JP2004094521A - Inquiry type learning method, learning device, inquiry type learning program, recording medium recorded with the program, recording medium recorded with learning data, inquiry type identification method and device using learning data, program, and recording medium with the program - Google Patents

Inquiry type learning method, learning device, inquiry type learning program, recording medium recorded with the program, recording medium recorded with learning data, inquiry type identification method and device using learning data, program, and recording medium with the program Download PDF

Info

Publication number
JP2004094521A
JP2004094521A JP2002253853A JP2002253853A JP2004094521A JP 2004094521 A JP2004094521 A JP 2004094521A JP 2002253853 A JP2002253853 A JP 2002253853A JP 2002253853 A JP2002253853 A JP 2002253853A JP 2004094521 A JP2004094521 A JP 2004094521A
Authority
JP
Japan
Prior art keywords
question
learning
program
question type
recording medium
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Granted
Application number
JP2002253853A
Other languages
Japanese (ja)
Other versions
JP4008313B2 (en
Inventor
Jun Suzuki
鈴木 潤
Yutaka Sasaki
佐々木 裕
Eisaku Maeda
前田 英作
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Nippon Telegraph and Telephone Corp
Original Assignee
Nippon Telegraph and Telephone Corp
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Nippon Telegraph and Telephone Corp filed Critical Nippon Telegraph and Telephone Corp
Priority to JP2002253853A priority Critical patent/JP4008313B2/en
Publication of JP2004094521A publication Critical patent/JP2004094521A/en
Application granted granted Critical
Publication of JP4008313B2 publication Critical patent/JP4008313B2/en
Anticipated expiration legal-status Critical
Expired - Lifetime legal-status Critical Current

Links

Images

Landscapes

  • Machine Translation (AREA)
  • Information Retrieval, Db Structures And Fs Structures Therefor (AREA)

Abstract

<P>PROBLEM TO BE SOLVED: To provide an inquiry type learning device for constituting a precise classifier for inquiry type identification. <P>SOLUTION: An identity characterizing an inquiry type is extracted from a learning sample of an inquiry, using a word attribute N-gram, by a feature extracting part 1. Then, an identity space is constituted based on the extracted identity by an identity vector preparing part 2 to convert each inquiry into an identity vector. At last, the classifier is constituted to identify the type of inquiry when an unknown inquiry is input, based on the identity vector, using a support vector machine, by a classifier constituting part 3. <P>COPYRIGHT: (C)2004,JPO

Description

【0001】
【発明の属する技術分野】
本発明は質問タイプ学習方法、装置、プログラム、および同プログラムを記録した記録媒体、ならびに質問タイプ同定方法、装置、プログラム、および同プログラムを記録した記録媒体に関する。
【0002】
【従来の技術】
近年、情報検索/抽出、計算機との対話、質問応答など、人間から計算機に対して自然文で質問して計算機に解答させる技術が使用される場面が増えてきている。質問タイプ同定とは、質問が与えられたときに、その質問が「何を訊いているのか」という意図を解析することである。質問タイプは、質問の意図を表すクラスであり、質問の意図を解析することは、質問タイプを同定する分類問題と捉えることができることから、質問タイプ同定と呼ぶ。
【0003】
計算機にとって人間が与える質問の意図を理解するのは非常に難しい問題である。人間と同様に、計算機にとっても、与えられた質問の意図を的確に理解できないと、質問に答えることはできない。
【0004】
従来、質問のパターンを限定し意味的制約のもとで解析を行う手法、人手によりルールを作成する方法などが提案されている。意味制約を用いる方法は、各質問タイプを表す表現を事前に決定しておき、その表現にあった質問しか受け付けられないという問題がある。ルールによる方法の場合は、質問文に出現する表現をルールに変換することで理論上は、多種多様な表現の質問を扱うことが可能であるが、ルールを作成するのに人手が非常にかかる点が問題となる。また、仮にコストをかけて作成する場合でも、様々なパラメータを調整し最適な同定ルールを決定するのは極めて困難である。
【0005】
このような問題を解決する一つの方法として、機械学習方法を用いて学習サンプルから自動で質問タイプの同定方法を学習する方法がある。機械学習方法では、その問題で分類すべき対象の特徴である「素性」を抽出し、その素性の集合を用いて分類器を学習するという手順で問題に適用する。
【0006】
【発明が解決しようとする課題】
このような機械学習方法では、素性の数が重要な要素となる。本発明で取り扱う質問タイプ同定問題は、質問の意図を判別する問題であるため、高精度な識別を行うためには多種多様な情報が必要となる。特に、質問文中にどのような単語が出現するかという情報と、出現した単語同士がどのような関係で使われているかという情報が重要になる。つまり、質問タイプ同定問題は、質問タイプを特徴付けるために必要な素性数は非常に多くなる。機械学習方法では、学習サンプルが無限に存在するならば、素性数が多くなっても問題にはならないが、実際には学習サンプルは限られた数しか存在しないため、学習サンプル数に対して素性数が多いと、学習サンプルに対して過学習してしまい、未知のデータに対しての精度が低くなるという問題がある。
【0007】
一般的に、機械学習方法をもちいる場合は、分類を行うのに最も有効であると考えられる必要最小限の素性のみを用いることが望ましい。しかし、質問タイプ同定では、素性の集合の中で、実際にどれが有効であるかを人間が直観的に判断するのは非常に難しい。このため、分類を行うのに最も有効であると考えられる必要最小限の素性を抽出するのは非常に困難であり、従来提案されている機械学習方法をそのまま適用すると高精度な分類器が学習できないという問題がある。
【0008】
本発明の目的は、質問タイプ同定のための高精度な分類器を構成する質問タイプ学習方法、装置、プログラム、同プログラムを記録した記録媒体および学習データを記録した記録媒体を提供することにある。
【0009】
本発明の他の目的は、前記分類器を用いて未知の質問から質問タイプを同定する質問タイプ同定方法、装置、プログラム、および記録媒体を提供することにある。
【0010】
【課題を解決するための手段】
まず、特徴抽出部により、質問の学習サンプルから質問タイプを特徴付ける素性を抽出する。質問タイプを特徴付ける素性を抽出する方法として、単語属性N−gramを用いる。単語属性とは、単語、品詞、意味情報のことであり、N−gramとは、それらの連鎖を意味する。
【0011】
この単語属性N−gramを用いることにより、質問タイプを特徴付ける素性を自動的に、かつ、網羅的に抽出することができる。例えば、人手により有効と考えられる素性を列挙する場合に比べてタイプを特徴付ける素性を効率的に抽出できる。
【0012】
表1に単語属性N−gramの例を示す。
【0013】
【表1】

