JP2004513458A - User-changeable translation weights - Google Patents

Abstract

In one aspect, the invention relates to a method that allows a user to view and change the weight associated with translating a source language string. The method includes displaying to a user a weight associated with the translation of the source language string, wherein the weight is used by a translation engine in selecting a translation, and the user is provided with a weight associated with the translation. Making it possible to make changes. In one embodiment, the method further comprises the step of allowing the user to reset the weight to a default value after the user changes the weight.

Description

本発明のシステムは、以下の文法規則のフォーマットを使用する。
【００８７】

この構文で、「＃」が前に付いている記号は、文構造の構文解析の目的では見えない仮想の記号であるが、いったん構文解析が得られるとサブ構造を構築するのに使用される。
【００８８】
このタイプの文法が与えられたとき、サブ構造の中で兄弟関係にあるノードの任意の列の中で、複数の構造変換を指定することができるようになる。これにより、文法規則に基づく構造変換機構は、テンプレート間構造変換機構のような強力さをある程度持つ機構に変わる。本発明のシステムは、上記の二番目のタイプの文法規則に基づいているけれども、自動的に一番目の形式に対応する文法規則に作成する。したがって、本システムは、第一の形式の文法規則を使用して文を構文解析し、第二の形式の文法規則を使用して構文解析構造を作ることができる。
【００８９】
構造変換には辞書制御構造変換処理１６４も含まれており、この処理は辞書１６１にアクセスして、構文解析ツリー上で文法規則制御構造変換処理が終わった後に同じツリー上で行われる。そして、生成規則制御構造変換処理によって、結果として得られた構文解析ツリーに生成規則を適用して、目的言語テキスト４１を供給する。
【００９０】
再び、図１および図２を参照すると、システムが上記プロセスにより、最適とランク付けした翻訳を取り出した後、その翻訳はディスプレイ２０を通してユーザに提示される。その後、ユーザは、その翻訳を採用するか、ユーザ入力装置２２を通して別構文解析システム３７を操作することにより編集するか、選択できる。編集作業中、ユーザは、翻訳された分の中で正しく翻訳された部分はそのままに保持しながら、その他の部分の再翻訳を要求することができる。この作業は、迅速に行える。なぜなら、システムはエキスパートのウエイト３１を持ったグラフを保持しているからである。
【００９１】
図１から図９を参照しながら、自動自然言語翻訳システムをある程度詳細に説明したので、これより、図１０から１６を参照して、本発明について説明する。本発明は、自動翻訳システムの様々なエキスパート（例えば、辞書見出し語蓋然性エキスパート１７２、統計エキスパート１７４、複合辞書エキスパート（図示せず））によって用いられるウエイトをユーザが視認することを可能とすることに関する。更に、本発明は、以降の翻訳に影響を与えるようにユーザがウエイトを変更可能とすることに関する。
【００９２】
図１０を参照すると、本発明の自動翻訳システム２００は、辞書ブラウザ２０５を含む。辞書ブラウザ２０５は、記憶装置１８と通信するハードウエアおよび／またはソフトウエア・コードであり、基本辞書４４、いずれかの専門用語辞書４６、いずれかのユーザが生成した辞書（図示せず）、および／または第三者が作成したものを含むいずれかの他の辞書、つまりユーザがアクセスするこれらの辞書のデータを獲得し操作する。辞書ブラウザ２０５は、ユーザからのコマンドおよび／またはデータを受信するためのユーザ入力装置２２と通信して、ユーザに辞書データを表示するためのディスプレイ２０と通信する。辞書ブラウザ２０５は、翻訳セッション中にユーザが辞書ブラウザ２０５を呼び出すまたは開始させることを可能とするための翻訳エンジン１６と通信状態にある。別の具体例では、辞書ブラウザ２０５は、翻訳エンジン１６に含まれるソフトウエア・コードの部分である。
【００９３】
図１１は、辞書ブラウザ２０５がディスプレイ２０に発生するグラフィカル・ユーザ・インタフェース（「ＧＵＩ」）の例示的な具体例を示す。このＧＵＩは、辞書見出し語フレーム２２５、選択基準フレーム２３０、およびウエイト表示フレーム２３５を含む。ＧＵＩは、辞書見出し語フレーム２２５に、ユーザが入力した選択基準フレーム２３０に表示されている基準に合致するソース言語の文字列（例えば語句、フレーズ）の全てを、各ソース言語文字列に関連するそれぞれの可能な翻訳と共に、ユーザに対して表示する。一具体例では、辞書見出し語フレーム２２５は、ソース言語文字列の可能な訳を、ソース言語文字列（例えばソース・ノード、親ノード）のインデントしたリスト（例えば、サブノード、子ノード）として表示する。この具体例では、ユーザは、ソース言語文字列の左側のマイナス記号ボックス２２８上でクリックすることで、翻訳のリストを隠すことができる。同様に、ユーザは、翻訳のリストが隠れている場合、ソース言語文字列の左側のプラス記号ボックス（図示せず）上でクリックすることで、それを表示することができる。
【００９４】
図１１において、辞書見出し語フレーム２２５は、ユーザがアクセスする全ての辞書２３０ｅの全ての見出し語を含み、これらの辞書は文字「ｂａｎｋ」２３０ｃではじまる（２３０ｂ）、全ての品詞２３０ｄのソース言語２３０ａの語句を有する。このＧＵＩによって、ユーザは、ドロップ・ダウン・メニューを用いて、各基準２３０ａ、２３０ｂ、２３０ｃ、２３０ｄ、２３０ｅを変更することができる。例えば、言語基準２３０ａを変更して、文字ボックス２３０ｃに入力された一つ（または複数）の文字で始まる（２３０ｂ）目的言語の翻訳を見つけることができる。また、言語基準２３０ａを変更して、文字ボックス２３０ｃに入力された一つ（または複数）の文字で始まる（２３０ｂ）ソース言語の語句または目的言語の翻訳のいずれかを見つけることができる。位置基準２３０ｂを変更して、文字ボックス２３０ｃに入力された一つ（または複数）の文字で終わる語句や、それに正確に一致するものや、大まかに一致するものを見つけたり、選択した辞書２３０ｅの全見出し語のリストを作ったりすることができる。文字ボックス２３０ｃは、ユーザが探している単語に含まれる一つ以上の文字をユーザが入力する場所である。ＰＯＳ基準２３０ｄによって、ユーザは、語句を特定の品詞に限定することができる。選択辞書基準２３０ｅによって、ユーザは、彼／彼女がアクセスする辞書の全てまたは特定のものを選ぶことができる。別の具体例では、辞書ブラウザ２０５は、翻訳エンジン１６から呼び出すことができる。この具体例では、翻訳エンジン１６は、辞書ブラウザ２０５に、自動翻訳の結果から選択された文字列を送り、それが文字ボックス２３０ｃに挿入される。
【００９５】
ウエイト表示フレーム２３５は、強調表示した翻訳２５０の異なる語形変化のウエイト２４５を表示する。図１１に示すように、ウエイト２４５ａ、２４５ｂ、２４５ｃ、２４５ｄ、２４５ｅの各々は等しい。これは、ソース文字ｂａｎｋに対する他動詞としての「ｄｅｐｏｓｉｔａｒ」の翻訳の優先度は、ソース文字の動詞が取り得る可能な語形変化（すなわち、不定詞、過去形、過去分詞、第三人称、現在分詞）の間で等しいということを表す。
【００９６】
また、ウエイト表示フレーム２３５は、サマリーウインドウ２４０も含む。概要ウインドウ２４０は、単一のバー２４１上に、強調表示された辞書見出し語２５０の語形変化の各々のウエイト２４５ａ、２４５ｂ、２４５ｃ、２４５ｄ、２４５ｅを表示する。この具体例では、全てのウエイトを、バー（例えば２４１）上の移動可能なビード（玉状のマーク）（例えば２４０ａ）として示す。バー２４１は、ウエイト２４５ａ、２４５ｂ、２４５ｃ、２４５ｄ、２４５ｅを配置することができる可能な値を含む範囲を表す。一具体例では、可能な値の範囲は標準化されている。標準化された範囲は、−１．０から＋１．０までである。辞書ブラウザ２０５がこの範囲を標準化する理由は、辞書ブラウザ２０５がアクセスする辞書（例えば、基本辞書４４、いずれかの専門用語辞書４６、いずれかのユーザが生成した辞書（図示せず）および／または第三者が作成したものを含むいずれかの他の辞書）は、様々な範囲（例えば、０から１００、−５０から０）のウエイトの見出し語を持ちえるからである。これらの様々な範囲を標準化することによって、ユーザは、一貫性のあるウエイト表示（例えば２４５ａ、２４５ｂ、２４５ｃ、２４５ｄ、２４５ｅ）を見て、さまざまな範囲を持つような異なる辞書の見出し語であっても、相対的なウエイト比較ができる。ウエイト２４５ａ、２４５ｂ、２４５ｃ、２４５ｄ、２４５ｅを変更するには、ユーザは、ビード（例えば２４０ａ）を所望の値に動かす。
【００９７】
図１１において、ウエイト２４５ａ、２４５ｂ、２４５ｃ、２４５ｄ、２４５ｅの全ては等しいので、サマリーウインドウ２４０においてウエイト２４５ａ、２４５ｂ、２４５ｃ、２４５ｄ、２４５ｅを表すビードは１つだけある。従って、ユーザが、サマリーウインドウ２４０のビード２４０ａをバー２４１に沿って動かすと、ウエイト２４５ａ、２４５ｂ、２４５ｃ、２４５ｄ、２４５ｅの全ての値が等しく変更される。一具体例では、ユーザがビード上でカーソルを移動させると、ＧＵＩは、小さい文字ボックス（例えばツールチップ）内に対応するウエイト２４５の数値表示をユーザに対して示す。サマリーウインドウ２４０は、ウエイト２４５の各々の複合値を表示するので、サマリーウインドウ２４０のビード２４０ａ上にカーソルを移動させた場合は、語形変化（すなわち、不定詞、過去形、過去分詞、第三人称、現在分詞）の各々と、それらに対応するウエイト２４５ａ、２４５ｂ、２４５ｃ、２４５ｄ、２４５ｅが、小さい文字ボックス（例えばツールチップ）内に表示される。
【００９８】
ウエイトを示すビード２４０ａを移動させてウエイトを等しく変更することに加えて、ユーザは、ウエイト２４５ａ、２４５ｂ、２４５ｃ、２４５ｄ、２４５ｅの各々の値を個々に変更することも可能である。図１２は、ユーザがウエイト２４５の各々を異なる値に個別にセットした例を示す。しかしながら、動詞ｂａｎｋの過去形および過去分詞形に関する２つのウエイト２４５ｂおよび２４５ｃは等しいということに留意されたい。これらの２つの語形変化のソース言語単語は同一であるので、２つのウエイト２４５ｂ、２４５ｃも同一でなければならない。サマリーウインドウ２４０は、ウエイト２４５ａ、２４５ｂ、２４５ｃ、２４５ｄ、２４５ｅの複合であり、従って、ウエイト２４５の４つの異なる値に対応する４つのビード２４０ａ、２４０ｂ、２４０ｃ、２４０ｄとして示す。ソーステキストが取る形に応じて、異なる翻訳が好ましいか、またはより一般的である場合は、異なるウエイトを用いることも可能である。ユーザは、翻訳している言語の種類に基づいて、これらの調整を行うことができる。
【００９９】
例えば、図１２は、ウエイト２４５を説明しており、ここでは現在分詞形「ｂａｎｋｉｎｇ」は、ウエイト２４５ｅがほぼ＋１．０であるので、「ｄｅｐｏｓｉｔａｒ」を用いて翻訳されることの優先度が極めて高いのに対し、不定詞形「ｂａｎｋ」は、ウエイト２４５ａがほぼ−１．０であるので、「ｄｅｐｏｓｉｔａｒ」を用いて翻訳されることの優先度が極めて低いことを示している。ウエイト２４５ａ、２４５ｂ、２４５ｃ、２４５ｄ、２４５ｅの各バー上の垂直線２５５は、そのウエイトの出荷時のデフォルト設定を表す。一具体例では、ユーザは、ウエイト２４５の全ての値を、いつでも出荷時のデフォルトに戻すことができる。
