, and significant judgment cannot be taken out well by combining these conditions. Moreover, considering the human life in general, it is often judged by combining a number of conditions that cannot be quantified, and it has been difficult to quantify this. In this respect as well, there was a big problem in putting the conventional expert system into practical use.
[0007]
An object of the present invention is to solve these problems, facilitate the input of certainty, make the construction of an expert system using this a practical level, and increase the reliability of system judgment.
[0008]
[Means for solving the problems and their functions and effects]
The expert apparatus of the present invention that solves at least a part of the above problems is as follows.
An expert device that constructs a knowledge base that accumulates propositions and stores the knowledge base in an available manner,
Input accepting means for accepting input from the input person regarding the proposition;
A proposition extraction means for extracting the proposition from the input of the input person;
A feature analysis means for analyzing the features provided in the received input;
Based on the analyzed feature, referring to data stored in association with the feature and the certainty factor in advance, and a certainty factor extracting means for extracting the certainty factor regarding the proposition as a numerical value;
A storage means for storing the extracted proposition and the extracted numerical value representing the certainty factor as a knowledge base;
The gist is that
[0009]
The invention of the reasoning method corresponding to such an expert device is as follows:
A method for inference by referring to a knowledge base in which propositions are accumulated,
Receiving input from the input person regarding the proposition via a keyboard or a voice recognition device;
The computer takes the proposition from the accepted input,
The computer analyzes the characteristics of the received input,
Based on the analyzed feature, referring to data stored in association with the feature and the certainty factor in advance, extract the certainty factor regarding the proposition as a numerical value,
Storing the extracted proposition and the extracted numerical value representing the certainty factor in a storage device as a knowledge base;
In response to an inquiry from the outside, the computer makes an inference by referring to the knowledge base and using the stored proposition and the numerical value representing the certainty factor.
Is the gist.
[0010]
According to the invention relating to the expert device and the inference method with the input of the certainty factor, when receiving an input related to a proposition (such as the proposition itself or an answer to the proposition), the characteristics of the input are analyzed and the data stored in advance is analyzed. Based on this, the certainty factor regarding the input proposition is extracted as a numerical value. Therefore, it is possible to input the certainty factor without any individual difference without requiring the input person to input the certainty factor numerically or to define a membership function. Moreover, no extra effort is required to input the certainty factor.
[0011]
Here, various methods can be considered for the input related to the proposition. For example, a sentence composed of a natural language may be input. When a sentence composed of a natural language is accepted, words constituting the sentence can be extracted by, for example, morphological analysis, and the certainty factor can be calculated using a predetermined word included in the sentence. As for the calculation of the certainty factor based on the word, for example, for the word representing the certainty factor, a numerical value for estimating the certainty factor is stored in advance, and this numerical value is read for the extracted word, thereby calculating the certainty factor. can do. For example, words that represent confidence are not only adverbs and conjunctions that indicate the level of "maybe", "approximately", "probably", "very", but adjectives such as "good", "bad", "great", and "unmotivated" And adjuncts, as well as particles such as “I think”, “I think” and “Kana”, auxiliary verbs, predicate verbs, etc. can be assumed. These can be collectively referred to as a relational expression or a framework expression for a narrative expression representing a proposition.
[0012]
Such relational expressions exist in any language. For example, taking Japanese as an example, the simple Japanese grammar framework gives only simple classifications such as listening and guessing. These expressions are not compatible with the norm of grammar, but rather are handled by comprehensive classification. Taking Japanese as an example, such an exhaustive collection and classification has already been done by Shuto et al. “Expression that gives a framework of Japanese sentence structure” (Kuniaki Suto, Toshiko Sugawara, Fukuoka University Research Institute Bulletin 63, (March 1983). Therefore, if you prepare a database, such as a dictionary, that gives certainty to each expression based on this exhaustive classification, you can extract words from sentences consisting of natural language using morphological analysis. Thus, it is easy to obtain the certainty by referring to the dictionary. Note that in languages such as English where words are independent in a sentence, the words can be extracted by simple filter processing without performing morphological analysis. Also, even in Japanese, it is easy to extract words if they are separated sentences.
[0013]
Thus, in order to calculate the certainty factor for the whole proposition using the certainty factor stored in advance for each word, the numerical value representing the certainty factor and the calculation method are stored for each word, and stored in the calculation of the certainty factor. The certainty factor expressed by the sentence may be calculated using a numerical value and a calculation method. Of course, if the words for storing numerical values are limited to specific words in advance (for example, words that generally increase the certainty level or words that generally decrease the certainty level), it is possible to limit the calculation method to one. .