Figure 2004094521
w,p,sはそれぞれ単語、品詞、意味情報を表し、質問はw,w,...,wで構成されているとする。この場合、抽出される特徴は、単語属性N−gramに基づき、表2のようになる。ここで、N−gramの要素が1つのものを1−gram、要素が2つの連鎖を2−gramと呼び、以下、要素3を3−gram、要素4を4−gram、...と呼ぶ。
【0014】
【表2】
Figure 2004094521
ここで、抽出された一つの単語属性の連鎖が一つの素性となる。
【0015】
次に、素性ベクトル作成部により、単語属性N−gramにより抽出された素性から素性空間を構成し、各質問を素性ベクトルに変換する。最後に、分類器構成部により、抽出された特徴と質問の学習サンプルから、大量の素性を用いても高精度な分類器を学習可能な統計的機械学習方法を用いて学習する。サンプル数に対して素性数が多い問題でも、高精度な分類器を学習できる機械学習手法としてSupport Vector Machine(参考文献:V.Vapnik, The Nature of Statistical Learning Theory. Spring−Verlag, New York, 1995. 参照)を用いる。
【0016】
学習時のSupport Vector Machineの入力は、クラスラベルと素性ベクトルのペアで表される。
【0017】
【数1】
Figure 2004094521
ここで、
【0018】
【外1】
Figure 2004094521
はi番目のサンプルの特徴ベクトルでn次元ベクトルであり、yはサンプルiのクラスを表すスカラー変数である。ただし、Support Vector Machineは二クラス分類器なので、分類すべきクラスは正例(+1)、負例(−1)の2つである。
【0019】
学習サンプルの例を表3に示す。このような学習サンプルは記録媒体に記録して質問タイプの同定の際に利用することが可能であり、データのみ流通させることができる。
【0020】
【表3】
Figure 2004094521
素性数nの場合は、n次元の素性空間を考えることにより、各質問から作成される素性ベクトルはn次元素性空間の一点を表すことになる。Support Vector Machineは、図3のように、n次元素性空間内で、2クラスを分離する超平面のうち、2クラス間のマージンが最大になる平面を最適識別平面として選択する学習手法である。
【0021】
また、Support Vector Machineは二クラス分類器であるので、多クラス分類手法を用いて適用する。多クラス分類手法としては、例えばone vs.rest法があり、分類対象となるクラスがn個存在するとき、任意の1クラスとそれ以外のn−1クラスを分類する分類器をn個作成することで、二クラス分類器で多クラスの分類を行う。この場合の識別境界は図4のようになる。
【0022】
最後に、以上のようにして作成された分類器を用いて、未知の質問が入力されたときに、該質問がどの質問タイプかを識別する。これは、未知の質問を表3と同様の方法で素性ベクトルに変換し、Support Vector Machineにより作成した分類器を用いて行う。識別ステップにより決定される質問タイプは、未知の質問から作成される素性ベクトルが素性空間上のどの位置に配置されるかで決定される。
【0023】
【発明の実施の形態】
次に、本発明の実施の形態について図面を参照して説明する。
【0024】
図1(1),(2)はそれぞれ本発明の一実施形態の質問タイプ学習装置、質問タイプ同定装置の構成図である。
【0025】
質問タイプ学習装置は特徴抽出部1と素性ベクトル作成部2と分類器構成部3を有している。質問タイプ同定装置は素性ベクトル作成部4と質問タイプ同定部5を有している。
【0026】
図2は本実施形態の質問タイプ学習装置と質問タイプ同定装置の処理の流れを示すフローチャートである。
【0027】
ステップ11に、特徴抽出部1は、質問の学習サンプルから質問タイプを特徴付ける素性を、単語N−gramを用いて抽出する。ステップ12に、素性ベクトル作成部2は単語属性N−gramにより抽出された素性から素性空間を構成し、各質問を素性ベクトルに変換する。ステップ13に、分類器構成部3は素性ベクトルからSupport Vector Machineを用いて、質問タイプを同定する分類器を構成する。
【0028】
未知の質問が入力されたとき、素性ベクトル作成部4で、未知の質問から質問タイプを特徴づける素性を、単語N−gramを用いて抽出し、抽出された素性を素性ベクトルに変換し、質問タイプ同定部5で、分類器構成部3で作成された分類器を用いて質問タイプを同定する。
【0029】
次に、本実施形態の具体例として、質問応答システムにおいて、質問タイプ同定問題を扱う場合を示す。
【0030】
質問応答システムの場合は、質問タイプは主に解答の種類の分類を表すタイプとなる。例えば、「人名」「地名」「人数」「日時」等である。
【0031】
まず、質問タイプ学習ステップを説明する。
【0032】
質問応答システムの場合の学習サンプルの例として、「日本の首都はどこですか」を考える。この質問は場所を訊いている質問なので、識別すべき質問タイプは「場所」となる。このように、学習サンプルは表4に示すように質問と識別すべき質問タイプのペアで与えられる。
【0033】
【表4】
Figure 2004094521
このような学習サンプルから、各質問タイプを特徴付ける素性を抽出する。まず、質問を単語単位に分割し、各単語に品詞、意味情報を付加する。抽出される単語属性N−gramとは、例えば「日本の首都はどこですか」という質問なら、図5で示すように各単語自身とその品詞、意味情報を要素とした組み合わせの連鎖である。考えられる全ての単語属性N−gramを抽出する。
【0034】
上記の「日本の首都はどこですか」という質問の場合、例えば表5のようになる。
【0035】
【表5】
Figure 2004094521
抽出された単語属性N−gramを素性として、素性空間を作成し、この素性空間を用いて統計的機械学習手法Support Vector Machineを用いて最適識別平面を決定し分類器を作成する。
【0036】
次に、質問タイプ同定ステップについて説明する。
【0037】
これは、Support Vector Machineにより作成された分類器を用いて、新たに入力される未知のサンプルの質問タイプを識別するステップである。
【0038】
入力された未知の質問に対して、素性ベクトルを作成し、出力される質問タイプをその質問の質問タイプと判定する。
【0039】
質問応答システムに適用する場合と同様に、対話システムや情報検索/抽出システムにも、それぞれのシステムに適した質問タイプを事前に決定すれば、それ以外は全く同様の方法で適用することができる。例えば、対話システムの場合で、スケジュール管理のタスクを取り扱っている場合は、質問タイプを「新規のスケジュール入力」「スケジュール変更」「スケジュール確認」等と決めることで適用することが可能である。また、情報検索/抽出システムの場合は、テキスト分類のカテゴリを質問タイプにすることで、検索範囲を限定することが可能になり、より高精度な検索が行えることができるようになる。
【0040】
なお、本発明の質問学習方法および質問タイプ同定方法は専用のハードウェアにより実現されるもの以外に、その機能を実現するためのプログラムを、コンピュータ読み取り可能な記録媒体に記録して、この記録媒体に記録されたプログラムをコンピュータシステムに読み込ませ、実行するものであってもよい。コンピュータ読み取り可能な記録媒体とは、フロッピーディスク、光磁気ディスク、CD−ROM等の記録媒体、コンピュータシステムに内蔵されるハードディスク装置等の記憶装置を指す。さらに、コンピュータ読み取り可能な記録媒体は、インターネットを介してプログラムを送信する場合のように、短時間の間、動的にプログラムを保持するもの(伝送媒体もしくは伝送波)、その場合のサーバとなるコンピュータシステム内部の揮発性メモリのように、一定時間プログラムを保持しているものも含む。
【0041】
【発明の効果】
以上説明したように、本発明によれば、質問タイプを特徴付ける単語の属性の構造を抽出する単語属性N−gramと、サンプル数に対して素性数が多い問題設定でも高精度な分類器を学習できる統計的機械学習手法SVM(Support Vector Machine)を組み合わせることにより、高精度な分類器を作成することが可能になり、これにより、質問のサンプルさえ作成すれば、質問タイプ同定のための高精度な分類器が作成可能となる。
【図面の簡単な説明】
【図1】本発明の一実施形態の質問学習装置と質問タイプ同定装置の構成図である。
【図2】図1中の質問学習装置と質問タイプ同定装置の処理を示すフローチャートである。
【図3】Support Vector Machineの概念図である。
【図4】多クラス分類により拡張されたSupport Vector Machineの概念図である。
【図5】単語属性N−gramの抽出例を示す図である。
【符号の説明】
1 特徴抽出部
2 素性ベクトル作成部
3 分類器構成部
4 素性ベクトル作成部
5 質問タイプ同定部
11〜15 ステップ[0001]
TECHNICAL FIELD OF THE INVENTION
The present invention relates to a question type learning method, a device, a program, and a recording medium on which the program is recorded, and a question type identification method, a device, a program, and a recording medium on which the program is recorded.
[0002]
[Prior art]
2. Description of the Related Art In recent years, the use of techniques for allowing a computer to answer a natural question to a computer, such as information search / extraction, dialogue with a computer, and answering a question, has been increasing. Question type identification, when given a question, is to analyze the intent of what the question is asking. The question type is a class representing the intention of the question, and analyzing the intention of the question can be regarded as a classification problem for identifying the question type, and is therefore called question type identification.