【０１００】
説明に役立つ実例として、図１３は、前置詞的なフレーズ「ｂｙ　ｔｈｅ　ｂａｎｋ」の自動翻訳の画面を示す。このフレーズを翻訳するために、自動翻訳システム２００は、図９に示し、関連する本文で説明したようなグラフを生成する。表１は、図９に示す前置詞的なフレーズの構成単語の各々について、可能な翻訳をリスト化したものである。図１に示された、各々の可能な翻訳に対する値は、その翻訳に関連したウエイトである。冠詞は、名詞の性別および数（すなわち単数または複数）に基づいてこれに一致するように選択されるので、冠詞はウエイトを有しない。
【０１０１】
【表１】

図１４は、名詞としての「ｂａｎｋ」に対する可能な翻訳が辞書見出し語フレーム２２５にリスト化されている画面を示す。これらの可能な翻訳は、それぞれ一つ以上の関連するウエイトを有する。例えば、強調表示された翻訳「ｂａｎｃｏ」２７０は、５つのウエイト２６０、２６１、２６２、２６３、２６４を有する。この翻訳が持ち得る５つの異なる意味があるので、５つのウエイトがある。これらの意味は、特性リスト２７５を用いて他のエキスパートに対して提示される。他のエキスパートは、これまで述べたように、これらの特性を、翻訳している他の単語および文の特性と比較して、可能な翻訳のスコアをつける。この例では、強調表示した翻訳「ｂａｎｃｏ」２７０に対する５つのウエイト２６０、２６１、２６２、２６３、２６４は全て等しいので、５つの特性リストの各々について、翻訳「ｂａｎｃｏ」を用いる優先度は同等であることを表す。このため、表１では、翻訳「ｂａｎｃｏ」のウエイトは、双方のコラムで同じである。ユーザが翻訳している言語の種類が、特性リストの１つに「ｂａｎｃｏ」を用いない場合、ユーザはそれに応じて一つ（または複数）のウエイトを調整することができる（例えば、使われない特定の特性リストに関連するウエイトを小さくし、および／または使われる可能性が高い他の特性リストのウエイトを全て大きくする）。
【０１０２】
図１５は、前置詞としての「ｂｙ」に対して可能な翻訳を辞書見出し語フレーム２２５にリスト化した画面を示す。これらの可能な翻訳は、それぞれ１つ以上の関連するウエイトを有する。例えば、強調表示された翻訳「ｐｏｒ」２８０は、６つのウエイト２８１、２８２、２８３、２８４、２８５、２８６を有する。この例では、強調表示された翻訳「ｐｏｒ」２８０に対する６つのウエイト２８１、２８２、２８３、２８４、２８５、２８６の全ては等しいので、６つの特性リストの各々について、翻訳「ｐｏｒ」が使われる優先度は同等であることを表す。このため、表１では、翻訳「ｐｏｒ」に対するウエイトは単一の値である。ユーザが翻訳している言語の種類が、特性リストの１つに「ｐｏｒ」を用いない場合、ユーザはそれに応じて一つ（または複数）のウエイトを調整することができる。
【０１０３】
再び表１を参照すると、名詞「ｂａｎｋ」の各翻訳の特性は、翻訳が属するコラム、すなわち川岸または金融機関のいずれか、を決定するのに役立つ。しかしながら、この例では、翻訳の文脈を与える（すなわち、他のエキスパートがある翻訳により高いスコアをつけるのに役立つ）他の単語または文が何もないので、自動翻訳システム２００は、翻訳が属するコラムには無関係に、最も高いウエイトを有する翻訳を選択する。他のエキスパートが、他の語句または文があるために、あるコラムを除外することが可能な場合、自動翻訳システム２００は、残りのコラムにおいて最も高い値を有する翻訳を選択する。言い換えると、他のエキスパートによるスコアが等しければ（例えば、意味的および文法的に同様な翻訳の選択肢として残っている翻訳のあるグループ）、このウエイトは、可能な翻訳の中から選択した一つの翻訳に決定する。表１に示すように、「ｂｙ」に対する翻訳「ｐｏｒ」は最も高い値を有し、「ｂａｎｋ」に対する翻訳「ｂａｎｃｏ」も最も高い値を有する。名詞「ｂａｎｃｏ」が男性名詞であるので、冠詞「ｅｌ」が選択される。このため、フレーズ「Ｂｙ　ｔｈｅ　ｂａｎｋ」を翻訳するために、図１３に示す翻訳「Ｐｏｒ　ｅｌ　ｂａｎｃｏ」が選択される。
【０１０４】
自動翻訳システム２００のユーザがこの翻訳に同意しない場合、ユーザはウエイトを調整して翻訳に影響を与えることができる。図１１に示すＧＵＩでは、ユーザは、ビード（例えば２４０ａ）をスライドさせてウエイトを調整する。ユーザが、語形変化の全てを調整して等しい値にしたい場合、ユーザはサマリーウインドウのビード２４０ａを用いて値を調整することができる。ユーザが、ウエイト２４５ａ、２４５ｂ、２４５ｃ、２４５ｄ、２４５ｅのうちの１つのみを調整したい場合、ユーザは特定の語形変化のウエイトに対応したビードのみを調整する。例えば、ユーザは、「ｂｙ　ｔｈｅ　ｂａｎｋ」に対する好ましい翻訳が「Ｊｕｎｔｏ　ａ　ｌａ　ｂａｎｄａ」であることを知っている。ユーザは、翻訳「ｊｕｎｔｏ　ａ」および「ｂａｎｄａ」のウエイトの各々を高くするように調整する（例えば、ビードを右側に動かす）。この結果の翻訳では、図１６に示すように、翻訳「ｊｕｎｔｏ　ａ」および「ｂａｎｄａ」が選ばれる。なぜなら、上述のように、他の全てが等しければ、最も高いウエイトを有する翻訳が選択され、ユーザが、好ましい翻訳のウエイトを他よりも高く設定したからである。名詞に対する好ましい翻訳が女性形であるので、ここでは冠詞「ｌａ」が使われている。ユーザが選択した値は、ユーザが好ましい選択を出現させたいと望む優先度の強さに依存する。ユーザが、他のエキスパートには無関係に好ましい翻訳を望むならば、ユーザは、好ましい翻訳のウエイトの各々を、範囲の最大値の近くに設定する（例えば標準化値＋１．０）。ユーザが他の翻訳よりもわずかに高い優先度しか望まず、他のエキスパートが翻訳に影響を与え続けることを望むならば、ユーザは、ウエイトの各々を、同じグループ（例えば同じコラム）内にある他の可能な翻訳よりもわずかに高く設定する。
【０１０５】
ユーザによるウエイト調整によって望む通りの好ましい翻訳が与えられない場合でも、ユーザの優先度に基づいた何らかの自動調整が行われ、最終的に、ユーザが望む所にウエイトを設定する。自動翻訳システム２００は、ユーザが別解釈を選択するたびにウエイトを調整する。自動翻訳システム２００は、ユーザが選択する別解釈のウエイトを高くする。別の具体例では、自動翻訳システム２００は、ユーザが置換したシステム２００の選択された翻訳のウエイトを低くする。一具体例では、増減値は、範囲の上限下限に対して、値がどこに位置するかに基づく。例えば、値が範囲の中心にある場合（例えば、範囲が−１．０から＋１．０のとき、０．０）、自動翻訳システム２００は、好ましい翻訳のウエイトの値を大幅に高くする（例えば範囲が−１．０から＋１．０の範囲で０．５）。値が範囲の終端に近く位置する場合（例えば範囲が−１．０から＋１．０のとき、０．８８）、自動翻訳システム２００は、好ましい翻訳のウエイトの値を少量だけ高くする（例えば範囲が−１．０から＋１．０のとき、０．０５）。この増減値は、自動翻訳システム２００による自動調整を最適化するように選択することができる。
【０１０６】
例えば、「ｂｙ　ｔｈｅ　ｂａｎｋ」の翻訳では、「ｊｕｎｔｏ」および「ｂａｎｄａ」のウエイトのユーザ調整は小さい増大であり、その結果の翻訳では、自動翻訳システム２００は、翻訳に「ｐｏｒ　ｅｌ　ｂａｎｃｏ」を選択する。ウエイトを再度調整する代わりに、ユーザは、別解釈のリストから、別解釈「ｊｕｎｔｏ　ａ　ｌａ　ｂａｎｄａ」を選択する。この別解釈の選択によって、自動翻訳システム２００は、「ｊｕｎｔｏ　ａ」および「ｂａｎｄａ」のウエイトを大きくし、および／または「ｐｏｒ」および「ｂａｎｃｏ」のウエイトを小さくする。増減の量は、いずれかの最適化アルゴリズムに従うようにすることができる。
【０１０７】
上述の機能および処理は、全て、汎用コンピュータと共に用いる様々なハードによって実現されているロジック設計および／またはプログラミング技法によって実現することができる。フローチャートに示したようなステップは、一般に、提示した順序で適用する必要はなく、ステップの組み合わせを組み合わせることも可能である。同様に、システムの機能性を、様々な方法でプログラムおよびデータ内に分散させることができる。更に、１つ以上の高級言語で文法および他の処理規則を開発すれば利点があるだろう。エンドユーザにはそれらをコンパイルしたフォーマットで供給できる。
【０１０８】
本明細書中に記載した機能性の全てを含めて、本明細書中に記載した自動自然言語翻訳システムの具体例のいずれも、汎用コンピュータ（例えばＡｐｐｌｅ　Ｍａｃｉｎｔｏｓｈ、ＩＢＭ　ＰＣまたは互換機、Ｓｕｎ　Ｗｏｒｋｓｔａｔｉｏｎ等）上で実行するディスケットまたは光コンパクト・ディスク（ＣＤ）等のコンピュータで読み取り可能な媒体上のコンピュータ・ソフトウエアとして提供することができる。ソフトウエアは、インターネットまたは他の何らかのネットワークを介して（例えばユーザに）送ることができ、必ずしもＣＤ等の物理的なデバイス上にある必要はない。また、ソフトウエアは、インターネットを介してユーザによってアクセスされ用いられるという点で、インターネットに基づくものとすることができる。
【０１０９】
当業者には、特許請求する本発明の精神および範囲から逸脱することなく、本明細書中に記載したものの変形、変更、および他の実施が想起されよう。従って、本発明は、前述の例示的な説明ではなく、特許請求の範囲の精神および範囲によって規定されるものとする。
【図面の簡単な説明】
【図１】
図１は、本発明の自然言語の自動翻訳を行うシステムのブロック図である。
【図２】
図２は、図１のシステムの全体的な機能を示すデータのフローチャートである。
【図３】
図３は、図１のシステムの動きを示すフローチャートである。
【図４】
図４は、図１のシステムのプリパーサー（予備的構文解析装置）の文末判定機能の動作を示すフローチャートである。
【図５】
図５は、図１のシステムのパーサー（構文解析装置）の動作を示すフローチャートである。
【図６】
図６は、図１のシステムの意味伝播の動作を示すフローチャートである。
【図７】
図７は、図１のシステムの構造変換の動作を示すフローチャートである。
【図８】
図８は、図１のシステムのエキスパート評価子のフローチャートである。
【図９】
図９は、「ｂｙ　ｔｈｅ　ｂａｎｋ」という句を例とした場合の、図１のシステムが使用するグラフのサンプルの図である。
【図１０】
図１０は、本発明による自然言語の自動翻訳を行うシステムの具体例の１つを示すブロック図である。
【図１１】
図１１は、ウエイトを示す、本発明の辞書ブラウザによって生成されたＧＵＩ（グライフィカル・ユーザ・インターフェース）の一例のスクリーンショットである。
【図１２】
図１２は、変更されたウエイトを示す、本発明の辞書ブラウザによって生成されたＧＵＩの一例のスクリーンショットである。
【図１３】
図１３は、本発明の自動翻訳システムによって生成された“Ｂｙ　ｔｈｅ　ｂａｎｋ”の翻訳結果の一例のスクリーンショットである。
【図１４】
図１４は、“ｂａｎｋ”の可能な翻訳のウエイトを示す、本発明の辞書ブラウザによって生成されたＧＵＩの一例のスクリーンショットである。
【図１５】
図１５は、“ｂｙ”の可能な翻訳のウエイトを示す本発明の辞書ブラウザによって生成されたＧＵＩの一例のスクリーンショットである。
【図１６】
図１６は、ユーザがウエイトを変更した後の、本発明の自動翻訳システムによって生成された“Ｂｙ　ｔｈｅ　ｂａｎｋ”の翻訳結果の一例のスクリーンショットである。[0001]
(Technical field)
The present invention relates generally to automatic natural language translation for translating one source text into another language. More specifically, the present invention relates to showing a weight value used for automatic translation to a user and permitting change.