[0014]
A numerical value representing the certainty factor assigned to each word may be a numerical value in the range of 0 to 1, and may include at least one of a comb method, a multiplication method, and an average value method as a calculation method. In particular, if you expand the range of words that store numerical values and also specify the calculation method for each word, you can make the calculation of certainty more flexible and calculate the certainty of the proposition expressed in natural language. Can be performed with high accuracy.
[0015]
In addition, since the word to which the certainty level is normally given is a word related to the above-described relational expression, it is naturally possible that the sentence indicating the proposition is composed only of words that do not store the certainty level. In such a case, that is, when the sentence does not include a word storing a numerical value representing the certainty level, the certainty level of the sentence can be set to a predetermined value, so that the certainty level can be easily input. It is useful in that.
[0016]
The Comb method in the calculation method is, for example, when confidence levels a and b (0 ≦ a, b ≦ 1) are given to two words, the confidence level C of a sentence in which these two appear as they are,
C = a + b−a * b
It is a method to ask for.
For example, when the certainty factors a and b are given to two words, the certainty factor C of a sentence in which these two appear as they are is as follows:
C = a * b
It is a method to ask for.
For example, when the certainty a and b are given to two words, the mean value method is a certainty C of a sentence in which these two appear as they are,
C = (a + b) / 2
(So-called “arithmetic average”).
Of course, the calculation method is not limited to these methods, and pays attention to the expression of confidence in a word, for example, a method for obtaining a synergistic average (C = √ (a * b)) in addition to an additive average, A method of obtaining an average value in consideration of a weighting count between words in a modification-modified relationship can also be employed.
[0017]
In addition, it is also possible to input the voice of the input person as the input related to the proposition, and calculate the certainty factor based on the characteristics of the voice. This is because it is known that a certain difference can be seen in the sound when talking about matters with certainty and matters with no certainty. As the characteristics of speech, the strength, length, fundamental frequency, accent, etc. of speech can be considered. At least one of them may be analyzed. Of course, it may be analyzed in a complex manner. In general, it is known that human beings are pronounced with strong, long and clear accents on matters that have certainty. This is a so-called heuristic rule that becomes apparent when the voices of many input users are processed. Therefore, when processing for one input person, it is more possible to perform voice input for a proposition whose confidence is known in advance and calibrate the relationship between the features on the voice and the confidence. desirable. It is known that the fundamental frequency of speech rises due to tension when the speaker lies or speaks with confidence that he or she is unsure. Therefore, it is possible to analyze the fundamental frequency of the voice of the input person through the input operation and extract the certainty factor by this. Of course, this information can also be used to correct the certainty factor stored for the words extracted by speech recognition.
[0018]
In the extraction of confidence based on speech, specifically, the strength, length, fundamental frequency, and accent of the speech are enhanced, the stronger the speech, the longer the speech lasts, and the lower the frequency, the more accented. As the degree of certainty increases, this can be done.
[0019]
The present invention can be understood as a method for constructing an expert system to which the above-described certainty factor input method is applied. That is, the expert system construction method of the present invention is
A method in which a computer constructs an expert system by accumulating a plurality of propositions together with the certainty of the input person for the propositions,
Receiving input from the input person regarding the proposition via a keyboard or a voice recognition device;
The computer takes the proposition from the accepted input,
The computer analyzes the characteristics of the received input,
Based on the analyzed feature, referring to data stored in association with the feature and the certainty factor in advance, extract the certainty factor regarding the proposition as a numerical value,
The extracted proposition and the extracted numerical value representing the certainty level are stored in a storage device as a database,
By giving a structure to each proposition stored in the database and making it a knowledge base
The gist is to build an expert system.
[0020]
According to such an expert stem construction method, just by performing input related to a proposition, the certainty about the proposition is extracted, the proposition and the certainty are stored in a database, and each proposition stored in this database has a structure. Can be constructed as an expert system. Therefore, it is possible to easily construct a highly reliable expert system.
[0021]
Each of these inventions can be understood as an invention of a program that is executed on a computer and realizes the above functions, and an invention as a recording medium on which such a program is recorded. The certainty factor input device or expert system may realize the above-described input acceptance, feature analysis, certainty factor extraction, etc. by executing a program on a computer, or a discrete circuit. You may implement | achieve from a structure. As the program, a well-known program language such as C language, Pascal, Fortran, Cobol, BASIC, or the like can be used, and an object-oriented program language, a language such as JavaScript, or the like can also be used. As the recording medium, various recording media such as a flexible disk, CD-ROM, DVD-ROM, and semiconductor memory (ROM, PROM, EEPROM, flash memory, etc.) can be used. Of course, these programs can be stored in a server placed on a network such as the Internet, and downloaded to a client computer for use.
