JP3817943B2 - Document summarization device and recording medium - Google Patents

Description

ここで、Score(relation-type)は主要格リストの順序を重みで置き換えたものである。即ち、Score(が格) ＞Score(連体修飾) の関係がある。
【０１４０】
そして、選択関係表生成手段２５は、トータルスコアの最も大きな２項関係を抽出して選択関係表に付加する。
なお、WcとWhの値を変えるのは、重要単語が係り側にある２項関係を優先するか、受け側にある２項関係を優先するかを決めるためである。重要単語として選ばれたもののスコアは、もう一方の単語のスコアより必ず高いので、重要単語が係り側にある２項関係を優先させたければWcの値をWhより高くすればよい。
【０１４１】
重要単語として選ばれたもののスコアは、全ての２項関係で共通なので、重要単語ではない方の単語のスコアを、それが係り側か受け側かで重みを変えればよい。また、Wc, WhとWlは、関係の重要さと、関係を持つ単語の重要さの何れをより重視するかを決めるためのものである。関係の重要さを全く考慮に入れないならば、Wl = 0とすればよい。
【０１４２】
以上のような処理を実行することにより、ある重要単語に対して同一の主要格を有する複数の２項関係が存在している場合には、設定に応じて重要度が高い２項関係が選択され、選択関係表に付加されることになる。
【０１４３】
第１および第２の実施の形態においては、ある重要単語に対して同じ主要格を持つ２項関係が複数ある場合に、関係表に出現する順序に応じてこれらの２項関係が選択されることになる。その場合、重要でない２項関係も選択され、要約文に付加されることになる。
【０１４４】
しかしながら、以上のような処理によれば、出現順序ではなく、重要度に応じて２項関係が選択されるので、重要な２項関係のみを抽出することが可能となる。
【０１４５】
以上のようにして生成された選択関係表は、補完関係付加手段１５に供給される。
補完関係付加手段１５は、選択関係表に記載されている２項関係同士の関係を補完する新たな２項関係を関係表から選択して選択関係表に付加するとともに、必須修飾句記録手段１７に記録されている対応表を参照して、必須格となる２項関係を関係表から選択して、選択関係表に新たに付加する。
【０１４６】
文合成手段１６は、補完関係付加手段１５から供給された選択関係表に記載されている２項関係を適宜接続することにより文を合成する。
以上の実施の形態によれば、不要に長い文が出力されることを防止することが可能となるとともに、重要な単語を優先的に結合した文を出力することが可能となる。
【０１４７】
続いて、図３６を参照して本発明の第４の実施の形態について説明する。
図３６は、本発明の第４の実施の形態の構成例を示す図である。この図において、図２の場合と対応する部分には同一の符号を付してあるのでその説明は省略する。なお、この実施の形態においては、図２の場合と比較して、文書記録手段３１、キーワード入力手段３２、および、文書検索手段３３が新たに付加されている。なお、その他の構成は、図２の場合と同様である。
【０１４８】
文書記録手段３１は、文書入力手段１１から入力された文書を記録する。キーワード入力手段３２からは、文書記録手段３１に記録されている文書を検索するためのキーワードが入力される。文書検索手段３３は、キーワード入力手段３２から入力されたキーワードを含んでいる文書を文書記録手段３１から検索して出力する。
【０１４９】
次に、以上の実施の形態の動作について説明する。
図３に示す文書を含む複数の文書が文書入力手段１１から入力されたとすると、これらの文書は文書記録手段３１に記録される。
【０１５０】
いま、キーワード入力手段３２からキーワードとして「中国」と「合弁」の２つの単語が入力されたとすると、文書検索手段３３は、これら２つのキーワードを双方とも含んでいる文書を文書記録手段３１から取得する。図３に示す文書は、「中国」と「合弁」という２つのキーワードを双方とも含んでいることから、この文書と、これらのキーワードを双方とも含む他の文書とが検出結果として出力される。なお、以下の処理では、簡単のために、図３に示す文書に対する処理についてのみ説明する。
【０１５１】
重要単語抽出手段１２は、図２の実施の形態の場合と同様の処理により、重要単語を抽出し、重要単語リストを生成する。また、重要単語抽出手段１２は、キーワード入力手段３２から入力されたキーワードを、重要単語として重要単語リストに付加する。なお、この実施の形態では、入力されたキーワードをＬｅｖｅｌ３の重要単語と置換する。従って、得られる重要単語リストは以下のようになる。
【０１５２】
Ｌｅｖｅｌ１ 富士山紡績、江山紡績
Ｌｅｖｅｌ２ 李白雲紡績廠、絹織物、合弁会社、社長、子会社
Ｌｅｖｅｌ３ 中国、合弁
関係表生成手段１３は、以上の重要単語を含む文の構文解析を行うことにより、各文の２項関係を抽出し、関係表を生成する。選択関係生成手段１４は、第１の実施の形態と同様の処理を行うことにより、関係表から所定の規則に応じて重要な２項関係を抽出し、選択関係表を生成する。
【０１５３】
補完関係付加手段１５は、選択関係表に記載されている２項関係同士の関係を補完する新たな２項関係を関係表から選択して選択関係表に付加するとともに、必須修飾句記録手段１７に記録されている対応表を参照して、必須格となる２項関係を関係表から選択して、選択関係表に新たに付加する。
【０１５４】
文合成手段１６は、補完関係付加手段１５から供給された選択関係表に記載されている２項関係を適宜接続することにより文を合成する。この結果得られる文を以下に示す。
【０１５５】
富士山紡績は中国の絹織物企業の李白雲紡績廠と提携。
富士山紡績は合弁会社を設立。
子会社の江山紡績（川崎竜之介社長）が三十％出資。
【０１５６】
このような合成文は、例えば、図示せぬ表示装置に供給され、ＣＲＴや印刷用紙に対して表示出力されることになる。
なお、このような処理は、文書検索手段３３によって取得された全ての文書に対して実行され、検索結果である複数の文書の要約を得ることができる。
【０１５７】
以上の実施の形態によれば、複数の文書から、所定のキーワードによって所望の文書を選択しようとする場合において、検索結果の文書の要約を作成して表示するようにしたので、目的となる文書を迅速に見つけることが可能となる。また、その場合に、合成される文中にキーワードを配置するようにしたので、操作者がうろ覚えの状態で目的となる文書を検索しているような場合においても、記憶が最も明確な単語を含む文が要約文として出力されることになるので、出力される文書が所望のものであるか否かを迅速に判断することが可能となる。
【０１５８】
なお、以上の実施の形態においては、入力されたキーワードをＬｅｖｅｌ３の重要単語と置換するようにしたが、例えば、他のレベルの重要単語と置換したり、所定のレベルの重要単語に付加するようにしてもよい。
【０１５９】
また、キーワードのみを重要単語として用いることも可能である。即ち、重要単語抽出手段１２によって重要単語を抽出せずに、キーワードとして入力された単語のみを重要単語とすることも可能である。
【０１６０】
その場合に得られる合成文を以下に示す。
富士山紡績は中国の企業の李白雲紡績廠と提携。
富士山紡績は合弁会社を設立。
【０１６１】
このような方法によっても、充分な要約文（合成文）を得ることができる。
また、第２の実施例と同様に、重要単語とキーワードの表示形式を変更するようにしてもよい。
【０１６２】
図３７は、重要単語には下線を、また、キーワードには上線を付加して表示した表示例である。このようにして表示することにより、合成された文中のどこにキーワードが配置されているかを迅速に知ることが可能となる。
【０１６３】
更にまた、キーワード入力手段３２から入力されたキーワードをシーソラス(Thesaurus) 展開して（入力されたキーワードに対応する類義語を新たなキーワードとして）検索処理を行い。取得された文書の要約を生成する場合において、シーソラス展開によって得られたキーワードを重要単語として追加したり、あるいは、生成された要約文を表示する場合に、シーソラス展開によって得られたキーワードの表示形式を変更するようにしてもよい。
【０１６４】
なお、上記の例では、関係表中の一定の条件を満たした関係を抽出し、選択関係表を生成しているが、選択関係表を用いなくても同様の処理を行うことができる。例えば、関係表中の一定の条件を満たした関係に対し、マークを付ければよい。このような例を第５の実施の形態として以下に説明する。
【０１６５】
図３８は、本発明の第５の実施の形態の構成例を示す図である。
文書入力手段４１は、処理の対象となる文書を入力する機能を備える。この文書入力手段４１から入力された文書４７は重要単語抽出手段４２に渡される。
【０１６６】
重要単語抽出手段４２は、文書内の各単語に重要度スコアを与える。スコアの計算方法としてはtf*IDF積などを用いる。この実施の形態では、全部の単語を抽出し、抽出した各単語に対して重要度のスコアを与える。
【０１６７】
関係表生成手段４３は、重要単語抽出手段４２によって得られた単語（全部または重要なもののみに限定してもよい）を含む文を解析し、単語間の２項関係を抽出する。解析方法は、構文解析や形態素解析結果に対するパターンマッチなどがある。解析結果に基づいて、関係表４８を生成する。
【０１６８】
関係選択手段４４は、関係表生成手段４３によって生成された関係表４８から、重要な単語を含む２項関係を所定の規則にしたがって抽出し、関係表４８中の関係に重要関係マークを与える。抽出規則としては、関係を構成する単語の重要度のスコアの合計が一定値以上のものをスコアの順に並べ、それらの上位から一定個数までを選択する。また、重要単語の数や、他の重要関係と同じ文中にあるか否かなどを抽出規則に含めてもよい。
【０１６９】
補完関係付加手段４５は、関係表４８中の重要関係マークが付与された関係群に対して、それらを補完する新たな関係に選択マークを与える。補完されるものは、必須格、必須修飾句などである。
【０１７０】
文合成手段４６は、関係表のマークのつけられた２項関係から、入力された文書４７の要約文４９を合成する。
次に、第５の実施の形態における動作について説明する。
【０１７１】
文書入力手段４１によって入力された文書４７は、重要単語抽出手段４２に渡される。重要単語抽出手段４２は、文書４７中に存在する単語全てに重要度スコアを付与する。ここでは、tf*IDF積を用いるものとする。なお、全ての単語に対してスコアを付与せずに、自立語のみに与えるようにしてもよい。また、自立語のうち、連体詞、副詞、代名詞を除いたものにしてもよい。
【０１７２】
全ての単語に対して重要度のスコアが付与されたら、関係表生成手段４３が単語間の２項関係を抽出し、それらの２項関係にスコア付けする。この例では、係り側の単語と受け側の単語とのスコアの合計を、２項関係に対するスコアとする。係り側あるいは受け側の名詞が複合名詞の場合、構成要素である名詞のスコアの合計を複合名詞のスコアとする。スコア付けに関しては、関係の種別に対してもスコアを与え、係り側単語スコア、受け側単語スコア、関係スコアのそれぞれに重み付けを行い、それらの総和を得る方法もある。計算方法の例としては、前述の式（１）に示した通りである。そして、スコア付けされた２項関係によって、関係表４８を生成する。
【０１７３】
次に、関係選択手段４４が、関係表４８の中からスコアの高い順に一定数の重要関係を選択する。この個数は予め与えるものとする。または、要約の全体長などを考慮して、その都度計算するようにしてもよい。関係選択手段４４は、関係表４８中の選択された重要関係に対してマークを付与する。
【０１７４】
補完関係付加手段４５と文合成手段４６とは、関係選択手段４４で選択された２項関係から、その時点で最もスコアの高い関係を順次ピックアップし、それらに重要関係を付加し、要約文を合成するプロセスを繰り返す。
【０１７５】
図３９は、第５の実施の形態における要約文合成処理を示すフローチャートである。
［Ｓ２０１］補完関係付加手段４５は、隣接する補完関係をまとめる。具体的には、マークの付けられた重要関係を１つピックアップする。そして、以下の条件を満たす重要関係を、ピックアップした重要関係に接続する。
【０１７６】
（１）距離０：重要関係中のどちらかの要素に対して、直接つながっている重要関係。
図４０は、距離０の補完関係を示す図である。図中、「Ａ」の要素５１と「Ｂ」の要素５２との関係が、ピックアップされた重要関係である。なお、この図では、係り受け関係の係り側から受け側に向かって矢印が示されている（図４１，４２においても同様）。この例では、「Ｃ１」の要素６１は、「Ａ」の要素５１に直接つながっているため、距離０で隣接する補完関係である。また、「Ｃ２」、「Ｃ３」の要素６２，６３は、「Ｂ」の要素５２に直接つながっているため、距離０で隣接する補完関係である。
【０１７７】
（２）距離１：重要関係中のどちらかの要素に対して、１つの基本関係（重要関係としてリストアップされなかったもの）を介してつながる重要関係。
図４１は、距離１の補完関係を示す図である。この図では、基本関係を太い矢印で示している（図４２においても同様）。この例では、「Ｄ１」の要素７１と「Ｅ１」の要素７２とで構成される重要関係は、「Ｅ１」の要素７２と「Ａ」の要素５１とが基本関係を有しているため、距離１で隣接する補完関係である。「Ｄ２」の要素７３と「Ｅ２」の要素７４とで構成される重要関係は、「Ｅ２」の要素７４と「Ｂ」の要素５２とが基本関係を有しているため、距離１で隣接する補完関係である。「Ｄ３」の要素７５と「Ｅ３」の要素７６とで構成される重要関係は、「Ｅ３」の要素７６と「Ｂ」の要素５２とが基本関係を有しているため、距離１で隣接する補完関係である。「Ｄ４」の要素７７と「Ｅ４」の要素７８とで構成される重要関係は、「Ｄ４」の要素７７と「Ｂ」の要素５２とが基本関係を有しているため、距離１で隣接する補完関係である。
【０１７８】
（３）距離２：ひとつの要素Ｅ（ふたつの基本関係）を介してつながる重要関係。
図４２は、距離２の補完関係を示す図である。この例では、「Ｇ１」の要素８１と「Ｈ１」の要素８２とで構成される重要関係は、受け側の要素８２が「Ｆ」の要素８０をはさんだ２つの基本関係を介して「Ｂ」の要素５２につながるため、距離２で隣接する補完関係である。「Ｇ２」の要素８３と「Ｈ２」の要素８４とで構成される重要関係は、係り側の要素８３が「Ｆ」の要素８０をはさんだ２つの基本関係を介して「Ｂ」の要素５２につながるため、距離２で隣接する補完関係である。「Ｇ３」の要素８５と「Ｈ３」の要素８６とで構成される重要関係は、受け側の要素８６が「Ｆ」の要素８０をはさんだ２つの基本関係を介して「Ｂ」の要素５２につながるため、距離２で隣接する補完関係である。
【０１７９】
なお、ここでは関係間の距離に応じて３種類の隣接重要関係をあげたが、これは増減可能である。一文を短くするためには短い距離のものしか選ばないようにすればよい。一文が長くてもよい場合は距離の大きいものまで含めるようにする。また、構成要素(単語)の長さや、生成したときの文長によって、どこまで含めるかをダイナミックに変えてもよい。
【０１８０】
図３９に戻り、フローチャートの続きを説明する。
［Ｓ２０２］補完関係付加手段４５は、必須基本関係を補足する。具体的には、各要素との間で必須格の２項関係を有する基本関係を加える。必須格は、「は」、「が」、「を」、「に」、名詞連続とする。そして、関係表中の必須補足関係に対して、選択マークを付与する。
［Ｓ２０３］文合成手段４６は、要約文の要素として確定した関係を、関係表から除去する。すなわち、ステップＳ２０１で選択した重要関係と、ステップＳ２０２で使った基本関係を関係表から除去する。
［Ｓ２０４］文合成手段４６は、文の合成を行う。具体的には、もともとあった助動詞、補助動詞、副助詞をつなげて文を合成し、合成した文を要約文リストに追加する。
［Ｓ２０５］文合成手段４６は、要約文リストの文長合計が一定値を超えたか否かを判断する。要約文リストの文長合計が一定値を超えたら場合には、ステップＳ２０１〜Ｓ２０４の繰り返し処理（ループ）を抜け、ステップＳ２０６に進む。そうでなければステップＳ２０１へ進み、再度繰り返し処理を行う。
【０１８１】
なお、重要関係が尽きた場合にも繰り返し処理を抜けステップＳ２０６に進むこととする。ただし、予め重要関係を多めに選ぶことにより、その終了条件を避けるものとする。
［Ｓ２０６］文合成手段４６は、ステップＳ２０４で得られた要約文を元々の出現順に並べ替える。
【０１８２】
このようにして、要約文が生成され、文合成手段４６が要約文を提示する。
次に、具体的な要約文生成例を示す。
図４３は、入力文書の例を示す図である。これは、「酒ビスケット」の紹介文である。このような文書４７が、文書入力手段４１から重要単語抽出手段４２に渡される。
【０１８３】
ここでは、上記のアルゴリズムを用いて、実際のＫＲ（キーリレーション関係）抽出データをもとにして、アルゴリズムのシミュレーションを行う。また、この例では、関係が理想通り抽出されていることを仮定する。
【０１８４】
まず、重要単語抽出手段４２が５個〜１０個の重要関係を選択する。ここでの単語のスコア付けはtf*IDF積を用いる。本来IDFは、その文書が属するある文書集合から計算するべきだが、ここでは、毎日新聞１９９４年の記事約１０万件から１万件をピックアップしたものを文書集合として計算したものを用いている。これは、Ｗｅｂ文書の集合を規定してDFを計算するのが困難であることと、新聞記事に偏りが少なく、一般に用いられる単語の頻度と大差ないと判断できるからである。
【０１８５】
次に、関係表生成手段４３が、各単語を含む文を解析し、単語間の２項関係を抽出する。そして、各２項関係に対するスコアを計算し、関係表が生成される。
図４４は、ピックアップされた関係から生成された関係表を示す図である。この関係表４８には、「関係名」、「第１の単語群」、「第２の単語群」、「出現数」、「第１の単語群スコア」、「第２の単語群スコア」、「全体スコア」、及び「マーク」の項目が設けられている。
【０１８６】
「関係名」には、２つの単語がどのような関係にあるのかが示されている。関係名「動た名」は、第１の単語群が動詞、第２の単語群が名詞であり、「た」という言葉によって接続されていることを示す。関係名「形動な名」は、第１の単語群が形容動詞、第２の単語群が名詞であり、「な」という言葉によって接続されていることを示す。関係名「に」は、第１の単語群と第２の単語群とが「に」という言葉によって接続されていることを示す。関係名「形い名」は、第１の単語群が形容詞、第２の単語群が名詞であり、「い」という言葉によって接続されていることを示す。関係名「複合」は、第１の単語群に設定されている単語が複合名詞であることを示す。関係名「の名」は、第１の単語群と第２に単語群とがともに名詞であり、「の」という言葉によって接続されていることを示す。関係名「ＮＮ」は、第１の単語群に設定されている単語が、名詞連続であることを示す。
【０１８７】
「第１の単語群」には、係り側の単語もしくは単語群が登録される。単語群として登録されるのは、係り側が複合名詞の場合や、名詞連続の場合である。また、関係名が「複合名詞」や「ＮＮ（名詞連続）」の単語群も、「第１の単語群」の項目に登録される。
【０１８８】
「第２の単語群」には、受け側の単語もしくは単語群が登録される。単語群として登録されるのは、受け側が複合名詞の場合や、名詞連続の場合である。
「出現数」には、入力された文書中に何回出現したのかが設定される。
【０１８９】
「第１の単語群スコア」は、「第１の単語群」に設定されている単語群のスコアである。複数の単語が設定されている場合には、各単語のスコアの合計である。
【０１９０】
「第２の単語群スコア」は、「第２の単語群」に設定されている単語群のスコアである。
「全体スコア」は、２項関係全体に対するスコアである。「第１の単語群スコア」の値と「第２の単語群スコア」の値とを合計することによって求められる。
【０１９１】
「マーク」は、関係選択手段４４によって選択された重要関係を示すフラグである。この項目に所定のマークが登録されていれば、重要な関係として選択されたことを示す。関係表生成手段４３によって関係表が生成された直後の段階では、重要関係の選択は行われていない。
【０１９２】
なお、この図では、「全体スコア」の高い順に重要関係が並べられている。
このような関係表が生成されると、関係選択手段４４によって全体スコアの高い重要関係が一定数選択される。
【０１９３】
図４５は、重要関係選択後の関係表を示す図である。このように、関係選択手段４４によって選択された２項関係に対して、「○」のマークが付与されている。
【０１９４】