[0002]
(Background information)
Various methods have been proposed for machine translation of natural languages. Typically, systems used for translation include a computer that receives input in one language, processes the received input, and outputs it in another language. This type of translation is not conventionally accurate, and requires a skilled operator to significantly modify the output. Translation work by conventional systems generally involves a structure conversion operation. The purpose of this structure transformation is to transform a parse tree with source language sentences (ie, a syntax structure tree) into a corresponding tree in the target language. Up to now, two types of structure conversion methods have been tried. That is, conversion based on grammar and conversion between templates.
[0003]
In the case of grammar-based transformations, the domain of the transformation is limited to the grammar rules used to obtain the source language parse tree (ie, the set of sub-nodes that are the direct children of a node) Is done. For example,
VP = VT01 + NP (The verb phrase [VerbPhrase]
One transitive verb [Transitive @ Verb]
And the noun phrase [NounPhrase] in this order
Lined up. )
And Japanese
1 + 2 => 2 + 1 (The order of VT01 and NP is reversed.)
When there is a rule, each parse tree of the source language to which this rule applies is transformed so that the order of verbs and objects is structurally reversed. In Japanese, the verb comes after the object. This method is very efficient in that it can quickly find where a particular transformation has taken place. This is because the particular transformation is performed exactly where the rules were used to obtain the source language parse tree. On the other hand, this method is weak in that its conversion domain is greatly limited as described above, and that natural language may require conversion rules that span nodes that are not children. It is also a conversion mechanism.
[0004]
In the conversion between templates, the structural conversion is specified in the form of an input / output (I / O) template or subtree. If an input template matches a structure tree, the portion of the structure tree that matches the template is converted as specified in the corresponding output template. While this is a very powerful conversion mechanism, it can come at the expense of finding out if an input template matches any part of a structure tree.
There is also known a method of performing a translation using a plurality of “experts” that perform evaluation of possible translation candidates and a process of selecting one of the candidates. Which translation is best is determined by the scores given by these experts for each of the possible translations. After the translation process is completed, the user views the translation selected by the automatic translation system. If the translation is incorrect, the user can correct the translation. After making some corrections, the automated system may also learn the user's preferences. For example, this is described in Takeda et al., US Pat. No. 5,826,220.
[0005]
(Summary of the Invention)
The automatic natural language translation system according to the present invention has many advantages when compared to conventional machine translation devices. After the system of the present invention automatically selects the most appropriate translation of the input text information and provides the user with an output (eg, a Japanese translation of the English input text), the user interacts with the system. Then, the displayed translation result can be edited or another translation can be automatically obtained. The user of the automatic natural language translation system of the present invention can improve productivity. This is because this system can automatically re-translate only the remaining portion while leaving the translation result portion determined to be good. This selective retranslation process is performed only on the parts that require retranslation, so that the user can save time and use the potentially abundant optimal rank that is high but incorrect translation. Freed from the tedious task of considering. Further, the system allows various fine-tuning of translation adjustments, and usually more of the final structure of the translation is generated by the system. Therefore, by using this system, potential mistakes by humans (users) are reduced, and the time required for reworking the sentence structure, personal match, tense change, and the like can be saved. The system effectively provides users with the benefits of extensive and accurate grammar and spelling knowledge.
[0006]
According to the automatic natural language translation system of the present invention, various processes for the ambiguity of the delimitation of sentences included in the source language and the powerful semantic propagation function make the translated sentence more accurate, and Rework can be reduced. The accumulated learning statistics further improve the translation accuracy by tailoring the preferred translation to a particular user site. The idiom processing method of the present system has an advantage that when there is a sentence including a word string constituting the idiom, it can be accurately translated without considering the meaning of the idiom. This system is not only efficient, but also has various functions such as long-range characteristic matching function. The structural balance expert and the peer structural expert of the system efficiently distinguish between intended and unintended parsing. Uppercase experts efficiently interpret correct uppercase words in sentences, and uppercase string processing efficiently processes proper nouns of compound words without completely ignoring interpretations as common nouns.
[0007]
The present invention is directed generally to an automatic natural language translation system that allows a user to view weight values used by one or more “experts” to select a particular translation from two or more alternatives. And The invention also relates generally to allowing a user to change the weight so that the automated system has at least a tendency to select any translation the user prefers. Before the system actually performs any translations, the user can view and / or adjust the weights so that the automated system will typically post-translate the user before "learning" user preferences. There is no need to perform error correction more than once.
[0008]
In another aspect, the invention relates to a method that allows a user to view and change the weight associated with translating a source language string. The method includes displaying to a user a weight associated with the translation of the source language string, the weight being used by the translation engine in selecting a translation, and the user changing the weight associated with the translation. And allowing the user to do so. In one embodiment, the method further includes the step of allowing the user to reset the weight to a default value after the weight is changed by the user.
[0009]
In another embodiment, the step of displaying further comprises displaying a plurality of weights to the user, each of the weights being associated with an inflection of the translation, and each of the weights being translated when selecting a translation. Used by the engine. In another embodiment, the enabling step further includes the step of allowing the user to individually change each of the one or more weights associated with the inflection of the translation. In another embodiment, the method further comprises displaying, in the summary window, all of the one or more weights associated with the inflection over a single display indicating a range of possible values.
[0010]
In another embodiment, the source language strings and translations are included in a translation dictionary. In another embodiment, the method further includes displaying the translation to the user in response to input from the user. In another embodiment, the input is at least one character in a source language. In another embodiment, the input is at least one character in the target language of the translation. In another embodiment, the method further includes displaying the translation to a user in response to input from the translation engine.
[0011]
In another embodiment, the method comprises selecting, by the translation engine, a translation from a plurality of possible translations of the source language string, wherein each of the plurality of possible translations is associated with a weight, The selecting further includes a step based at least in part on the weight. In another embodiment, the method includes receiving a source language string and determining a plurality of possible translations by at least one expert system, wherein each of the plurality of translations is a source language string. And the step of selecting the translation having the highest weight from the plurality of possible translations.
[0012]
In another embodiment, the method further comprises providing a graphical user interface for displaying weights to a user. In another embodiment, the method includes providing a graphical user interface for displaying to a user at least one source language dictionary entry that meets established criteria; and at least one source language dictionary entry. Allowing the user to hide and display one or more translations of the word. In another embodiment, the method further comprises adjusting the weight in response to and adapted to operation by the user of the graphical user interface.
[0013]
In another embodiment, the graphical user interface displays to the user the current value of the weight for a range of possible values. In another embodiment, the range of possible values is standardized. In another embodiment, the method comprises the steps of: allowing a user to select a translation from a plurality of possible translations; and providing, by the translation engine, at least one of a plurality of translation weights in response to the user selection. Adjusting one of the two. In another embodiment, the adjusting step further comprises increasing a weight associated with the selected translation. In another embodiment, the adjusting step further comprises reducing a weight associated with the translation replaced by the user selection.
[0014]
In another embodiment, the present invention relates to a machine translation system that allows a user to view and change the weight associated with a source language translation, the system comprising: A dictionary browser configured to provide data to be used for display to a user, wherein the weight is used by a translation engine in selecting a translation and allows a user to change the weight associated with the translation. A dictionary browser configured to do so. The system also includes a translation dictionary that communicates with the dictionary browser to receive and transmit data, the translation dictionary including at least one translation and an associated weight, and associated with the data received from the dictionary browser. Includes a translation dictionary that changes weights. In addition, the system communicates with a dictionary browser, receives data from a user, and transmits data to the dictionary browser. A user input device communicates with the dictionary browser, receives data from the dictionary browser, and receives an image of the data. And a display device for displaying to the user.
[0015]
In one embodiment, the dictionary browser is further configured to allow the user to reset the weight to a default value after the weight is changed by the user. In another embodiment, the dictionary browser is further configured to provide data used to display one or more weights to a user, each of the weights being associated with an inflection of the source language string, Are used by the translation engine in selecting a translation. In another embodiment, the dictionary browser is further configured to allow a user to individually change each of the one or more weights associated with the inflection of the translation.
[0016]
In another embodiment, the system displays, in a summary window, data for displaying all of one or more weights associated with the inflection over a single display showing a range of possible values. It is configured to provide to the device. In another embodiment, the dictionary browser is further configured to provide data used to display the translation in response to input received from the user. In another embodiment, the input is at least one character in a source language. In another embodiment, the input is at least one character in the target language of the translation.
[0017]
In another embodiment, the dictionary browser is further configured to provide a graphical user interface for displaying a weight to the translation to a user. In another embodiment, the dictionary browser is further configured to provide a graphical user interface for displaying to the user at least one source language dictionary entry that meets established criteria, and wherein at least one Allows a user to hide and display one or more translations of a source language dictionary headword. In another embodiment, the dictionary browser is further configured to adjust the weight in response to and adapted to operation by a user of the graphical user interface.
[0018]
In another embodiment, the graphical user interface is configured to display to the user the current value of the weight for a range of possible values. In another embodiment, the range of possible values is standardized. In another embodiment, the system further includes a translation engine that communicates with the dictionary browser and selects a translation based at least in part on the weight of the translation.
[0019]
In another embodiment, the translation engine is further configured to receive a source language string, and configured to determine a plurality of possible translations by at least one expert system, wherein each translation identifies a source language. It is configured to translate equally and to select the translation with the highest weight from a plurality of possible translations. In another embodiment, the translation engine is further configured to allow a user to select a translation from a plurality of possible translations, and the dictionary browser is further configured to respond to the user selection with a plurality of possible translations. It is configured to adjust at least one of the translation weights. In another embodiment, the dictionary browser is further configured to increase a weight associated with the selected translation. In another embodiment, the dictionary browser is further configured to reduce a weight associated with the translation replaced by the user selection.