[0003]
It is a very difficult problem for a computer to understand the intent of a human question. Like a human, a computer cannot answer a question unless it properly understands the intent of the given question.
[0004]
Conventionally, there have been proposed a method of performing analysis under a semantic constraint by limiting a question pattern, and a method of manually creating rules. The method using the semantic constraint has a problem that an expression representing each question type is determined in advance, and only a question corresponding to the expression is accepted. In the case of the rule-based method, it is theoretically possible to handle questions with a wide variety of expressions by converting expressions that appear in the question sentence into rules, but it takes a lot of labor to create rules The point becomes a problem. Further, even if it is created at a high cost, it is extremely difficult to adjust various parameters and determine an optimal identification rule.
[0005]
One method of solving such a problem is to automatically learn a question type identification method from a learning sample using a machine learning method. In the machine learning method, "features" which are features of a target to be classified in the problem are extracted and applied to the problem by a procedure of learning a classifier using a set of the features.
[0006]
[Problems to be solved by the invention]
In such a machine learning method, the number of features is an important factor. Since the question type identification problem dealt with in the present invention is a problem of determining the intention of a question, a wide variety of information is required to perform highly accurate identification. In particular, information about what words appear in the question sentence and information about how the appearing words are used in each other are important. In other words, the question type identification problem requires a very large number of features to characterize the question type. In the machine learning method, if the number of training samples is infinite, it does not matter if the number of features increases, but since there are only a limited number of training samples, If the number is large, there is a problem in that the learning sample is over-learned, and the accuracy for unknown data is reduced.
[0007]
Generally, when using a machine learning method, it is desirable to use only the minimum necessary features considered to be the most effective for performing classification. However, in question type identification, it is very difficult for a human to intuitively judge which is actually effective in a set of features. For this reason, it is very difficult to extract the minimum necessary features considered to be the most effective for performing classification, and if the conventionally proposed machine learning method is applied as it is, a high-precision classifier will be trained. There is a problem that can not be.
[0008]
An object of the present invention is to provide a question type learning method, a device, a program, a recording medium recording the program, and a recording medium recording learning data, which constitute a high-precision classifier for question type identification. .
[0009]
Another object of the present invention is to provide a question type identification method, an apparatus, a program, and a recording medium for identifying a question type from an unknown question using the classifier.
[0010]
[Means for Solving the Problems]
First, a feature extracting unit extracts features that characterize a question type from a learning sample of the question. A word attribute N-gram is used as a method for extracting a feature that characterizes a question type. The word attributes are words, parts of speech, and semantic information, and the N-gram means a chain of them.
[0011]
By using this word attribute N-gram, it is possible to automatically and comprehensively extract features that characterize the question type. For example, a feature that characterizes a type can be extracted more efficiently than a case where features that are considered to be effective manually are enumerated.
[0012]
Table 1 shows an example of the word attribute N-gram.
[0013]
[Table 1]
Figure 2004094521
w, p, s each represent word, part of speech, the meaning information, the question w 1, w 2,. . . , And it is composed of a w n. In this case, the extracted features are as shown in Table 2 based on the word attribute N-gram. Here, one N-gram element is called 1-gram, and a chain of two elements is called 2-gram. Hereinafter, element 3 is 3-gram, element 4 is 4-gram,. . . Call.