[0022]
DETAILED DESCRIPTION OF THE INVENTION
Embodiments of the present invention and examples thereof will be described below.
(1) First embodiment:
FIG. 1 is a block diagram showing an embodiment of the present invention. In this embodiment, the proposition is input via a sentence input unit M1 such as a keyboard. The input sentence is a so-called natural language sentence, and the morpheme analysis unit M2 performs processing for extracting words t1, t2,... Constituting the sentence. When the sentence input unit M1 such as a keyboard inputs a sentence with kana-kanji conversion, the sentence input unit M1 and the morpheme analysis unit M2 are integrally configured so that when inputting with kana-kanji change. It is easy to extract a word. The word taken out by the morphological analysis is sent to the lexical analyzer M3, where the certainty factor is analyzed. The lexical analyzer M3 refers to a database (hereinafter referred to as a certainty factor dictionary) DDB in which the certainty factor for each word is prepared in advance, and the certainty factor d (t) of the word constituting the sentence and the certainty factor aggregation method (hereinafter referred to as the certainty factor method). , Simply referred to as a calculation method). The certainty factor and the calculation method for each word thus extracted are output to the certainty factor totaling unit M4, where the certainty factor is calculated based on the calculation method given for each word.
[0023]
The certainty factor is calculated as follows. Here, the sentence “The weather seems to be a bit broken” will be explained as an example. When four words “weather”, “slightly”, “crash”, and “likely” are extracted from this sentence, a certainty factor is given to “weather” and “crash” when the confidence dictionary DDB is referenced. It is assumed that the certainty factor 0.4 is given for “little” and the certainty factor 0.5 is given for “like”. Further, it is assumed that an average value method is given as a calculation method for “a little”, and a multiplication method is given for the calculation method. In this case, the certainty factor x2 for the proposition “The weather is going to collapse” uses the calculation method of the word placed behind in the sentence,
x2 = 0.4 * 0.5 = 0.2
It becomes.
[0024]
Note that if the input sentence has one word with certainty, such as “the weather seems to break down”, b = 1 in the division method, and the certainty of the word is the certainty x1 of the whole proposition. Calculated. In addition, when three or more words with certainty are found, such as “I heard that the weather seems to be a bit crushed”, these may be calculated in order. In the above sentence, if “heard” is confidence 0.5 and the calculation method is multiplication,

It becomes. The total certainty calculated in this way is collected by the database construction unit M5 and stored together with the proposition itself in the proposition database SDB. In this way, the proposition database SDB which is the basis of the expert system is prepared.
[0025]
(2) Second embodiment:
As shown in FIG. 2, the second embodiment of the present invention will be described by comparison with the components of the first embodiment. The voice input unit M11 is replaced with a morphological analysis unit instead of the sentence input unit M1. A voice analysis unit M12 is provided instead of M2, and a voice analysis unit M13 is provided instead of the lexical analysis unit M3. The rest of the configuration is almost the same as in the first embodiment. Here, the voice input unit M11 is a circuit made up of a microphone and an amplifier that does not perform voice recognition but takes in the intensity and length of the sound. The voice analysis unit M2 is a circuit that recognizes the voice intensity in five levels and the length in three levels. The speech analysis unit M13 is a circuit that obtains the confidence level of the input person regarding the proposition by referring to the confidence level dictionary DDB from the recognized voice strength V and length L. At this time, the word may be input by voice recognition, or may be input from a keyboard or the like as in the first embodiment. As in the first embodiment, if the certainty factor x is obtained from the voice strength V and length L for each word, this is aggregated and stored as a database.
[0026]
Even with such a configuration, it is possible to extract the certainty factor based on the input from the input person about the proposition based on the features (in this embodiment, the strength and length of the voice), and store this in the database. You can do it.
[0027]
Both of the above-described embodiments only include the degree of confidence in the database and the storage of propositions, but once these information are stored as a database, by giving a structure to each proposition stored in this database, It is easy to construct this as an expert system.
[0028]
In order to further clarify the embodiment of the present invention described above, examples will be described in detail below.