次に、補完関係付加手段４５と文合成手段４６とによって、図３９に示した処理が行われる。
まず、１回目のループについて説明する。
【０１９５】
ステップＳ２０１において、補完関係付加手段４５が、全体スコアがもっとも高い重要関係を選択する。ここで選択する重要関係は、
● こめ−た−酒，ビスケット
である。そして、隣接する重要関係をまとめる。隣接する重要関係は、
● 風雅な−酒，ビスケット
である。この重要関係は、「風雅」という単語が「酒，ビスケット」に直接接続されているため、距離０の隣接関係である。
【０１９６】
ステップＳ２０２において、必須格を持つ基本関係を補足する。この例では「こめ」に対して、「を格」の追加をする。その結果、
●ロマン−を−こめ
の基本関係が追加される。関係表４８中のこの基本関係に対して、選択マークを付与する。
【０１９７】
ステップＳ２０３において、文合成手段４６により、重要関係が除去される。ここで用いた３つの関係がリストから除去される。
ステップＳ２０４において、文合成手段４６により文が合成される。この１回目のループで得られる表現は、
「ロマンをこめた風雅な酒ビスケット」
となる。
【０１９８】
以後、同様にループを進める。
２回目のループでは、
重要関係：「五合，升−に−納め」
隣接する重要関係：なし
必須格補足：なし（「納め−て−お納めします」）
表層表現：「五合升に納めて」
が得られる。
【０１９９】
３回目のループでは
重要関係：「高（い）−特選,吟醸,酒」
隣接する重要関係：なし
必須格補足：「香り（が）高い」
表層表現：「香り高い特選吟醸酒」
関係を５つとった場合はここで終わりとなる。
【０２００】
４回目のループでは
重要関係：「神奈川−の−洋菓子,ウエダ」
隣接する重要関係：なし
必須格補足：「ウエダが−創作（した）」
表層表現：「神奈川の洋菓子ウエダが創作した」
５回目のループでは
重要関係：「ビスケット生地を−浸し」
隣接する重要関係：「銘酒−の−香り」
（「香り−に−浸し」（全体スコア、１１位）を介してつながっている）
必須格補足：なし
表層表現は「ビスケット生地を銘酒の香りに浸しました」である。
【０２０１】
以上でピックアップした関係は、次のようになる。
「ロマンをこめた風雅な酒ビスケット」
「五合升に納めて」
「香り高い特選吟醸酒」
「神奈川の洋菓子ウエダが創作した」
「ビスケット生地を銘酒の香りに浸しました」
抽出された文を、元の文書における出現順に並べ直すことで、要約文が生成される。
【０２０２】
図４６は、生成された要約文を示す図である。このような要約文が文合成手段４６から出力される。
以上のように、関係表に対して選択された旨のマークを付与しておけば、選択関係表を別途設けなくてもよくなる。
【０２０３】
次に、第６の実施の形態について説明する。第６の実施の形態では、要約文として使用することが決定した関係を削除するのではなく、スコアを落とした状態で存続させておく。
【０２０４】
図４７は、第６の実施の形態の構成例を示す図である。この図において、文書入力手段４１と重要単語抽出手段４２については、第５の実施の形態の同名の構成要素を同じ機能を有しているため、図３９の同名の構成要素と同じ符号を付して説明を省略する。
【０２０５】
関係表生成手段４３ａは、基本的に第５の実施の形態の関係表生成手段４３（図３９に示す）と同じ機能を有しているとともに、さらに次の機能を有している。すなわち、文合成手段４６ａより再スコアリング要求を受けると、要約文として選択された単語のスコアを下げ、関係表のスコアに関する値を再計算する。
【０２０６】
関係選択手段４４ａは、関係表生成手段４３ａによって関係表が生成されたとき、及び関係表生成手段４３ａによって再スコアリングが行われたときに、その時点で最も全体スコアの高い重要関係を選択する。
【０２０７】
補完関係付加手段４５ａは、関係選択手段４４ａが選択した関係に対する必須関係を追加する。
文合成手段４６ａは、関係選択手段４４ａが選択した関係と、補完関係付加手段４５ａによって付加された必須関係とに基づいて、表層文を生成する。そして、表層文全体の長さが閾値を超えていなければ、再スコアリング要求を関係表生成手段４３ａに送る。再スコアリング要求には、新たに追加された表層文を構成する単語の情報が含まれる。表層文全体の長さが閾値を超えていれば、生成されている表層文を、元の文書４７への出現順に並べ、要約文４９ａを生成する。
【０２０８】
次に、第６の実施の形態による要約文の生成手順を説明する。
なお、関係表生成までのアルゴリズムは第５の実施の形態と同様である。そこで、関係表生成後の処理について以下に説明する。
【０２０９】
図４８は、第６の実施形態における要約文合成処理を示すフローチャートである。
［Ｓ２１１］関係表が生成されたら、関係選択手段４４ａが重要度スコアの最も高い２項関係を１つだけ選択する。
［Ｓ２１２］補完関係付加手段４５ａが、必須関係を追加する。初期段階では、「が」、「は」、「を」、「に」など固定的なものとする。なお、ここでは予め重要関係を抜き出していないので、隣接重要関係を補完するステップはない。
［Ｓ２１３］補完関係付加手段４５ａが、関係の補充（曖昧性の保存）を行う。「Ａ→ＢのＣ」のような形態の場合であって「Ａ→Ｂ」または「Ａ→Ｃ」のみが抽出されている場合は、関係「ＢのＣ」を加える。
［Ｓ２１４］文合成手段４６ａは、枝数などをカウント、または、表層文生成を仮適用して表層文全体の長さが、予め設定してある閾値を超えたか否かを判断する。閾値を超えた場合にはステップＳ２１６に進み繰り返し処理（ループ）を抜ける。そうでなければステップＳ２１５に進む。
［Ｓ２１５］関係表生成手段４３ａは、再スコアリングを行う。すなわち、使った単語のスコアを落とす。ここでは各単語のスコアを１／５にしている。その後、新しい単語のスコアから新しい関係のスコアを再計算する。その後ステップＳ２１１の重要関係の選択に戻り、新しくスコアが最大になった関係を選択するところからこの手順を繰り返す。
［Ｓ２１６］文合成手段４６ａは、要約表層文を生成し、列挙する。この際に、出現順に並べ直すことにより要約文を生成する。そして、処理を終了する。
【０２１０】
次に、具体的な要約文の生成例について説明する。
１回目のループでの処理を、図４３で示したサンプルの文書４７を通して説明する。
【０２１１】
ステップＳ２１１では、関係選択手段４４ａが、重要度スコアのもっとも高い２項関係を選択する。ここでは
● こめ（た）−酒，ビスケット
が選択される。このとき、関係選択手段４４ａは、関係表中の選択した重要関係のマークを「１−Ｃ」とする。このマークは、第１回目のループで得られるコアの関係（ステップＳ２１１で選択された重要関係）であることを意味している。
【０２１２】
ステップＳ２１２では、補完関係付加手段４５ａにより、必須関係が付与される。ここでは、「こめる」の「を」格
● ロマン−を−こめた
が付与される。そして、補完関係付加手段４５ａは、関係表中の必須格を持つ基本関係に対して、「１−Ａ」のマークを付与する。このマークは、第１回目のループで得られる付加関係（ステップＳ２１２で付加された基本関係）であることを意味している。
【０２１３】
ステップＳ２１３では、補完関係付加手段４５ａにより、曖昧性の保存が行われる。「Ａ→ＢのＣ」のような形態の場合で「Ａ→Ｂ」または「Ａ→Ｃ」のみが抽出されている場合は、関係「ＢのＣ」を加える。ここでは、該当する例はない。
【０２１４】
ステップＳ２１４では、文合成手段４６ａによって、終了条件を満たしたか否かが判断される。この段階では、枝数２、要約文字数は１３（「ロマンをこめた酒ビスケット」）であり、まだ終了条件を満たしていないものとする。
【０２１５】
ステップＳ２１５では、関係表生成手段４３ａにより、再スコアリングが行われる。すなわち、使った単語のスコアを落とし、新しい単語のスコアから新しい関係のスコアを再計算する。
【０２１６】
図４９は、１回目のループ終了後の関係表を示す図である。この関係表４８ａでは、１回目のループで選択された重要関係である「こめ−た−酒−ビスケット」には「１−Ｃ」のマークが付与されている。そして、この重要関係に含まれる単語「こめ」、「酒」、「ビスケット」のスコアが１／５にされ、関係表のスコアが再計算されている。その結果、最初に１位だった「こめ−た−酒−ビスケット」と、「風雅−な−酒−ビスケット」のスコアは、それぞれ「5.15/5+9.68/5=2.97」、「4+9.68/5=5.94」となっている。図中、下線が引かれているのが、値の変更されたスコアである。このため、最初は３位のスコアだった「五合升−に−納め（て）」（スコア13.44）が最大のスコアとなる。
【０２１７】
２回目以降も同様な操作を続ける。ここでは、あらかじめ終了条件を決めずに何回か繰り返してみる。
２回目のループを行う。全体スコアが「１３．４４」である「五合升−に−納め（て）」が選択され、枝数は３となり、要約文字数は７文字追加で２２となる。
【０２１８】
図５０は、２回目のループ終了後の関係表を示す図である。この関係表４８ａでは、新たに「五合升−に−納め（て）」に対して「２−Ｃ」のマークが付与されている。
【０２１９】
３回目のループを行う。「高い−特選，吟醸，酒」が選択され、「香り−高い」が追加される。表層表現は「香り−高い−特選，吟醸，酒」である。枝数は５となり、要約文字数は９文字追加で２９となる。
【０２２０】
図５１は、３回目のループ終了後の関係表を示す図である。この関係表４８ａでは、新たに「高い−特選，吟醸，酒」に対して「３−Ｃ」のマークが付与されている。また、図示されていないが、「香り−高い」の基本関係に対して「３−Ａ」のマークが付与されている。
【０２２１】
４回目のループを行う。「神奈川−の−洋菓子，ウエダ」が選択され、「ウエダが−創作（した）」が追加される。枝数は７となり、要約文字数は１５文字追加で４２となる。
【０２２２】
最終的に得られる要約は、
「ロマンをこめた酒ビスケット」
「五合升−に−納めて」
「香り−高い−特選，吟醸，酒」
「神奈川−の−洋菓子,ウエダ ウエダが−創作した」
である。この表層文を、元文書での出現順に並べ替えることで、要約文が生成される。図５２は、第６の実施の形態によって生成された要約文４９ａを示す図である。
【０２２３】
ところで、上記の例では、重要語の関係を関係表によって管理していたが、関係表に替えて解析木を用いることもできる。そして、解析木に対してマークを付けることができる。
【０２２４】
図５３は、マーク付けが行われた解析木を示す図である。これは、第６の実施の形態において１回目のループが終了した時点での解析木を示している。この図において、太い線で結ばれた単語は、選択された２項関係であることを示している。また、その線の近傍に示された記号は、選択マークを示している。マークの意味は、第６の実施の形態と同様である。太線の矩形で囲まれた単語は、選択された単語を示している。
【０２２５】
このように、解析木にマークを施すことによって、要約文を構成する関係を抽出していくことができる。
なお、上記の処理機能は、コンピュータによって実現することができる。その場合、文書要約装置が有すべき機能の処理内容は、コンピュータで読み取り可能な記録媒体に記録されたプログラムに記述されており、このプログラムをコンピュータで実行することにより、上記処理がコンピュータで実現される。
【０２２６】
コンピュータで読み取り可能な記録媒体としては、磁気記録装置や半導体メモリ等がある。市場に流通させる場合には、ＣＤ−ＲＯＭ(Compact Disk Read Only Memory) やフロッピーディスク等の可搬型記録媒体にプログラムを格納して流通させたり、ネットワークを介して接続されたコンピュータの記憶装置に格納しておき、ネットワークを通じて他のコンピュータに転送することもできる。
【０２２７】
コンピュータで実行する際には、コンピュータ内のハードディスク装置等にプログラムを格納しておき、メインメモリにロードして実行する。
【０２２８】
【発明の効果】
以上説明したように本発明では、文書入力手段からは、処理の対象となる文書を入力し、重要単語抽出手段は、文書入力手段から入力された文書から、重要な単語を抽出し、関係表生成手段は、重要単語抽出手段によって抽出された重要な単語を含む文の構造を解析し、その文に含まれている単語同士の２項関係を解析して関係表を生成し、関係選択手段は、関係表生成手段によって生成された関係表から、重要な単語を含む２項関係を所定の規則に従って選択し、文合成手段は、関係選択手段によって選択された２項関係から文を合成するようにしたので、入力された文書から短くて的確な要約を作成することが可能となる。
【図面の簡単な説明】
【図１】 本発明の文書要約装置の原理を説明する原理図である。
【図２】 本発明の第１の実施の形態の構成例を示す図である。
【図３】 図２に示す実施の形態において処理される文書の一例である。
【図４】 図２に示す重要単語抽出手段によって、図３に示す文書から抽出された重要単語リストの一例を示す図である。
【図５】 図２に示す関係表生成手段によって、重要単語が含まれている文に対して構文解析がなされた場合に得られる結果の一例を示す図である。
【図６】 図５に示す構文解析の結果得られる関係表の一例を示す図である。
【図７】 図３に示す文書に対して付与された文番号の一例を示す図である。
【図８】 図２に示す実施の形態において実行される処理の一例を説明するフローチャートである。
【図９】 図８に示す「重要単語を含む２項関係の抽出処理」の一例を説明するフローチャートである。
【図１０】 図８に示す「直接関係の探索処理」の一例を説明するフローチャートである。
【図１１】 図１０に示す処理の結果生成される選択関係表の一例を示す図である。
【図１２】 図９に示す「同じ文にある関係の探索処理」の一例を説明するフローチャートである。
【図１３】 図１２に示す処理の結果生成される選択関係表の一例を示す図である。
【図１４】 図９に示す「主要格の探索処理」の一例を示すフローチャートである。
【図１５】 主要格リストの一例である。
【図１６】 生成される選択関係表の一例を示す図である。
【図１７】 生成される選択関係表の一例を示す図である。
【図１８】 追加される２項関係を示す図である。
【図１９】 図８に示す「必須格の補足処理」の結果生成される選択関係表の一例を示す図である。
【図２０】 図８に示す「表層句の生成処理」の一例を説明するフローチャートである。
【図２１】 図２０に示す「選択されたアークの係り側のノードに対する探索処理」の一例を説明するフローチャートである。
【図２２】 図２０に示す「選択されたアークの受け側のノードに対する探索処理」の一例を説明するフローチャートである。
【図２３】 図２０に示す処理によって、選択関係表に記載されている２項関係が合成されて得られた文の一例を示す図である。
【図２４】 図２０に示す処理によって、選択関係表に記載されている２項関係が合成されることによって得られた文の他の一例を示す図である。
【図２５】 図２０に示す処理によって、選択関係表に記載されている２項関係が合成されることによって得られた文の更に他の一例を示す図である。
【図２６】 図２に示す実施の形態において処理される文書の一例を示す図である。
【図２７】 図２６に示す文書から抽出された重要単語の一例を示す図である。
【図２８】 図２６に示す文書から合成された文の一例を示す図である。
【図２９】 図２に示す実施の形態において処理される文書の一例を示す図である。
【図３０】 図２９に示す文書から抽出された重要単語の一例を示す図である。
【図３１】 図２９に示す文書から合成された文の一例を示す図である。
【図３２】 本発明の第２の実施の形態の構成例を示す図である。
【図３３】 図３２に示す実施の形態において図３に示す文書を処理した結果、表示装置に表示される文の表示例を示す図である。
【図３４】 図３２に示す実施の形態において図３に示す文書を処理した結果、表示装置に表示される文の他の表示例を示す図である。
【図３５】 本発明の第３の実施の形態の構成例を示す図である。
【図３６】 本発明の第４の実施の形態の構成例を示す図である。
【図３７】 図３６に示す実施の形態において図３に示す文書を処理した結果、表示装置に表示される文の表示例を示す図である。
【図３８】 本発明の第５の実施の形態の構成例を示す図である。
【図３９】 第５の実施の形態における要約文合成処理を示すフローチャートである。
【図４０】 距離０の補完関係を示す図である。
【図４１】 距離１の補完関係を示す図である。
【図４２】 距離２の補完関係を示す図である。
【図４３】 入力文書の例を示す図である。
【図４４】 ピックアップされた関係から生成された関係表を示す図である。
【図４５】 重要関係選択後の関係表を示す図である。
【図４６】 生成された要約文を示す図である。
【図４７】 本発明の第６の実施の形態の構成例を示す図である。
【図４８】 第６の実施の形態における要約文合成処理を示すフローチャートである。
【図４９】 １回目のループ終了後の関係表を示す図である。
【図５０】 ２回目のループ終了後の関係表を示す図である。
【図５１】 ３回目のループ終了後の関係表を示す図である。
【図５２】 第６の実施の形態によって生成された要約文を示す図である。
【図５３】 マーク付けが行われた解析木を示す図である。
【符号の説明】
１ 文書入力手段
２ 重要単語抽出手段
３ 関係表生成手段
４ 選択関係生成手段
５ 補完関係付加手段
６ 文合成手段
１１ 文書入力手段
１２ 重要単語抽出手段
１３ 関係表生成手段
１４ 選択関係生成手段
１５ 補完関係付加手段
１６ 文合成手段
１７ 必須修飾句記録手段
２１ 重要単語数制限手段
２２ 表示形式変更手段
２３ 表示装置
２５ 選択関係表生成手段
３１ 文書記録手段
３２ キーワード入力手段
３３ 文書検索手段[0001]
BACKGROUND OF THE INVENTION
The present invention provides a document summarization device, and Recording medium To the body In particular, a document summarization device that creates and outputs a summary of an input document, and A recording medium recording a program for causing a computer to execute such processing To the body Related.
[0002]
[Prior art]
When a document search is performed, it is necessary to select a compatible document from the search results. For this purpose, the contents of each document included in the search result must be grasped. However, when a large number of documents are extracted as search results, it is not easy to read the description of each document and understand the contents of each document.
[0003]
There are two types of methods for allowing the reader to grasp the document contents in a short time: (1) reducing the reading amount and (2) supporting fast reading.
As a method for shortening the former document, the following methods can be cited.
(A) Keyword enumeration
(B) Important sentence pickup
(C) Sentence shortening
First, the “keyword enumeration” method is intended to make the contents of a document roughly understood by enumerating important words such as frequent words in the document. Examples of such a method include “Cutting, et al.,“ Scatter / Gather: A Cluster-based Approach to Browsing Large Document Collections ”SIGIR-92, 1992”. In this method, a document is expressed as a collection of frequently used keywords and a part of article titles.