[0020]
In another aspect, the invention relates to an article of manufacture having computer readable program means for enabling a user to view and change the weight of a translation. The article of manufacture is computer readable program means for displaying to a user a weight associated with the translation of the source language string, wherein the weight is used by a translation engine in selecting a translation. Computer readable program means for allowing a user to change the weight associated with the translation.
[0021]
The foregoing and other objects, features, features and advantages of the invention will become more apparent from the following detailed description of the invention and the appended claims.
[0022]
The same reference numbers in different drawings generally identify the same parts, even in different drawings. The drawings are not necessarily to scale, emphasis instead generally being placed upon illustrating the principles of the invention.
[0023]
(Detailed description of the invention)
First, an outline of the automatic natural language translation system of the present invention will be described without referring to the drawings. After an overview, a description will be given with reference to various drawings.
[0024]
An automatic natural language translation system can translate a source natural language into a destination natural language. In one embodiment, the system translates English into Japanese. In another embodiment, the system translates from Japanese to English. In yet another embodiment, the system performs a translation from English to another language or from another language to English. Other languages include Spanish, Chinese, German, French, Portuguese, Korean and several other languages, and there are dictionaries in these languages that can be accessed. The system includes means for receiving and storing the source natural language, a translation engine for creating a translation of the target language, means for displaying the translation result to the user, providing the user with another translation result, and Means for displaying. In one specific example of this system, the translation engine includes a preparser (preliminary parser), a parser (parser), a graph generator, an evaluator, a graph scorer, a parser extractor, and a structure converter. including. The preparser examines the input text and resolves where the boundaries of the input sentence are ambiguous. Thereafter, the preparser displays the input text in the form of a parsing chart composed of dictionary entry words. The parser parses the chart to obtain possible syntactic classifications for the input text. The graphing device generates a graph of the possible parsing of the input text based on the parsing chart. The graph includes nodes and subnodes associated with possible interpretations of the input text. An evaluation function consisting of a series of experts evaluates the graph of possible interpretations and adds expert weights to the nodes and subnodes of the graph. The graph scorer uses expert weights to derive the scores of the subnodes, and then associates the top N scores with each node. The parser extractor assigns a parse tree structure to the preferred interpretation determined by the graph scorer. The structure conversion device performs a structure conversion relating to the syntax analysis tree structure in order to obtain a translation in the target language.
[0025]
The following three paragraphs describe the following: (A) how the graph scorer combines the expert weights to calculate the final weight score for each subnode; (b) how the graph scorer combines the subnode scores to produce the final node score. (C) how linguistic information is propagated through the tree of nodes and subnodes.
[0026]
To calculate the final weight score for each subnode, the graph scorer associates a constant value with each subnode. An analysis of the linguistic information associated with each subnode determines the score of the subnode. See FIG. 8 for an example. Here, a series of expert evaluators examine the linguistic information stored at each node and subnode. The graph scorer adds the individual weight scores for each expert to get the final weight average for a particular node or subnode. Combining multiple weight scores into one weight average score is a standard task in computer science. One method that can be used is to multiply the result of each expert by a constant (weight) assigned to that expert. The weight assigned to each expert is a matter determined by the designer. The designer determines the priority (weight) assigned to each expert. The weighted average is the sum of each number multiplied by a constant and a series of those numbers. For example, the following equation is obtained.
Average weight = (w₁) (X₁) + (W₂) (X₂) +. . . + (W_n) (X_n)
However, weight w₁, W₂,. . . , W_nAre non-negative numbers and sum to 1. See, for example, Spiegel, "Theory and Problems 76 of Probability and Statistics" (1975, McGraw-Hill, Inc.), which describes the use of weighted averages for expected statistical values.
[0027]
To sum the subnode scores to get a final node score, the graph scorer can propagate the subnode scores from the bottom portion of the graph to the top portion. For a graph in which each node has a set of N scores, one or more propagation methods can be determined. One technique that can be used to propagate subnode scores is a storage method that is some kind of dynamic programming to solve optimization problems. Solutions to the optimization problem can involve many possible numbers (results). The goal is to find the optimal value. The algorithm used for the optimization solves each subproblem only once and stores the results, so that the answer does not have to be recalculated each time the same problem is encountered. For a more detailed description that applies to optimization problems, see, for example, Cormen et al., "Invitation to Algorithms", McGraw-Hill, 1990, pp. 301-314. Pages 301, 302, and 312 of this "Invitation to Algorithm" describe one method that can be used to propagate subnode score information in a graph.
[0028]
When propagating linguistic information in the tree, the part that propagates semantics of the system operates to convey semantic information from smaller internal components to larger components. Semantic propagation applies to the four classes of syntactic classification used in the parsing process (SEMNP, SEMVP, SEMADJ and VERB). Before semantic propagation takes place, the linguistic information stored in the node must be analyzed. Analysis of semantic information stored in the node checks which noun-type element and verb-type element of the grammatical rule, and applies which selective restriction slot of the verb-type element of the grammatical rule to which noun-type object Guided by a set of rules that guess. Gerald Gazda, in his book "Natural Language Processing of Prologues" (Addison Wesley Publishing Co., 1989), stores in nodes of non-ring-shaped graphs that have similar orientations to the graphs disclosed herein. Describes a set of rules that can be used to analyze the semantic information being performed. Gazda describes the use of property matching to match information about neighboring nodes. Gazda states that property matching involves the following equation:
[0029]
"A property that appears on one node must be the same as a property that appears on another node. The latest research shows that the details of a class of properties are related to the parent classification and their properties. When appearing on a child showing a morpheme that does, the details of its properties hypothesize the principle of being equal, this child being called the "head" of the phrase. Most phrases have only one head. Thus, for example, a verb phrase inherits the tense of that verb. Because the verb is the "head" of the verb phrase. There is no easy way to specify this principle in detail, using the notational resources we have used so far, throughout the grammar. However, if one assumes that the relevant properties can all be found on one branch of the DAG, rule by rule can show the effect of the principle very clearly. Let's call the label at the head of this branch. Then we can write the normal VP rule as:
[0030]
VP ---> V NP PP
<Head of V> = <Head of VP>
In this case, the value of the “head” property of V and the value of the “head” property on the parent VP must be the same. "
The rules discussed in Gazda can be readily applied to each class of syntax disclosed herein. The linguistic information assigned to each node using Gazda's rule can be propagated through the tree by storage techniques.
[0031]
Here, to summarize the contents of the above three paragraphs, the weighted average is one method of determining the subnode score, and each subnode score uses a well-known storage technique applied to the optimization problem. The method described in Gazda's book can be used to analyze the linguistic information stored at each node, and this linguistic information can Can be used to propagate through the parsing chart.
[0032]
In an automatic natural language translation system, an automatic retranslation function can be executed after the first automatic translation is completed. That is, after the system automatically selects the most appropriate possible translation of the input text information and provides it to the user as an output result (preferably, a Japanese translation of the input English text, or from Japanese to English Translation), the user can interact with the system to modify the displayed translation or automatically obtain another translation.
[0033]
The automatic natural language translation system uses a language model that decomposes one sentence into lower character strings. A substring is one or more words that appear in the order specified as part of the sentence. For example, the lower-order character string "The @ man @ is @ happy" includes "The", "The @ man", "man @ is @ happy.", "Is" and "The @ man @ is @ happy", but "is @ man" , "Man @ man" and "The @ is" are not included.
[0034]
Different language models define substrings in different ways and at different levels of detail. For example, in the sentence "They @ wow1d @ 1ike @ an @ arrow.", "An @ arrow" is usually classified as a noun phrase (NP). There are also classification models for "an @ arrow" with syntactic properties (e.g., singular noun phrases) and literary properties (weapons). If the meaning of this phrase is ambiguous, there are more than one way to classify it, for example, "an @ arrow" can also mean an arrow-shaped symbol. When language models provide a way to resolve ambiguity, they typically resolve ambiguity by combining smaller units into larger units. When evaluating larger units, these models consider only the information contained in the larger units.
[0035]
As a specific example of this system, the semantic property (symbol or weapon) of "an @ arrow" is used in evaluating the verb phrase "1ike @ an @ arrow" in the sentence "They @ wow1d @ 1ike @ an @ arrow." Is done. On the other hand, if the syntax of the phrase "an @ arrow" is in the sentence "He @ shot @ it @ with @ an @ arrow.", The semantic characteristic of "an @ arrow" is the verb phrase "shot @ it @ with @ an @ arrow." Not used when evaluating.
[0036]
Exported properties exist for any substring of a sentence (interpreted substring), which is interpreted in one way against a particular language model. Exported properties are all properties used to evaluate the combination of one interpreted substring with other units that form a larger substring. An export is an interpreted substring that is interpreted together with the exported properties. Properties that are included in the interpreted substring, but not exported, are called substructures.
[0037]
The system parser contains a grammar database. The parser uses grammar rules to find all possible sentence interpretations. The grammar database has X = A1A2. . . It consists of a series of context-independent phrase structure rules expressed in the form An. X is A1, A2,. . . , An and is referred to as one of the highest-order nodes that is subordinate to the lower nodes (subnodes) from A1 to An.
[0038]
The grapher of the system graphically represents the many possible interpretations of a sentence. Each node in the graph corresponds to the export of a certain substring. As a specific example of the system, one export is represented by one node. The graph includes arcs (arc-like lines) emanating from a plurality of nodes tied to one export, the arcs representing a substructure of the export based on the application of grammar rules. The graph represents at least two types of arcs used in the standard version of Chomsky. That is, (1) a single arc pointing to one different export of the same substring, (2) two arcs containing a set of pointers to two exports (the substrings of those exports are , Concatenated to form the lower character string of the original export). Chomsky's non-normal form grammar graph includes one arc with N sets of pointers pointing to N exports.
[0039]
The graph includes one starting point export S from which all parts of the graph can be reached by following a series of arcs. A starting point export corresponds to an entire sentence.
[0040]
Multiple arcs start from one node only if one or more exports form the same export. (A set of pointers in two arcs is not considered multiple arcs in this sense.) Only if the export is an element of multiple exports does the multiple arcs form one node Will point to. Nodes without arcs correspond to dictionary headwords assigned to lower character strings.
[0041]
Multiple language experts assign a numerical score to the set of exports. The language expert applies this score to each node of the graph. In one embodiment of the system, the score matrix (each element of the matrix is a weight that multiplies by a particular expert score) is a fixed-length "N" floating-point number for any sentence. ".
[0042]
The score is evaluated by a scoring module built into the graphing engine and / or parser. A score is calculated for all exports that are a component of a single top export. The score of the top export is calculated as the sum of the exports that make it the top-level export and the scores of all the experts to be combined (for example, the score given by the structural balance expert). Is done.
[0043]
The order in which nodes are scored by tracing nodes is a standard, bottom-up tracing graph search algorithm. In this algorithm, the scored nodes are marked and will not be scored again. During the scoring process, the scoring module evaluates dictionary headword nodes before evaluating any nodes in higher units. One score is obtained from each dictionary entry word.
[0044]
If there are multiple ways to do one export, you will get multiple scores. That is, if there are k ways to export, there are k possible scores. The multiple scores are processed as follows.
[0045]
(1) In a single-element rule, each of the k lower-order export scores is added to the expert value applied to the single-element rule and the resulting k number of The score vector is related to the parent export.