(3) Device configuration in the embodiment:
Next, examples of the present invention will be described. FIG. 3 performs a certainty factor input according to the first embodiment of the present invention. In the example shown in FIG. 3, an example is shown in which the program for realizing the certainty factor input of the present invention is installed in the database server 20 connected to the network 10 such as the Internet. The database server 20 can be used as a stand-alone computer, but can be used as a server from other clients as described below. That is, users of a large number of computers 30 and 40 connected to the network 10 take out each person's knowledge as a rule (proposition) and input it to the database server 20 via the network 10. A large-scale expert system is to be constructed in the database server 20. Since the input portion is the same for both the database server 20 and the computers 30 and 40, the database server 20 will be described below as an example. When the expert system is completed, the database server 20 that receives the inquiry from the computers 30 and 40 via the network 10 performs inference using the expert system built therein, and the inference result is obtained from the network 10. It is used in such a form that it is returned to the client computer 30 or the like through Any expert system can be applied, but in the following description, an expert system for diagnosing a disease will be described as an example.
[0029]
The internal configuration of the database server 20 will be described with reference to FIG. The database server 20 includes a network interface (NT-I / F) 21 that controls data exchange with the network 10 via a modem or a router 18, a CPU 22 that performs processing, a ROM 23 that stores processing programs and fixed data, RAM 24 as a work area, timer 25 for managing time, display circuit 26 for managing display on the monitor 29, hard disk (HD) 27 storing text data as a database, keyboard 11, mouse 12, and microphone 13 interface Input interface (I / F) 28 and the like. Although the hard disk 27 is described as being fixed, it may be removable, or a removable storage device (for example, CD-ROM, CD-R, CD-RW, DVD-ROM, DVD-RAM, A flexible disk or the like can also be used together. In this embodiment, the processing program of the database server 20 is stored in the ROM 23. However, the processing program may be stored in the hard disk 27 and expanded and executed on the RAM 24 at startup. Alternatively, it may be read from the above-described removable recording medium. Further, it may be executed by reading from another server via the network 10.
[0030]
The hard disk 27 stores a kana-kanji conversion dictionary DIC. This kana-kanji conversion dictionary is referred to when a kana character string input from the keyboard 11 is analyzed and converted into a kana-kanji character string. In this dictionary DIC, in addition to word reading, notation, and grammatical information, as illustrated in FIG. 4, as shown in FIG. 4, a certainty factor data CNF and a calculation method indicating a certainty factor aggregation method are stored. Has been. Of course, these data are not given to all words, but are given only to words that may affect the certainty factor.
[0031]
In the database server 20 shown in FIG. 3, a database is constructed on the hard disk 27. This database DB is basically constructed as a set of propositions consisting of text data. This proposition is input as a sentence input from the keyboard 11. In constructing the database DB, a sentence input from the keyboard may be registered as it is, or may be registered after performing a certain standardization. By standardizing, it is possible to unify the variability of terms and notation that are likely to occur when the creators of sentences are different, etc., and simplify the rules (structure) when searching expert systems using the database DB Can be
[0032]
The standardization of this sentence will be briefly described. When registering a sentence to be described later on the hard disk 27, the database server 20 first performs a morphological analysis on the sentence, as shown in FIG. 5, and recognizes that the sentence is composed of words having grammatical information. After that, the following four levels of standardization are performed on sentences composed of these words. (A) Character standardization processing (standardization of characters to be replaced with predetermined characters),
It is a fluctuation of the character level notation, unifying the fluctuation of the notation that does not change the concept at all such as the presence or absence of a long sound of the katakana ending, the katakana full-width / half-width.
(B) Notation unification process (process to unify fluctuation of notation to a predetermined notation),
It is a fluctuation of word level notation, and unifies multiple kanji notations that have almost no different meaning.
(C) Independent word processing (processing for replacing independent words with other independent words in accordance with predetermined replacement criteria),
It is a variation of the independence word notation, and it replaces words that may have slightly different meanings to unify the independence words used.
[0033]
In addition to these standardizations, (D) ancillary word processing (processing for fluctuations in phrase level notation) is also conceivable, but the standard word for adjunct word processing has an effect on certainty such as unification of sentence ending expressions. May adjust elements. Therefore, in this embodiment, standardization processing for attached words is not performed.
[0034]
The standardization process described above is normally performed by referring to a standardized dictionary provided in the hard disk 27. The standardized dictionary includes a dictionary corresponding to the above standardization process. After performing morphological analysis, the dictionary is referred to for each word, and a plurality of notations and a plurality of words are registered for each word. This is replaced with one predetermined notation, one word.