[0004]
According to such a method, it is possible to infer rough contents from the displayed frequent keywords.
Next, in the method based on “important sentence pickup”, an important sentence is determined based on frequent words and sentence patterns, and a summary is generated by picking them up and connecting them. For example, Japanese Patent Application Laid-Open No. 02-093866 discloses a method of picking up a sentence including keywords that are registered in advance and that are semantically important. Japanese Patent Laid-Open No. 06-348696 discloses a method of scoring each word in a document and picking up a certain number of sentences with a high total score. According to such a method, since a summary in which important sentences are listed is output, more specific contents can be obtained as compared with a case where only keywords are displayed.
[0005]
In the last “sentence shortening” method, the following techniques have been proposed.
That is, Japanese Patent Application Laid-Open No. 5-101120 discloses a document abstract generation apparatus that shortens sentences by deciding the types of dependency to be omitted and those not to be deleted, and deleting clauses determined to be omitted. ing. For example, “continuous modifier, ga, ni, kara” is left, and “adverb, lantern modifier,” is deleted.
[0006]
Japanese Patent Application Laid-Open No. 6-035961 discloses a document summarizing apparatus that deletes modifiers and shortens sentences.
Furthermore, Japanese Patent Application Laid-Open No. 8-212228 discloses a summary sentence creating apparatus that creates a summary sentence by combining high importance elements. In this summary sentence creation device, the importance is determined as follows using the part-of-speech information obtained by morphological analysis. That is, the importance of nouns and verbs is high, and the importance of adverbs is low. Also, using parsing technology, the main case, objective case, and predicate are highly important, and other cases are low. Furthermore, the importance of the main sentence is high and the importance of the subordinate is low. Then, a summary sentence is created by deleting words that are determined to be less important.
[0007]
Furthermore, Japanese Patent Laid-Open No. 6-259423 discloses an automatic summary sentence creation method in which an unnecessary paragraph is deleted from an important paragraph after it has been selected. In this summary sentence automatic creation method, unnecessary portions are selected by pattern matching. Then, the phrase in parentheses, the sentence after the paraphrase word such as “that is”, and the sentence after the word that draws an example such as “for example” are deleted.
[0008]
According to such a method, the document from which the less important part is deleted is displayed, so that the outline of the document can be quickly understood.
Next, as a method for supporting “speed reading”, the following techniques are disclosed.
[0009]
That is, in the document speed reading support display method disclosed in Japanese Patent Application Laid-Open No. 05-158401, the important part and the other part are separated, and the parts other than the important part are reduced or displayed. According to such a method, by tracing and reading important parts (parts written in kanji, noun / verb parts, and search keyword parts), the meaning can be understood quickly compared to the normal display method. It becomes possible to do.
[0010]
[Problems to be solved by the invention]
However, the method based on keyword enumeration has a problem that it is difficult to understand the exact contents because the relationship between displayed frequent keywords is not shown.
[0011]
Also, in the method using important sentence pickup, sentences containing many words with high appearance frequency are extracted as they are, so that the abstracts generated tend to be long sentences. Therefore, there is a problem that it takes time to read.
[0012]
Furthermore, the method using sentence shortening has a problem that there is a possibility that an important part may be excessively deleted because each method determines whether or not to delete only by the type of dependency. In other words, it is often impossible to determine whether or not a word represents important information in the sentence simply by using a dependency relationship. Therefore, in some cases, important information may not be displayed. there were.
[0013]
Furthermore, since the entire document is displayed in the method for supporting speed reading, there is a problem that it is difficult to use the method for selecting a specific document from a plurality of documents, for example.
[0014]
The present invention has been made in view of these points, and provides a document summarization apparatus capable of accurately summarizing documents without deleting necessary information. With the goal .
[0015]
Another object of the present invention is to provide an optimum document summarizing apparatus for specifying a desired document from a plurality of documents.
[0016]
[Means for Solving the Problems]
In the present invention, in order to solve the above problems, in a document summarizing apparatus for creating and outputting a summary of an input document, a document input unit for inputting a document to be processed, and an input from the document input unit Important word extracting means for extracting an important word from a document, and a sentence structure including the important word extracted by the important word extracting means, and a binary relation between the words included in the sentence Generated by the relation table generating means for analyzing the relationship and generating the relation table Said A relation selection means for selecting a binary relation including an important word from a relation table according to a predetermined rule; A complementary relationship adding means for selecting a new binary relation for complementing the binary relation selected by the relation selecting means from the relation table; The relation selection means And the complementary relationship adding means And a sentence synthesizing unit for synthesizing a sentence from the binary relation selected by (1).
[0017]
Here, the document to be processed is input from the document input means. Then, important words are extracted from the document input from the document input means by the important word extraction means. Next, the relation table generating means analyzes the structure of the sentence including the important word extracted by the important word extracting means, analyzes the binary relation between the words included in the sentence, and generates the relation table. Is done. Then, the relation selection means selects a binary relation including an important word from the relation table generated by the relation table generation means according to a predetermined rule. Furthermore, a new binary relationship that complements the binary relationship selected by the relationship selecting unit is selected from the relationship table by the complementary relationship adding unit. Then, by means of sentence synthesis, relation selection means And complementary relationship adding means Sentences are synthesized from the binary relations selected in.
[0018]
DETAILED DESCRIPTION OF THE INVENTION
Hereinafter, embodiments of the present invention will be described with reference to the drawings.
FIG. 1 is a principle diagram for explaining the principle of the document summarizing apparatus of the present invention.
[0019]
In this figure, a document to be processed is input from the document input means 1. The important word extraction unit 2 extracts an important word from the document input from the document input unit 1 with reference to the appearance frequency. The relation table generating unit 3 analyzes the structure of a sentence including an important word extracted by the important word extracting unit 2, extracts a binary relation between words included in the sentence, and generates a relation table. . The selection relation table generation means 4 extracts a binary relation including an important word from the relation table generated by the relation table generation means 3 according to a predetermined rule, and generates a selection relation table. The complementary relationship adding means 5 acquires a new binary relationship that complements the binary relationship included in the selection relationship table from the relationship table and adds it to the selection relationship table. The sentence synthesizing means 6 synthesizes and outputs a sentence that is a summary of the document input from the binary relation included in the selection relation table to which the new binary relation is added by the complementary relation adding means 5.