[0046]
(2) In the rule composed of two elements, for example, it is assumed that the left child has g scores and the right child has h scores. At that time, by adding each score of the right child to each score of the left child, and further adding the expert value applied to the rule consisting of two elements, a total of g × h scores are calculated. Will be. If the value of g × h exceeds N, only the top N scores are kept with the parent node.
[0047]
(3) If an export for a node can be made in multiple ways, up to N scores are added to the score list for that node, and only the top scores are kept.
[0048]
At the completion of the score calculation, the above method guarantees that: Each export associates its node with a set of g (g ranges from 1 to N) scores; the g scores are the g most likely (language) to create an expert. Represents the method (for the model): it contains the attributes of all substructures not shown in the export. In special cases, such as the root node S, this scoring method yields the g most likely ways to form the whole sentence.
[0049]
Each score in the above score list has an associated pointer. The pointers provide information indicating which scores in the lower export score list have been summed and combined to produce a higher level score. By tracking each pointer, the g most likely interpretations of the sentence can be extracted as an unambiguous parse tree.
[0050]
The automatic natural language translation system will be described in more detail with reference to FIGS. Thereafter, various improvements of the present invention will be described with reference to FIGS.
[0051]
Referring to FIGS. 1 and 2, the automatic natural language translation system 10 of the present invention includes an input interface 12, a translation engine 16, a storage device 18, a user input device 22, a display 20, and an output interface 14. The input interface is adapted to receive a series of texts written in a source language such as English or Japanese. The input interface may include a keyboard, a voice interface or a digital electronic interface such as a modem or serial input. The translation engine translates the source language using the data in the storage device. The translation engine can be made entirely of hardware implemented logic circuits, or can include one or more processing units and associated storage instructions. The translation engine can include the elements and portions thereof described below: a preparser 24, a parser 26, a graph generator 28, a parser / translator evaluator 30, a parser extractor 32, a structure converter 34, and A user interface 42 including another parsing system 37. The structure conversion device can include a structure conversion device 36 by grammar rule control, a structure conversion device 38 by dictionary control, and a structure conversion device 40 by generation rule control. Storage device 18 may include one or more disks (eg, hard disks, floppy disks and / or optical disks) and / or memory storage devices (eg, RAM). These storage devices can store all or some of the elements described below. That is, a basic dictionary 44, a technical term dictionary 46, a dictionary created by the user, a grammar rule 48, a generation rule 50, a semantic characteristic tree 52, a structural characteristic tree 54, and a graph 56. The storage device 18 is also used to store all kinds of information that is used or useful for performing translations, such as input text information written in source natural language, output text information written in destination language, and The information includes one or more dictionaries, domain keywords, and grammar rules. User input interface 22 includes a keyboard, mouse, touch screen, light pen or other user input device and is available to a user of the system. The display can be a computer display, a printer or other type of display, or it can be another device to inform a user. The output interface 14 exchanges the final translation of the source text in the target language, for example, in Japanese, Spanish or English. The interface may include an electronic interface such as a printer, display, audio interface, modem or serial line, or may include other devices for sending text to an end user.
[0052]
As an operation of the specific example of the translation system of the present invention, as shown in FIGS. 1, 2 and 3, the preparser 24 first performs a preliminary parsing operation (step 102) on the source text 23. This operation includes resolving the ambiguity of the end-of-sentence boundary of the source text, and creates the structural analysis chart 25 including the dictionary headword. Thereafter, the parser 26 parses the chart created by the preparser (step 104), and obtains a parse chart 27 describing the possibility of the syntax. The graph creating device 28 creates a possible interpretation graph 29 based on the structural analysis chart obtained in the structural analysis step (step 106). The evaluator 30, which accesses a series of experts 43, evaluates the stored interpretation graph (step 108) and adds expert weights 31 to the graph. The graph scorer 33 scores the nodes and associates the top N (eg, 20) scores 35 with each node. The parser extractor 32 assigns a parsing tree structure 39 to this preferred interpretation (step 110). Thereafter, the structure conversion device 34 accessing the conversion table 58 performs a structure conversion process (step 112) on the tree in order to obtain the translation 41 that has become the target language. The user can interact with another parsing system 37 to obtain other translations.
[0053]
Referring to FIG. 4, the system of the present invention performs a preliminary parsing process by dividing the input string into tokens that include individual punctuation marks and groups of characters that form words (step 114). The presence of spaces affects the interpretation of characters at this level. For example, “−” in “x − y” is a dash, while “−” in “xy” is a hyphen.
[0054]
The preparser then combines the tokens into a phrase (step 116). At this level, the preparser recognizes special structures (eg, Internet addresses, telephone numbers and social security numbers) as one unit. The preparser also performs dictionary lookups to separate groups. For example, when "re-enact" is listed as "reenact" in the dictionary, it is one word, but when it is not listed, it is three separate words (eg, "re", "-", “Encat”).
[0055]
In the next preliminary parsing stage, a determination is made as to where the sentence ends (step 118). During this process, the preparser accesses the base dictionary, the terminology dictionary, and the on-board user-created dictionary to determine the possible end of each sentence (ie, after each word in the source text). Perform a series of steps. The preparser need not perform the steps in the particular order as presented, and the steps can be implemented as a sequence of ordered rules, or can be coded and implemented in hardware.
[0056]
Referring to FIG. 5, when there is a word string of a symbol that cannot be parsed such as a series of dashes “−−−−−”, the preparser interprets and records itself as one “sentence”. Not targeted (step 120). The preparser requests two consecutive carriage returns at the end of the sentence (step 122). If the first letter of the next word is lowercase, the preparser does not consider it the end of a sentence (step 124). If a sentence starts on a new line and is short (eg, a title), the preparser considers it a single sentence.
[0057]
The preparser considers a terminator (.), Question mark (?), Or exclamation point (!) As the end of a sentence, except in certain cases, such as including closing parentheses and closing quotes (step 128). For sentences ending with "." ","? "", Etc., the preparser uses virtual punctuation after the quotes, in addition to the punctuation before the quotes. Another example of punctuation that may be needed for "?" Is shown below.
The @ question @ was @ "What \ do \ you \ want?"
Did he ask ask the question "What do you want?"
Are you Concerted about “theotherpeop1e”?
In the case of English, each sentence above is likely to end with "?". The presence of virtual punctuation added by the preparser indicates that there is something like a question mark before the quotation mark, or nothing at all. Behind the quotes are something like a terminator or question mark. The grammatical structure of the rest of the sentence allows the most appropriate choice to be made in subsequent processing steps.
[0058]
The preparser also uses some additional methods in the preliminary parsing of the terminator (steps 130, 132, 134, 136 and 138). Some of the abbreviations included in the dictionary are marked at the beginning of the sentence as being never used, and some are marked at the end of the sentence as being never used (step 130). These rules are always respected. For example, "Ltd" is not used at the beginning of a sentence, and "Mr" is not used at the end of a sentence. The preparser also does not consider the end of a sentence to be the end of a sentence, unless the next word is a grammatically frequent word such as "the" or "in" (step 132). If the word before the stop is in one of the dictionaries, the sentence ends at the stop (step 134). If the word before the stop is not in the dictionary, has a stop in the word (eg, IBM), and the next word is not in the dictionary as a lowercase letter, Or, if the next word itself is uppercase, the sentence is considered not to end at that terminator (step 136). Otherwise, the terminator indicates the end of the sentence (step 138).
[0059]
Referring again to FIGS. 2 and 3, once the sentence boundaries are specified by the preparser, the parser places the phrases of the sentence into a syntactical classification and calculates their possible syntactic interpretations 25 of the sentence. The grammar rule of the grammar database is applied to the phrase (step 104). The grammar rules 48 can be implemented as a series of computer-readable rules that represent the grammatical restrictions of the language. In English, there are hundreds of such rules, and these rules can apply to hundreds of syntactic classes. To reduce the computational burden of this process, the different possible meanings of a word are ignored.
[0060]
In the next step (step 106), the graphing device uses the dictionary to extend the parser results, captures different meanings of the word, and represents all semantic interpretations of the sentence, directional and unclosed Create a shape graph. This graph is created with the help of a series of semantic propagation processes described later. These processes are performed based on a series of created grammar rules, and in some cases, access a semantic property tree to obtain semantic information. The semantic property tree is a tree structure including the semantic classification. The tree is roughly organized from abstract to concrete, and the decision on how the two words are related in meaning is determined by how far apart the tree is in the tree. Each level within can be done from two perspectives, somewhere. For example, “cat” and “dog” are more relevant than “cat” and “pudding”. Thus, the former two are only slightly separated in the tree. "Animal" and "cat" are examples of words located at different levels of the tree. This is because "animal" is a more abstract word than "cat".
[0061]
Referring to FIG. 9, the graph of this figure includes a node 80 and its subnodes 82, 84, 86, which are represented by pointers (arrows) 88, 89, 90, 91 indicating various types of relationships. It links to 80. The first type of relationship shown in this graph is that a node representing a phrase has pointers to nodes of words and subphrases constituting the phrase. For example, the node 84 representing “the @ bank” is linked by pointers 92 and 93 to the words “the” 94 and “bank” 95 that constitute it. The second type of relationship shown in this graph is when the phrase interpretation has pointers to more than one way to make the same higher-level construct from lower-level parts. For example, a node 80 representing the phrase "by @ the @ bank" has two source interpretation locations 81,83, from which pointers 88 and 89, 90, and 91 indicate the respective components. In this example, the different constituent words have different sub-nodes 84 and 86, each of which has a different meaning for the phrase "the @ bank". The structure of the graph is defined by the results of the parsing operation and is limited by the syntax of the source sentence. The nodes of this graph are associated with storage locations of semantic information accumulated during the process of semantic propagation.
[0062]
The part that propagates the meaning of the system serves to propagate semantic information from smaller components to larger components made up of them. The semantic information is classified into four classes of syntactic classification used in the initial parsing processing. The four classes are SEMNP (including noun-type objects and prepositional phrases), SEMVP (verb-type objects having a normal subject), SEMADJ (adjective), and VERB (often a vocabulary that takes an object). Verb-type verb). Other syntax classifications are ignored by certain rules. The grammar rule setter can also ignore the absolute processing by marking the rule with a particular mark. These special orders come first.
[0063]
The way in which semantic properties propagate in a system has two aspects. The first aspect is to examine the noun-type and verb-type components of the grammar rules to find out which optional restriction slots of the verb-type component apply to the noun-type object. It is a set of rules that can be done. For example, the rule for the verb phrase of the sentence "I @ perpended @ him @ to @ go." Is roughly VP = VT11 + NP + VP (where VP is a verb phrase, VT11 is a type of a transitive verb, and NP is a noun phrase). . One example of a default rule is that if the verb takes an object, the first NP to the right of the verb must apply a selection restriction. Another rule states that VP restrictions on the subject must apply to the first NP to the left of the VP. Together, these rules ensure that the semantic preference of both "persume @ him" and "him @ go" is evaluated. As mentioned earlier, these rules reflect the complex grammar of English, so there may be very few of these.
[0064]
Referring to FIG. 6, the semantic propagation process involves copying the selection restrictions from the SEMVP to the statement (step 140). If SEMNP is used as a representation of the position, its validity is evaluated for a semantic constant defining a good position (step 142). If a rule involves the union of two SEMNPs (detected by making a logical AND of the syntactic properties), the graph creator calculates the logical product of the semantic properties and calculates the semantic distance Apply to expert (step 144).