[0035]
Next, a process when inputting a certainty factor in the present embodiment will be described. FIG. 6 is a flowchart showing a processing program for inputting the certainty factor. As shown in the figure, this program is executed when a proposition (sentence) is input. When this process is started, a process of inputting a sentence from the keyboard 11 is first performed (steps S100 and S105). Input is performed in sentence units, that is, in units until a phrase is input. When a sentence is input, a sentence standardization process is performed (step S110). The standardization process is described above or a process for unifying terms. For example, the process of unifying “having fever”, “having a net”, “having a net” into one standard notation “having fever”. Thereafter, it is checked whether or not the inputted sentence is appropriate as a proposition (step S120).
[0036]
The sentence entered here is finally used to construct an expert system and diagnose the disease name. Therefore, a sentence that is meaningless as a proposition (for example, “Illness is dangerous” or “Illness is illness”) or an inconsistent sentence (for example, “high fever and normal heat” Etc.) are eliminated at this stage.
[0037]
If it is determined that the input sentence is appropriate as a proposition (step S120), a process for analyzing the morpheme is performed on the input sentence (step S130). If morphological analysis is used, words constituting a sentence can be easily cut out. For such morphological analysis, a technique established as a kana-kanji conversion technique such as a minimum cost method can be used. FIG. 7 shows a case where the input sentence n is “Slightly fever occurs when tuberculosis occurs”. If the sentence n is morphologically analyzed, words such as “tuberculosis”, “dato”, “generally”, “slight fever”, “ga”, “out”, and “toru” can be extracted as shown in the figure.
[0038]
Next, with respect to these words, the process of reading the certainty factor data CNF and the calculation method is performed with reference to the dictionary DIC (step S140). In this embodiment, the certainty factor data is stored in the kana-kanji conversion dictionary DIC, but may be prepared as another dictionary so as to give only the certainty factor CNF and the calculation method. In the example shown in FIG. 4, a certainty factor “0.6” is found for “approximately”, a Comb method is found as the calculation method, a certainty factor “0.8” is found for “Taruru”, and a multiplication method is found as the calculation method. It was. These certainty figures are, for example, “slightly”, “slightly”, “more”, “approximately”, “normal”, “probably”, “pretty”, “very”, “very”, .1) ”“ Slightly (0.2) ”“ More (0.3) ”“ About (0.4) ”“ Normal (0.5) ”“ Maybe (0.6) ”“ Much (0.7 ) ”,“ Very (0.8) ”,“ Very (0.9) ”, and so on. In this example, the value is set in increments of 0.1, but values may be given in increments of 1/20 or 1/100, such as 0.15. In addition, an optimum calculation method is designated for each word. In general, the Comb method is used for a type of word that increases the certainty level (for example, “one layer” or “very”), and the multiplication method is used for a type of word that significantly reduces the certainty level (for example, “heard” or “like”). Also, the average value method tends to be designated for each type of word (for example, “on average”, “normally”, etc.) whose relation to other words is small relative to the certainty factor.
[0039]
Next, using the obtained certainty factor, a process of calculating the entire certainty factor of the sentence n is performed (step S150). In this embodiment, the certainty factor is calculated using a calculation method designated for the word “Taru”, that is, a multiplication method. Details of this method have already been described and will not be repeated, but in this example, the overall confidence xn of sentence n is
xn = 0.6 * 0.8 = 0.48
It became. If none of the words appearing in the sentence has a certainty level, the certainty level of the sentence is set to 0.5. For example, in the case of a proposition sentence such as “Weather breaks down”, since there is no word for adjusting the certainty factor, the certainty factor of the whole sentence is set to 0.5. Next, a process for storing the certainty factor xn of the entire sentence together with the sentence n and storing it in the hard disk 27 is performed (step S160). Thus, the certainty factor input process in this embodiment is completed.
[0040]
It supplements about the calculation method of reliability. It is assumed that the certainty factor and the calculation method for each word are registered in the dictionary DIC as follows.

[0041]
In this case, the certainty factor of each next sentence is calculated as follows.
Example sentence 1: “The weather collapses.”
Certainty factor 0.5 (because there is no word with certainty factor in the sentence).
Example sentence 2: “The weather seems to be a little crumbled.”
Certainty factor x2 = 0.4 * 0.5 = 0.2
Example sentence 3: “I heard that the weather seems to be a little bit broken.”