[0020]
Next, a configuration example of the embodiment of the sentence summarizing apparatus of the present invention will be described with reference to FIG. In this figure, a document to be processed is input from the document input means 11. The important word extraction unit 12 extracts an important word from the document input from the document input unit 11 with reference to the appearance frequency. As an important word extraction method, a tf * IDF (tf: term frequency, IDF: inverse document frequency) product (details will be described later) is used.
[0021]
The relation table generation unit 13 analyzes the structure of a sentence including an important word extracted by the important word extraction unit 12 and extracts a binary relation between words included in the sentence to generate a relation table. To do. As an analysis method, generally known syntax analysis, pattern matching according to morphological analysis, or the like can be used.
[0022]
The selection relationship table generation unit 14 extracts a binary relationship including an important word from the relationship table generated by the relationship table generation unit 13 according to a predetermined rule, and generates a selection relationship table. The selection method is based on the number of important words included in the binary relation, whether or not the main case is included, whether or not it exists in the same sentence as the important binary relation, and the like.
[0023]
The complementary relationship adding means 15 acquires a new binary relationship that complements the binary relationship included in the selection relationship table from the relationship table and adds it to the selection relationship table. As a method for selecting a new binary relation, an isolated relation or an essential case recorded in the essential modifier recording means 17 is referred to.
[0024]
The sentence synthesizing unit 16 synthesizes a sentence that is a summary of the input document from the binary relationship included in the selection relationship table to which the new binary relationship is added by the complementary relationship adding unit 15.
[0025]
The essential modifier recording means 17 is a correspondence between a verb and a mandatory case (subject, subject, subject, counter case, noun continuation, etc. (“ha”, “ga”, “o”, “ni”, etc.)) A table is recorded, and necessary information is supplied in response to a request from the complementary relationship adding means 15.
[0026]
Next, the operation of the above embodiment will be described.
From the document input means 11, for example, a document as shown in FIG. The important word extraction unit 12 extracts an important word from the document input from the document input unit 11 with reference to the appearance frequency. Note that there is a method of using a score based on the tf * IDF product as a method of avoiding the influence of words that are frequently used in many documents and frequently appear in many documents. The details are described in “G. Salton“ Automatic Text Processing, ”Addison-Wesley (1989)”.
[0027]
FIG. 4 shows a list of important words extracted by the important word extracting unit 12 from the document of FIG. In this example, Level 1 to Level 3 are classified in order of appearance frequency of words, and Level 1 having the highest appearance frequency includes “Mt. Fuji” and “Eyama”.
[0028]
The relation table generation means 13 performs syntax analysis of the sentence including the important word extracted by the important word extraction means 12, and analyzes the binary relation (dependency relation) of the words included in the target sentence. . Then, a relation table is generated according to the obtained analysis result. Note that as a syntax analysis method, a method of applying pattern matching to a morphological analysis result may be used.
[0029]
A specific example of the parsing result is shown in FIG. In this example, the word surrounded by a solid line is an important word. Further, words surrounded by a broken line are words other than important words. That is, “Mt. Fuji Spinning”, “Alliance”, “Li Baiyun Spinning”, “Joint Venture”, “China”, “Silk Fabric” are important words, and others are not important words.
[0030]
Based on the analysis results as described above, the relation table generating unit 13 extracts two words (binary relations) corresponding to the modification side and the reception side of the modification relation, and generates the relation table shown in FIG. To do.
[0031]
In the relationship table shown in FIG. 6, the relationship between the dependency side, the reception side, and the dependency relationship is displayed in a list. The sentence numbers shown in this figure are numbers assigned to the respective sentences as shown in FIG. This sentence number is assigned a different sentence number for a heavy sentence or a compound sentence. For example, the sentence “Mt. Fuji Spinning Establishes a Joint Venture” in the second line is a heavy sentence, so the sentence number “1-1” is assigned to the first half of the sentence, The sentence number “1-2” is assigned to the latter half of the sentence.
[0032]
Note that in the table shown in FIG. Actually, fields where usage, auxiliary verbs, and the like are described are provided on each of the engaging side and the receiving side.
Next, operations of the selection relationship table generation unit 14, the complementary relationship addition unit 15, the sentence synthesis unit 16, and the essential modifier phrase recording unit 17 will be described with reference to the flowchart shown in FIG.
[0033]
When this flowchart is started, the following processing is executed.
[S1] The selection relation table generation unit 14 executes a binary relation extraction process including important words.
[S2] The complementary relationship adding means 15 executes a relationship analysis process for the isolated relationship (described later).
[S3] The supplementary relationship adding unit 15 refers to the information recorded in the essential modifier recording unit 17 and executes a supplementary process for the essential case.
[S4] The sentence synthesizing unit 16 appropriately synthesizes the binary relations included in the selection relation table and generates a summary sentence as a processing result.
[0034]
Each of the above processes is a subroutine process, and each process will be described in detail below.
When the process of step S1 is started, the following process shown in FIG. 9 is executed. In the following processing, from the binary relations described in the relational table, the relational side or at least one of the receiving side is an important word, and the important relation is selected from the selected binary relations. The binary relation is extracted and added to the selection relation table.
[S1-0] The selection relation table generation unit 14 sets all the contents of the selection relation table to an empty (NULL) state.
[S1-1] The selection relationship table generation unit 14 executes a direct relationship search process. Details of this processing will be described later with reference to FIG.
[S1-2] The selection relation table generation unit 14 executes a binary relation extraction process in the same sentence. Details of this process will be described later with reference to FIG.
[S1-3] The selection relation table generation unit 14 executes a search process for the main case. Details of this processing will be described later with reference to FIG.
[0035]
When the process of step S1-1 shown in FIG. 9 is started, the flowchart shown in FIG. 10 is executed. In this process, out of the binary relations described in the relational table, those that include important words on both the dependency side and the receiving side are selected and added to the selection relational table.
[S11] The selection relation table generating unit 14 acquires one binary relation from the relation table. For example, a binary relation consisting of “Mt. Fuji” and “Alliance” is acquired from the relationship table shown in FIG.
[S12] The selection relation table generating unit 14 determines whether or not both the relational side and the receiving side of the acquired binary relation are important words. As a result, if both are important words, the process proceeds to step S13, and if not, the process proceeds to step S15.
[0036]
In the example of FIG. 6, “Fujisan Spinning” and “Alliance” are both important words, and therefore YES is determined, and the process proceeds to step S13.
[S13] The selection relation table generating unit 14 adds the acquired important word to the selection relation table.
[0037]
In the example of FIG. 6, “Fujisan Spinning” and “Alliance” are added to the selection relationship table.
[S14] The selection relation table generating unit 14 deletes important words corresponding to the dependency side and the reception side from the important word list.
[0038]
In the example of FIG. 6, “Mt. Fuji” and “Provision” are deleted from the important word list.
[S15] The selection relation table generating unit 14 determines whether or not an unprocessed binary relation remains in the relation table. As a result, if an unprocessed binary relation remains, the process returns to step S11, and if not, the process ends.
[0039]
When the above processing is applied to the example of FIG. 6, the selection relationship table shown in FIG. 11 is obtained. In this example, the relationship between “Mt. Fuji” and “Alliance” included in the sentence number 1-1 is a “subject” relationship, and this item has been acquired in the process of step S1-1. It is shown. Similarly, the relationship between “Shanghai spinning machine B” and “partnership” and “subsidiary” and “Jiangshan spinning” are shown.
[0040]
Here, the item “obtained step” at the right end of the table indicates that the above binary relationship was obtained in step S1-1.
Before this process is executed, the important words listed in the important word list are “Fujisan Spinning”, “Jiangshan Spinning”, “Shanghai Spinning B”, “Silk Fabric”, “Joint Venture”, “President”, “Subsidiary”, “Alliance”, “Capital”, “China”. When the process of step S1-1 is executed, “Fujisan Spinning”, “Alliance”, “Shanghai Spinning B”, “Subsidiary”, and “Jiangshan Spinning” are selected. As a result, these items are important. It will be deleted from the word list. Therefore, after the process of step S1-1 is executed, “silk fabric”, “joint venture”, “president”, “capital”, and “China” remain in the important word list. .
[0041]
When the process of step S1-1 shown in FIG. 9 is completed, step S1-2 is subsequently executed.
FIG. 12 is a flowchart showing details of the process in step S1-2. When this process is started, the following process is executed. In this process, both the dependency side and the receiving side are present in the same sentence (sentence having the same sentence number) as the binary relation that is an important word, and either the dependency side or the receiving side is an important word. A binary relation is selected and added to the selection relation table.
[S31] The selection relation table generation unit 14 determines whether there is an unprocessed relation among the binary relations acquired in the process of step S1-1 in FIG.
[0042]
Now that the processing has just started, there are three unprocessed binary relationships in the selection relationship table shown in FIG. 11, so that the determination is YES and processing proceeds to step S32.
[S32] The selection relation table generating unit 14 acquires only one binary relation acquired in step S1-1 from the selection relation table.
[0043]
In this example, a binary relation consisting of “Mount Fuji” and “Alliance” is extracted from the selection relationship table shown in FIG.
[S33] The selection relation table generating unit 14 determines whether or not a binary relation R having the same sentence number as the acquired binary relation exists in the relation table. As a result, when it exists, it progresses to step S34, and when that is not right, it returns to step S31.
[0044]
In the present example, the sentence number of the binary relation consisting of “Mt. Fuji” and “Alliance” is “1-1”, and the binary relation (for example, “silk fabric” and “enterprise”) having the same sentence number. And “China” and “Company”) are present in the relational table, it is determined YES and the process proceeds to step S34.
[S34] The selection relation table generating means 14 determines whether or not there is a binary relation S including only one important word in the binary relation R, and if it exists, the process proceeds to step S35. If not, the process returns to step S31.
[0045]
In this example, the two relations of “silk fabric” and “company” and “China” and “company” contain the important words “silk fabric” and “China”, respectively. Proceed to S35.
[S35] The selection relation table generating unit 14 adds the binary relation S to the selection relation table.
[0046]
In the present example, two binary relationships of “silk fabric” and “company” and “China” and “company” are added to the selection relationship table.
[S36] The selection relation table generating unit 14 deletes the important words included in the binary relation S from the important word list.
[0047]
In this example, “silk fabric” and “China” are deleted from the important word list. As a result, only “joint venture”, “president”, and “capital” exist in the important word list.
[0048]
When the above processing is repeated, in the third processing, the same sentence number (=) as the binary relationship consisting of the third “subsidiary” and “Eyama spinning” in the selection relationship table shown in FIG. Two binary relationships consisting of “Ryunosuke Kawasaki”, “President”, “Capital” and “650 million yen” having 3) are added to the selection relationship table. Then, the important words “President” and “Capital” included in these are deleted from the important word list. As a result, only the “joint company” remains in the important word list.
[0049]
FIG. 13 is a diagram showing an example of the selection relationship table generated by the above processing. From the comparison with FIG. 11, it can be seen that the fourth to seventh binary relations are newly added by the above processing.
[0050]
The item “obtained step” at the right end of the table indicates that the above binary relation was obtained in step S1-2.
Next, with reference to FIG. 14, the details of the main case search process (the process of step S1-3) shown in FIG. 9 will be described. This process selects a binary relation corresponding to the main case from one of the relation side and the receiving side that is an important word and adds it to the selection relation table. When this flowchart is started, the following processing is executed.
[S51] The selection relation table generating unit 14 refers to the important word list and determines whether or not an unprocessed important word remains. As a result, if unprocessed important words remain, the process proceeds to step S52, and if not, the process ends.
[0051]
In the present example, as a result of the process of FIG. 12, “Joint Venture Company” still remains in the important word list, and thus the process is executed. That is, in the first process, since the above-mentioned important word remains, it is determined as YES, and the process proceeds to step S52.
[S52] The selection relation table generating unit 14 acquires one important word K from the important word list.
[0052]
In this example, “joint venture” is selected as the important word K.
[S53] The selection relation table generating unit 14 determines whether or not the binary relation R including the important word K exists in the relation table. If it exists, the process proceeds to step S54, and if not, the process proceeds to step S56. Proceed to
[0053]
In this example, since there is a binary relationship (binary relationship consisting of “joint company” and “establishment”) corresponding to the important word “joint company”, it is determined as YES and the process proceeds to step S54.
[S54] The selection relation table generating means 14 determines whether or not the relation of the binary relation R is described in the main case list (see FIG. 15) stored in a memory or the like (not shown). As a result, if it is described in the main case list, the process proceeds to step S55, and if not, the process proceeds to step S56.
[0054]
In the current example, the binary relationship consisting of “joint venture” and “establishment” is “Okaku”, which is described in the third place in the main case list shown in FIG. It determines with YES and progresses to step S55.
[S55] The selection relation table generating unit 14 adds the binary relation R to the selection relation table.
[0055]
In the present example, “joint venture” and “establishment” are added to the selection relationship table.
[S56] The selection relation table generating unit 14 deletes the important word K from the important word list.
[0056]
As a result of the above processing, a selection relationship table as shown in FIG. 16 is generated. In this selection relationship table, “Joint Venture” and “Establishment” and “(Target)” indicating these relationships are shown in the last row as the binary relationship obtained by the processing of Step S1-3 described above. Is described.
[0057]
It should be noted that the above binary relation selection may be executed in the order of the major case list, that is, according to the importance of the major case. The relation of the binary relation selected in that case is as follows.
[0058]
Is the case (subject) → is the case (theme) → is the case (target) → is the case (opposite) → ...
The case is analyzed using a pattern match from a noun sequence. For example, if the pattern is “person name“ president ””, it is interpreted that this “person name” represents “president”. On the other hand, in “Subsidiary Eyama Boseki”, “No” is equivalent, but here we do not interpret it. This is because even if the interpretation of “no” is left ambiguous, there is no harm in the summary.
[0059]
When the above process ends, the process returns to the flowchart of FIG. 8 and the process of step S2 is executed.
In this process, a binary relationship connecting an isolated relationship and another binary relationship is extracted. Here, the isolated relationship means the binary relationship obtained in step 1-2 in FIG. A new binary relation is extracted from the relational table for connection with the binary relation obtained in step 1-1 of FIG. Such processing can be achieved by graph search.
[0060]
Since the analysis result shown in the relation table can be regarded as two words (nodes) connected by a relation (arc) (graph), the above-described processing can be executed by a graph search. Since this isolated relation (isolated arc) satisfies the condition that it is in the same sentence as the binary relation already included in the selection relation table (binary relation acquired in step S1-1), this search must be successful. To do.
[0061]
For example, in the example shown in FIG. 16, first, a binary relation consisting of “capital” and “650 million yen” is acquired.
Next, there is a two-term relationship that consists of “650 million yen” and “fund”, with the relationship of “650 million yen” on the related side, and the relationship is “ Selected.
[0062]
Subsequently, a two-term relationship having a relationship to “fund”, that is, a relationship consisting of “Esanbo” and “fund” is selected.
Then, two sets of binary relations of “650 million yen” and “investment” and “Eyama spinning” and “investment” are added to the selection relation table.