[0014]
[Table 2]
Figure 2004094521
Here, a chain of one extracted word attribute is one feature.
[0015]
Next, the feature vector creation unit forms a feature space from the features extracted by the word attribute N-gram, and converts each question into a feature vector. Finally, the classifier constructing unit learns from the extracted feature and question learning samples using a statistical machine learning method capable of learning a highly accurate classifier even using a large amount of features. Support Vector Machine (Reference: V. Vapnik, The Nature of Statistical Learning Theory. Spring-Nearling, Spring-New Year, 95-year-old) ) Is used.
[0016]
The input of the Support Vector Machine at the time of learning is represented by a pair of a class label and a feature vector.
[0017]
(Equation 1)
Figure 2004094521
here,
[0018]
[Outside 1]
Figure 2004094521
Is the feature vector of the i-th sample, which is an n-dimensional vector, and y i is a scalar variable representing the class of sample i. However, since Support Vector Machine is a two-class classifier, the class to be classified is a positive example (+1) and a negative example (-1).
[0019]
Table 3 shows examples of learning samples. Such a learning sample can be recorded on a recording medium and used for identification of a question type, and only data can be distributed.
[0020]
[Table 3]
Figure 2004094521
In the case of the feature number n, by considering an n-dimensional feature space, the feature vector created from each question represents one point of the n-th elemental space. The Support Vector Machine is a learning method for selecting, as an optimal discrimination plane, a plane having a maximum margin between two classes among hyperplanes separating two classes in an n-order elemental space as shown in FIG. .
[0021]
Further, since Support Vector Machine is a two-class classifier, it is applied using a multi-class classification method. As the multi-class classification method, for example, one vs. one. There is a rest method, and when there are n classes to be classified, by creating n classifiers for classifying any one class and the other n-1 classes, a two-class classifier can be used for multi-class classification. Perform classification. The identification boundary in this case is as shown in FIG.
[0022]
Finally, when an unknown question is input using the classifier created as described above, the type of the question is identified. This is performed by converting an unknown question into a feature vector in the same manner as in Table 3, and using a classifier created by Support Vector Machine. The question type determined by the identification step is determined by where in the feature space the feature vector created from the unknown question is located.
[0023]
BEST MODE FOR CARRYING OUT THE INVENTION
Next, embodiments of the present invention will be described with reference to the drawings.
[0024]
FIGS. 1A and 1B are configuration diagrams of a question type learning device and a question type identification device according to an embodiment of the present invention, respectively.
[0025]
The question type learning device includes a feature extraction unit 1, a feature vector creation unit 2, and a classifier configuration unit 3. The question type identification device has a feature vector creation unit 4 and a question type identification unit 5.
[0026]
FIG. 2 is a flowchart showing the processing flow of the question type learning device and the question type identification device of the present embodiment.
[0027]
In step 11, the feature extraction unit 1 extracts features that characterize the question type from the learning sample of the question using the word N-gram. In step 12, the feature vector creation unit 2 constructs a feature space from the features extracted by the word attribute N-gram, and converts each question into a feature vector. In step 13, the classifier constructing unit 3 constructs a classifier for identifying the question type using the Support Vector Machine from the feature vector.
[0028]
When an unknown question is input, the feature vector creating unit 4 extracts features that characterize the question type from the unknown question using the word N-gram, converts the extracted features into feature vectors, The type identification unit 5 identifies the question type using the classifier created by the classifier construction unit 3.
[0029]
Next, as a specific example of the present embodiment, a case where a question type identification problem is handled in a question answering system will be described.
[0030]