Certainty factor x3 = (0.4 * 0.5) * 0.5 = 0.1
Example sentence 4: “The weather is going to collapse quite violently.”
Certainty factor x2 ′ = 0.8 + 0.9−0.8 * 0.9 = 0.98
Example sentence 5: “The weather seems to collapse quite violently.”
Certainty factor x3 ′ = 0.98 * 0.5 = 0.495
[0042]
According to the certainty factor input method of the present embodiment described above, the certainty factor and the calculation method prepared in advance in the dictionary DIC for the word included in the sentence input from the keyboard 11 are obtained, and the certainty factor of the entire sentence based on this. Can be given. Therefore, the input user does not need to perform troublesome processing such as inputting a certainty factor for each proposition (sentence) and defining a membership function. In addition, since the certainty of a sentence can be set through a natural Japanese sentence that is normally used, there is an advantage that individual differences and the like due to numerical input are less likely to occur. Furthermore, since not only the certainty factor is given for each word but also a calculation method for counting the certainty factors is given for each word, the certainty factor can be determined in detail. Further, in this embodiment, since standardization processing is performed on the input sentence, the sentence as the proposition is not slightly different depending on the input person.
[0043]
(4) Construction of expert system:
By the method described above, the proposition and the certainty factor are stored in the hard disk 27 as a database. Then, next, the method of comprising this as an expert system is demonstrated. FIG. 8 is a schematic configuration diagram of an expert system to be realized. As shown in the figure, this expert system includes a knowledge base AB, an inference mechanism AN, a knowledge acquisition support module AS, an inference process explanation module AE, a user interface UI, and the like. The processing in which the proposition and the certainty factor are input in the first embodiment described above corresponds to the knowledge acquisition support module AS, and as a result, the obtained proposition and certainty factor are stored in the knowledge base AB. In the present embodiment, a production system is basically used, and the knowledge base AB has a suspicion of tuberculosis such as "If you have a runny nose, you may have a cold." The knowledge base AB is configured as a type of knowledge that “XX means XX”.
[0044]
The inference mechanism AN performs inference using the knowledge base thus accumulated. The inference is performed while appropriately using forward or backward inference using conditions input via the user interface UI. As will be described in a second embodiment to be described later, the user interface accepts what a user answers to a question issued from the inference mechanism AN, and passes this to the inference mechanism AN. For example, “Question: What are the symptoms?” “Answer: I have a runny nose and I have fever.” “Question: Is my runny nose?” “Answer: Yes.” “Question: I have a fever. Through the exchanges such as “?” And “Answer: It is very expensive”, the user interface UI gives the inference mechanism AN the propositions that are the basis of inference: “I have a runny nose” and “I have a fever”. Pass the certainty of the user.
[0045]
The inference mechanism AN performs inference using these propositions and certainty factors, and outputs a conclusion to the inference process explanation module AE. The inference process explanation module AE presents the inference result to the user via the user interface UI. For example, from the above questions and answers, it is judged that the confidence of the proposition of high fever is high, and an inference result such as “You are a cold symptom, but there is a possibility of influenza.” For example, the reasoning process is displayed on the display.
[0046]
(4) Second embodiment:
Next, a second embodiment of the present invention will be described. The second embodiment is a part in which the user interface UI passes the user's answer to the inference mechanism AN in the case where the database accumulated by the processing of the first embodiment is constructed as the knowledge base AB in the expert system by the above-described method. Applied. In other words, the reliability input of the second embodiment uses basically the same hardware configuration as that of the first embodiment described above, and the inference mechanism AN determines that the user asks "Question: Do you have heat?" This is for inputting the certainty of the answer that has been answered. In the second embodiment, the certainty factor is input using the user's voice.
[0047]
FIG. 9 is a flowchart showing an outline of processing in the second embodiment. This processing routine is started when the expert system is already prepared and the user uses the expert system. When this routine is started, first, a message “Please input symptoms” is displayed on the monitor 29 (step S200). The user views this display and inputs a symptom such as “there is fever” using the keyboard 11 (step S210). When the user input is received, this sentence is analyzed, and the inference mechanism AN repeats the question (step S230) until the inference condition is sufficient (step S220). For example, the question “How hot is it?” Is output. In response to this, the user may answer “very hot” from the keyboard 11, but in this embodiment, the user is prompted to input an answer using the microphone 13 (step S240). Therefore, when the user replies “hot” or “hot” to the microphone 13, the user inputs this (step S 250), and analyzes the features of the user's pronunciation (step S 260).