[0063]
As a result of the above processing, as shown in FIG. 18, the binary relationship consisting of “capital” and “650 million yen” that are isolated, and “subsidiary” extracted in the processing of step S1-1. And a binary relationship (“650 million yen” and “investment”, and “Eyama spinning” and “investment”) are extracted and included in the selection relationship table. It will be newly added.
[0064]
When the same processing is repeated, the “silk fabric” and “company” acquired in step S1-2 and the “shanghai spinning machine B” acquired in step S1-1 and “partnership” are linked. ”And“ Shanghai Spinning B ”are acquired and added to the selection relationship table.
[0065]
Furthermore, “Ryonosuke Kawasaki” and “President” acquired in Step S1-2 and “Subsidiary” and “Eyama Spinning” acquired in Step S1-1 are connected to “Eyama Spinning” and “President”. Binary relation is acquired and added to the selection relation table.
[0066]
An example of the selection relationship table generated by the above processing is shown in FIG. In this example, binary relations newly added by the above-described processing are shown in the ninth to twelfth.
[0067]
When the above process ends, the process of step S3 in FIG. 8 is executed.
In step S <b> 3, the complementary relationship adding means 15 acquires the essential case essential for the verb included in the selection relationship table from the relationship table and adds it to the selection relationship table. The required cases are subject, subject, subject, opposite case, and continuous nouns.
[0068]
Currently, there are three types of verbs listed in the selection relationship table: “join”, “establish”, and “fund”. The complementary relationship adding means 15 refers to the correspondence table showing the correspondence relationship between the verb and its essential case recorded in the essential modifier recording means 17 and searches for the essential case of these verbs from the relation table.
[0069]
Among the essential cases corresponding to these verbs, those included in the relational table are as follows.
“Mt. Fuji”, “Establish” (ha (theme))
“Thirty percent”, “Invest in” (-(Target))
Here, “-” indicates that the case particle is omitted.
[0070]
These essential cases are added to the selection relationship table by the complementary relationship adding means 15.
FIG. 19 is a diagram showing an example of the selection relationship table generated by the above processing. In the example of this figure, the aforementioned new binary relationship is added to the 13th and 14th.
[0071]
Note that nounary concepts (and sub-variant nouns) may be searched from the relational table and transferred to the selective relational table if their destination does not exist in the selective relational table.
Further, the above process may not be performed on the noun concept.
[0072]
In addition, although what was recorded in the essential modifier recording means 17 was used for an indispensable case, it can also make a user selectable, for example. Moreover, you may make it set about each verb. For example, “partner” may have “to (joint)” and “de (region)” in addition to the subject.
[0073]
Furthermore, “thing”, “thing”, “time”, “method”, “means”, etc., which have a low specific ability of meaning (having a broad meaning) by themselves, have their modifiers (“no” or other May be added (not present in the example of FIG. 3).
[0074]
When the above process ends, the process of step S4 in FIG. 8 is executed.
FIG. 20 is a flowchart illustrating details of the process in step S4 of FIG. In this process, the binary relations included in the selection relation table generated by the above process are appropriately connected to generate a sentence (summary sentence) as a processing result. When the flowchart shown in FIG. 20 is started, the following processing is executed.
[S71] The sentence synthesizing unit 16 acquires one binary relation from the selection relation table.
[0075]
Now, for example, a binary relationship consisting of “Mt. Fuji” and “Alliance”, which is the first binary relationship in the selection relationship table shown in FIG. 19, is acquired.
[S72] The sentence composition unit 16 deletes the acquired item from the selection relation table.
[0076]
Now, the binary relation consisting of “Mt. Fuji” and “Alliance” is deleted from the selection relation table.
[S73] The sentence synthesizing unit 16 generates a graph instance from the acquired binary relation.
[0077]
Now, an important word “Mt. Fuji” and “affiliation” are set as two nodes, and a graph instance with a relation “ha (theme)” added to an arc connecting these nodes is generated. At this time, an appearance sentence number (a number incremented sequentially from 1) is assigned to the arc, and is used when outputting the generated graph (details will be described later).
[S74] The sentence synthesizing unit 16 selects an arc to which no survey end mark indicating that the survey has ended is added in the graph instance.
[0078]
In the present example, an arc connecting “Mt. Fuji” and “partnership” to which no survey end mark has been added is selected.
[S75] The sentence synthesizing unit 16 executes a search process for the node on the side of the selected arc. This process is a subroutine, the details of which will be described later.
[0079]
In the present example, “Mount Fuji”, which is a node on the side of the engagement, is selected, and search processing for this node is executed.
[S76] The sentence synthesizing unit 16 executes a search process for the node on the receiving side of the selected arc. This process is a subroutine, the details of which will be described later.
[0080]
In the present example, “partnership” as the receiving node is selected, and search processing for this node is executed.
[S77] The sentence synthesizing unit 16 adds a survey end mark to the selected arc.
[0081]
In this example, the survey end mark is added to the arc connecting “Mt. Fuji” and “Alliance”.
[S78] The sentence synthesizing unit 16 determines whether or not survey end marks have been added to all arcs. As a result, when the survey end mark is added to all arcs, the process proceeds to step S79, and when not added, the process returns to step S74.
[0082]
In the present example, when a new arc is added to the graph composed of “Mt. Since it is not performed, it is determined as NO and the process returns to step S74.
[S79] The sentence synthesizing unit 16 determines whether or not there is still a binary relation in the selection relation table. As a result, if there is still a binary relationship in the selection relationship table, the process returns to step S71, and if not, the process ends.
[0083]
For example, if the processing for the first two-term relationship (“Mt. Fuji” and “Alliance”) in the selection relationship table is completed, a sentence centered on these is generated (see FIG. 23). In this case, since the other two-term relationship still remains, YES is determined in step S79, the process returns to step S71, and the same processing as described above is repeated. And when the process with respect to all the binary relations described in the selection relation table is complete | finished, it determines with NO and complete | finishes a process.
[0084]
Next, with reference to FIG. 21 and FIG. 23, the details of the processing of step S75 and step S76 shown in FIG. 20 will be described.
FIG. 21 is a flowchart showing details of the process in step S75 shown in FIG. When this flowchart is started, the following processing is executed.
[S91] The sentence synthesizing unit 16 determines whether or not the selected relation table has a binary relation that has the same word on the receiving side as the node on the relational side of the acquired binary relation. As a result, if there is a relationship on the receiving side that is the same word as the dependency side node, the process proceeds to step S92, and if not, the process returns to the process of FIG.
[0085]
At present, there is no binary relation on the receiving side that is the same word as the “relevant” node “Mt. Fuji”, so the determination is NO and processing returns to FIG.
[S92] The sentence synthesizing unit 16 extracts from the selection relation table the binary relation that the receiving side node has the same word as the relation side node, and then deletes it from the selection relation table.
[S93] The sentence synthesizing unit 16 determines whether or not the extracted binary relation and the relation of the node previously added to the relation side node are the same. As a result, if they are the same, the process returns to step S91, and if not, the process proceeds to step S94.
[0086]
For example, in the previous processing, when “Yamaguchi Prefecture” (of (attribute / part)) is added to “Mt. Fuji Spinning”, “SME” (of (attribute / part)) is extracted. In this case, since “Yamaguchi Prefecture's SME Fujisan Spinning” is redundant, the addition of “SMEs” mentioned later will be cancelled.
[S94] The sentence synthesizing unit 16 adds the extracted binary relation to the relation side node.
[0087]
As a result of the above processing, the binary relation having the same word as the node on the side of the selected arc on the receiving side is extracted from the selection relation table and then sequentially added to the graph instance.
[0088]
FIG. 22 is a flowchart showing details of the process in step S76 shown in FIG. When this flowchart is started, the following processing is executed.
[S111] The sentence synthesizing unit 16 determines whether or not a binary relation that has the same word as the received node of the acquired binary relation on the related side or the receiving side exists in the selection relation table. . As a result, if there is a binary relation having the same word as that of the receiving node on the engaging side or the receiving side, the process proceeds to step S112, and if not, the process returns to the process of FIG.
[0089]
Now, as the binary relations that have the same word as the partner node “affiliation” on the recipient side, there is “Li Baiyun Spinning 廠” and “affiliation” shown second in the selection table. In step S111, it is determined as YES and the process proceeds to step S112.
[S112] The sentence synthesizing unit 16 extracts from the selection relation table a binary relation having the same word as that of the receiving node in the related or receiving node, and then deletes it from the selection relation table.
[0090]
In this example, “Li Baiyun Spinning” and “Alliance” are extracted from the selection relationship table and then deleted.
[S113] The sentence synthesizing unit 16 determines whether or not the extracted binary relation and the relation of the arc previously added to the receiving node are the same. As a result, if they are the same, the process returns to step S111, and if not, the process proceeds to step S114.
[0091]
In this example, since no node has been added to the receiving node, it is determined as NO and the process proceeds to step S114.
[S114] The sentence synthesizing unit 16 adds the extracted binary relation to the receiving node.
[0092]
In this example, as shown in FIG. 23, “Li Baiyun Spinning” is added to the node “affiliation”.
When the processes in steps S74 to S78 shown in FIG. 20 are repeated, in the second process, an arc (an arc added in the first process) connecting “Li Baiyun Spinning” and “Alliance” is processed. Is selected as the target. Then, as a result of the processing in step S75, “company” is added to “Li Baiyun spinning mill” (see FIG. 23). Further, in the third process, an arc connecting “company” and “Li Baiyun spinning mill” is selected as a process target. Similarly, “China” and “silk fabric” are added as a result of the process of step S75. Since no connectable binary relation exists in the selection relation table, a new node is not added, and as a result, a new arc (an arc without a survey end mark) does not exist, and step In S78, it is determined as YES, and the process proceeds to Step S79.
[0093]
In step S79, since the binary relation still remains in the selection relationship table, it is determined YES and the process returns to step S71.
Then, the same processing as described above is repeated, and graphs as shown in FIGS. 24 and 25 are generated. In the example of FIG. 24, a graph in which “Mt. Fuji”, “Establishment”, and “Joint Venture” are connected is shown. In the example shown in FIG. 25, “capital”, “650 million yen”, “subsidiary”, “Eyama spinning”, “investment”, “Ryunosuke Kawasaki”, “president”, and “30” A graph in which “%” is connected is shown.
[0094]
When the above processing is completed, the sentence synthesizing unit 16 defines the output order for the plurality of generated graphs. Specifically, in the process of step S73, the graph with the smallest appearance number assigned to each arc of the graph is set as the graph number, and the graph is sorted in ascending order of the number.
[0095]
Then, the sentence synthesizing unit 16 performs a process of deleting the duplicated important word. That is, when a word at a node of each graph is described in the important word list, a mark is assigned to the important word. If there is a graph in which all the important words of the node are marked, the graph is deleted as redundant.
[0096]
The reason why such a process is performed is that a graph composed only of important words that have already been used often overlaps with the content of the existing sentence. Note that all graphs may be output without executing such processing.
[0097]
Next, the sentence synthesizing unit 16 performs a process of synthesizing a sentence from the above graph.
That is, when the graph includes a verb, the words of each node are arranged in the following order.
[0098]
[Subject] [Subject] [Subject] [Confrontation] [Other Cases] * [Attachment]
* Indicates self
For nouns, the words of each node are arranged in the following order.
[0099]
[Commodity modification] [Attribute / part] [Noun continuation] [Same rank] * [Adjunct] ["(" Supplement ")"]
In addition, each case shall be [involvement side] [surface layer name]. These may be further detailed by subdivision of verbs and nouns. With the above processing, for example, a sentence “Mt. Fuji Spinning is affiliated with Li Baiyun Spinning Co., Ltd., a Chinese silk fabric company” is generated from the graph of FIG.
[0100]
Also, from the graph in FIG. 24, “Fujisan Spinning established a joint venture.”
Furthermore, from the graph of FIG. 25, the sentence “Capital of 650 million yen is invested 30% by subsidiary Eyama Boseki (President Ryunosuke Kawasaki)” is generated.
[0101]
Then, the sentence synthesizing unit 16 outputs the generated three sentences to a display device such as a CRT (Cathode Ray Tube) display or a printer. As a result, the following sentence is output on the CRT or printing paper.
[0102]
Fuji Spinning has partnered with silk fabric company Lee Baiyun Spinning Co., Ltd.
Fujisan Spinning established a joint venture.
The capital of 650 million yen was invested 30% by subsidiary subsidiary Eyama Boseki (President Ryunosuke Kawasaki).
[0103]
Next, the results of processing other documents according to the above embodiment will be described. As a document to be processed, the one shown in FIG. 26 is used.
FIG. 27 is a list of important words extracted from the document shown in FIG. 26 by the important word extracting means 12. In this example, “Nakamura Chemical Industry”, “Yamada Chemical Industry”, “Chemical”, “Nakamura Chemical”, and others are extracted as important words.
[0104]
FIG. 28 shows the generated summary sentence. In this example, the following summary sentence is output.
Yamada Chemical Co., Ltd. established a 50-50 joint venture with Nakamura Chemical Industry.
[0105]
The new company is “Nakamura Chemical”.
In the future, the company will merge with Yamada Group agrochemical company.
Research and development of agricultural chemicals.
[0106]
Decline in exports due to yen appreciation.
Industry reorganization will also increase.
Next, the result of processing another document according to the above embodiment will be described. As a document to be processed, the one shown in FIG. 29 is used.
[0107]
FIG. 30 is a list of important words extracted from the document shown in FIG. 29 by the important word extracting means 12. In this example, “juicy meat”, “Fujio Yamamizu”, “meat”, “station building” and others are extracted as important words.
[0108]
FIG. 31 shows the generated summary sentence. In this example, the following summary sentence is output.
Tokyo Shibuya station building.
[0109]
Meat discount counter, juicy meat opens.
President Fujio Yamamizu is excited.
40% for beef and 30 to 40% for chicken are cheap.
[0110]
Mainly in shopping centers (SC), station buildings, and department stores in the Tokyo metropolitan area.
The ability to attract customers at the food department increased.
According to the above embodiment, it is possible to accurately generate a document summary with a small number of characters.
[0111]
For example, it becomes easier to grasp the outline in comparison with the keyword enumeration. In other words, in keyword enumeration, since keywords are displayed independently, the relationship between keywords cannot be grasped, but in the present invention, the relationship between keywords and keywords can be grasped, so what is roughly described? I can know.
[0112]
Further, in the present invention, since the length of one sentence is shorter than the important sentence pickup (+ deletion), the contents can be quickly grasped. That is, in the important sentence pickup, the length of one sentence tends to be long (the processing result of a plurality of samples is about 50 characters on average), and it takes time to grasp the outline. Even when unnecessary elements are deleted, the average is about 35 characters. On the other hand, the average number of characters in one sentence according to the present invention is about 17 characters, which is about half the length of important sentence pickup (+ deletion).
[0113]
Furthermore, there is a feature that the sentence is cut better than the important sentence pickup (+ deletion). That is, in the important part pick-up (+ deletion), the predicate of the sentence tends to remain as it is, and information that is not very meaningful for document selection tends to remain. For example, when the important part pick-up process is performed on the document of FIG. 26, the following output is obtained.