[0065]
In examining the rules that govern the propagation of semantic properties, the grapher may use a "head" of SEMNP that is propagated to a higher level (eg, it may be a part of one SEMNP that contains more words). Is found, the graph creation device also propagates the semantic property (step 146). However, if the “head” is a word indicating a part (eg, “portion”, “part”), the “head” propagates left or right from the SEMNP. SEMVP and SEMADJ are propagated in the same way, with the only exception being that they contain semantic situations that indicate parts. Adjectives are part of the SEMVP class in this sense.
[0066]
If the SEMVP is created from a rule that includes VERB, the graphing device propagates the subject restriction of VERB upward unless the VP is passive. In the case of the passive voice, the first object restriction of VERB is propagated (step 150). In the case of a rule that includes a SEMVP, the grapher attempts to apply the SEMVP selection restrictions to NPs encountered when moving left from the SEMVP (step 152). In all cases involving rules involving SEMADJ, the grapher will attempt to apply the SEMADJ selection restrictions to all SEMNPs encountered when moving right from SEMADJ, and if unsuccessful, the left Attempt to move in the direction (step 154).
[0067]
For all remaining VERB object selection restrictions (which are not propagated upwards because they are passive), the grapher instead applies them sequentially to the SEMNPs encountered on the right side of the VERB. (Step 156). With all these rules, the verb selection restriction runs out as soon as it applies to something. In all rules leading up to this rule, SEMNP is not exhausted, even if something is applied. Starting from this rule, SEMNP is exhausted. Finally, if a rule produces a SEMVP, the graphing device determines whether a previously unused SEMVP or SEMADJ is included, and if so, moves it up. (Step 158).
[0068]
The system also performs linguistic trait feature matching. Linguistic characteristics are the characteristics of words and other components. Syntax property matching is used by the parser, and semantic property matching is used by the graph creator. However, the same approach is used for both. For example, "the" has a syntax property of plural, while "he" has a syntax property of singular. In the characteristic matching, grammar rules are marked so that the grammatical rules are applied only when certain conditions of the characteristics of the word are satisfied. For example, assume that there are the following rules.
[0069]
S = NP ｛＠｝ + VP ｛＠｝
Here, the symbol ＠ means that the characteristics of the numbers of NP and VP must match. Therefore, this rule assumes that "key @ are" or "he @ is" is correct, but does not recognize "key @ is" or "he @ are".
[0070]
The characteristic match restrictions are divided into those for “between short distances” and those for “between long distances”. The processing over long distances is calculated when the grammar is compiled, not when actually processing the sentence. The long-distance processing flow to be executed is coded as an instruction byte.
[0071]
The computation of long-range property processing must start with an n-element rule (ie, a rule that may have more than one input to its right). The system then assigns codes to the various binary element rules so that the set of properties propagates between the rules in the correct manner. By decomposing the n element rules into binary element rules, the parsing task is greatly simplified, but since the system retains the path information of the characteristic set spanning the binary element rules, The system retains the power of long-range processing.
[0072]
In the system of the present invention, the dictionary may also contain "phrases" of words, while also having a representation of the individual words that make up those phrasals. These two forms will eventually compete with each other and the best one will remain. For example, “b1ack @ sheep” is registered in the dictionary in the sense of a surplus. However, in some cases, the phrase "b1ack @ sheep" may mean black sheep. Because both forms are retained, translations that are not considered idioms are also selected as correct translations.
[0073]
Idioms also belong to another category. For example, the system can use three types of classification:
[0074]
Almighty: United States of of America
Priority: long @ ago
Normal: b1ack @ sheep
Almighty idioms take precedence over possible interpretations of any of the words making up the word sequence. Preferred phrases have precedence over other configurations belonging to the same general type and using the exact same words. Ordinary idioms have the same status and compete with other headwords.
[0075]
The resulting graph is evaluated by an expert (step 108, FIG. 3), which provides a score that indicates the likelihood of a correct interpretation of the graph. The system of the present invention includes a scoring method that applies to all parts of a sentence of any length, not just the entire sentence. An important factor in using a graph is that the subtree is fully scored and analyzed only once, even if it appears in so many sentences. For example, in the phrase "Near \ the \ bank \ there \ is \ a \ bank.", The phrase "Near \ the \ bank" has at least two meanings, but it is determined once that the most appropriate interpretation of this phrase is once. Only done. The phrase "the @ is @ a @ bank" can similarly be interpreted in two ways, but only one decision is made as to which of these two interpretations is most appropriate. Thus, the sentence can be interpreted in four different meanings, but the subphrase is scored only once. Another feature of this graph is that each node has easily accessible information about the length of that part of the sentence. With this feature, the upper N most suitable interpretations of any lower character strings of the English sentence can be performed without re-analyzing the English sentence.
[0076]
In a single run, a total of only the top N optimal analyzes of the sentence are obtained each time (N is some number of the order of 20), but by using the graph, the system allows the system to: The result of the user's selection for the smaller component can be taken in, and the top N optimal analyzes respecting the user's selection can be performed. All of these analyzes can be performed quickly without re-analyzing the sentence or scoring any lower-order character strings.
[0077]
Referring to FIG. 8, the processing of the expert evaluator 30 is performed based on various factors that characterize each translation, and these factors are handled by various experts. The rule probability expert 170 evaluates the average relative frequency of the grammar rules used to obtain the initial source language parse tree. The selection restriction expert 178 evaluates the degree of semantic matching of the obtained translation. The dictionary headword probabilities expert 172 evaluates the average relative frequency of a particular "speech" of a word in the sentence used to obtain the initial source language parse tree. The statistical expert 174 evaluates the average relative frequency of a particular paraphrase selected for a translation. In one embodiment, the arcs of dictionary headword probabilities expert 172 and statistics expert 174 are combined to create a composite dictionary expert (not shown).
[0078]
Some of the experts (e.g., dictionary entry probabilities expert 172, statistics expert 174, composite dictionary expert (not shown)) evaluate using the weights of the possible translations of the dictionary entry. This weight indicates the preference of a dictionary entry to appear in the translation. The weight corresponds to a plurality of inflections that a dictionary entry word can have. For example, dictionary entries may have singular and plural forms. In some languages, dictionary entries have masculine, feminine, or neutral forms. A dictionary entry may also have different inflections based on its part of speech (eg, in the case of a verb, infinitive, past, past participle, third person, and / or present participle).
[0079]
The system automatically determines the English "part of speech" (POS) for individual English words, phrases, and word groups. When the system translates a sentence, it automatically determines the part of speech and usually makes the right decision. However, sometimes the sentence to be translated and the whiteness are ambiguous. When words or phrases that can be interpreted as different parts of speech are included, a plurality of different but "correct" interpretations will be obtained in the sentence in which those words or phrases appear. The user of the system may ignore the part of speech that the system automatically determines and instead manually set any part of speech for a word, phrase or group of words. For example, in the English text of the source "John sawaw a boyy with a a telescope.", The system user can specify "a @ boy @ with @ a @ telescope" as a noun phrase, and consequently refer to the sentence as "the boy has a telescope. "And John did use a telescope to see the boy." In cases where forcing the system to override the POS rules would result in worse results than better translations, the user should either reduce the manual POS designation as much as possible, or place a more restrictive manual POS designation. You can deal with it by making it loose. Noun phrases are less restrictive than nouns, and groups are the least restrictive POS designations. The table below shows the various possible POS designations.
[0080]
Part of speech (POS)
noun
Noun phrase
Verb (transitive or intransitive)
Verb phrase
adjective
Adjective phrase
adverb
Adverb phrase
preposition
Prepositional phrase
conjunction
group
English
Part-of-speech settings for "adjective phrases" and "adverb phrases" are effective when certain English sentences have different meanings depending on how the system interprets a particular prepositional phrase. For example, the sentence "We \ need \ a \ book \ on \ the \ fourth \ of \ Jully" may be interpreted as "on \ the \ fourth \ of \ July" if it is interpreted as having an adjective meaning. It means "I want a book on the day." However, if "on \ the \ fourth of of July" is interpreted as an adverb phrase, it means "We want a book on July 4." If the user thinks that the system has automatically assigned an incorrect POS to "on \ the \ four \ of \ July", the user will be able to use the "on \ the \ fourth \ of \ Jully" sentence "We \ need \ a \ book \ on". Can be manually set to another POS. If the user does not want the system to translate a particular word, phrase or group of words from English to Japanese, such a word, phrase or group of words can be set to a POS of "English". The user can remove one or more POS settings, whether the settings were made automatically by the system or manually by the user.
[0081]
The system maintains a history of usage statistics for translations at one or more levels for each customer site. For example, the system retains statistics at the level of the surface form of the phrase (the frequency at which "leaving" was used as a transitive or intransitive verb) and also at the semantic level (meaning "leave after", Or does it mean "departed" from?) In the latter type, the numbers of appearances of “leave”, “leaves”, “left”, and “leaving” are accumulated. The system may also keep a distinction between the usage statistics used in the last few sentences and the usage statistics occurring at any time at the customer site. Further, the system can distinguish between cases where the user has intervened and instructed to use a particular meaning of the phrase and cases where the system has used the particular meaning of the phrase without the user's confirmation. .
[0082]
The structural balance expert 182 is a feature related to the length of a sentence component and is based on features common to English and many other European languages. In some (but not all) constructs, sentences with heavy (long) elements to the left of light elements are not preferred in these languages. For example,
Mary \ hit \ Bill \ with \ a \ broom.
(The left is light and the right is heavy)
Mary \ hit \ with \ a \ bloom \ Bill.
(Heavy on the left and light on the right)
Mary \ with \ a \ broad \ a \ dog \ that \ tried to \ bite \ her.
(Heavy on the left, heavier on the right)
When there are two parsers of a sentence, one contains a string "heavy on the left and light on the right" and has a structure that tries to avoid such a string, and the other does not. , The former is deemed not to represent the intended interpretation of the sentence. This expert is an effective way to distinguish between intended and unintended parsing.
[0083]
In the equivalence structure of the pattern "A \ of \ B \ and \ C", it is determined whether the intended interpretation is "A \ of \ B \ and \ C" or "A \ of \ B \ and \ C". It can be difficult. The peer structure expert 180 measures the semantic distance between BCs and the semantic distance between ACs to determine which peer mode will combine the two elements that are semantically closer. This expert accesses the semantic property tree during processing. This expert is also an effective way to distinguish between intended and unintended parsing of a sentence.
[0084]
Many English words have potential ambiguity in interpreting them as common nouns or proper nouns. The capitalization expert 176 uses the position of the capitalization in the sentence to determine the likelihood that capitalization is significant, for example, the following sentence:
Brown \ is \ my \ first \ choice.
My first first choice is is Brown.
The first sentence is inherently ambiguous in meaning, but in the second sentence, "Brown" is much more likely to be a person's name than a color name. The expert will determine whether words starting with uppercase letters are at the beginning of a sentence or not at the beginning of words (as in the example above), whether words appearing in uppercase are included in the dictionary, and words starting with lowercase letters are registered in the dictionary. Factors are taken into account. This expert is an effective way to correctly interpret capitalized words in sentences.
[0085]
When a sentence includes a word sequence of words that start with an uppercase letter, the entire word sequence can be treated as a proper noun or as a series of common noun sequences. Since the system of the present invention uses uppercase word string processing, it prefers the former interpretation. Further, when the word string itself cannot be parsed by ordinary grammatical rules, the word string can be processed as a single noun phrase that has not been analyzed and not translated. This processing has been proved to be a very effective means for processing proper nouns of compound words, and interpretation of ordinary nouns with low appearance levels can be processed without completely ignoring them.