[0048]
Here, the voice analysis is performed on the following two points. One is the voice strength, and the other is the voice length. The strength of the voice can be easily recognized by the absolute value of the voice input from the microphone 13, but the absolute value of the voice is likely to change depending on the distance and angle between the microphone and the speaker. Therefore, in this embodiment, the strength of the voice is determined by detecting the strength ratio of the voice between the consonant part and the vowel part. This is because so-called strong words have strong consonant sounds relative to vowel parts. The vowel part can be detected by recognizing the formant structure included in the speech. In order to recognize the formant structure included in the speech, the speech may be Fourier transformed to determine whether a peak exists in a specific frequency region. The sound intensity of the vowel part specified in this way (absolute value of the crest value) and the sound intensity of the consonant part existing before and after this (normally recognized as an area having a high frequency band by Fourier transform) By adopting the ratio, it is possible to recognize the emphasis of the consonant part.
[0049]
On the other hand, the voice length is detected as follows. For example, comparing “hot” with “hot”, it can be seen that the longer the duration of “tsu”, the more confident or more stressed the speaker is. Therefore, it is detected whether or not there is a part in which such a syllable continues for a predetermined time. In general, in the case of Japanese, since only the vowel part can be continued, the voice input from the microphone 13 is Fourier transformed, the formant structure is found, the vowel section is specified, and this section is used for the normal speaker's story. If there is a part that is twice or more as seen from the pitch, this is regarded as emphasis.
[0050]
Next, a process of giving a certainty factor is performed based on the analyzed voice feature (step S270). In this case, a table or map for inputting the strength of the voice and the length of the emphasized vowel may be prepared and used to give certainty. FIG. 10 shows an example of a table that gives a certainty factor based on the strength of voice and the length of a specific vowel. The user interface UI passes the given certainty factor x to the inference mechanism AN. The inference mechanism AN receives the input proposition (“there is heat”) and the certainty factor x about this, and advances the inference with reference to the knowledge base AB. The above process is repeated until the reasoning is advanced enough to draw a conclusion.
[0051]
When the inference mechanism AN determines that the inference has been completed (step S220), the result is displayed on the monitor 29 (step S280), and an explanation of the inference process is added if necessary (steps S290 and S295). This routine ends.
[0052]
According to the second embodiment described above, the user does not have to input the certainty factor for the input proposition numerically, but can answer the question verbally and in ordinary Japanese. Can be entered. Therefore, no troublesome work is required for inputting the certainty factor, and there is no adverse effect that a high certainty factor is put in order to input individual differences or numerical values in the numerical value input. The expert system can input not only a simple proposition such as “I have a fever” but also the confidence of the input person for this proposition. Can be shown.
[0053]
In the above embodiment, the proposition is based on the input of a sentence from the keyboard 11. However, the proposition may be input using voice recognition. Furthermore, some options may be displayed on the monitor 29 and selected by using the mouse 12 for input. For example, in response to an input “There is fever”, “high fever” is displayed in a button shape with letters of different sizes, and the input person selects one to accept the certainty level. Also good. Alternatively, when the character color is changed and a red character is selected, it is possible to make an association such that the certainty factor is high. Whichever method is used, there is no change in extracting the certainty factor by analyzing the characteristics of the input made by the input person. In the second embodiment, the certainty factor may be determined based on the strength of the accent, the level of the fundamental frequency, or the like instead of, or together with, the strength or weakness of the voice.
[0054]
As mentioned above, although the Example of this invention and some modifications were demonstrated, this invention is not limited to the said Example and modification at all, and it implements with a various aspect in the range which does not deviate from the summary. Of course, For example, in the above embodiment, an expert system is created by constructing a database, but it may be used only for inputting a certainty factor and seasoning a computer-human interaction. In computers and robots that can communicate with human beings, it is extremely important to understand the remarks of human beings who interact with them. In addition to semantic analysis, we evaluate the confidence level of speakers and respond to them. Changing is extremely effective in realizing a natural man-machine interface. In addition, it is applied to an automatic answering system for telephones, etc., and is configured to determine the urgency level of a complaint from the voice of a user who calls a telephone number for customer service, etc., and to connect to an operator, emergency telephone number such as 110 The configuration used for judging the urgent nature of the requirements of the person who has made a call (as a result, mischievous calls can be dealt with by an automatic response), etc.
[Brief description of the drawings]
FIG. 1 is a block diagram showing a first embodiment of the present invention.