[0114]
Yamada Chemical Co., Ltd. announced that it has agreed with Nakamura Chemical Industries to establish a 50-50 joint venture in the agrochemical business.
In the processing result of the present embodiment, the following sentence is obtained.
[0115]
Yamada Chemical Co., Ltd. established a 50-50 joint venture with Nakamura Chemical Industry.
As is clear by comparing these, it can be seen that the points of the processing results according to the present embodiment are accurately extracted.
[0116]
Furthermore, in the important sentence pickup (+ deletion), there are cases where the meaning becomes unknown due to excessive deletion, but according to the present embodiment, such a problem can be solved. For example, when the conventional important sentence pick-up (+ deletion) processing is performed for the sentence “The biggest factor is the deterioration in export profitability due to the rapid appreciation of the yen.” . This is clearly too much deletion. On the other hand, according to the present embodiment, since the supplement processing of the main case for each word is executed, it is possible to obtain an appropriate summary that “the cause is the deterioration in export profit due to the strong yen”.
[0117]
Furthermore, according to the present embodiment, since a single phrase can be composed of a plurality of sentences, it is possible to prevent defocusing due to scattered meanings when originally composed of short sentences. It becomes.
[0118]
In the case where the original document presents the meaning by a combination of short sentences, the context is lost when the important sentence is picked up from there as in the prior art. Even in such a case, according to the present embodiment, the context can be complemented.
[0119]
For example, from the two sentences, “Meat discount counter, juicy meat department” and “Food department's ability to attract customers since juicy meat opened,” in this embodiment, “Meat discount counter, It is possible to synthesize and output a sentence "Juicy meat is a store." As a result, it is possible to generate an accurate summary even in a document whose meaning is formed by stacking short sentences.
[0120]
Next, with reference to FIG. 32, a second embodiment of the present invention will be described.
FIG. 32 is a diagram illustrating a configuration example of the second exemplary embodiment of the present invention. In this figure, parts corresponding to those in FIG. 2 are denoted by the same reference numerals, and description thereof is omitted. In this embodiment, important word number limiting means 21, display format changing means 22, and display device 23 are newly added. Other configurations are the same as those in FIG.
[0121]
The important word number limiting means 21 controls the number of important words extracted by the important word extracting means 12. In addition, the display format changing unit 22 changes the display format of the portion corresponding to the important word when the sentence synthesized by the sentence synthesizing unit 16 is displayed on the display device 23.
[0122]
Next, the operation of the above embodiment will be briefly described.
Assume that the example sentence shown in FIG. 3 is input to the document input unit 11. Then, the important word extracting unit 12 extracts an important word according to the control of the important word number limiting unit 21. If the number of important words to be extracted is limited (set) to 7, the important word extracting means 12 will extract the following important words.
[0123]
Level 1 Fujisan spinning, Eyama spinning
Level2 Li Baiyun spinning silk, silk fabric, joint venture, president, subsidiary
The relation table generating means 13 extracts the binary relations between the sentences including the above important words, thereby generating a relation table. The selection relationship generating unit 14 extracts the important binary relationship from the relationship table according to a predetermined rule by performing the processing described in the first embodiment, and generates a selection relationship table.
[0124]
The complementary relationship adding unit 15 selects a new relationship that complements the relationship between the binary relationships described in the selection relationship table from the relationship table, adds it to the selection relationship table, and records it in the essential modifier recording unit 17. Referring to the correspondence table, a binary relation that becomes an essential case is selected from the relation table, and is newly added to the selection relation table.
[0125]
The sentence synthesizing unit 16 synthesizes a sentence by appropriately connecting the binary relations described in the selection relation table supplied from the complementary relation adding unit 15. The resulting sentence is shown below.
[0126]
Fuji Spinning has partnered with silk fabric company Lee Baiyun Spinning Co., Ltd.
Fujisan Spinning established a joint venture.
Subsidiary Eyama Boseki (President Ryunosuke Kawasaki) invested 30%
Then, the display format changing unit 22 changes the display format of the important words included in the above sentence, and outputs it to the display device 23.
[0127]
FIG. 33 is a display example of a sentence displayed on the display device 23 at this time. In this example, an underline is added to a portion corresponding to an important word. By displaying in this way, it becomes possible to quickly grasp important parts in the sentence.
[0128]
FIG. 34 shows an example of another display format. In this example, the font size of the important word is changed. By doing so, it is possible to make it easier to grasp a visually important part as compared with the case of FIG.
[0129]
Although not shown in the figure, the display color of important words can be changed.
Subsequently, a third embodiment of the present invention will be described with reference to FIG.
[0130]
FIG. 35 is a diagram illustrating a configuration example of the third embodiment of the present invention. In this figure, parts corresponding to those in FIG. 2 are denoted by the same reference numerals, and description thereof is omitted. In this embodiment, the operation of the selection relationship table generating unit 25 is different from that in FIG. Other configurations are the same as those in FIG.
[0131]
In this embodiment, the selection relation table generating unit 25 selects the binary relation corresponding to a certain important word in the search process of the main case shown in step S1-3 of FIG. A score having a high score is selected with reference to a score (to be described later).
[0132]
In addition, the selection relation table generating unit 25 extracts other binary relations included in the same sentence as the binary relation in which both the dependency side and the receiving side are important words in the process of step S1-2. In this case, the binary relation is extracted within a range not exceeding the preset upper limit number.
[0133]
Next, the operation of the above embodiment will be briefly described.
When a predetermined document is input to the document input unit 11, the important word extraction unit 12 extracts an important word.
[0134]
The relation table generation unit 13 extracts the binary relations of the sentences by syntactic analysis of the sentences including the important words, and generates a relation table.
In step S35 shown in FIG. 12, the selection relationship table generating unit 25 adds the binary relationship S to the selection relationship table, and if there are a plurality of S, the two terms in a range not exceeding the predetermined upper limit value. Add a relationship after selecting it.
[0135]
In that case, the graph search in step S3 is interrupted at the search depth (predetermined upper limit value). Note that the relationship for which the search is interrupted is not added to the selection relationship table. The important words that should have been added here are added separately in the process of step S1-3.
[0136]
Instead of interrupting the search, it is possible to search for all connection candidates and select the one with the smallest number of hops (number of arc connections).
According to the processing as described above, it is possible to prevent an unnecessarily long sentence of the synthesis result. That is, in the example of the first and second embodiments, in step S2, the number of important words described in the selection relationship table (binary relationship in which both the dependency side and the reception side are important words) is reached. There are cases where things must be connected. This means that it results in a long phrase.
[0137]
However, in the present embodiment, since the number of binary relations connected to reach the selected important word is limited in step S1-2, the generated sentence is prevented from becoming unnecessarily long. be able to.
[0138]
Further, the selection relation table generating unit 25 executes the following process instead of the main case search process shown in FIG. In this embodiment, it is assumed that a score in the document is described for each word in the relation table. Further, it is assumed that the relation table has a column for entering the score of the whole relation. That is, the relationship table includes the following items.
1 sentence number
2 Related words
3 Dependent Word Score (complement-score)
4 recipient words
5 Receiving word score (head-score)
6 relation-type
7 Total score
The selection relation table generating unit 25 extracts a binary relation having the selected important word on the side of the relation or the receiver from the relation table including the above items. Then, the total score of each binary relation is calculated. The calculation formula for the total score is shown below.
[0139]
[Expression 1]

Here, Score (relation-type) is obtained by replacing the order of the main case list with the weight. That is, there is a relationship of “Score”> Score (complex modification).
[0140]
Then, the selection relationship table generating unit 25 extracts the binary relationship having the largest total score and adds it to the selection relationship table.
The reason why the values of Wc and Wh are changed is to determine whether priority is given to the binary relationship in which the important word is on the dependency side or the binary relationship on the reception side. Since the score of an important word is always higher than the score of the other word, the value of Wc may be set higher than Wh if priority is given to the binary relation where the important word is on the side of the dependency.
[0141]
Since the score of the word selected as the important word is common to all the binary relations, the weight of the word that is not the important word may be changed depending on whether it is the dependency side or the reception side. Wc, Wh and Wl are for determining which of the importance of the relationship and the importance of the word having the relationship is more important. If you do not take into account the importance of the relationship, you can set Wl = 0.
[0142]
By executing the processing as described above, if there are multiple binary relationships having the same major case for a certain important word, a binary relationship having a high importance is selected according to the setting. And added to the selection relationship table.
[0143]
In the first and second embodiments, when there are a plurality of binary relations having the same major case for a certain important word, these binary relations are selected according to the order of appearance in the relation table. It will be. In that case, an unimportant binary relation is also selected and added to the summary sentence.
[0144]
However, according to the processing as described above, the binary relation is selected according to the importance, not the order of appearance, so it is possible to extract only the important binary relation.
[0145]
The selection relationship table generated as described above is supplied to the complementary relationship adding means 15.
The complementary relationship adding unit 15 selects a new binary relationship that complements the relationship between the binary relationships described in the selection relationship table from the relationship table, adds the selected relationship to the selection relationship table, and also includes the essential modifier recording unit 17. Referring to the correspondence table recorded in (2), a binary relation as an essential case is selected from the relation table and newly added to the selection relation table.
[0146]
The sentence synthesizing unit 16 synthesizes a sentence by appropriately connecting the binary relations described in the selection relation table supplied from the complementary relation adding unit 15.
According to the above embodiment, it is possible to prevent an unnecessarily long sentence from being output and to output a sentence in which important words are preferentially combined.
[0147]
Subsequently, a fourth embodiment of the present invention will be described with reference to FIG.
FIG. 36 is a diagram illustrating a configuration example of the fourth exemplary embodiment of the present invention. In this figure, parts corresponding to those in FIG. 2 are denoted by the same reference numerals, and description thereof is omitted. In this embodiment, compared with the case of FIG. 2, document recording means 31, keyword input means 32, and document search means 33 are newly added. Other configurations are the same as those in FIG.
[0148]
The document recording unit 31 records the document input from the document input unit 11. A keyword for searching for a document recorded in the document recording unit 31 is input from the keyword input unit 32. The document search means 33 searches the document recording means 31 for a document containing the keyword input from the keyword input means 32 and outputs it.
[0149]
Next, the operation of the above embodiment will be described.
If a plurality of documents including the document shown in FIG. 3 are input from the document input unit 11, these documents are recorded in the document recording unit 31.
[0150]
Assuming that two words “China” and “Joint” are input as keywords from the keyword input means 32, the document search means 33 acquires a document containing both of these two keywords from the document recording means 31. To do. Since the document shown in FIG. 3 includes both of the two keywords “China” and “Joint”, this document and another document including both of these keywords are output as detection results. In the following processing, only the processing for the document shown in FIG. 3 will be described for simplicity.
[0151]
The important word extracting means 12 extracts important words and generates an important word list by the same processing as in the embodiment of FIG. The important word extraction unit 12 adds the keyword input from the keyword input unit 32 to the important word list as an important word. In this embodiment, the input keyword is replaced with an important word of Level3. Therefore, the important word list obtained is as follows.
[0152]
Level 1 Fujisan spinning, Eyama spinning
Level2 Li Baiyun spinning silk, silk fabric, joint venture, president, subsidiary
Level3 China, joint venture
The relation table generating means 13 extracts the binary relation of each sentence by performing syntax analysis of the sentence including the above important words, and generates a relation table. The selection relationship generation unit 14 performs the same processing as in the first embodiment, thereby extracting an important binary relationship from the relationship table according to a predetermined rule, and generates a selection relationship table.
[0153]
The complementary relationship adding unit 15 selects a new binary relationship that complements the relationship between the binary relationships described in the selection relationship table from the relationship table, adds the selected relationship to the selection relationship table, and also includes the essential modifier recording unit 17. Referring to the correspondence table recorded in (2), a binary relation as an essential case is selected from the relation table and newly added to the selection relation table.
[0154]
The sentence synthesizing unit 16 synthesizes a sentence by appropriately connecting the binary relations described in the selection relation table supplied from the complementary relation adding unit 15. The resulting sentence is shown below.
[0155]
Fuji Spinning has partnered with Chinese silk fabric company Li Baiyun Spinning Co., Ltd.
Fujisan Spinning established a joint venture.
Subsidiary Eyama Boseki (President Ryunosuke Kawasaki) invested 30%.
[0156]
Such a synthesized sentence is supplied to, for example, a display device (not shown), and displayed on a CRT or printing paper.
Such processing is executed for all documents acquired by the document search means 33, and a summary of a plurality of documents as search results can be obtained.
[0157]
According to the above embodiment, when a desired document is selected from a plurality of documents using a predetermined keyword, a summary of the search result document is created and displayed. Can be found quickly. In that case, since the keyword is arranged in the sentence to be synthesized, even if the operator is searching for the target document in the state of being profound, the word that is most clearly stored is included. Since the sentence is output as a summary sentence, it is possible to quickly determine whether or not the output document is a desired one.
[0158]
In the above embodiment, the input keyword is replaced with an important word of Level 3. For example, it is replaced with an important word of another level or added to an important word of a predetermined level. It may be.
[0159]
It is also possible to use only keywords as important words. That is, it is possible to make only the word inputted as a keyword an important word without extracting the important word by the important word extracting means 12.
[0160]
The synthesized sentence obtained in that case is shown below.
Fujisan has partnered with Chinese company Li Baiyun Spinning.
Fujisan Spinning established a joint venture.
[0161]
A sufficient summary sentence (composite sentence) can also be obtained by such a method.
Further, as in the second embodiment, the display format of important words and keywords may be changed.
[0162]
FIG. 37 shows a display example in which important words are displayed with an underline and keywords with an upper line. By displaying in this way, it is possible to quickly know where the keyword is arranged in the synthesized sentence.
[0163]
Further, the keyword input from the keyword input means 32 is expanded as a thesaurus (with a synonym corresponding to the input keyword as a new keyword) to perform a search process. When generating the summary of the acquired document, add the keyword obtained by the thesaurus expansion as an important word, or display the keyword obtained by the thesaurus expansion when displaying the generated summary sentence May be changed.
[0164]
In the above example, the relationship satisfying a certain condition in the relationship table is extracted and the selection relationship table is generated. However, the same processing can be performed without using the selection relationship table. For example, a mark may be attached to a relationship that satisfies a certain condition in the relationship table. Such an example will be described below as a fifth embodiment.
[0165]
FIG. 38 is a diagram illustrating a configuration example of the fifth embodiment of the present invention.
The document input unit 41 has a function of inputting a document to be processed. The document 47 input from the document input unit 41 is transferred to the important word extraction unit 42.
[0166]
The important word extraction means 42 gives an importance score to each word in the document. The tf * IDF product is used as the score calculation method. In this embodiment, all the words are extracted, and an importance score is given to each extracted word.
[0167]
The relation table generating unit 43 analyzes a sentence including the word (all or only important ones) obtained by the important word extracting unit 42 and extracts a binary relation between the words. Analysis methods include syntax analysis and pattern matching for morphological analysis results. Based on the analysis result, the relation table 48 is generated.