[0086]
Referring to FIG. 7, the machine translation system of the present invention uses a structure conversion mechanism 162 controlled by a grammar rule, which has the efficiency of a structure conversion method based on a simple grammar rule. In addition, it has the power close to the structure conversion method between templates. The method relies on using grammar rules 160 that can specify non-horizontal composite structures. Below is the format of the rules used by other translation systems,

The system of the present invention uses the following grammar rule format.
[0087]

In this syntax, the symbols preceded by a "#" are virtual symbols that are not visible for the purpose of parsing sentence structures, but are used to build substructures once the parse has been obtained. .
[0088]
Given this type of grammar, it is possible to specify multiple structural transformations in any column of sibling nodes in the substructure. As a result, the structure conversion mechanism based on the grammar rules is changed to a mechanism having a certain degree of strength, such as a structure conversion mechanism between templates. Although the system of the present invention is based on the second type of grammar rule described above, it automatically creates a grammar rule corresponding to the first type. Thus, the system can parse the sentence using the first type of grammar rules and create a parsing structure using the second type of grammar rules.
[0089]
The structure conversion also includes a dictionary control structure conversion process 164, which is performed on the same tree after accessing the dictionary 161 and completing the grammar rule control structure conversion process on the parse tree. Then, the target language text 41 is supplied by applying the generation rule to the resulting parse tree by the generation rule control structure conversion process.
[0090]
Referring again to FIGS. 1 and 2, after the system retrieves the translation ranked as optimal by the above process, the translation is presented to the user through display 20. The user can then choose to adopt the translation or edit it by operating another parsing system 37 through the user input device 22. During the editing operation, the user can request the retranslation of the other portions while keeping the correctly translated portion of the translated portion as it is. This can be done quickly. This is because the system holds a graph with expert weights 31.
[0091]
Having described the automatic natural language translation system in some detail with reference to FIGS. 1 to 9, the present invention will now be described with reference to FIGS. The present invention relates to allowing a user to view the weights used by various experts of an automated translation system (eg, dictionary headword probability expert 172, statistical expert 174, composite dictionary expert (not shown)). . Further, the invention relates to allowing a user to change the weight to affect subsequent translations.
[0092]
Referring to FIG. 10, the automatic translation system 200 of the present invention includes a dictionary browser 205. The dictionary browser 205 is hardware and / or software code that communicates with the storage device 18 and includes a basic dictionary 44, any terminology dictionary 46, any user-generated dictionary (not shown), and And / or obtain and manipulate any other dictionaries, including those created by third parties, i.e., those dictionaries that are accessed by the user. Dictionary browser 205 communicates with user input device 22 for receiving commands and / or data from the user, and with display 20 for displaying dictionary data to the user. Dictionary browser 205 is in communication with translation engine 16 to allow a user to invoke or initiate dictionary browser 205 during a translation session. In another embodiment, dictionary browser 205 is a piece of software code included in translation engine 16.
[0093]
FIG. 11 shows an exemplary embodiment of a graphical user interface (“GUI”) generated by the dictionary browser 205 on the display 20. This GUI includes a dictionary entry word frame 225, a selection criterion frame 230, and a weight display frame 235. The GUI includes, in the dictionary entry word frame 225, all of the source language strings (eg, phrases, phrases) that match the criteria displayed in the selection criteria frame 230 input by the user, in association with each source language string. Display to the user with each possible translation. In one embodiment, the dictionary headword frame 225 displays the possible translations of the source language string as an indented list (eg, subnodes, child nodes) of the source language string (eg, source node, parent node). . In this example, the user can hide the list of translations by clicking on the minus sign box 228 to the left of the source language string. Similarly, if the list of translations is hidden, the user can display it by clicking on the plus sign box (not shown) to the left of the source language string.
[0094]
In FIG. 11, a dictionary headword frame 225 includes all headwords of all dictionaries 230e accessed by the user, these dictionaries begin with the characters "bank" 230c (230b), and the source language 230a of all parts of speech 230d. With the phrase This GUI allows the user to change each criterion 230a, 230b, 230c, 230d, 230e using a drop down menu. For example, the language criterion 230a can be changed to find a translation in the target language that starts with one (or more) characters (230b) entered in the character box 230c. Also, the language criterion 230a can be changed to find either a source language phrase or a target language translation (230b) that starts with one (or more) characters entered in the character box 230c. Change the position reference 230b to find words that end with one (or more) characters entered in the character box 230c, those that exactly match it, those that roughly match it, and those in the selected dictionary 230e. You can even make a list of all headwords. The character box 230c is a place where the user inputs one or more characters included in the word that the user is searching for. The POS criterion 230d allows a user to limit a phrase to a particular part of speech. The selected dictionary criteria 230e allows the user to select all or specific dictionaries that he / she accesses. In another embodiment, dictionary browser 205 can be invoked from translation engine 16. In this specific example, the translation engine 16 sends the character string selected from the result of the automatic translation to the dictionary browser 205, and the character string is inserted into the character box 230c.
[0095]
The weight display frame 235 displays different inflection weights 245 of the highlighted translation 250. As shown in FIG. 11, each of the weights 245a, 245b, 245c, 245d, and 245e is equal. This is because the priority of the translation of "depositar" as a transitive verb to the source character bank is based on the possible inflections (ie, infinitive, past tense, past participle, third person, present participle) that the source character verb can take. Represents equality between
[0096]
The weight display frame 235 also includes a summary window 240. The summary window 240 displays, on a single bar 241, the weights 245a, 245b, 245c, 245d, 245e of each of the inflections of the highlighted dictionary entry 250. In this example, all weights are shown as movable beads (eg, ball-shaped marks) (eg, 240a) on a bar (eg, 241). Bar 241 represents a range that includes possible values where weights 245a, 245b, 245c, 245d, 245e can be placed. In one embodiment, the range of possible values is standardized. The standardized range is from -1.0 to +1.0. The reason that the dictionary browser 205 standardizes this range is that the dictionary browser 205 accesses the dictionary (eg, the basic dictionary 44, any terminology dictionary 46, any user-generated dictionary (not shown), and / or This is because any other dictionaries, including those created by third parties, can have a range of weight entries (eg, 0-100, -50-0). By standardizing these various ranges, the user would see consistent weight indications (eg, 245a, 245b, 245c, 245d, 245e) and see different dictionary headwords with different ranges. Even so, relative weight comparisons can be made. To change the weights 245a, 245b, 245c, 245d, 245e, the user moves a bead (eg, 240a) to a desired value.
[0097]
In FIG. 11, since the weights 245a, 245b, 245c, 245d, and 245e are all equal, there is only one bead representing the weights 245a, 245b, 245c, 245d, and 245e in the summary window 240. Thus, when the user moves the bead 240a of the summary window 240 along the bar 241, all the values of the weights 245a, 245b, 245c, 245d, 245e are changed equally. In one embodiment, when the user moves the cursor over the bead, the GUI shows the user a numerical display of the corresponding weight 245 in a small character box (eg, a tooltip). Since the summary window 240 displays the composite value of each of the weights 245, when the cursor is moved over the bead 240a of the summary window 240, the inflection (ie, infinitive, past tense, past participle, third person, Each of the present participles) and their corresponding weights 245a, 245b, 245c, 245d, 245e are displayed in a small character box (eg, a tooltip).
[0098]
In addition to moving the weight bead 240a to change the weights equally, the user can also individually change the value of each of the weights 245a, 245b, 245c, 245d, 245e. FIG. 12 shows an example in which the user individually sets each of the weights 245 to a different value. Note, however, that the two weights 245b and 245c for the past and past participle of the verb bank are equal. Since the source language words of these two inflections are identical, the two weights 245b, 245c must also be identical. The summary window 240 is a composite of the weights 245a, 245b, 245c, 245d, 245e and is therefore shown as four beads 240a, 240b, 240c, 240d corresponding to the four different values of the weight 245. Depending on the form that the source text takes, different weights may be used if different translations are preferred or more general. The user can make these adjustments based on the type of language being translated.
[0099]
For example, FIG. 12 illustrates a weight 245, where the present participle "banking" has a very high priority of being translated using "depositar" because the weight 245e is approximately +1.0. On the other hand, the infinitive form "bank" indicates that the priority of translation using "depositar" is extremely low because the weight 245a is approximately -1.0. A vertical line 255 on each bar of weights 245a, 245b, 245c, 245d, 245e represents the factory default settings for that weight. In one embodiment, the user can return all values of the weight 245 to factory defaults at any time.
[0100]
As an illustrative example, FIG. 13 shows a screen for automatic translation of the prepositional phrase "by the bank". To translate this phrase, automatic translation system 200 generates a graph as shown in FIG. 9 and described in the associated text. Table 1 lists possible translations for each of the constituent words of the prepositional phrase shown in FIG. The value for each possible translation shown in FIG. 1 is the weight associated with that translation. The article has no weight because the article is selected to match based on the gender and number (ie, singular or plural) of the noun.
[0101]
[Table 1]

FIG. 14 shows a screen where possible translations for “bank” as a noun are listed in the dictionary headword frame 225. Each of these possible translations has one or more associated weights. For example, the highlighted translation "banco" 270 has five weights 260, 261, 262, 263, 264. There are five weights because there are five different meanings that this translation can have. These meanings are presented to other experts using the property list 275. Other experts compare these characteristics to the characteristics of other words and sentences that are being translated, and score possible translations, as described above. In this example, the five weights 260, 261, 262, 263, 264 for the highlighted translation "banco" 270 are all equal, so the priority of using the translation "banco" is equal for each of the five property lists. It represents that. Thus, in Table 1, the weight of the translation "banco" is the same in both columns. If the type of language that the user is translating does not use "banco" in one of the property lists, the user can adjust one (or more) weights accordingly (e.g., Reduce the weight associated with a particular property list and / or increase the weight of all other property lists that are likely to be used).
[0102]
FIG. 15 shows a screen listing possible translations for the preposition “by” in the dictionary entry word frame 225. Each of these possible translations has one or more associated weights. For example, the highlighted translation “por” 280 has six weights 281, 282, 283, 284, 285, 286. In this example, the priority for which the translation "por" is used for each of the six property lists, because all six weights 281, 282, 283, 284, 285, 286 for the highlighted translation "por" 280 are equal. The degrees represent equality. Thus, in Table 1, the weight for the translation "por" is a single value. If the type of language the user is translating does not use "por" in one of the property lists, the user can adjust one (or more) weights accordingly.
[0103]
Referring again to Table 1, the characteristics of each translation of the noun "bank" help determine the column to which the translation belongs, ie, either a bank or a financial institution. However, in this example, since there are no other words or sentences that give the context of the translation (ie, help other experts score higher translations), the automated translation system 200 may use the column to which the translation belongs. Regardless of the translation, the translation with the highest weight is selected. If another expert can exclude one column because of another phrase or sentence, the automatic translation system 200 selects the translation with the highest value in the remaining columns. In other words, if the scores by other experts are equal (for example, a group of translations that remain as semantically and grammatically similar translation options), this weight is assigned to one of the possible translations To decide. As shown in Table 1, the translation "por" for "by" has the highest value, and the translation "banko" for "bank" also has the highest value. Since the noun “banco” is a masculine noun, the article “el” is selected. Therefore, in order to translate the phrase "By @ the @ bank", the translation "por @ el @ banko" shown in FIG. 13 is selected.