In the case of a question answering system, the question type is a type mainly representing the type of answer. For example, "person name", "place name", "number of people", "date and time", etc.
[0031]
First, the question type learning step will be described.
[0032]
As an example of a learning sample for the question answering system, consider "Where is the capital of Japan?" Since this question is a question asking for a place, the question type to be identified is “place”. In this way, the learning samples are given as pairs of questions and question types to be identified as shown in Table 4.
[0033]
[Table 4]
Figure 2004094521
From such a learning sample, features that characterize each question type are extracted. First, a question is divided into words, and parts of speech and semantic information are added to each word. The extracted word attribute N-gram is, for example, a question of "Where is the capital of Japan?", As shown in FIG. 5, a chain of combinations of each word and its part of speech and semantic information as elements. Extract all possible word attributes N-gram.
[0034]
In the case of the question "Where is the capital of Japan" above, for example, it is as shown in Table 5.
[0035]
[Table 5]
Figure 2004094521
A feature space is created using the extracted word attribute N-gram as a feature, and an optimal discrimination plane is determined by using the feature space and a statistical machine learning method Support Vector Machine to create a classifier.
[0036]
Next, the question type identification step will be described.
[0037]
This is a step of using a classifier created by the Support Vector Machine to identify the question type of a newly input unknown sample.
[0038]
A feature vector is created for the input unknown question, and the output question type is determined as the question type of the question.
[0039]
As in the case of applying to a question answering system, if a question type suitable for each system is determined in advance, it can be applied to an interactive system and an information search / extraction system in exactly the same manner. . For example, in the case of an interactive system that handles a task of schedule management, it can be applied by determining the question type as "input a new schedule", "change schedule", "confirm schedule", or the like. In the case of the information search / extraction system, by setting the category of the text classification to the question type, the search range can be limited, and a more accurate search can be performed.
[0040]
It should be noted that the question learning method and the question type identification method of the present invention are not limited to those realized by dedicated hardware, and a program for realizing the functions is recorded on a computer-readable recording medium. May be read by a computer system and executed. The computer-readable recording medium refers to a recording medium such as a floppy disk, a magneto-optical disk, a CD-ROM, or a storage device such as a hard disk device built in a computer system. Further, the computer-readable recording medium is one that dynamically holds the program for a short time (transmission medium or transmission wave), such as a case where the program is transmitted via the Internet, and serves as a server in that case. It also includes those that hold programs for a certain period of time, such as volatile memory inside a computer system.
[0041]
【The invention's effect】
As described above, according to the present invention, a word attribute N-gram for extracting the structure of the attribute of a word characterizing a question type, and a highly accurate classifier can be learned even in a problem setting having a large number of features with respect to the number of samples. By combining a statistical machine learning technique SVM (Support Vector Machine) that can be used, a high-precision classifier can be created. As a result, if only a sample of a question is created, a high-precision classifier for question type identification can be obtained. Classifier can be created.
[Brief description of the drawings]
FIG. 1 is a configuration diagram of a question learning device and a question type identification device according to an embodiment of the present invention.
FIG. 2 is a flowchart showing a process of a question learning device and a question type identification device in FIG. 1;
FIG. 3 is a conceptual diagram of a Support Vector Machine.
FIG. 4 is a conceptual diagram of a Support Vector Machine extended by multi-class classification.
FIG. 5 is a diagram showing an example of extracting a word attribute N-gram.
[Explanation of symbols]
DESCRIPTION OF SYMBOLS 1 Feature extraction part 2 Feature vector creation part 3 Classifier construction part 4 Feature vector creation part 5 Question type identification part 11-15 steps