FIG. 2 is a block diagram showing a second embodiment of the present invention.
FIG. 3 is an explanatory diagram illustrating a hardware configuration according to the embodiment of this invention.
FIG. 4 is an explanatory diagram for explaining registration contents of a kana-kanji conversion dictionary DIC.
FIG. 5 is an explanatory diagram showing an outline of standardization processing;
FIG. 6 is a flowchart showing a certainty factor input processing routine in the first embodiment;
FIG. 7 is an explanatory diagram showing an example of a certainty factor input.
FIG. 8 is an explanatory diagram showing a schematic configuration of an expert system in an embodiment.
FIG. 9 is a flowchart showing a certainty factor input processing routine in the second embodiment;
FIG. 10 is an explanatory diagram illustrating the relationship between voice characteristics and certainty.
FIG. 11 is an explanatory diagram showing an example of a membership function.
[Explanation of symbols]
10 ... Network
11 ... Keyboard
12 ... Mouse
13 ... Microphone
18 ... Router
20 ... Database server
22 ... CPU
23 ... ROM
24 ... RAM
25 ... Timer
26. Display circuit
27 ... Hard disk
29 ... Monitor
30, 40 ... Computer
AB ... Knowledge Base
AE ... Inference process explanation module
AN ... Inference mechanism
AS ... Knowledge acquisition support module

Claims (7)

An expert device that constructs a knowledge base that accumulates propositions and stores the knowledge base in an available manner,
Input accepting means for accepting input from the input person regarding the proposition;
A proposition extraction means for extracting the proposition from the input of the input person;
A feature analysis means for analyzing the features provided in the received input;
Based on the analyzed feature, referring to data stored in association with the feature and the certainty factor in advance, and a certainty factor extracting means for extracting the certainty factor regarding the proposition as a numerical value;
An expert apparatus comprising: storage means for storing the extracted proposition and the extracted numerical value representing the certainty factor as a knowledge base. The expert device according to claim 1,
The input receiving means is means for inputting a sentence composed of a natural language as an input related to the proposition,
The feature analysis means is means for extracting words constituting the accepted sentence;
The certainty factor extraction means stores in advance a numerical value for estimating a certainty factor for each of a plurality of predetermined words, and when the extracted word is included in the predetermined plurality of words, the extracted word An expert apparatus comprising a certainty factor calculation means for reading the numerical value stored for the word and calculating the certainty factor of the proposition including the word using the numerical value. The expert device according to claim 2,
The certainty factor calculating means includes:
Storing a numerical value representing a certainty factor and a calculation method for each of the predetermined plurality of words;
In the calculation of the certainty factor, the expert device is a means for calculating the certainty factor that constitutes the sentence and expresses the proposition including the word using the stored numerical value and the calculation method. The expert device according to claim 1,
The input receiving means is means for inputting an input person's voice as input related to the proposition,
The feature analysis means is means for extracting features provided in the received voice,
The certainty factor extracting means is means for calculating a certainty factor by referring to data stored in association with the extracted voice feature and the certainty factor in advance. The expert device according to claim 4,
The feature analysis means is an expert device comprising means for analyzing at least one of the strength, length, fundamental frequency, and accent of the voice as the voice feature. A method for inference by referring to a knowledge base in which propositions are accumulated,
Receiving input from the input person regarding the proposition via a keyboard or a voice recognition device;
The computer takes the proposition from the accepted input,
The computer analyzes the characteristics of the received input,
Based on the analyzed feature, referring to data stored in association with the feature and the certainty factor in advance, extract the certainty factor regarding the proposition as a numerical value,
Storing the extracted proposition and the extracted numerical value representing the certainty factor in a storage device as a knowledge base;
An inference method in which a computer makes an inference with reference to the knowledge base, using the stored proposition and a numerical value representing a certainty factor, in response to an inquiry from the outside. A method in which a computer constructs an expert system by accumulating a plurality of propositions together with the certainty of the input person for the propositions,
Receiving input from the input person regarding the proposition via a keyboard or a voice recognition device;
The computer takes the proposition from the accepted input,
The computer analyzes the characteristics of the received input,
Based on the analyzed feature, referring to data stored in association with the feature and the certainty factor in advance, extract the certainty factor regarding the proposition as a numerical value,
The extracted proposition and the extracted numerical value representing the certainty level are stored in a storage device as a database,
An expert system construction method for constructing an expert system by giving a structure to each proposition stored in the database to form a knowledge base.

Images