[0168]
The relationship selection unit 44 extracts a binary relationship including an important word from the relationship table 48 generated by the relationship table generation unit 43 according to a predetermined rule, and gives an important relationship mark to the relationship in the relationship table 48. As extraction rules, words having a total importance score that is greater than or equal to a certain value are arranged in order of score, and from the top of them, a certain number is selected. In addition, the number of important words and whether or not they are in the same sentence as other important relationships may be included in the extraction rule.
[0169]
The complementary relationship adding means 45 gives a selection mark to a new relationship that complements the relationship group to which the important relationship mark in the relationship table 48 is added. What is complemented are mandatory cases, mandatory modifiers, etc.
[0170]
The sentence synthesizing means 46 synthesizes the summary sentence 49 of the input document 47 from the binary relations marked in the relation table.
Next, the operation in the fifth embodiment will be described.
[0171]
The document 47 input by the document input unit 41 is transferred to the important word extraction unit 42. The important word extraction unit 42 assigns importance scores to all the words existing in the document 47. Here, the tf * IDF product is used. In addition, you may make it give only an independent word, without giving a score with respect to all the words. Moreover, you may make it the thing which excluded the conjunction, adverb, and pronoun among independent words.
[0172]
When importance scores are assigned to all the words, the relation table generating unit 43 extracts the binary relations between the words, and scores those binary relations. In this example, the sum of the scores of the engaging word and the receiving word is used as the score for the binary relation. In the case where the noun on the clerk side or the receiver side is a compound noun, the score of the compound noun is the sum of the scores of the constituent nouns. With respect to scoring, there is a method in which a score is also given to the type of relationship, and each of the dependency-side word score, the receiving-side word score, and the relationship score is weighted to obtain a total sum thereof. An example of the calculation method is as shown in the above formula (1). Then, a relation table 48 is generated based on the scored binary relations.
[0173]
Next, the relationship selecting means 44 selects a certain number of important relationships from the relationship table 48 in descending order of score. This number is given in advance. Alternatively, it may be calculated each time in consideration of the overall length of the summary. The relationship selection unit 44 gives a mark to the selected important relationship in the relationship table 48.
[0174]
The complementary relation adding means 45 and the sentence synthesizing means 46 sequentially pick up the relations with the highest score at that time from the binary relations selected by the relation selecting means 44, add the important relations to them, and add the summary sentences. Repeat the synthesis process.
[0175]
FIG. 39 is a flowchart showing summary sentence synthesis processing in the fifth embodiment.
[S201] The complementary relationship adding unit 45 collects adjacent complementary relationships. Specifically, one important relationship with a mark is picked up. Then, an important relationship that satisfies the following conditions is connected to the selected important relationship.
[0176]
(1) Distance 0: An important relationship that is directly connected to either element in the important relationship.
FIG. 40 is a diagram illustrating a complementary relationship of distance 0. In the figure, the relationship between the element 51 of “A” and the element 52 of “B” is an important relationship that has been picked up. In this figure, an arrow is shown from the dependency side of the dependency relationship toward the reception side (the same applies to FIGS. 41 and 42). In this example, since the element 61 of “C1” is directly connected to the element 51 of “A”, they are in a complementary relationship adjacent to each other with a distance of zero. In addition, since the elements 62 and 63 of “C2” and “C3” are directly connected to the element 52 of “B”, they are in a complementary relationship adjacent to each other with a distance of zero.
[0177]
(2) Distance 1: An important relationship connected to one of the elements in the important relationship through one basic relationship (not listed as an important relationship).
FIG. 41 is a diagram illustrating the complementary relationship of the distance 1. In this figure, the basic relationship is indicated by a thick arrow (the same applies to FIG. 42). In this example, the important relationship constituted by the element 71 of “D1” and the element 72 of “E1” has a basic relationship between the element 72 of “E1” and the element 51 of “A”. The complementary relationship is adjacent at a distance of 1. The important relationship constituted by the element 73 of “D2” and the element 74 of “E2” is adjacent to the distance 1 because the element 74 of “E2” and the element 52 of “B” have a basic relationship. It is a complementary relationship. The important relationship constituted by the element 75 of “D3” and the element 76 of “E3” is adjacent to the distance 1 because the element 76 of “E3” and the element 52 of “B” have a basic relationship. It is a complementary relationship. The important relationship constituted by the element 77 of “D4” and the element 78 of “E4” is adjacent to the distance 1 because the element 77 of “D4” and the element 52 of “B” have a basic relationship. It is a complementary relationship.
[0178]
(3) Distance 2: An important relationship connected through one element E (two basic relationships).
FIG. 42 is a diagram illustrating a complementary relationship of the distance 2. In this example, the important relationship constituted by the element 81 of “G1” and the element 82 of “H1” is “B” via two basic relationships in which the receiving element 82 sandwiches the element 80 of “F”. ”Is connected to the element 52 of“ ”. The important relationship constituted by the element 83 of “G2” and the element 84 of “H2” is the element 52 of “B” via two basic relations in which the element 83 on the engagement side sandwiches the element 80 of “F”. Therefore, the complementary relationship is adjacent at a distance of 2. The important relationship constituted by the element 85 of “G3” and the element 86 of “H3” is the element 52 of “B” through two basic relationships in which the element 86 on the receiving side sandwiches the element 80 of “F”. Therefore, the complementary relationship is adjacent at a distance of 2.
[0179]
Although three types of adjacent important relationships are given here according to the distance between the relationships, this can be increased or decreased. To shorten a sentence, only select a short distance. If a sentence can be long, include text with a large distance. Further, the extent of inclusion may be dynamically changed according to the length of the component (word) and the sentence length when it is generated.
[0180]
Returning to FIG. 39, the continuation of the flowchart will be described.
[S202] The complementary relationship adding means 45 supplements the essential basic relationship. Specifically, a basic relationship having an indispensable binary relationship with each element is added. The required cases are “ha”, “ga”, “ha”, “ni”, and a noun sequence. A selection mark is assigned to the essential supplementary relationship in the relationship table.
[S203] The sentence synthesizing unit 46 removes the relation established as an element of the summary sentence from the relation table. That is, the important relationship selected in step S201 and the basic relationship used in step S202 are removed from the relationship table.
[S204] The sentence synthesizing unit 46 synthesizes sentences. Specifically, a sentence is synthesized by connecting auxiliary verbs, auxiliary verbs, and auxiliary particles that were originally present, and the synthesized sentence is added to the summary sentence list.
[S205] The sentence synthesizing unit 46 determines whether the total sentence length of the summary sentence list exceeds a certain value. If the total sentence length of the summary sentence list exceeds a certain value, the process repeats (loop) from steps S201 to S204, and proceeds to step S206. Otherwise, the process proceeds to step S201 and repeats the process again.
[0181]
It should be noted that even when the important relationship is exhausted, the process repeats and the process proceeds to step S206. However, by selecting a large number of important relationships in advance, the end condition is avoided.
[S206] The sentence synthesizing unit 46 rearranges the summary sentences obtained in step S204 in the original appearance order.
[0182]
In this way, a summary sentence is generated, and the sentence synthesizing means 46 presents the summary sentence.
Next, a specific summary sentence generation example is shown.
FIG. 43 is a diagram illustrating an example of an input document. This is an introductory text about “Sake Biscuits”. Such a document 47 is transferred from the document input means 41 to the important word extraction means 42.
[0183]
Here, using the above algorithm, an algorithm simulation is performed based on actual KR (key relation relation) extracted data. In this example, it is assumed that the relationship is extracted as ideal.
[0184]
First, the important word extracting means 42 selects 5 to 10 important relationships. The scoring of words here uses the tf * IDF product. Originally, IDF should be calculated from a certain document set to which the document belongs, but here, a document set obtained by picking up about 10,000 articles from about 100,000 articles of the Mainichi Shimbun 1994 is used. This is because it is difficult to calculate a DF by defining a set of Web documents, and there is little bias in newspaper articles, and it can be determined that the frequency of commonly used words is not very different.
[0185]
Next, the relation table generating unit 43 analyzes a sentence including each word and extracts a binary relation between the words. Then, a score for each binary relation is calculated, and a relation table is generated.
FIG. 44 is a diagram showing a relation table generated from the picked-up relation. The relationship table 48 includes “relation name”, “first word group”, “second word group”, “number of appearances”, “first word group score”, “second word group score”. , “Overall score”, and “Mark” items are provided.
[0186]
“Relationship name” indicates the relationship between the two words. The relation name “moved name” indicates that the first word group is a verb and the second word group is a noun, which are connected by the word “ta”. The relation name “adjective name” indicates that the first word group is an adjective verb, the second word group is a noun, and is connected by the word “na”. The relation name “ni” indicates that the first word group and the second word group are connected by the word “ni”. The relation name “form name” indicates that the first word group is an adjective and the second word group is a noun, and is connected by the word “i”. The relation name “composite” indicates that the word set in the first word group is a compound noun. The relation name “no name” indicates that the first word group and the second word group are both nouns and are connected by the word “no”. The relation name “NN” indicates that the word set in the first word group is a noun sequence.
[0187]
In the “first word group”, a word or a word group on the side of the engagement is registered. The word group is registered when the noun is a compound noun or a noun continuation. In addition, a word group whose relation name is “compound noun” or “NN (noun continuous)” is also registered in the item of “first word group”.
[0188]
In the “second word group”, the receiving word or word group is registered. The word group is registered when the recipient is a compound noun or when the noun is continuous.
The “number of appearances” is set to how many times it appears in the input document.
[0189]
The “first word group score” is a score of the word group set in the “first word group”. When a plurality of words are set, it is the total score of each word.
[0190]
The “second word group score” is a score of the word group set in the “second word group”.
“Overall score” is a score for the entire binary relation. It is obtained by adding the value of “first word group score” and the value of “second word group score”.
[0191]
The “mark” is a flag indicating the important relationship selected by the relationship selecting unit 44. If a predetermined mark is registered in this item, it indicates that it has been selected as an important relationship. At the stage immediately after the relationship table is generated by the relationship table generation means 43, the important relationship is not selected.
[0192]
In this figure, important relationships are arranged in descending order of the “overall score”.
When such a relationship table is generated, the relationship selection means 44 selects a certain number of important relationships having a high overall score.
[0193]
FIG. 45 is a diagram showing a relationship table after selecting an important relationship. As described above, the mark “◯” is given to the binary relation selected by the relation selecting unit 44.
[0194]
Next, the process shown in FIG. 39 is performed by the complementary relationship adding means 45 and the sentence synthesizing means 46.
First, the first loop will be described.
[0195]
In step S201, the complementary relationship adding unit 45 selects an important relationship having the highest overall score. The important relationship to choose here is
● Rice cakes, biscuit
It is. Then, the adjacent important relationships are summarized. The adjacent important relationship is
● Elegant-Sake, Biscuits
It is. This important relationship is an adjacent relationship with a distance of 0 because the word “Fuya” is directly connected to “Sake, Biscuits”.
[0196]
In step S202, the basic relationship having the essential case is supplemented. In this example, “same” is added to “rice”. as a result,
● Romance
Basic relationship is added. A selection mark is given to this basic relationship in the relationship table 48.
[0197]
In step S203, the sentence synthesizing means 46 removes the important relationship. The three relationships used here are removed from the list.
In step S204, the sentence synthesizing means 46 synthesizes the sentence. The expression obtained in this first loop is
"Romantic sake biscuits with romance"
It becomes.
[0198]
Thereafter, the loop proceeds in the same manner.
In the second loop,
Important relationship: “Gogo, 升 -ni-paid”
Adjacent important relationship: None
Mandatory case supplement: None ("Pay-Take-Pay")
Surface expression: “Pay in Gogo”
Is obtained.
[0199]
In the third loop
Important relationship: “High-Specialties, Ginjo, Sake”
Adjacent important relationship: None
Mandatory supplement: “High fragrance”
Surface expression: “Scented specialty ginjo sake”
If you have five relationships, it ends here.
[0200]
In the fourth loop
Important relationship: “Kanagawa-no-Western confectionery, Ueda”
Adjacent important relationship: None
Mandatory case supplement: "Ueda-Creation"
Surface expression: “Created by Kanagawa's Western confectionery Ueda”
In the fifth loop
Important relationship: “Dip biscuit dough”
Adjacent important relationship: “Sake-no-aroma”
(Connected through “scent-dipping” (overall score, 11th place))
Required case supplement: None
The surface expression is "I dip the biscuit dough in the scent of a fine wine".
[0201]
The relationship picked up is as follows.
"Romantic sake biscuits with romance"
"Pay to Gogoen"
"Scented specialty ginjo sake"
"Created by Kanagawa's Western confectionery Ueda"
“I dipped the biscuit dough in the scent of a fine wine”
A summary sentence is generated by rearranging the extracted sentences in the order of appearance in the original document.
[0202]
FIG. 46 is a diagram showing the generated summary sentence. Such a summary sentence is output from the sentence synthesizing means 46.
As described above, if a mark indicating that the selection is made is given to the relationship table, it is not necessary to provide a selection relationship table separately.
[0203]
Next, a sixth embodiment will be described. In the sixth embodiment, the relationship determined to be used as a summary sentence is not deleted, but is kept in a state where the score is lowered.
[0204]
FIG. 47 is a diagram illustrating a configuration example of the sixth embodiment. In this figure, the document input means 41 and the important word extraction means 42 have the same functions as the constituent elements of the same name in the fifth embodiment, and therefore the same reference numerals as those of the constituent elements of the same name in FIG. Therefore, the description is omitted.
[0205]
The relation table generation means 43a has basically the same function as the relation table generation means 43 (shown in FIG. 39) of the fifth embodiment, and further has the following functions. That is, when a re-scoring request is received from the sentence synthesizing unit 46a, the score of the word selected as the summary sentence is lowered, and the value related to the score of the relation table is recalculated.
[0206]
The relationship selection unit 44a selects an important relationship having the highest overall score at that time when the relationship table is generated by the relationship table generation unit 43a and when the re-scoring is performed by the relationship table generation unit 43a. .
[0207]
The complementary relationship adding unit 45a adds an essential relationship to the relationship selected by the relationship selecting unit 44a.
The sentence synthesizing unit 46a generates a surface sentence based on the relationship selected by the relationship selecting unit 44a and the essential relationship added by the complementary relationship adding unit 45a. If the length of the entire surface sentence does not exceed the threshold value, a rescoring request is sent to the relation table generating unit 43a. The re-scoring request includes information on the words constituting the newly added surface sentence. If the length of the entire surface sentence exceeds the threshold, the generated surface sentence is arranged in the order of appearance in the original document 47, and the summary sentence 49a is generated.
[0208]
Next, a summary sentence generation procedure according to the sixth embodiment will be described.
The algorithm up to the generation of the relation table is the same as that in the fifth embodiment. Therefore, the process after the generation of the relation table will be described below.
[0209]
FIG. 48 is a flowchart showing a summary sentence synthesis process in the sixth embodiment.
[S211] When the relation table is generated, the relation selection means 44a selects only one binary relation having the highest importance score.
[S212] The complementary relationship adding unit 45a adds the essential relationship. In the initial stage, “GA”, “HA”, “ON”, “NI”, etc. are fixed. Here, since the important relationship is not extracted in advance, there is no step for complementing the adjacent important relationship.