Claims (9)

事前に用意された質問文のサンプルからそれぞれの質問タイプを特徴付ける素性を、単語属性N−gramを用いて抽出するステップと、
抽出された素性から素性空間を構成し、各質問を素性ベクトルに変換するステップと、
前記素性ベクトルからSupport Vector Machineを用いて、未知の質問が入力されたときどの質問タイプかを識別する分類器を構成するステップを有する質問タイプ学習方法。
Extracting a feature characterizing each question type from a sample of a prepared question sentence using a word attribute N-gram;
Constructing a feature space from the extracted features and converting each question to a feature vector;
A question type learning method, comprising: using a Support Vector Machine from the feature vector to configure a classifier for identifying which question type is input when an unknown question is input.
請求項1に記載の質問タイプ学習方法で得られた素性ベクトルと、正例と負例からなるクラスのペアからなる学習サンプルが記録されている記録媒体。A recording medium on which is stored a feature vector obtained by the question type learning method according to claim 1 and a learning sample including a pair of a class including a positive example and a negative example. 事前に用意された質問文のサンプルからそれぞれの質問タイプを特徴付ける素性を、単語属性N−gramを用いて抽出する手段と、
抽出された素性から素性空間を構成し、各質問を素性ベクトルに変換する手段と、
前記素性ベクトルからSupport Vector Machineを用いて、未知の質問が入力されたときどの質問タイプかを識別する分類器を構成する手段を有する質問タイプ学習装置。
Means for extracting a feature characterizing each question type from a sample of question sentences prepared in advance using a word attribute N-gram;
Means for constructing a feature space from the extracted features and converting each question into a feature vector;
A question type learning apparatus, comprising: means for configuring a classifier for identifying which question type is input when an unknown question is input, using a Support Vector Machine from the feature vector.
請求項1に記載の質問タイプ学習方法をコンピュータに実行させるためのプログラム。A program for causing a computer to execute the question type learning method according to claim 1. 請求項4に記載のプログラムを記録した記録媒体。A recording medium on which the program according to claim 4 is recorded. 未知の質問から質問タイプを特徴づける素性を、単語属性N−gramを用いて抽出するステップと、
抽出された素性を素性ベクトルに変換するステップと、
前記素性ベクトルから、請求項1に記載の分類器を用いて質問タイプを同定するステップを有する質問タイプ同定方法。
Extracting a feature characterizing the question type from the unknown question using the word attribute N-gram;
Converting the extracted features into feature vectors;
A question type identification method comprising a step of identifying a question type from the feature vector using the classifier according to claim 1.
未知の質問から質問タイプを特徴づける素性を、単語属性N−gramを用いて抽出する手段と、
抽出された素性を素性ベクトルに変換する手段と、
前記素性ベクトルから、請求項1に記載の分類器を用いて質問タイプを同定する手段を有する質問タイプ同定装置。
Means for extracting a feature characterizing a question type from an unknown question using a word attribute N-gram;
Means for converting the extracted features into feature vectors,
A question type identification device comprising: means for identifying a question type from the feature vector using the classifier according to claim 1.
請求項6に記載の質問タイプ同定方法をコンピュータに実行させるためのプログラム。A program for causing a computer to execute the question type identification method according to claim 6. 請求項8に記載のプログラムを記録した記録媒体。A recording medium on which the program according to claim 8 is recorded.
JP2002253853A 2002-08-30 2002-08-30 Question type learning device, question type learning program, recording medium recording the program, recording medium recording a learning sample, question type identification device, question type identification program, recording medium recording the program Expired - Lifetime JP4008313B2 (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
JP2002253853A JP4008313B2 (en) 2002-08-30 2002-08-30 Question type learning device, question type learning program, recording medium recording the program, recording medium recording a learning sample, question type identification device, question type identification program, recording medium recording the program