[0104]
If the user of the automatic translation system 200 does not agree with the translation, the user can adjust the weight to affect the translation. In the GUI shown in FIG. 11, the user adjusts the weight by sliding a bead (for example, 240a). If the user wants to adjust all of the inflections to equal values, the user can adjust the values using the summary window bead 240a. If the user wants to adjust only one of the weights 245a, 245b, 245c, 245d, 245e, the user adjusts only the bead corresponding to the particular inflection weight. For example, the user knows that the preferred translation for “by the bank” is “Junto a la banka”. The user adjusts the weight of each of the translations "junto @ a" and "banda" to be higher (eg, moves the bead to the right). In the translation of this result, as shown in FIG. 16, the translations “junto @ a” and “banda” are selected. This is because, as described above, if all others are equal, the translation with the highest weight is selected and the user has set the weight of the preferred translation higher than the others. The article "la" is used here because the preferred translation for the noun is the feminine form. The value selected by the user depends on the strength of the priority that the user desires to make the preferred selection appear. If the user desires a preferred translation independent of other experts, the user sets each of the preferred translation weights near the maximum of the range (eg, standardized value +1.0). If the user only wants a slightly higher priority than the other translations and the other experts want to keep affecting the translation, the user will place each of the weights in the same group (eg in the same column) Set slightly higher than other possible translations.
[0105]
Even if the user does not give the preferred translation as desired due to the weight adjustment, some automatic adjustment based on the user's priority is performed, and finally the weight is set where the user desires. The automatic translation system 200 adjusts the weight each time the user selects another interpretation. The automatic translation system 200 increases the weight of another interpretation selected by the user. In another embodiment, the automatic translation system 200 lowers the weight of the selected translation of the system 200 replaced by the user. In one embodiment, the increase or decrease is based on where the value is located relative to the upper and lower limits of the range. For example, if the value is at the center of the range (e.g., 0.0 when the range is from -1.0 to +1.0), the automatic translation system 200 may significantly increase the value of the preferred translation weight (e.g., 0.5 in the range of -1.0 to +1.0). If the value is near the end of the range (e.g., 0.88 when the range is -1.0 to +1.0), the automatic translation system 200 increases the value of the preferred translation weight by a small amount (e.g., the range). Is -1.0 to +1.0, 0.05). This increase or decrease can be selected to optimize automatic adjustment by the automatic translation system 200.
[0106]
For example, in the translation of "by @ the @ bank", the user adjustment of the weight of "junto" and "banda" is a small increase, and in the resulting translation, the automatic translation system 200 selects "por @ el @ banko" for the translation. I do. Instead of adjusting the weight again, the user selects another interpretation "junto @ a @ la @ banda" from the list of alternatives. Depending on the choice of this alternative interpretation, the automatic translation system 200 may increase the weight of “junto @ a” and “banda” and / or reduce the weight of “por” and “bando”. The amount of increase or decrease can be according to any optimization algorithm.
[0107]
All of the functions and processes described above can be implemented by logic design and / or programming techniques implemented by various hardware used with general-purpose computers. Steps such as those shown in the flowcharts generally do not need to be applied in the order presented, and combinations of steps can be combined. Similarly, the functionality of the system can be distributed among programs and data in various ways. In addition, it would be advantageous to develop grammar and other processing rules in one or more high-level languages. End users can be supplied in a compiled format.
[0108]
Any of the specific examples of the automatic natural language translation systems described herein, including all of the functionality described herein, may be implemented on general purpose computers (eg, Apple Macintosh, IBM PC or compatibles, Sun Worksstation, etc.). ) May be provided as computer software on a computer readable medium such as a diskette or an optical compact disc (CD). The software can be sent (eg, to the user) over the Internet or some other network, and need not necessarily be on a physical device such as a CD. Also, the software may be Internet-based in that it is accessed and used by users via the Internet.
[0109]
Variations, modifications, and other implementations of what has been described herein will occur to those skilled in the art without departing from the spirit and scope of the claimed invention. Accordingly, the invention is to be defined not by the preceding illustrative description but instead by the spirit and scope of the following claims.
[Brief description of the drawings]
FIG.
FIG. 1 is a block diagram of a system for automatically translating a natural language according to the present invention.
FIG. 2
FIG. 2 is a data flow chart showing the overall function of the system of FIG.
FIG. 3
FIG. 3 is a flowchart showing the operation of the system of FIG.
FIG. 4
FIG. 4 is a flowchart showing the operation of the sentence end determination function of the preparser (preliminary syntax analyzer) of the system of FIG.
FIG. 5
FIG. 5 is a flowchart showing the operation of the parser (syntax analysis device) of the system of FIG.
FIG. 6
FIG. 6 is a flowchart showing the semantic propagation operation of the system of FIG.
FIG. 7
FIG. 7 is a flowchart showing the structure conversion operation of the system of FIG.
FIG. 8
FIG. 8 is a flowchart of the expert evaluator of the system of FIG.
FIG. 9
FIG. 9 is a diagram of a sample graph used by the system of FIG. 1 for the example of the phrase "by @ the @ bank".
FIG. 10
FIG. 10 is a block diagram showing one specific example of a system for automatically translating a natural language according to the present invention.
FIG. 11
FIG. 11 is a screen shot of an example of a GUI (graphical user interface) generated by the dictionary browser of the present invention, showing weights.
FIG.
FIG. 12 is a screen shot of an example of a GUI generated by the dictionary browser of the present invention showing the changed weights.
FIG. 13
FIG. 13 is a screen shot of an example of the translation result of “By @ the \ bank” generated by the automatic translation system of the present invention.
FIG. 14
FIG. 14 is a screen shot of an example of a GUI generated by the dictionary browser of the present invention, showing possible translation weights for "bank".
FIG.
FIG. 15 is a screen shot of an example of a GUI generated by the dictionary browser of the present invention showing possible translation weights of "by".
FIG.
FIG. 16 is a screen shot of an example of the translation result of “By @ the \ bank” generated by the automatic translation system of the present invention after the user changes the weight.

Claims (39)

A method that allows a user to view and change the weight associated with translating a source language string,
Displaying the weight associated with the translation of the source language string to the user, wherein the weight is used by a translation engine in selecting the translation;
Allowing the user to change the weight associated with the translation;
A method comprising: 2. The method of claim 1, further comprising the step of allowing a user to reset the weight to a default value after the weight is changed by the user. The displaying further includes displaying a plurality of weights to the user, wherein each of the weights is associated with an inflection of the translation, and wherein each of the weights is associated with the translation engine when selecting the translation. 2. The method of claim 1, wherein the method is used by: 4. The method of claim 3, wherein the enabling further comprises enabling the user to individually change each of the one or more weights associated with the inflection of the translation. 4. The method of claim 3, further comprising displaying all of the one or more weights associated with the inflection over a single display showing a range of possible values in a summary window. The method of claim 1, wherein the source language string and the translation are included in a translation dictionary. The method of claim 1, further comprising displaying the translation to the user in response to input from the user. The method of claim 7, wherein the input is at least one character of the source language. The method of claim 7, wherein the input is at least one character in the target language of the translation. The method of claim 1, further comprising displaying the translation to the user in response to input from the translation engine. Selecting, by the translation engine, a translation from a plurality of possible translations of the source language string, wherein each of the plurality of possible translations is associated with a weight; 2. The method of claim 1, further comprising the step of: Receiving the source language string,
Determining a plurality of possible translations by at least one expert system, each translation in the plurality translating the source language string equally;
Selecting the translation with the highest weight from the plurality of possible translations;
2. The method of claim 1, further comprising: The method of claim 1, further comprising providing a graphical user interface for displaying the weight to the user. Providing a graphical user interface for displaying to the user at least one source language dictionary headword that meets established criteria;
Allowing a user to hide and display one or more translations of the at least one source language dictionary entry word;
14. The method of claim 13, further comprising: 14. The method of claim 13, further comprising adjusting the weight in response to and adapted to operation of the graphical user interface by the user. 14. The method of claim 13, wherein the graphical user interface displays the current value of the weight to the user for a range of possible values. 17. The method of claim 16, wherein the range of possible values is standardized. Allowing the user to select the translation from a plurality of possible translations;
Adjusting at least one of the weights of the plurality of translations in response to the user selection by the translation engine;
2. The method of claim 1, further comprising: 19. The method of claim 18, wherein the adjusting further comprises increasing a weight associated with the selected translation. 19. The method of claim 18, wherein said adjusting further comprises reducing a weight associated with said translation replaced by said user selection. A machine translation system that allows a user to visually recognize and change weights related to translation of a source language,
A dictionary browser configured to provide data used to display to the user the weight associated with the translation of the source language string, wherein the weight is a translation engine when selecting the translation. A dictionary browser used by the user and configured to allow the user to change the weight associated with the translation;
A translation dictionary for receiving and transmitting data in communication with the dictionary browser, the translation dictionary comprising at least one translation and associated weights, and changing the associated weights in response to data received from the dictionary browser. Translation dictionary
A user input device that is in communication with the dictionary browser, receives data from the user, and transmits the data to the dictionary browser;
A display device that is in communication with the dictionary browser, receives data from the dictionary browser, and displays a display of the data to the user;
A system comprising: 22. The system of claim 21, wherein the dictionary browser is further configured to allow a user to reset the weight to a default value after a user changes the weight. The dictionary browser is further configured to provide data used to display one or more weights to the user, each of the weights being associated with an inflection of the source language string, and 22. The system of claim 21, wherein each is used by the translation engine in selecting the translation. 24. The system of claim 23, wherein the dictionary browser is further configured to allow the user to individually change each of the one or more weights associated with the inflection of the translation. A summary window configured to provide the display device with data for displaying all of the one or more weights associated with the inflection over a single display showing a range of possible values. 24. The system of claim 23, wherein 22. The system of claim 21, wherein the dictionary browser is further configured to provide data used to display the translation in response to input received from the user. 27. The system of claim 26, wherein said input is at least one character of said source language. 27. The system of claim 26, wherein the input is at least one character in the target language of the translation. 22. The system of claim 21, wherein the dictionary browser is further configured to provide a graphical user interface for displaying the weight for the translation to the user. The dictionary browser is further configured to provide a graphical user interface for displaying at least one source language dictionary entry matching the established criteria to the user, the at least one source language dictionary. 30. The system of claim 29, wherein the system allows the user to hide and display one or more translations of headwords. 30. The system of claim 29, wherein the dictionary browser is further configured to adjust the weight in response to and adapted to operation of the graphical user interface by the user. 30. The system of claim 29, wherein the graphical user interface is configured to display a current value of the weight to the user for a range of possible values. 33. The system of claim 32, wherein the range of possible values is standardized. 22. The system of claim 21, further comprising a translation engine in communication with the dictionary browser to select the translation based at least in part on the weight of the translation. The translation engine is further configured to receive a source language string, configured to determine a plurality of possible translations by at least one expert system, each translation translating the source language equally; 35. The system of claim 34, wherein the system is configured to select a translation having the highest weight from the plurality of possible translations. The translation engine is further configured to allow the user to select the translation from a plurality of possible translations, and the dictionary browser is further configured to respond to the user selection with the plurality of possible translations. 35. The system of claim 34, wherein the system is configured to adjust at least one of the weights of the translation. 37. The system of claim 36, wherein the dictionary browser is further configured to increase a weight associated with the selected translation. 37. The system of claim 36, wherein the dictionary browser is further configured to reduce a weight associated with the translation replaced by the user selection. An article of manufacture comprising computer readable program means for allowing a user to view and change the weight of translation,
Computer readable program means for displaying to the user the weight associated with the translation of the source language string, wherein the weight is used by a translation engine in selecting the translation;
Computer readable program means for allowing the user to change the weight associated with the translation;
A product comprising:

Images