[S213] The complementary relationship adding unit 45a performs relationship supplementation (preservation of ambiguity). In the case of a form such as “C of A → B” and only “A → B” or “A → C” is extracted, the relationship “C of B” is added.
[S214] The sentence synthesizing unit 46a counts the number of branches or the like, or provisionally applies surface sentence generation, and determines whether or not the length of the entire surface sentence exceeds a preset threshold value. If the threshold value is exceeded, the process proceeds to step S216 to exit from the repeated process (loop). Otherwise, the process proceeds to step S215.
[S215] The relation table generating unit 43a performs re-scoring. That is, the score of the word used is dropped. Here, the score of each word is set to 1/5. The new relationship score is then recalculated from the new word score. Thereafter, the procedure returns to the selection of the important relationship in step S211, and this procedure is repeated from the point of selecting the relationship having the highest score.
[S216] The sentence synthesizing unit 46a generates and lists the summary surface sentence. At this time, summary sentences are generated by rearranging them in the order of appearance. Then, the process ends.
[0210]
Next, a specific summary sentence generation example will be described.
The processing in the first loop will be described through the sample document 47 shown in FIG.
[0211]
In step S211, the relationship selecting unit 44a selects a binary relationship having the highest importance score. here
● Rice-Sake, Biscuits
Is selected. At this time, the relationship selecting unit 44a sets the selected important relationship mark in the relationship table to “1-C”. This mark means the core relationship (the important relationship selected in step S211) obtained in the first loop.
[0212]
In step S212, the essential relationship is given by the complementary relationship adding means 45a. Here, “Komeru” “O” is rated
● Full of romance
Is granted. Then, the complementary relationship adding unit 45a gives the mark “1-A” to the basic relationship having the required case in the relationship table. This mark means an additional relationship (basic relationship added in step S212) obtained in the first loop.
[0213]
In step S213, ambiguity is preserved by the complementary relationship adding means 45a. In the case of a form such as “C of A → B”, when only “A → B” or “A → C” is extracted, the relationship “C of B” is added. There is no corresponding example here.
[0214]
In step S214, the sentence synthesizing unit 46a determines whether the end condition is satisfied. At this stage, it is assumed that the number of branches is 2 and the number of summary characters is 13 (“Liquor biscuit with romance”), and the end condition is not yet satisfied.
[0215]
In step S215, the re-scoring is performed by the relationship table generating unit 43a. That is, the score of the used word is dropped, and the score of the new relationship is recalculated from the score of the new word.
[0216]
FIG. 49 is a diagram showing a relationship table after the end of the first loop. In this relationship table 48a, “Kome-Ta-Sake-Biscuits”, which is an important relationship selected in the first loop, is given a “1-C” mark. The scores of the words “Kome”, “Sake”, and “Biscuit” included in this important relationship are reduced to 1/5, and the scores in the relationship table are recalculated. As a result, the scores of “Kome-Ta-Sake-Biscuits” and “Fenga-Na-Sake-Biscuits” which were first place were “5.15 / 5 + 9.68 / 5 = 2.97”, “4 + 9.68”, respectively. /5=5.94 ". In the figure, the underlined score is a score whose value has been changed. For this reason, the highest score is “Gogo-Nen-Tame” (score 13.44), which was the third highest score.
[0217]
The same operation is continued after the second time. Here, it repeats several times without ending conditions in advance.
Perform a second loop. “Gogo-Nen-Tame” is selected with an overall score of “13.44”, the number of branches is 3, and the number of summary characters is 22 by adding 7 characters.
[0218]
FIG. 50 is a diagram showing a relation table after the end of the second loop. In the relationship table 48a, a mark “2-C” is newly added to “Five Goes-to-pay”.
[0219]
Perform the third loop. “High-specialized, ginjo, sake” is selected, and “fragrance-high” is added. The surface expression is “fragrance-high-specialized, ginjo, sake”. The number of branches is 5, and the number of summary characters is 29 with 9 characters added.
[0220]
FIG. 51 is a diagram showing a relationship table after the end of the third loop. In the relationship table 48a, a mark “3-C” is newly added to “high-specialized, ginjo, sake”. Although not shown, the mark “3-A” is given to the basic relationship “scent-high”.
[0221]
Perform the fourth loop. “Kanagawa-no-Western confectionery, Ueda” is selected, and “Ueda-creation” was added. The number of branches is 7, and the number of summary characters is 42 by adding 15 characters.
[0222]
The final summary is
"Liquor biscuit with romance"
“Pay in Gogo--”
"Aroma-high-specialties, ginjo, sake"
"Kanagawa-Western confectionery, Ueda Ueda-created"
It is. A summary sentence is generated by rearranging the surface sentence in the order of appearance in the original document. FIG. 52 is a diagram showing a summary sentence 49a generated according to the sixth embodiment.
[0223]
By the way, in the above example, the relationship between the important words is managed by the relation table, but an analysis tree can be used instead of the relation table. A mark can be attached to the parse tree.
[0224]
FIG. 53 is a diagram illustrating an analysis tree that has been marked. This shows an analysis tree at the time when the first loop is completed in the sixth embodiment. In this figure, words connected by a thick line indicate that the selected binary relationship is established. Moreover, the symbol shown in the vicinity of the line has shown the selection mark. The meaning of the mark is the same as in the sixth embodiment. A word surrounded by a bold rectangle indicates the selected word.
[0225]
In this way, by adding marks to the parse tree, it is possible to extract the relationships that make up the summary sentence.
The above processing functions can be realized by a computer. In this case, the processing contents of the functions that the document summarizing apparatus should have are described in a program recorded on a computer-readable recording medium, and the above processing is realized by the computer by executing the program by the computer. Is done.
[0226]
Examples of the computer-readable recording medium include a magnetic recording device and a semiconductor memory. When distributing to the market, store the program on a portable recording medium such as a CD-ROM (Compact Disk Read Only Memory) or floppy disk, or store it in a computer storage device connected via a network. In addition, it can be transferred to another computer through the network.
[0227]
When executed by a computer, the program is stored in a hard disk device or the like in the computer, loaded into the main memory and executed.
[0228]
【The invention's effect】
As described above, in the present invention, the document to be processed is input from the document input unit, the important word extraction unit extracts an important word from the document input from the document input unit, and the relation table. The generating means analyzes the structure of the sentence including the important word extracted by the important word extracting means, analyzes the binary relation between the words included in the sentence, generates a relation table, and the relation selecting means Selects a binary relation including an important word from the relation table generated by the relation table generating means according to a predetermined rule, and the sentence synthesizing means synthesizes a sentence from the binary relation selected by the relation selecting means. As a result, a short and accurate summary can be created from the input document.
[Brief description of the drawings]
FIG. 1 is a principle diagram illustrating the principle of a document summarizing apparatus according to the present invention.
FIG. 2 is a diagram illustrating a configuration example of a first embodiment of the present invention.
3 is an example of a document processed in the embodiment shown in FIG. 2. FIG.
4 is a diagram showing an example of an important word list extracted from the document shown in FIG. 3 by the important word extracting means shown in FIG. 2;
FIG. 5 is a diagram illustrating an example of a result obtained when syntax analysis is performed on a sentence including an important word by the relation table generation unit illustrated in FIG.
6 is a diagram showing an example of a relation table obtained as a result of the syntax analysis shown in FIG. 5. FIG.
7 is a diagram showing an example of sentence numbers assigned to the document shown in FIG. 3; FIG.
FIG. 8 is a flowchart illustrating an example of processing executed in the embodiment shown in FIG.
FIG. 9 is a flowchart illustrating an example of “binary relationship extraction processing including important words” illustrated in FIG. 8;
FIG. 10 is a flowchart illustrating an example of “direct relationship search processing” illustrated in FIG. 8;
11 is a diagram showing an example of a selection relationship table generated as a result of the processing shown in FIG.
FIG. 12 is a flowchart for explaining an example of “relation processing for relationships in the same sentence” shown in FIG. 9;
13 is a diagram showing an example of a selection relationship table generated as a result of the processing shown in FIG.
FIG. 14 is a flowchart showing an example of “main case search processing” shown in FIG. 9;
FIG. 15 is an example of a main case list;
FIG. 16 is a diagram illustrating an example of a selection relationship table that is generated.
FIG. 17 is a diagram showing an example of a selection relationship table that is generated.
FIG. 18 is a diagram showing a binary relation to be added.
FIG. 19 is a diagram showing an example of a selection relationship table generated as a result of the “essential case supplement processing” shown in FIG. 8;
FIG. 20 is a flowchart for explaining an example of “surface phrase generation processing” shown in FIG. 8;
FIG. 21 is a flowchart for explaining an example of “search processing for a node on the side of a selected arc” shown in FIG. 20;
FIG. 22 is a flowchart for explaining an example of “search processing for a node on the receiving side of a selected arc” shown in FIG. 20;
FIG. 23 is a diagram illustrating an example of a sentence obtained by combining the binary relations described in the selection relation table by the process illustrated in FIG. 20;
24 is a diagram showing another example of a sentence obtained by combining the binary relations described in the selection relation table by the process shown in FIG.
FIG. 25 is a diagram showing still another example of a sentence obtained by combining the binary relations described in the selection relation table by the process shown in FIG. 20;
FIG. 26 is a diagram showing an example of a document processed in the embodiment shown in FIG.
FIG. 27 is a diagram illustrating an example of important words extracted from the document illustrated in FIG.
FIG. 28 is a diagram showing an example of a sentence synthesized from the document shown in FIG.
FIG. 29 is a diagram showing an example of a document processed in the embodiment shown in FIG. 2;
30 is a diagram showing an example of important words extracted from the document shown in FIG. 29. FIG.
FIG. 31 is a diagram showing an example of a sentence synthesized from the document shown in FIG. 29;
FIG. 32 is a diagram illustrating a configuration example of a second embodiment of the present invention.
33 is a diagram showing a display example of a sentence displayed on the display device as a result of processing the document shown in FIG. 3 in the embodiment shown in FIG. 32;
34 is a diagram showing another display example of a sentence displayed on the display device as a result of processing the document shown in FIG. 3 in the embodiment shown in FIG. 32;
FIG. 35 is a diagram illustrating a configuration example of a third embodiment of the present invention.
FIG. 36 is a diagram illustrating a configuration example of a fourth embodiment of the present invention.
FIG. 37 is a diagram showing a display example of a sentence displayed on the display device as a result of processing the document shown in FIG. 3 in the embodiment shown in FIG. 36;
FIG. 38 is a diagram illustrating a configuration example of a fifth embodiment of the present invention.
FIG. 39 is a flowchart showing summary sentence synthesis processing in the fifth embodiment;
FIG. 40 is a diagram illustrating a complementary relationship of distance 0.
FIG. 41 is a diagram illustrating a complementary relationship of distance 1;
FIG. 42 is a diagram illustrating a complementary relationship of distance 2;
FIG. 43 is a diagram illustrating an example of an input document.
FIG. 44 is a diagram showing a relationship table generated from a picked-up relationship.
FIG. 45 is a diagram showing a relationship table after selection of an important relationship.
FIG. 46 is a diagram showing a generated summary sentence.
FIG. 47 is a diagram illustrating a configuration example of a sixth embodiment of the present invention.
FIG. 48 is a flowchart showing summary sentence synthesis processing in the sixth embodiment;
FIG. 49 is a diagram showing a relational table after the end of the first loop.
FIG. 50 is a diagram showing a relation table after the end of the second loop.
FIG. 51 is a diagram showing a relationship table after the end of the third loop.
FIG. 52 is a diagram illustrating a summary sentence generated according to the sixth embodiment.
FIG. 53 shows an analysis tree that has been marked.
[Explanation of symbols]
1 Document input means
2 Important word extraction means
3 Relational table generation means
4 Selection relationship generation means
5 Complementary relationship adding means
6 sentence composition means
11 Document input means
12 Important word extraction means
13 Relational table generation means
14 Selection relationship generation means
15 Complementary relationship adding means
16 sentence composition means
17 Required modifier recording means
21 Means to limit the number of important words
22 Display format change means
23 Display device
25 Selection relation table generation means
31 Document recording means
32 Keyword input means
33 Document search means

Claims (8)

In a document summarization apparatus that creates and outputs a summary of an input document,
A document input means for inputting a document to be processed;
Important word extraction means for extracting important words from the document input from the document input means;
A relation table generating unit that analyzes a structure of a sentence including an important word extracted by the important word extraction unit, analyzes a binary relation between words included in the sentence, and generates a relation table;
Relation selection means for selecting a binary relation including an important word from the relation table generated by the relation table generation means according to a predetermined rule;
A complementary relationship adding means for selecting a new binary relation for complementing the binary relation selected by the relation selecting means from the relation table;
Sentence synthesizing means for synthesizing a sentence from the binary relation selected by the relation selecting means and the complementary relation adding means ;
A document summarization apparatus comprising: 2. The document summary according to claim 1, wherein the complementary relationship adding means selects a new binary relationship connecting the isolated binary relationship of the sentence and the main binary relationship of the sentence from the relationship table. apparatus. The complementary relationship adding means selects a new binary relationship connecting the isolated binary relationship of the sentence and the main binary relationship of the sentence from the relationship table , and a predetermined number of 2 from the selected binary relationship. 2. The document summarizing apparatus according to claim 1, wherein only the term relation is further selected. It further has an essential modifier recording means for recording the type of modifier required for each word,
  2. The document summarization apparatus according to claim 1, wherein the complementary relationship adding unit selects a new binary relationship with reference to information recorded in the essential modifier recording unit. In a computer-readable recording medium on which a document summarization program for creating and outputting a summary of an input document is recorded,
  The computer,
  Document input means for inputting a document to be processed,
  Important word extraction means for extracting important words from the document input from the document input means;
  A relation table generating means for analyzing a structure of a sentence including an important word extracted by the important word extracting means and analyzing a binary relation between words included in the sentence to generate a relation table;
  A relation selection means for selecting a binary relation including an important word from the relation table generated by the relation table generation means according to a predetermined rule;
  A complementary relation adding means for selecting a new binary relation from the relation table to complement the binary relation selected by the relation selecting means;
  Sentence synthesizing means for synthesizing a sentence from the binary relation selected by the relation selecting means and the complementary relation adding means;
  A computer-readable recording medium on which a document summarizing program is recorded. 6. The document summary according to claim 5, wherein the complementary relationship adding means selects a new binary relationship connecting the isolated binary relationship of the sentence and the main binary relationship of the sentence from the relationship table. A computer-readable recording medium on which a program is recorded. The complementary relationship adding means selects a new binary relationship connecting the isolated binary relationship of the sentence and the main binary relationship of the sentence from the relationship table, and a predetermined number of 2 from the selected binary relationship. 6. The computer-readable recording medium recording the document summarizing program according to claim 5, wherein only the term relation is further selected. Further causing the computer to function as essential modifier recording means for recording the type of modifier required for each word;
  6. The document summarization program according to claim 5, wherein said complementary relation adding means selects a new binary relation with reference to information recorded in said essential modifier recording means. Computer-readable recording medium.

Images