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
JP2002253853A JP4008313B2 (en) 2002-08-30 2002-08-30 Question type learning device, question type learning program, recording medium recording the program, recording medium recording a learning sample, question type identification device, question type identification program, recording medium recording the program

Publications (2)

Publication Number Publication Date
JP2004094521A true JP2004094521A (en) 2004-03-25
JP4008313B2 JP4008313B2 (en) 2007-11-14

Family

ID=32059744

Family Applications (1)

Application Number Title Priority Date Filing Date
JP2002253853A Expired - Lifetime JP4008313B2 (en) 2002-08-30 2002-08-30 Question type learning device, question type learning program, recording medium recording the program, recording medium recording a learning sample, question type identification device, question type identification program, recording medium recording the program

Country Status (1)

Country Link
JP (1) JP4008313B2 (en)

Cited By (9)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
JP2006039862A (en) * 2004-07-26 2006-02-09 Mitsubishi Electric Corp Data classification apparatus
JP2006113746A (en) * 2004-10-13 2006-04-27 Hewlett-Packard Development Co Lp Document classification apparatus, method and program
JP2006244262A (en) * 2005-03-04 2006-09-14 Nec Corp Retrieval system, method and program for answer to question
JP2008134889A (en) * 2006-11-29 2008-06-12 National Institute Of Information & Communication Technology Opinion collection system, opinion collection method, and opinion collection program
US7590603B2 (en) * 2004-10-01 2009-09-15 Microsoft Corporation Method and system for classifying and identifying messages as question or not a question within a discussion thread
KR101064617B1 (en) 2009-02-27 2011-09-15 고려대학교 산학협력단 Method and apparatus for classifying multivariate stream data
JP2015102914A (en) * 2013-11-21 2015-06-04 日本電信電話株式会社 Method for learning incomprehensible sentence determination model, and method, apparatus and program for determining incomprehensible sentence
WO2019198386A1 (en) * 2018-04-13 2019-10-17 国立研究開発法人情報通信研究機構 Request rephrasing system, method for training of request rephrasing model and of request determination model, and conversation system
WO2024047929A1 (en) * 2022-08-29 2024-03-07 株式会社日立製作所 Company evaluation processor system

Families Citing this family (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
JP7084158B2 (en) 2018-02-23 2022-06-14 トヨタ自動車株式会社 Information processing methods, programs, information processing devices, and information processing systems

Cited By (14)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
JP2006039862A (en) * 2004-07-26 2006-02-09 Mitsubishi Electric Corp Data classification apparatus
JP4536445B2 (en) * 2004-07-26 2010-09-01 三菱電機株式会社 Data classification device
US7590603B2 (en) * 2004-10-01 2009-09-15 Microsoft Corporation Method and system for classifying and identifying messages as question or not a question within a discussion thread
JP4713870B2 (en) * 2004-10-13 2011-06-29 ヒューレット−パッカード デベロップメント カンパニー エル.ピー. Document classification apparatus, method, and program
JP2006113746A (en) * 2004-10-13 2006-04-27 Hewlett-Packard Development Co Lp Document classification apparatus, method and program
JP2006244262A (en) * 2005-03-04 2006-09-14 Nec Corp Retrieval system, method and program for answer to question
JP2008134889A (en) * 2006-11-29 2008-06-12 National Institute Of Information & Communication Technology Opinion collection system, opinion collection method, and opinion collection program
KR101064617B1 (en) 2009-02-27 2011-09-15 고려대학교 산학협력단 Method and apparatus for classifying multivariate stream data
JP2015102914A (en) * 2013-11-21 2015-06-04 日本電信電話株式会社 Method for learning incomprehensible sentence determination model, and method, apparatus and program for determining incomprehensible sentence
WO2019198386A1 (en) * 2018-04-13 2019-10-17 国立研究開発法人情報通信研究機構 Request rephrasing system, method for training of request rephrasing model and of request determination model, and conversation system
JP2019185521A (en) * 2018-04-13 2019-10-24 国立研究開発法人情報通信研究機構 Request paraphrasing system, request paraphrasing model, training method of request determination model, and dialog system
JP7149560B2 (en) 2018-04-13 2022-10-07 国立研究開発法人情報通信研究機構 Request translation system, training method for request translation model and request judgment model, and dialogue system
US11861307B2 (en) 2018-04-13 2024-01-02 National Institute Of Information And Communications Technology Request paraphrasing system, request paraphrasing model and request determining model training method, and dialogue system
WO2024047929A1 (en) * 2022-08-29 2024-03-07 株式会社日立製作所 Company evaluation processor system

Also Published As

Publication number Publication date
JP4008313B2 (en) 2007-11-14

Similar Documents

Publication Publication Date Title
CN109766540B (en) General text information extraction method and device, computer equipment and storage medium
CN111026842B (en) Natural language processing method, natural language processing device and intelligent question-answering system
WO2018207723A1 (en) Abstract generation device, abstract generation method, and computer program
CN104503998B (en) For the kind identification method and device of user query sentence
CN109190110A (en) A kind of training method of Named Entity Extraction Model, system and electronic equipment
JP4904496B2 (en) Document similarity derivation device and answer support system using the same
JP6832501B2 (en) Meaning generation method, meaning generation device and program
CN112052324A (en) Intelligent question answering method and device and computer equipment
US20220414463A1 (en) Automated troubleshooter
CN111144102B (en) Method and device for identifying entity in statement and electronic equipment
CN112541337A (en) Document template automatic generation method and system based on recurrent neural network language model
CN111143531A (en) Question-answer pair construction method, system, device and computer readable storage medium
CN108536673B (en) News event extraction method and device
JP4008313B2 (en) Question type learning device, question type learning program, recording medium recording the program, recording medium recording a learning sample, question type identification device, question type identification program, recording medium recording the program
CN114647713A (en) Knowledge graph question-answering method, device and storage medium based on virtual confrontation
CN114358017A (en) Label classification method, device, equipment and storage medium
JP5230927B2 (en) Problem automatic creation apparatus, problem automatic creation method, and computer program
CN111680501B (en) Query information identification method and device based on deep learning and storage medium
US11599580B2 (en) Method and system to extract domain concepts to create domain dictionaries and ontologies
CN115017271B (en) Method and system for intelligently generating RPA flow component block
JP6775465B2 (en) Dialogue rule collation device, dialogue device, dialogue rule collation method, dialogue method, dialogue rule collation program, and dialogue program
CN111401069A (en) Intention recognition method and intention recognition device for conversation text and terminal
JP6586055B2 (en) Deep case analysis device, deep case learning device, deep case estimation device, method, and program
JP2001022727A (en) Method and device for classifying and learning text and storage medium storing text classifying and learning program
JP2003263441A (en) Keyword determination database preparing method, keyword determining method, device, program and recording medium

Legal Events

Date Code Title Description
A621 Written request for application examination

Free format text: JAPANESE INTERMEDIATE CODE: A621

Effective date: 20040728

RD03 Notification of appointment of power of attorney

Free format text: JAPANESE INTERMEDIATE CODE: A7423

Effective date: 20040728

RD04 Notification of resignation of power of attorney

Free format text: JAPANESE INTERMEDIATE CODE: A7424

Effective date: 20040728

RD04 Notification of resignation of power of attorney

Free format text: JAPANESE INTERMEDIATE CODE: A7424

Effective date: 20050614

A131 Notification of reasons for refusal

Free format text: JAPANESE INTERMEDIATE CODE: A131

Effective date: 20070627

A521 Request for written amendment filed

Free format text: JAPANESE INTERMEDIATE CODE: A523

Effective date: 20070724

TRDD Decision of grant or rejection written
A01 Written decision to grant a patent or to grant a registration (utility model)

Free format text: JAPANESE INTERMEDIATE CODE: A01

Effective date: 20070822

A61 First payment of annual fees (during grant procedure)

Free format text: JAPANESE INTERMEDIATE CODE: A61

Effective date: 20070829

FPAY Renewal fee payment (event date is renewal date of database)

Free format text: PAYMENT UNTIL: 20100907

Year of fee payment: 3

R150 Certificate of patent or registration of utility model

Ref document number: 4008313

Country of ref document: JP

Free format text: JAPANESE INTERMEDIATE CODE: R150

Free format text: JAPANESE INTERMEDIATE CODE: R150

FPAY Renewal fee payment (event date is renewal date of database)

Free format text: PAYMENT UNTIL: 20100907

Year of fee payment: 3

FPAY Renewal fee payment (event date is renewal date of database)

Free format text: PAYMENT UNTIL: 20110907

Year of fee payment: 4

FPAY Renewal fee payment (event date is renewal date of database)

Free format text: PAYMENT UNTIL: 20120907

Year of fee payment: 5

FPAY Renewal fee payment (event date is renewal date of database)

Free format text: PAYMENT UNTIL: 20130907

Year of fee payment: 6

S531 Written request for registration of change of domicile

Free format text: JAPANESE INTERMEDIATE CODE: R313531

R350 Written notification of registration of transfer

Free format text: JAPANESE INTERMEDIATE CODE: R350

EXPY Cancellation because of completion of term